content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Nanomechanical resonators and nanoelectromechanical systems are
widely used in a multitude of applications such as ultrafast
actuation and sensing at the zeptogram and sub-attonewton
scales~\cite{Roukes2007,cooling,Ekinci,Cap1,Cap3}. The
extraordinary sensitivity that such NEMS provide rely mainly upon
their very high oscillating pulsations and quality factors $Q$.
However, while $Q$ factors in the tens or hundreds of thousands
can be obtained in vacuum and/or cryogenic environment, these
values degrade dramatically in liquid and gas phases meaning that
much more work is still to be done to reach the technological
level (see however~\cite{Ekinci, rugar01}). It is therefore
necessary to characterize more precisely the viscous forces
exerted by fluids on the vibrating motion of
NEMS and this constitutes the motivation for the present work.\\
Over the last years, micro- and nano-fluidics experiments
\cite{Tabeling} reported that the physical properties of fluids
flowing into or around confined systems, such as
nano-channels~\cite{nanoc}, are strongly modified compared with
those encountered at the micro and macro scales. In particular, it
has been shown that the usual no-slip boundary conditions, which
are universally used since the $19^{th}$ century to model the
behavior of Newtonian fluids at a solid interface, break down at
the nano scale~\cite{Lauga}. Such modifications of boundary
conditions are also expected to have a huge impact on NEMS
dynamics since properties known for Microelectromechanical systems
(MEMS)~\cite{Green,Naik,Paul,Dorignac,Basak,Tung} cannot simply be
scaled down to the nano realm.\\ In a recent work, we investigated
the importance of gas damping on a thermally actuated microlever
as it is gradually approached towards an infinite wall in
parallel geometry~\cite{nous}. The experiment performed at room
temperature in air showed that the sub-Angstrom lever oscillation
amplitude, i.e., in the direction perpendicular to the parallel
planes, is completely frozen as the gap $d$ is progressively
decreased from 20 $\mu$m to 400 nm. Moreover, the friction force
recorded was much larger that predicted by the Navier-Stokes
hydrodynamical equations solved together with the no-slip boundary
conditions. Instead, the reported results are qualitatively and
quantitatively well understood if one accepts the perfect-slip
boundary conditions for which friction at the lever/gas interface
is prohibited.\\
The aim of the present article is twofold. First,
we study theoretically the motion of the Newtonian fluid, i.e.
air, around the oscillating micro plate. Starting with the
linearized Navier-Stokes equation we analytically show that
perfect slip boundary conditions lead to the correct dynamical
behavior reported experimentally in ref.~\cite{nous}. We compare
our findings with the other more traditional approaches based on
the no-slip boundary conditions and show that they necessarily
conflict with the experimental facts. During the analysis we also
discuss the different possible boundary conditions and in
particular the impact of the slip length of the beam dynamical
behavior. The second goal of this paper is to show the important
implications that our findings may have on the engineering and
architecture of future NEMS operating in gaseous environment.
Here, on the basis of the results obtained in ref.~\cite{nous} for
a model system we discuss precisely the existence of a critical
overdamped regime for NEMS oscillating in fluids and study the
influence of materials, geometrical, and intrinsic dynamical
properties on this regime.\\
The paper is organized as follows. In
Sec.~II we discuss the general characteristics of the dynamic
problem in conjunction with Navier-Stokes equations and boundary
conditions. In Sec.~III the static regime corresponding to various
slip lengths and valid for small gap values is studied
analytically and compared (successfully) with the experimental
results reported in ref.~\cite{nous}. In sec.~IV we briefly discuss the possible microscopic mechanisms involved in order to explain the reported results. In particular, we compare the calculations obtained here for the plane-plane geometry to those already obtained in the sphere-plane configuration by Taylor and Vinogradova~\cite{Vinogradova,Vinogradova2}. Finally, in Sec.~V we
discuss the consequences of our findings for NEMS dynamics. A
summary is given in Sec.~VI.
\section{Mechanical oscillator in a newtonian fluid}
We consider a thin silicon commercial cantilever used in atomic
force microscopy (AFM) for liquid imaging~\cite{nous}. This system
is modelled as a parallelepiped with length $L$, width $w$ and
thickness $t$. The lever is clamped by one end and oscillates
mainly along the $z$ direction (see Fig.~1).
\begin{figure}[hbtp]
\includegraphics[width=0.7\columnwidth]{Fig1}\\% Here is how to import EPS art
\caption{Scheme of the one end-clamped parallelepiped cantilever
used in ref.~\cite{nous}. The lever has dimensions $t$, $L$, and
$w$ and is vibrating in the $z$ direction at a distance $d$ from
the substrate. }
\end{figure}
In vacuum this beam may be viewed as a harmonic oscillator with an
intrinsic resonance pulsation $\omega_0$ and effective mass $m$
(internal losses can be fairly neglected in the following
analysis). In the fluid, the viscous force acting on the lever is
characterized by a dissipative coefficient $\gamma$ connecting the
viscous force $F_z$ normal to the lever to the velocity $U$ of the
lever along the same direction: $F_z=-\gamma\cdot U$. At short
distance, i.e, in the non retarded regime, $\gamma$ is given
by~\cite{nous}:
\begin{equation}
\gamma= \frac{2 \eta Lw}{d},
\end{equation}
where $\eta$ is the (dynamic) fluid viscosity and $d$ the distance
between the lever and the substrate (see Fig.~1). As we will now
show this law results directly from the perfect slip boundary
conditions applied to the Navier-Stokes equation, and it differs
from the usual $\gamma\simeq \eta w L^3/d^3$ behavior deduced by
considering the
no slip conditions~\cite{nous,landau}.\\
To derive this result we start from the non-linear Navier-Stokes equations for an incompressible fluid
\begin{equation}
\rho \left[ \partial_t \vec v+(\vec v \cdot\vec{\nabla}) \vec v\right] = \eta \vec{\nabla} ^2 \vec v - \vec{\nabla} p \qquad
\label{eq:ns}
\end{equation}
where $\vec v$ is the local fluid velocity, $\rho$ its density,
and $p$ the gas pressure. To calculate the gas flow around the
lever we take into account specifical properties of the system
under study. First, the importance of the non-linear term $(\vec v
\cdot\vec{\nabla}) \vec v$ can be estimated from the knowledge of
the Reynolds number $R_e:=v \cdot X/\nu=\rho v\cdot X/\eta$
calculated for a characteristic length $X$ and fluid velocity $v$
($\nu=\eta/\rho=1.5\times 10^{-5}$ m$^{2}$/s is the kinematic
viscosity). In the present context, a correct order of magnitude
of the velocity in the fluid is given by the lever velocity
components $U_x$, $U_z:=U$ along the x and z direction
respectively. An important related feature is that we have here
$U_x\ll U$. Indeed, writing $\theta \simeq \delta z/L$ (with
$\delta z=0.05$ nm the typical lever oscillation amplitude) the
main angle shown by the lever with the $x$ axis we get the
estimation $U_x/U\simeq \theta \simeq 10^{-6} $ which implies that
the motion is mainly vertical. To simplify our analysis we
therefore assume that the lever is a horizontal plate vibrating
along the z direction. The lever oscillating at the frequency
$\omega/(2\pi)=50$ $k$Hz one obtains $U\simeq\omega \delta
z/(2\pi)\simeq 2.5 \times10^{-6}$ m/s and thus with $X:=d=50$
$\mu$m
\begin{eqnarray}
R_e=\frac{\omega \delta z \cdot d}{2\pi\nu}\simeq 0.8\times 10^{-5}\ll 1.
\end{eqnarray}
As a consequence of this ultra-small Reynolds number we will
completely neglect nonlinearity in the rest of this work. As a
corollary of this analysis we also deduce the Mach number
$M=v/c\sim U/c\simeq 10^{-8}\ll 1$ ($c$ sound velocity in air).
The vanishing value of $M$ justifies the fluid incompressibility hypothesis $\vec{\nabla}\cdot \vec v=0$.\\
The second question that we should deal with concerns the
amplitude of the dynamical term $\rho \partial_t \vec
v=-i\omega_0\rho\vec v$ in Eq.~\ref{eq:ns}. Since we are concerned
with harmonic oscillations we introduce a second Reynolds number:
\begin{eqnarray}
R'_e=\frac{\omega_0 \cdot d^2}{\nu}\simeq 2\times 10^{-2}\cdot
(d[\mu m])^2,
\end{eqnarray}
where $d[\mu m]$ is the measure of $d$ in micrometers. Clearly $R'_e$ (and therefore $\rho \partial_t \vec v$) is negligible as far as $d \ll\sqrt{2\nu/\omega_0}\simeq 10$ $\mu$m, i.e. as far as the gap $d$ is smaller than the boundary layer thickness $\delta_B$ (see below for discussion). In the present work we will only consider this static regime and neglect the dynamic term.\\
In order to solve the linearized Navier-Stokes equation
\begin{equation}
\label{eq:nsL} \eta \vec{\nabla} ^2
\vec v - \vec{\nabla} p\simeq 0,
\end{equation}
one must provide the precise boundary conditions for the fluid velocity at the solid interfaces. The condition
for the normal component of the velocity is quite obvious~\cite{landau,batchelor}. Indeed, since the fluid cannot
go through a solid interface the fluid velocity $v_z$ must equal the velocity of the plate $U$ at $z=d$ and must
also vanish along the surface $z=0$~\cite{landau,batchelor}. However, the precise form of the conditions for the tangential
components $v_x$ and $v_y$ is not so natural and is a subject of debates and controversies since the birth
of hydrodynamics~\cite{Lauga,Loeb,Neto,Charlaix}. The problem was already well addressed by Navier~\cite{Navier} who introduced the two most known
possibilities which are respectively the no-slip and perfect slip boundary conditions. Following
the no-slip hypothesis the fluid velocity along the interface must equal the in-plane velocity of the solid boundary. In the present case the no-slip conditions lead to $v_x=v_y=0$ along the substrate plane $z=0$ and $v_x=U_x\simeq0$, and $v_y=0$ along the cantilever interface at $z=d$. The no-slip hypothesis is well documented in the literature and experimentally justified at the macroscale~\cite{Lauga,Loeb,Neto,Charlaix}. It leads however to increasing difficulties and contradictions in the micro and nanofluidic regime where fundamental and statistical properties of molecules such as the mean free path and the surface roughness cannot be ignored~\cite{Lauga,Loeb,Neto,Charlaix,Tabeling}.\\
The second extreme possibility, the perfect slip boundary
conditions, suppose having $\partial_zv_i=0$ (from now on we use
the notation $i=x,y$) along the solid interface. These are
reminiscent for conditions on the viscous stress tensor
$\sigma_{iz}=\eta(\partial_zv_i+\partial_iv_z)=0$ meaning that no
tangential friction is allowed between fluid and solid. We however
point out that the equivalence between the relations
$\sigma_{iz}=0$ and $\partial_zv_i=0$ assume the additional
conditions $\partial_iv_z=0$ which implies that $v_z$ does not
depend on $x,y$ in the vicinity of the solid boundary (this is
obviously true on the interface provided that the boundary
conditions on $v_z$ are fulfilled). Beyond these two extreme cases
more realistic approaches were proposed to take into account a
possible partial slip at the boundary. In particular
Navier~\cite{Lauga,Loeb,Neto,Charlaix,Tabeling,Navier} suggested
the existence of a surface friction force law
$\sigma_{iz}n_z=\kappa v_i$ where $\vec n=n_z \hat{z}$ is a unit
vector normal to the surface and oriented outwardly from the solid
to the fluid and $\kappa$ a dissipative coefficient. With the same
assumptions as before this leads to
\begin{equation}
n_z\partial_zv_i=\kappa v_i/\eta=v_i/b,\label{slipl}
\end{equation}
where $b$ is the so called slip length. This law has been
considerably studied in the recent years with the advent of micro
and nanofluidics~\cite{Lauga,Neto,Charlaix,Tabeling}. Eq.~\ref{slipl} has been confirmed many times
in particular for liquid flows in nanochannels~\cite{Neto}. However, the
measurements of the associated slip length $b$ reveals a broad
spectrum of values which specifically depend on the system
considered~\cite{Tabeling,Charlaix}. This shows that only a microscopic analysis could lead
to a better understanding of the phenomenon. Keeping this point
for latter discussions we will here apply the law given by
Eq.~\ref{slipl} to our problem and see how it compares with the
experimental results.
\section{The static limit }
In the present analysis we consider the static regime valid for $d
\ll \delta_B:=\sqrt{2\nu/\omega_0}$. We must therefore solve the
system of coupled equations $\eta \vec{\nabla} ^2 \vec v =
\vec{\nabla} p $, $\vec{\nabla}\cdot \vec v=0$ together with the
conditions given by Eq.~\ref{slipl}. The problem is
reminiscent for the one solved by Reynolds in the case of two
parallel disks in a dissipative fluid. Reynolds considered two
disks approaching each other with a constant velocity $\pm U$
along their common axis of symmetry. However, despite geometry
differences, Reynolds considered specifically the case of no-slip
boundary conditions (which were universally accepted at that time)
and not the more general Eq.~\ref{slipl}. Using the same
approximations than Reynolds we here suppose $\partial_x
v_i,\partial_y v_i\ll\partial_z v_i$, $\partial_x v_z,\partial_y
v_z\ll\partial_z v_z$, and $\partial_z p\simeq0$ which are
standard in lubrication theory. We then have
\begin{eqnarray}
\eta\partial^2_zv_i\simeq\partial_i p.
\end{eqnarray}
After integration with respect to $z$ this leads to
\begin{equation}
v_i(x,y,z)=\frac{1}{2\eta}\partial_i
p(x,y)z^2+\alpha_i(x,y)z+\beta_i(x,y).
\end{equation}
To be general we are introducing two \emph{a priori} different
slip lengths $b_0$ and $b_1$ for the interfaces at $z=0$ and $z=d$,
respectively. This hypothesis implies
\begin{eqnarray}
v_i(z=0)=b_0\partial_z v_i(z=0), & v_i(z=d)=-b_1\partial_z v_i(z=d)\nonumber,
\end{eqnarray} and therefore
\begin{equation}
v_i(x,y,z)=\frac{1}{2\eta}\partial_i
p(x,y)[z^2-d\frac{2b_1+d}{b_0+b_1+d}(z+b_0)].
\end{equation}
The fluid incompressibility relation $\partial_x v_x+\partial_y
v_y+\partial_zv_z=0$ and the boundary conditions for $v_z$ at
$z=0$ give the expression: $v_z(x,y,z)=-\int_0^z(\partial_x
v_x+\partial_y v_y) dz$, i.e.,
\begin{eqnarray}
v_z=\frac{1}{2\eta}(\partial_x^2+\partial_y^2)p(x,y)[d\frac{2b_1+d}{b_0+b_1+d}(\frac{z^2}{2}+b_0z)-\frac{z^3}{3}].\nonumber\\
\end{eqnarray}
At $z=d$ we have the boundary condition $v_z=U$ and we deduce
\begin{eqnarray}
(\partial_x^2+\partial_y^2)p(x,y)=\frac{2\eta U}{[(-\frac{1}{3}+\frac{1}{2}\frac{2b_1+d}{b_0+b_1+d})d^3+\frac{2b_1+d}{b_0+b_1+d}b_0 d^2]}\nonumber\\
\end{eqnarray}
and
\begin{eqnarray}
v_z(x,y,z)=\frac{U[-\frac{z^3}{3}+d\frac{2b_1+d}{b_0+b_1+d}(\frac{z^2}{2}+b_0z)]}{[(-\frac{1}{3}+\frac{1}{2}\frac{2b_1+d}{b_0+b_1+d})d^3+\frac{2b_1+d}{b_0+b_1+d}b_0
d^2]}. \label{eq:vzc}
\end{eqnarray}
In particular, in the limit $b_0=b_1\rightarrow +\infty$ we have
$v_z=Uz/d$ whereas for $b_0=b_1=0$ we obtain
$v_z=6U(-z^3/3+z^2d/2)/d^3$. It should be observed that the
solution for $v_z$ in the perfect slip limit looks like the well
known solution of the Couette problem ~\cite{landau,batchelor} for
the permanent fluid motion between two plates in relative and
uniform motion along the $x$ direction. It is worth noting that
the direction of the fluid motion and the boundary conditions used
are however completely different in these two problems (indeed, in
the Couette problem we assume the no-slip boundary conditions and
neglect $v_z$~\cite{landau,batchelor}).
Here due to the presence of terms with $v_z$ our approach deviates from this standard result.\\
The normal force exerted by the fluid located between $z=0$ and
$z=d$ on each surface element $dxdy$ of the lever is given by
\begin{equation}
dF_z=[-\sigma_{zz}+p]|_{z=d}dxdy=[-2\eta\partial_zv_z|_{z=d}+p(x)]dxdy.
\end{equation}
$\sigma_{zz}=2\eta\partial_zv_z$ is the fluid stress tensor along
the $z$ direction and therefore $-\sigma_{zz}|_{z=d}dxdy$ is the
dissipative part of the resulting force $dF_z$. Using
Eq.~\ref{eq:vzc} we obtain
\begin{eqnarray}
-\sigma_{zz}|_{z=d}=\frac{2\eta
U[d^2-\frac{2b_1+d}{b_0+b_1+d}(d^2+b_0
d)]}{[(-\frac{1}{3}+\frac{1}{2}\frac{2b_1+d}{b_0+b_1+d})d^3+\frac{2b_1+d}{b_0+b_1+d}b_0
d^2]}.
\end{eqnarray}
Interestingly this contribution is independent of $x,y$ and of the
lateral boundary shape $(C)$ associated with the lever and the
substrate. The resulting dissipative force
$-\int_{(S)}\sigma_{zz}|_{z=d}dxdy$ on the plate of surface $S$ is
consequently
\begin{eqnarray}
F_z^{\textrm{dissip.}}=\frac{2\eta
SU[d^2-\frac{2b_1+d}{b_0+b_1+d}(d^2+b_0
d)]}{[(-\frac{1}{3}+\frac{1}{2}\frac{2b_1+d}{b_0+b_1+d})d^3+\frac{2b_1+d}{b_0+b_1+d}b_0
d^2]}.
\end{eqnarray}
The second contribution to the force is associated with the volume
pressure $p(x,y)$ that is solution of the 2D Poisson equation
$(\partial_x^2+\partial_y^2)p(x,y)=A$ where $A$ is a constant (see
Eq.~11). Lateral boundary conditions along $(C)$ are here playing
explicitly a role in the analysis. Let us suppose for example that
the substrate is infinitely extended in the $z=0$ plane and that
the lever is delimited by a rectangular boundary of length $L$ and
width $w$ in the $z=d$ plane. A lengthy Fourier analysis (see
appendix) shows that the pressure field can be expanded as:
\begin{eqnarray}
p(x,y)=p_0+\frac{4AL^2}{\pi^3}\sum_n\frac{1}{n^3}\sin{(\pi
nx/L)}[\alpha_n e^{+\pi ny/L}\nonumber\\ +\beta_n e^{-\pi n
y/L}-1],
\end{eqnarray}
where $n$ are odd integers and $\alpha_n,\beta_n$ are constant
coefficients (see appendix). To obtain this formula the boundary
condition $p=p_0$ ($p_0$ is the atmospheric pressure) was imposed
on the rectangular contour. Similarly, if we consider the 1D
problem with $L$ finite, $w=+\infty$, and with $p$ a function of
$x$ only we deduce after a direct integration
\begin{equation}
p(x)=p_0+\frac{A}{2}(x^2-xL),\label{truc}
\end{equation} where we used the boundary conditions $p(0)=p(L)=p_0$. Equivalently, this
result could also be obtained from the Fourier expansion
$p(x)=p_0-\frac{4AL^2}{\pi^3}\sum_n\frac{1}{n^3}\sin{(\pi nx/L)}$.
Finally, as a last example, we consider the circular plate of
radius $R$. Assuming the radial symmetry this problem leads to the
solution
\begin{equation}
p(r)=p_0+\frac{A}{4}(r^2-R^2),\label{retruc}
\end{equation}
where $r$ is the radial coordinate and $p(R)=p_0$ along the
circular contour of radius $R$. More generally, by using the
uniqueness theorem it is easily shown that the pressure $p(x,y)$
can always be written
\begin{equation}
p(x,y)=p_0+A\cdot g(x,y),
\end{equation}
where $g(x,y)$ is the solution of
$(\partial_x^2+\partial_y^2)g(x,y)=1$ which is univocally
determined by the boundary condition $g(x,y)=0$ along the
contour. The resulting vertical force due to the pressure field on
the plate of surface $S$ is thus given by
\begin{eqnarray}
F_z^{\textrm{pressure}}=\int\int_{(S)} (p(x,y)-p_0)dxdy\nonumber\\=A\int\int_{(S)} g(x,y)dxdy=AS\langle g\rangle,
\end{eqnarray}
where the term $-p_0$ equilibrating the pressure $+p_0$ from
Eq.~19 is due to the force exerted by the fluid on the second side
of the cantilever. Considering the examples quoted previously we
find explicitly
\begin{eqnarray}
F_z^{\textrm{pressure}}=-\frac{8AL^3w}{\pi^4}\sum_n\frac{1}{n^4}[1\nonumber\\-\frac{2L}{\pi
nw}\frac{(e^{+\pi n w/(2L)}-e^{-\pi n w/(2L)})^2}{e^{+\pi n
w/L}-e^{-\pi n w/L}}],
\end{eqnarray}
for the rectangular plate, and
\begin{eqnarray}
F_z^{\textrm{pressure}}=-\frac{8AL^3w}{\pi^4}\sum_n\frac{1}{n^4}=-\frac{AL^3w}{12},
\end{eqnarray}
for the 1D plate (note that $w$ is a finite width in the $y$
direction introduced for reasons of dimensionality and that Eq.~22
is the limit of Eq.~21 for $w\rightarrow+\infty$). Similarly, for
the disk we deduce
\begin{eqnarray}
F_z^{\textrm{pressure}}=-\frac{\pi AR^4}{8}.
\end{eqnarray}
These forces due to pressure can be compared with the dissipative
contribution given by Eq.~15 which can equivalently be expressed
as
\begin{eqnarray}
F_z^{\textrm{dissip.}}=AS[d^2-\frac{2b_1+d}{b_0+b_1+d}(d^2+b_0 d)],
\end{eqnarray}
and leads to the total force $F_z=F_z^{\textrm{pressure}}+F_z^{\textrm{dissip.}}$:
\begin{eqnarray}
F_z=AS[\langle g\rangle+d^2-\frac{2b_1+d}{b_0+b_1+d}(d^2+b_0 d)],
\end{eqnarray}
with
\begin{eqnarray}
A=\frac{2\eta
U}{[(-\frac{1}{3}+\frac{1}{2}\frac{2b_1+d}{b_0+b_1+d})d^3+\frac{2b_1+d}{b_0+b_1+d}b_0
d^2]}.
\end{eqnarray}
In the usual limit of vanishing slip lengths
$F_z^{\textrm{dissip.}}=0$ and we therefore have
\begin{eqnarray}
F_z=\frac{12\eta US\langle g\rangle}{d^3},
\end{eqnarray} which is the generalization of the Reynolds formula for an arbitrary plate shape in the no-slip limit.
Inversely, in the limit $b_0=b_1$ infinite, i.e, perfect slip, we obtain $p(x,y)=p_0$, $F_z^{\textrm{pressure}}=0$ and thus
\begin{eqnarray}
F_z=-\frac{2\eta US}{d},
\end{eqnarray} which is the result used in ref.~\cite{nous} in the perfect slip
limit. It is worth noting that this formula is only dependent on
the surface $S$ of the vibrating plate and not on its exact
geometry. It implies that the result obtained in ref.~\cite{nous}
should be very robust with respect to geometry considerations.
\begin{figure}[hbtp]
\includegraphics[width=8.5cm]{Fig2}\\% Here is how to import EPS art
\caption{Evolution of the beam quality factor $Q$ as a function of the gap $d$ for different slip lengths $b$ and for the lever mechanical properties given in ref.~\cite{nous}. The experimental data are in good agreement only with $b> 500$ $\mu m$ (red curve). }
\end{figure}
We point out that this perfect-slip limit can be directly obtained by
solving the equation $\eta\partial^2_zv_z=0$ with the boundary
conditions for $v_z$. Indeed, the direct integration gives $v_z=U
z/d$. The pressure $p=p_0$ can be deduced
similarly~\cite{remarque}.\\
It is worth noting
that in the derivation of all these formulas the Reynolds
assumption $\partial_zp=0$ implies also
$\eta\partial^2_zv_z=\partial_z p=0$. However, from Eq.~12 we
deduce
\begin{eqnarray}
\eta\partial^2_zv_z=-\frac{A}{2}[z-\frac{2b_1+d}{b_0+b_1+d}\frac{d}{2}].
\end{eqnarray}
In the perfect slip limit (where $A\rightarrow 0$) the
Reynolds condition is automatically fulfilled, i.e., the solution is self-consistent. However, in general
$\eta\partial^2_zv_z\neq0$ and this limits the validity of the previous results. The Reynolds hypothesis can nevertheless be
justified, for all practical purposes, if
$|\partial^2_zv_z|\ll|\partial^2_zv_i|$, that is, if
$|z-\frac{2b_1+d}{b_0+b_1+d}\frac{d}{2}|\ll|\partial_ig|$. Since
$|\partial_ig|\sim L$, where $L$ is a typical lateral dimension of
the lever, the Reynolds assumption is in general justified for
$L\gg d$.\\
To study the influence of the slip length on the dynamic properties of the oscillator considered in ref.~\cite{nous} we will now consider a simplification.
Such a simplified force expression is obtained if the condition
$b_0=b_1=b$ is fulfilled.
Considering for example the rectangular plate (in the limit given
by Eq.~22) we get for the total force
\begin{eqnarray}
F_z=-\eta ULw\frac{L^2+12bd}{d^3+6bd^2}.
\end{eqnarray}
From this follows the damping constant $\gamma:=-F_z/U$. To
compare this result with the experiment we show on Fig.~2 the
evolution of the quality factor $Q=k/(\omega_0\gamma)$ of the
oscillating beam calculated for a stiffness $k=0.0396$ $N/m$ and a
pulsation $\omega_0=2\pi\times 50$ $kHz$ considered in
ref.~\cite{nous} and for different slip lengths $b$. Clearly, the
agreement with the experimental data is very good in the limit $d
\lesssim \delta_B$ (i.e, far from the observed saturation at large
gaps) for giant slip lengths such as $b> 500$ $\mu m$ (red
curve). Oppositely the reported results are in total conflict with
the standard no-slip prediction $Q\sim d^3$ (blue curve). While
the present study focussed on the static regime, it is interesting
to remark from dimensional analysis that one should expect to
observe saturation around $\delta_B\simeq 10$ $\mu$m (i.e.,
$\gamma_{\textrm{lim}}=2\eta S/\delta_B$). The value observed
experimentally corresponds to $\delta_B\simeq 25$ $\mu$m, which is
of the same order of magnitude but nevertheless significantly
larger. The difference could be imputed to geometry
considerations, i.e., to the fact that the dynamics of the fluid
around the lever should be strongly influenced by the finite size
of the system under study (indeed, in the regime where the
boundary layer $\delta_B$ plays explicitly a role, retardation
should be taken into account). A different explanation could be
that the model of perfect slip breaks down at large gap. A
preliminary analysis in that direction shows that if we conserve
the inertial term in the Navier-Stokes equations we indeed obtain
a saturation regime due to an additional damping occurring on the
length scale $b\sim \delta_B$ and this even if $b$ is infinite in the static
regime.
\section{Discussion}
Fundamentally, the existence of a giant slip length regime is very
surprising and interesting and should therefore be discussed
carefully. Here, we will only review some results which, we think,
are important to justify microscopically the results
discussed in this article and in ref.~\cite{nous}. We remind that from a microscopic point of view, the slippage
coefficient is actually linked to the very nature of the
interaction between the oscillating surface and the air molecules.
Historically, the first theoretical analysis of this phenomenon
goes back to J.~C.~Maxwell and to its kinetic theory of
gases~\cite{Maxwell}. Following this approach one can indeed
distinguish between a specular channel of interaction, for which
the molecules are colliding elastically with the surface, and a
channel of interaction for which molecules are reflected
diffusively from the
wall~\cite{Maxwell,Loeb,Fichman,Arya,Guo,Sharipova,Lauga}. This
second channel is linked to multiple collisions between molecules
and also to adsorption by the surface. The slip length $b$ in this
statistical model is given by the Maxwell
formula~\cite{Maxwell,Loeb,Fichman,Arya,Guo,Sharipova}:
\begin{equation}
b\simeq\frac{2}{3}\bar{\lambda}\frac{2-p_d}{p_d},
\end{equation}
where $\bar{\lambda}$ is the typical mean free path of gas
molecules (i.e., $\bar{\lambda}\simeq 60$ nm for air in ambient
conditions) and $p_d$ the tangential momentum accommodation
coefficient, i.e., the fraction of those molecules hitting the
surface which are reflected diffusively. Clearly, if $p_d$
vanishes then the slip length is infinite. This suggests that in
the working regime of our mechanical oscillator the molecules are
mainly reflected specularly. Furthermore, recent analysis based on
the fluctuation dissipation theorem and the Green-Kubo
relationship emphasize the importance of several other
microscopical parameters on the molecular dynamics close to a
surface~\cite{Lauga,Charlaix,Neto}. Altogether these studies
provide an estimate for the slip length given by:
\begin{eqnarray}
b\sim\frac{ k_B T\eta D}{C_{\bot}\rho\sigma\epsilon^2},
\end{eqnarray}
where $D$ is the fluid diffusion coefficient, $\sigma,\epsilon$
are respectively a typical length and energy characterizing the
molecular interaction, and $C_{\bot}$ a coefficient measuring the
roughness (larger $C_{\bot}$ mean larger roughness). This shows
the importance of surface roughness but also of surface defects
and nano-structuration~\cite{Charlaix} for the physics of slippage
at solid-gas interfaces. Future work should investigate the
effects of theses parameters on the damping coefficient $\gamma$.
It is worth noting that past studies on the slippage at a
solid-fluid interface mainly focussed on liquids for which the
mean free path is much smaller ($\bar{\lambda}\sim$ 1 nm) and for
which interactions between molecules are much stronger. The
existence of a partial slip regime implies in those cases to work
in the realm of nanofluidics (e.g., nanochannels) with separating
gap $d$ well below the micrometer range~\cite{Neto,Tabeling} or
with AFMs in contact mode~\cite{Neto,Aime}.
Here oppositely, we consider gases and we can define a Knudsen
coefficient $K_n=\bar{\lambda}/d\sim 0.001-0.06$ which corresponds
to a regime of transition flow \cite{Tabeling,Charlaix,Lauga}
occurring at large gap values $d$ (it is indeed well known that
important deviations to the no slip boundary conditions appear for
$K_n\sim 10^{-3}-10^{-2}$~\cite{Tabeling}).\\
Another relevant length in the analysis is the vibration amplitude
of the lever which we reported in ref.~\cite{nous} to be $\delta z\simeq 0.05$ nm. This is actually a very small amplitude
which is attainable experimentally mainly because of the thermal excitation mode and high sensitivity of the optical detection setup used in \cite{nous}. This value for $\delta z$ is also comparatively smaller that those attainable with actuated AFM~\cite{Maali}. Working in this regime where $\delta z/\bar{\lambda}\simeq10^{-3}\ll 1$ could therefore lead to new physics and we expect that further studies in this direction will be done in a close future (e.g., to compare the effect of the vibration amplitude and of the excitation modes on the micro- and nano-lever dynamics).\\
In this context it is worth mentioning that the theoretical model developed in this article for the plane-plane configuration constitutes the equivalent of the Vinogradova formula obtained for the sphere-plane configuration~\cite{Vinogradova,Vinogradova2,Maali,Charlaix,Lauga}. The analysis of Vinogradova generalizes the result obtained by Taylor~\cite{Neto} in the no slip limit and which predicts a friction force $F_z=-6\pi\eta R^2 U/d$ for a sphere of radius R moving along the $z$ axis with the velocity $U$ perpendicularly to the interface $z=0$ and separated from this plane by a minimal gap $d$. It is interesting to point out that the no-slip condition predicts the same law $F_z\propto 1/d$ as in the plane-plane configuration with the perfect slip condition (i.e., Eq.~28) but with a different numerical value for $\gamma_{\textrm{Taylor}}:=6\pi\eta R^2/d$ (compare Eq.~1). The Vinogradova model predicts oppositely $F_z\simeq 2\pi\eta R^2U\ln{(6b/d)}/b$ in the limit $b\rightarrow + \infty$ of the perfect slip.
\begin{table}
\begin{tabular}{l||c|c}
& plane-plane & sphere-plane\\
\hline
\hline
$b=0$ & $\gamma=-\eta\langle g\rangle S/d^3$ & $\gamma=6\pi\eta R^2 /d$\\
\hline
$b\rightarrow +\infty$ & $\gamma=2\eta S/d$ & $\gamma=2\pi\eta R^2\ln{(d/(6b))}/b$\\
\end{tabular}
\caption{Table summarizing the asymptotic viscous force regimes for both the plane-plane and sphere-plane configuration and for the no-slip and perfect slip cases. }
\end{table}
The different regimes of force depending of the value for the slippage length $b$ and from the geometry considered are summarized in Tab.~1.\\
The present problem reminds us a very known similar difficulty encountered in experiments for measuring the Casimir force in the sphere-plane or plane-plane configuration. It is worth pointing out however, that in the Casimir effect~\cite{Genet} the force in the plane-plane configuration varies as $F_z\sim \hbar S/d^4$ whereas it varies as $F_z\sim \hbar R/d^3$ for the sphere-plane configuration ($\hbar$ is the planck constant). Beside the important difference in the power law behavior in $d$ between these expressions and those predicted by hydrodynamics it is interesting to observe that in the perfect slip limit the viscous force decays slowly when $d$ increases whereas the same force vanishes in the sphere-plane as predicted by the Vinogradova formula for $b=+\infty$. Therefore, in order to observe the perfect slip regime the plane-sphere configuration would be much more demanding than the plane-plane configuration studied in this work and in \cite{nous} since it implies that one should consider very small gaps $d$ to obtain a finite effect with the sphere. This could have implication in optical near-field microscopy where viscosity is suggested as a possible mechanism to justify the shear force applied on the tip probes~\cite{Karrai} (see also \cite{Brun1} and \cite{Brun2} for experimental demonstrations in high vacuum at cryogenic temperatures). Additionally, this sensitivity of the force behavior with $b$ in the sphere-plane configuration could be used to probe more precisely the value of the slippage length than in the plane-plane geometry. Oppositely, the giant slip effect studied in this paper could thus be a specificity of the planar geometry and is expected therefore to have a huge impact on the NEMS dynamics which are mostly developed with such geometry.
\section{Implications for NEMS architectures and dynamics}
This brings us to the second point that we shall now
discuss in this section, that is the implication of our results
for NEMS and MEMS engineering. Owing to the formula given by Eq.~1
the effect of dissipation on the beam motion decays very slowly
with $d$ and in particular we see that the quality factor
$Q:=m\omega_0/\gamma$ decreases linearly with $d$.
\begin{figure}[hbtp]
\includegraphics[width=8.2cm]{Fig3}\\% Here is how to import EPS art
\caption{\label{fig:Dcrit2} Evolution of critical gap as a
function of frequency $f_0=\omega_0/(2\pi)$ for a Si (black line),
SiC (red line), GaAs (green line), and Au (blue line) cantilever,
respectively. The black dashed line corresponds to the limit
associated with the boundary layer. Experimental data points for
the lever studied in ref.~\cite{nous} at the internal resonance
frequency 50 kHz are also shown for distance $d$ close to the
overdamping regime (black crosses). Red crosses and circles
correspond to characteristics of NEMS realized with a distance to
the substrate constantly equal to $d=400 nm$. Such NEMS can
approximately be modeled as the one shown in Fig.~1.}
\end{figure}
Additionally, when $d$ decreases the pulsation at resonance
\begin{equation}
\omega_{\textrm{reson.}}=\sqrt{\omega^2_0-\frac{1}{2}\left(\frac{\gamma}{m}\right)^2}
\end{equation}
is progressively down shifted~\cite{nous}. This occurs until
$\omega_{\textrm{reson.}}=0$, i.e. from Eqs.~1,33 when the critical
distance $d_c$ given by
\begin{equation}
d_c= \frac{\sqrt{2}\eta Lw}{m\omega_0},
\end{equation} is reached. For $d \leq d_c$ the lever motion
is consequently overdamped (nonlinear effect are also expected in this limit ). To describe quantitatively the
importance of such a regime on beam dynamics we remind that for the
levers considered here we have:
$\omega_0=\sqrt{\frac{E}{12\rho}}r_0^2\cdot \frac{t}{L^2}$ and
$m=\rho t L w$ where $\rho$ is the bulk density, $E$ is the
Young's modulus, and $r_0\simeq 1.875104$~\cite{Cleland}.\\
Fig.~3 shows the variations of $d_c$ as a function of
$f_0=\omega_0/(2\pi)$ for a thin lever with $t=180$ nm (same as in
ref.~\cite{nous}) and for different commonly used materials. For
low frequencies in the 10 kHz range and below, the overdamping
regime appears already at large separation distance $d_c\geq$
0.1-1 $\mu$m. Oppositely, for very high oscillator frequencies in
the 100 kHz range and beyond we have $d_c \lesssim 10-100$ nm and
the overdamping regime becomes a fundamental issue only at the
nanoscale. For comparison we show on the same graph the
experimental data points taken from ref.~\cite{nous} and
corresponding to working distances $d$ which are decreasing until
the overdamped regime at $d_c$ is reached. Additionally, we show
also the physical characteristics recorded (i.e., internal
resonance frequency $f_0$, and fixed distance gap $d=400$ nm with
the substrate) of typical Si made NEMS. These NEMS can be with a
good approximation described with the simple geometry considered
here. Clearly, working with such NEMS in a gaseous environment may
strongly affect their dynamics and only for very high $f_0$ could
the overdamping regime actually be overcome. This is indeed
confirmed for those NEMS annexed by a red cross in Fig.~3 which we
studied experimentally by using the same optical method as
described in ref.\cite{nous}. The experiment showed that NEMS with
such gaps do not resonate in air at room temperature confirming
therefore the role played by overdamping. For the NEMS indicated
by a red circle, \emph{i.e.}, with frequency in the 100 kHz and
MHz range we were out of the detection sensitivity of our setup
and no data were available.\\
\begin{figure}[hbtp]
\includegraphics[width=1\columnwidth]{Fig4}\\% Here is how to import EPS art
\caption{\label{fig:Dcrit2} Evolution of the critical gap as a
function of the ratio $t/L$ for the same lever materials as in
Fig.~2. Colors lines and data points have the same meaning as in
Fig.~2.}
\end{figure}
As a complementary analysis we show in Fig.~4 the explicit
dependence of $d_c$ on the aspect ratio $t/L$ for different
materials. In the typical range of aspect ratio considered the
overdamping regime covers distance gaps $d$ going from the
micrometer range to the nanoscale and, therefore, cannot be
neglected. Again, this fact is confirmed by comparing these graphs
with
available experimental data (see Fig.~4 and compare with Fig.~3).\\
Furthermore, it is also useful to remind once again that for a large gap
$d$ the important length scale is the boundary layer thickness
$\delta_B$ which characterizes the spatial region surrounding the
lever for which viscosity has an impact on the fluid
dynamics~\cite{landau,batchelor}. For $d\geq \delta_B$ the
substratum lies outside this layer and dissipation must
saturate~\cite{landau,batchelor}, as reported in ref.~\cite{nous}.
Comparisons with values for $d_c$ (see Fig.~3) show that the
overdamping regime is always reached for gaps smaller than
$\delta_B$. The overdamping regime appears consequently as
a robust limitation which should affect the design of any NEMS
working in fluids. The results obtained here for a particular
lever geometry are expected to be very general as soon as the beam
geometrical dimensions are larger than $\delta_B$.
However, when dimensions are smaller, boundary effects due to the
finite size of the system should explicitly be taken into account
in the definition of $d_c$.
\section{Conclusion}
In this article we studied the linearized Navier-Stokes equation in the static regime to describe the damping mechanism of oscillating micro plate in air close to a substrate. We considered the influence of the slip length and showed that results reported in ref.~\cite{nous} are only compatible with very large slip length in the range $b> 500\mu$m. We discussed the implication of this mechanism on the oscillation properties of NEMS and showed that an overdamping behavior represents a fundamental mechanism for such systems. We expect that this work could have important consequences for NEMS engineering.
\begin{acknowledgments}
This research was partly supported by a ``Carnot-NEMS''
collaborative grant between CEA-LETI and Institut N\'eel.
\end{acknowledgments}
|
\section{Introduction}
The study of pairing phenomena in nuclear physics has been an important topic for many years
and different theoretical approaches have been developed. On one side, the seniority scheme, developed by Racah and collaborators \cite{ra,rt}, allows for the study of pairing in the shell-model framework. Modern developments of this approach led to the so-called exact pairing formalism (see, for example, \cite{zv} and references therein). On the other side, the Bardeen--Cooper--Schrieffer (BCS) theory of superconductivity \cite{bcs1,bcs2}, originally developed for large, macroscopic systems, led Bohr, Mottelson and Pines \cite{bmp} to point out the possible usefulness of the the BCS ideas for describing pairing phenomena in the finite nuclear systems. Further work by Belayev along this line showed that pairing correlations affect practically all low-energy properties of nuclei \cite{be}. Since the early approaches, a great amount of work has been done on the effects of pairing in nuclei (see, e. g. \cite{bb} and references therein), however there are problems which deserve further attention, especially when finite-size effects become important. One of these topics, which is of current interest both in the physics of nuclei \cite{vv} and of other mesoscopic systems \cite{cks}, is the issue of possible collective phenomena associated with the pairing interaction. Here we address this problem by using a semiclassical approximation for the equations of motion, but taking into due account the finite size of the many-body system.
Recently, Urban and Schuck \cite{us} have used a semiclassical approach to study the dynamics of trapped systems made of atomic fermions. These mesoscopic systems can contain a very large number of particles and can be viewed as laying between the macroscopic systems described by the BCS theory and the "small" nuclear systems. The semiclassical approximation of \cite{us} leads to simplified equations, thus allowing for the solution of problems that would become unmanageable in a fully quantum approach for a large number of constituents.
The semiclassical approximation is valid for small or mesoscopic systems with a characteristic size $R$, provided $p_F R/\hbar>>1$, where $p_F$ is the Fermi momentum. This is equivalent to the condition that $\hbar\omega_0<<\epsilon_F$, where $\omega_0$ is a frequency characteristic of the classical orbits and $\epsilon_F$ the Fermi energy. We consider systems which, in the absence of pairing, are described by the hamiltonian $h_0=p^2/2m+V_0({\bf r})$. Except for special cases, the classical orbits in the potential $V_0({\bf r})$ are not exactly periodic, they are rather multiply periodic (see, for example, \cite{gol} p.463), leading to eigenfrequencies $\omega_{\bf n}=\sum_\alpha n_\alpha\omega_\alpha$ , which are harmonics of the fundamental frequencies $\omega_\alpha\sim\omega_0$ \cite{bdd}. In nuclei, these eigenfrequencies are related to the frequencies of giant resonances .
Since in heavy nuclei the condition $p_F R/\hbar>>1$ is well satisfied,
in \cite{abdm} we have used an approach similar to that of Ref. \cite{us} to study the linear response of (heavy) nuclei, with the aim of developing a simplified tool for the study of pairing effects in low-energy nuclear excitations. In that paper a drastic approximation has been made: the pairing field $\Delta({\bf r},{\bf p},t)$ has been approximated with a constant phenomenological parameter $\Delta$ (both in the statics and dynamics).
It is well known that such an approximation violates the continuity equation and consequently introduces spurious contributions into the density strength function and its energy-weighted sum rule (EWSR). This problem was solved in \cite{abdm} by means of an appropriate prescription which enforces upon the density fluctuations the constraint coming from the continuity equation, however a prescription is not very satisfactory from a theoretical point of view and here we want to improve on that approach by introducing better kinetic equations that do not need any prescription in order to satisfy the continuity equation and the EWSR. The resulting equations allow also for the study of collective effects associated with the pairing interaction.
Both here and in Ref. \cite{abdm}, pairing is included in semiclassical equations of motion as in Ref. \cite{us}. We focus on weak pairing where the pairing gap $\Delta<<\hbar \omega_0$. This paper is concerned with the frequency response of nuclei and in our formalism the pairing gap is a frequency. In the limit of small $\hbar$ the gap may be small, while the frequency $\Delta/\hbar$ remains finite. Mesoscopic systems with $\Delta>>\hbar\omega_0$ can also be studied with the formalism of this paper, but they are not discussed here.
While in \cite{abdm} the (complex) time-dependent Wigner-transformed pairing field $\Delta({\bf r},{\bf p},t)$ was approximated with a real, constant, phenomenological parameter $\Delta$, here we study the improved approximation in which, for small density fluctuations
\begin{equation}
\Delta({\bf r},{\bf p},t)=\Delta_0({\bf r},{\bf p})+\delta\Delta({\bf r},{\bf p},t)\approx\Delta+\delta\Delta^r({\bf r},t)+i\delta\Delta^i({\bf r},t)\,.
\end{equation}
Thus, in the present approximation, the pairing field is allowed to oscillate and to become complex, the possible momentum dependence of the complex fluctuations is neglected though, in order to simplify the theory. The static pairing field $\Delta_0({\bf r},{\bf p})$ is approximated with the same phenomenological constant used in \cite{abdm}, hence our approach is not fully self-consistent, however, the pairing-field fluctuations $\delta\Delta^r({\bf r},t)$ and $\delta\Delta^i({\bf r},t)$ are derived from self-consistent relations. We study the new set of equations of motion that arise in this improved approximation and determine the effects of the additional terms on the density response function of nuclei.
\section{Basic equations and approximations}
\subsection{Equilibrium properties}
We assume that our system is saturated both in spin and isospin space, so we do not need to introduce explicitly the spin and isospin variables and the physics can be discussed, semiclassically, in terms of the two
equilibrium phase-space distributions
$\rho_0({\bf r},{\bf p})$, and $\kappa_0({\bf r},{\bf p})$ which, according to Ref. \cite{rs}, p.550, are given by\footnote{note the opposite sign of our function $\kappa_0$, compared to the function $\kappa$ in Refs. \cite{us} and \cite{rs}}
\begin{eqnarray}
\label{rhonot}
\rho_0({\bf r},{\bf p})&=&{\frac{1}{2}} {\Big(}1-\frac{h_0({\bf r},{\bf p})-\mu}{E({\bf r},{\bf p})}{\Big)}\,,\\
\label{knot}
\kappa_0({\bf r},{\bf p})&=&-\frac{1}{2}\frac{\Delta_0({\bf r},{\bf p})}{E({\bf r},{\bf p})}\,,
\end{eqnarray}
with the quasiparticle energy
\begin{equation}
\label{qpe}
E({\bf r},{\bf p})=\sqrt{\Delta^2_0({\bf r},{\bf p})+(h_0({\bf r},{\bf p})-\mu)^2}\,.
\end{equation}
The chemical potential $\mu$ is determined by the number of nucleons $A$ through the relation
\begin{equation}
\label{anorm}
A=\frac{4}{(2\pi\hbar)^3}\int d{\bf r} d{\bf p}\rho_0({\bf r},{\bf p})\,.
\end{equation}
The equilibrium hamiltonian
\begin{equation}
\label{hnot}
h_0({\bf r},{\bf p})=\frac{p^2}{2m}+V_0({\bf r})
\end{equation}
contains the (Hartree) mean field $V_0({\bf r})$, which should be evaluated self-consistently, however in the following we use a phenomenological potential well instead. Like in \cite{abdm}, we approximate the
static nuclear mean field with a spherical square-well potential of radius $R$, this choice allows us to take into account finite-size effects and, at the same time, to recover the simplicity of homogeneous systems.
In the following, we also approximate the equilibrium pairing field $\Delta_0({\bf r},{\bf p})$ with the phenomenological parameter $\Delta$, which, in heavy nuclei takes values between 1 and 1.5 MeV (\cite{bm}, p. 170). As a consequence of this approximation, both the static and dynamic equations are considerably simplified. The static distributions, in particular, become a function of the particle energy $\epsilon=h_0({\bf r},{\bf p})$ alone, like in an infinite homogeneous system:
\begin{eqnarray}
&&\rho_0(\epsilon)=\frac{1}{2}{\Big (}1-\frac{\epsilon-\mu}{E(\epsilon)}{\Big )}\,,\\
\label{kzeroeps}
&&\kappa_0(\epsilon)=-\frac{\Delta}{2E(\epsilon)}\,,\\
&&E(\epsilon)=\sqrt{\Delta^2+(\epsilon-\mu)^2}\,,\\
\label{v2}
&&\rho_0'(\epsilon)=\frac{d\rho_0(\epsilon)}{d\epsilon}=-\frac{1}{2}\frac{\Delta^2}{E^3(\epsilon)}\,,\\
\label{kp}
&&\kappa_0'(\epsilon)=\frac{d\kappa_0(\epsilon)}{d\epsilon}=\frac{\Delta(\epsilon-\mu)}{2E^3(\epsilon)}\,.
\end{eqnarray}
\subsection{Dynamical equations for small amplitudes}
The dynamical equations of motion are \cite{us,abdm}
\begin{eqnarray}
i\hbar\frac{\partial\rho}{\partial t}&=&i\hbar\{h,\rho\}-2i{\rm Im}[\Delta^*({\bf r},{\bf p},t)\kappa]+i\hbar{\rm Re}\{\Delta^*({\bf r},{\bf p},t),\kappa\}\,,\\
i\hbar\frac{\partial\kappa}{\partial t}&=&2(h-\mu)\kappa-\Delta({\bf r},{\bf p},t)(2\rho_{ev}-1)+i\hbar\{\Delta({\bf r},{\bf p},t),\rho_{od}\}\,,
\end{eqnarray}
with
\begin{eqnarray}
\rho_{ev}&=&{\frac{1}{2}}[\rho({\bf r},{\bf p},t)+\rho({\bf r},-{\bf p},t)]\,\\
\rho_{od}&=&{\frac{1}{2}}[\rho({\bf r},{\bf p},t)-\rho({\bf r},-{\bf p},t)]\,.
\end{eqnarray}
The equations of motion are gauge-invariant in a self-consistent theory, because multiplying both $\kappa({\bf r},{\bf p},t)$ and $\Delta({\bf r},{\bf p},t)$ by a common phase factor $\exp(i\chi)$ leaves the equations unchanged. Here we make a particular choice of gauge
where the equilibrium quantities $\kappa_0$ and $\Delta_0$ are both real. In the small-amplitude limit, we write the equations of motion for $\rho$ and $\kappa$ in terms of their deviations from equilibrium. To first order, $\delta\kappa^r$gives the change in magnitude of $\kappa$ and $\delta\kappa^i\approx\kappa_0\chi$ is proportional to the change $\chi$ in the phase of $\kappa$.
Thus, we assume that, at time $t=0$, the system is perturbed by a weak external driving field of the kind
\begin{equation}
\label{eff}
\delta h({\bf r},t)=\eta\delta(t) Q({\bf r})\,,
\label{vext}
\end{equation}
where $\delta(t)$ is a Dirac $\delta$-function in time and $\eta$ is a parameter specifying the strength of the external field, then at time $t>0$, we have to deal with the quantities
\begin{eqnarray}
\label{roev}
\rho^{ev}({\bf r},{\bf p},t)&=&\rho_0(\epsilon)+\delta \rho^{ev}({\bf r},{\bf p},t)\,,\\
\rho^{od}({\bf r},{\bf p},t)&=&\delta\rho^{od}({\bf r},{\bf p},t)\,,\\
\kappa({\bf r},{\bf p},t)&=&\kappa_0(\epsilon)+\delta\kappa^r({\bf r},{\bf p},t)+i\delta\kappa^i({\bf r},{\bf p},t)\,,\\
\label{delta}
\Delta({\bf r},t)&=&\Delta+\delta\Delta^r({\bf r},t)+i\delta\Delta^i({\bf r},t)\,.
\end{eqnarray}
In the present improved approximation, the dynamic equations derived in \cite{abdm}
become (cf. Eqs. (43--46) of \cite{abdm})
\begin{eqnarray}
\label{even3}
&&i\hbar\partial_t\delta\rho^{ev}=i\hbar\{h_0,\delta\rho^{od}\}-2i\Delta[\delta\kappa^i+\frac{\delta\Delta^i}{2E(\epsilon)}]\,\\
\label{odd3}
&&i\hbar\partial_t\delta\rho^{od}=i\hbar\{h_0,\delta\rho^{ev}\}+i\hbar\{\delta h,\rho_0\}{+i\hbar\{\delta\Delta^r,\kappa_0\}}\,,\\
\label{fast3}
&&-\hbar\partial_t\delta\kappa^i=2(\epsilon-\mu)[\delta\kappa^r+\frac{\delta\Delta^r}{2E(\epsilon)}]+2\kappa_0\delta h-2\Delta\delta\rho^{ev}\,,\\
\label{fast4}
&&\hbar\partial_t\delta\kappa^r=2(\epsilon-\mu)[\delta\kappa^i+\frac{\delta\Delta^i}{2E(\epsilon)}]\,.
\end{eqnarray}
In Ref. \cite{abdm} the fourth equation of motion (EOM) has been replaced by a supplementary condition enforced by the Pauli principle (cf. Eq. (54) of \cite{abdm}). This was done in order to simplify the formalism. Here instead, we use the four EOM (\ref{even3}--\ref{fast4}) as our staring point and will show that the two methods give similar results.
In Eqs. (\ref{even3}--\ref{fast4}) terms of order $\hbar^2$ or higher, have been neglected.
\subsection{Self-consistency and continuity equation}
Compared to \cite{abdm}, the present approach contains the two extra unknown functions $\delta\Delta^r({\bf r},t)$ and $\delta\Delta^i({\bf r},t)$, hence we need two additional equations in order to determine these quantities.
In a self-consistent theory, the changes in the pairing field are related to the changes in the anomalous density. There are many possible choices for the self-consistency relation. Two minimum requirements are that the total particle number should be conserved and that the particle density and current density should satisfy a continuity equation. A choice that satisfies both requirements is obtained from the self-consistency relation (gap equation) written in the form (cf. Eq. (11) of \cite{us}, see also \cite{pb}):
\begin{equation}
\label{gap}
g\int d{\bf p} {\Big (}\frac{\kappa({\bf r},{\bf p},t)}{\Delta({\bf r},t)}+\frac{1}{p^2/m}{\Big)}=1\,.
\end{equation}
Here $g$ is a parameter determining the strength of the pairing interaction. We have assumed that the ${\bf p}$-dependence of the dynamic pairing field can be neglected.
By differentiating Eq. (\ref{gap}), we get the first-order relation
\begin{equation}
\label{var}
\int d{\bf p} {\Big (}\delta\kappa({\bf r},{\bf p},t)-\kappa_0({\bf r},{\bf p})\frac{\delta\Delta({\bf r},t)}{\Delta}{\Big)}=0\,,
\end{equation}
where $\kappa_0({\bf r},{\bf p})$ and $\Delta$ are real equilibrium quantities, while $\delta\kappa$ and
$\delta\Delta$ are their complex fluctuations.
The real and imaginary parts of Eq. (\ref{var}) give the two independent relations:
\begin{eqnarray}
\label{sc1}
\int d{\bf p} {\Big (}\delta\kappa^r({\bf r},{\bf p},t)-\kappa_0({\bf r},{\bf p})\frac{\delta\Delta^r({\bf r},t)}{\Delta}{\Big)}&=&0\,,\\
\label{sc2}
\int d{\bf p} {\Big (}\delta\kappa^i({\bf r},{\bf p},t)-\kappa_0({\bf r},{\bf p})\frac{\delta\Delta^i({\bf r},t)}{\Delta}{\Big)}&=&0\,.
\end{eqnarray}
These conditions, based on Eq. (\ref{gap}), take into account the residual pairing interaction in a self-consistent way.
Note that, since the conditions (\ref{sc1}, \ref{sc2}) do not depend on $g$, they are valid both for weak and strong pairing. For our purpose, we do not need further information about the pairing interaction.
Using Eq. (\ref{kzeroeps}) gives
\begin{eqnarray}
\label{sc11}
\int d{\bf p} {\Big (}\delta\kappa^r({\bf r},{\bf p},t)+\frac{\delta\Delta^r({\bf r},t)}{2E(\epsilon)}{\Big)}&=&0\,,\\
\label{sc22}
\int d{\bf p} {\Big (}\delta\kappa^i({\bf r},{\bf p},t)+\frac{\delta\Delta^i({\bf r},t)}{2E(\epsilon)}{\Big)}&=&0\,.
\end{eqnarray}
Self-consistency conditions similar to these have been used by the authors of Ref. \cite{cks} in their quantum calculations for infinite homogeneous systems.
It is important to check that, even with the approximate form (\ref{sc11}, \ref{sc22}) of the self-consistency relations, the continuity equation is still satisfied.
Integrating Eq. (\ref{even3}) over ${\bf p}$ gives
\begin{equation}
i\hbar\;\;\partial_t\int d{\bf p}\delta\rho^{ev}({\bf r},{\bf p},t)=i\hbar\int d{\bf p}\{h_0,\delta\rho^{od}\}-2i\Delta\int d{\bf p}[\delta\kappa^i+\frac{\delta\Delta^i}{2E(\epsilon)}]\,.
\end{equation}
The last integral vanishes because of Eq. (\ref{sc22}), giving \begin{equation}
\label{almost}
\partial_t\delta\varrho({\bf r},t)=-\partial_{{\bf r}}\cdot{\bf j}({\bf r},t)+\frac{4}{(2\pi\hbar)^3}\int d{\bf p}\partial_{{\bf r}}h_0\cdot\partial_{{\bf p}}\delta\rho^{od}\,,
\end{equation}
with the density fluctuation
\begin{equation}
\label{denf}
\delta\varrho({\bf r},t)=\frac{4}{(2\pi\hbar)^3}\int d{\bf p}\delta\rho^{ev}({\bf r},{\bf p},t)
\end{equation}
and the current density
\begin{equation}
{\bf j}({\bf r},t)=\frac{4}{(2\pi\hbar)^3}\int d{\bf p} \frac{{\bf p}}{m}\delta\rho^{od}({\bf r},{\bf p},t)\,.
\end{equation}
For a Hamiltonian of the kind (\ref{hnot}), the integral in Eq. (\ref{almost}) can be shown to vanish and the continuity equation is satisfied. This is a crucial difference between the present approach and that of \cite{abdm}.
\section{Fourier expansions}
First of all, we take the Fourier transform in time of the various time-dependent fluctuations in Eqs. (\ref{roev}--\ref{delta}). Since these quantities are nonvanishing only for $t>0$, we suppose that $\omega$ has a vanishingly small positive imaginary part $i\varepsilon$, to insure convergence of the integrals
$f(\omega)=\int_{-\infty}^\infty dt e^{i\omega t}f(t)$. In the following, the imaginary part of $\omega$ will not be written explicitly.
Then, following \cite{abdm}, we introduce the following Fourier expansions based on the method of action-angle variables (${\bf \Phi}$ are the angle variables, while ${\bf I}$ are the action variables):
\begin{eqnarray}
\delta h({\bf r},\omega)&=&\sum_{\bf n}\delta h_{\bf n}({\bf I},\omega)e^{i{\bf n}\cdot {\bf \Phi}}\,,\\
\delta\rho({\bf r},\pm{\bf p},\omega)&=&\sum_{\bf n}\delta \rho^\pm_{\bf n}({\bf I},\omega)e^{i{\bf n}\cdot {\bf \Phi}}\,,\\
\delta\rho^{ev}_{\bf n}&=&{\frac{1}{2}}[\delta\rho^+_{\bf n}+\delta\rho^-_{\bf n}]\,\\
\delta\rho^{od}_{\bf n}&=&{\frac{1}{2}}[\delta\rho^+_{\bf n}-\delta\rho^-_{\bf n}]\,\\
\delta\kappa^{r,i}({\bf r},{\bf p},\omega)&=&\sum_{\bf n}\delta\kappa^{r,i}_{\bf n}({\bf I},\omega)e^{i{\bf n}\cdot {\bf \Phi}}\,,\\
\delta\Delta^{r,i}({\bf r},\omega)&=&\sum_{\bf n}\delta\Delta^{r,i}_{\bf n}({\bf I},\omega)e^{i{\bf n}\cdot {\bf \Phi}}\,.
\end{eqnarray}
Moreover we use the relation
\begin{equation}
\{f,h_0\}=\sum_{\bf n}i\omega_{\bf n}f_{\bf n}e^{i{\bf n}\cdot {\bf \Phi}}\,,
\end{equation}
where $\omega_{\bf n}=\sum_\alpha n_\alpha \omega_\alpha$ are the eigenfrequencies of the uncorrelated system \cite{bdd}, to obtain from the dynamic equations (\ref{even3}--\ref{fast4}) the system of algebraic equations
\begin{eqnarray}
\label{coe1}
\hbar\omega\delta\rho_{\bf n}^{ev}&=&\hbar\omega_{\bf n}\delta\rho^{od}_{\bf n}-2i\Delta[\delta\kappa^i_{\bf n}+\frac{\delta\Delta^i_{\bf n}}{2E}]\,,\\
\label{coe2}
\hbar\omega\delta\rho_{\bf n}^{od}&=&\hbar\omega_{\bf n}\delta\rho^{ev}_{\bf n}-\hbar\omega_{\bf n}\rho_0'\delta h_{\bf n}-\hbar\omega_{\bf n}\kappa_0'\delta\Delta^r_{\bf n}\,,\\
\label{coe3}
i\hbar\omega\delta\kappa^i_{\bf n}&=&2(\epsilon-\mu)[\delta\kappa^r_{\bf n}+\frac{\delta\Delta^r_{\bf n}}{2E}]+2\kappa_0\delta h_{\bf n}-2\Delta\delta\rho^{ev}_{\bf n}\,,\\
\label{coe4}
-i\hbar\omega\delta\kappa^r_{\bf n}&=&2(\epsilon-\mu)[\delta\kappa^i_{\bf n}+\frac{\delta\Delta^i_{\bf n}}{2E}]\,.
\end{eqnarray}
From Eqs. (\ref{coe1}, \ref{coe2}) we get \footnote { From now on, we use units $\hbar=c=1$. }
\begin{equation}
(\omega^2-\omega^2_{\bf n})\delta\rho^{ev}_{\bf n}=-\omega^2_{\bf n}\rho_0'\delta h_{\bf n}-\omega_{\bf n}^2\kappa_0'\delta\Delta^r_{\bf n}-
2i\omega\Delta\delta\kappa^i_{\bf n}-\frac{2i\omega\Delta}{\Omega}\delta\Delta^i_{\bf n}\,,
\end{equation}
while Eqs.(\ref{coe3}, \ref{coe4}) give
\begin{equation}
\delta\kappa^i_{\bf n}=\frac{\omega_\mu}{\omega^2-\omega_\mu^2} {\Big (} \omega_\mu\frac{\delta\Delta^i_{\bf n}}{\Omega}-i\omega\frac{\delta\Delta^r_{\bf n}}{\Omega}+i\omega\frac{2\Delta}{\omega_\mu} \delta\rho^{ev}_{\bf n}-i\omega\frac{2\kappa_0}{\omega_\mu}\delta h_{\bf n}{\Big )}\,.
\end{equation}
By combining the last two equations, we obtain
\begin{eqnarray}
\label{delkai}
\delta\kappa^i_{\bf n}&=&A_{\bf n}(\omega)\frac{\delta\Delta^i_{\bf n}}{\Omega}-i\omega\omega_\mu A_{\bf n}(\omega)\frac{\delta\Delta^r_{\bf n}}{\Omega^3}
+2i\omega\Delta A_{\bf n}(\omega)\frac{\delta h_{\bf n}}{\Omega^3}\,,\nonumber\\
\end{eqnarray}
with
\begin{eqnarray}
A_{\bf n}(\omega)&=&\frac{\omega^2}{D_{\bf n}(\omega)}-1\,,\\
D_{\bf n}(\omega)&=&\omega^2-\omega_\mu^2-4\Delta^2\frac{\omega^2}{\omega^2-\omega_{\bf n}^2}=\omega^2-\Omega^2-4\Delta^2\frac{\omega_{\bf n}^2}{\omega^2-\omega_{\bf n}^2}\,,
\end{eqnarray}
and
\begin{eqnarray}
\Omega&=&2E=\sqrt{4\Delta^2+\omega_\mu^2}\,,\\
\omega_\mu&=&2(\epsilon-\mu)\,,\\
\label{ombar}
\bar\omega_{\bf n}^2&=&\omega_{\bf n}^2+\Omega^2\,.
\end{eqnarray}
In a similar way, we get also
\begin{equation}
\delta\kappa^r_{\bf n}=\frac{i\omega\omega_\mu}{\omega^2}[1+A_{\bf n}(\omega)]\frac{\delta\Delta^i_{\bf n}}{\Omega}+\omega_\mu^2A_{\bf n}(\omega)\frac{\delta\Delta^r_{\bf n}}{\Omega^3}-2\Delta\omega_\mu A_{\bf n}(\omega)\frac{\delta h_{\bf n}}{\Omega^3}\,
\end{equation}
and
\begin{eqnarray}
\label{delron}
\delta\rho^{ev}_{\bf n}&=&{\Big (} \frac{\bar\omega_{\bf n}^2(\omega^2-\frac{\omega_{\bf n}^2\omega_\mu^2}{\bar\omega_{\bf n}^2})}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}{\Big )}\frac{4\Delta^2}{\Omega^3}\delta h_{\bf n}\\
&-&{\Big (}\frac{\bar\omega_{\bf n}^2(\omega^2-\frac{\omega_{\bf n}^2\omega_\mu^2}{\bar\omega_{\bf n}^2})}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}{\Big )}\frac{2\Delta\omega_\mu}{\Omega^3}\,\delta\Delta^r_{\bf n}
-{\Big (}\frac{\omega^2}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}{\Big )}\frac{2i\omega\Delta}{\Omega}\delta\Delta^i_{\bf n}\,.\nonumber
\end{eqnarray}
By using the last two equations, it can be easily checked that, for the mode ${\bf n}=(0,0,0)$, the following relation is satisfied:
\begin{equation}
\label{pauli}
\delta\kappa^r_{\bf n}(\omega)=-\frac{\omega_\mu}{2\Delta}\delta\rho^{ev}_{\bf n}(\omega)\,.
\end{equation}
In \cite{abdm} the equation of motion (\ref{coe4}) was replaced by this relation, which is a condition required by the Pauli principle. Thus we see that, for the mode ${\bf n}=0$, the Pauli principle constraint (\ref{pauli}) is authomatically satisfied by the solution of the equations of motion (\ref{coe1}--\ref{coe4}).
The more general case ${\bf n}\neq0$ is not so immediate and it is discussed in the Appendix.
\section{Constant-$\Delta$ approximation}
The first term in Eq. (\ref{delron}) gives the constant-$\Delta$ part of the density response. By using Eq. (\ref{ombar}), it can be written as
\begin{equation}
\label{delrocd}
\delta\rho^{cd}_{\bf n}={\Big (} \frac{\omega_{\bf n}^2(\omega^2-\frac{\omega_{\bf n}^2\omega_\mu^2}{\bar\omega_{\bf n}^2})}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}{\Big )}\frac{4\Delta^2}{\Omega^3}\delta h_{\bf n}+{\Big (} \frac{(\omega^2-\frac{\omega_{\bf n}^2\omega_\mu^2}{\bar\omega_{\bf n}^2})}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}{\Big )}\frac{4\Delta^2}{\Omega}\delta h_{\bf n}\,.
\end{equation}
The first part of this expression contains a factor $1/\Omega^3$ which, as a function of $\epsilon$, has a very narrow peak at the Fermi surface when $\Delta$ is small compared to $\mu$. The second part instead contains a factor $1/\Omega$, which is also peaked at the Fermi surface, but is more spread out. This second term gives an unwanted contribution to the energy-weighted sum rule \cite{abdm}. In the following we will check explicitly that the self-consistent pairing field fluctuations $\delta\Delta^i_{\bf n}$ restore the sum rule to its correct value.
In the constant-$\Delta$ approximation the eigenfrequencies of the system are determined by the poles in Eq. (\ref{delrocd}), that is, by the condition
\begin{equation}
\label{cond}
(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)=\omega^4-\bar\omega^2_{\bf n}\omega^2+\omega_{\bf n}^2\omega_\mu^2=0\,,
\end{equation}
with solutions
\begin{equation}
\label{ompm}
\omega^2_\pm({\bf n},{\bf I})={\frac{1}{2}}\bar\omega^2_{\bf n}({\bf I}){\Big (} 1\pm\sqrt{1-4\frac{\omega^2_{\bf n}({\bf I})\omega_\mu^2(\epsilon)}{\bar\omega^4_{\bf n}({\bf I})}} {\Big )}\,.
\end{equation}
These expressions look rather unusual, however they are closely related to the eigenfrequencies found in \cite{abdm} and also, as will be shown in the following, to the energy of two-quasiparticle excitations.
First of all we notice that, when ${\bf n}=0$, the eigenfrequencies $\omega_+$ coincide with the eigenfrequencies $\bar\omega_{\bf n}$ found in \cite{abdm} since
\begin{equation}
\omega^2_+({\bf n}=0,{\bf I})=\Omega^2(\epsilon)\,,
\end{equation}
while
\begin{equation}
\omega^2_-({\bf n}=0,{\bf I})=0\,.
\end{equation}
Moreover, for the modes ${\bf n}\neq0$, we have that
\begin{eqnarray}
\omega^2_+({\bf n},{\bf I}){\Big |}_{\epsilon=\mu}&=&\bar\omega^2_{\bf n}({\bf I})\,,\\
\omega^2_-({\bf n},{\bf I}){\Big |}_{\epsilon=\mu}&=&0\,.
\end{eqnarray}
As a general rule, the range of particle energies which is most important is that within a distance of a few times $\Delta$ from the Fermi surface $\mu$, hence $\omega_\mu$ is of order $\Delta$ and the dimensionless parameter
\begin{equation}
\alpha_{\bf n}({\bf I})=4\frac{\omega^2_{\bf n}({\bf I})\omega_\mu^2(\epsilon)}{\bar\omega^4_{\bf n}({\bf I})}<<1\,,
\end{equation}
both for small ($\omega_{\bf n}({\bf I})>>\Delta$) and large ($\omega_{\bf n}({\bf I})<<\Delta$) systems. In these cases, by expanding the square root in Eq. (\ref{ompm}), we have that
\begin{eqnarray}
\label{ompa}
\omega^2_+({\bf n},{\bf I}){\Big |}_{\epsilon\approx\mu}&\approx&\bar\omega^2_{\bf n}({\bf I})\,,\\
\label{omma}
\omega^2_-({\bf n},{\bf I}){\Big |}_{\epsilon\approx\mu}&\approx&\frac{\omega^2_{\bf n}({\bf I})\omega_\mu^2(\epsilon)}{\bar\omega^2_{\bf n}({\bf I})}\,.
\end{eqnarray}
This is an excellent approximation in the first term of (\ref{delrocd}) because of the factor $\frac{1}{\Omega^3(\epsilon)}$. It is interesting to note that the factor $(\omega^2-\frac{\omega_{\bf n}^2\omega_\mu^2}{\bar\omega_{\bf n}^2})$ in the numerator of $\delta\rho^{cd}_{\bf n}$ approximately cancels the poles $(\omega^2-\omega_-^2({\bf n},{\bf I}))$. The same cancellation occurs in the coefficient of $\delta\Delta^r_{\bf n}$ in Eq. (\ref{delron}), while it does not occur for the coefficient of $\delta\Delta^i_{\bf n}$.
In conclusion, in the constant-$\Delta$ approximation, the solutions of the equations of motion (\ref{coe1}--\ref{coe4})
have two branches, corresponding to the eigenfrequencies $\omega_+({\bf n},{\bf I})$ and $\omega_-({\bf n},{\bf I})$. These solutions do not satify the constraint (\ref{pauli}) exactly when ${\bf n}\neq0$, but, as shown in the Appendix, one branch does, approximately. Since $\alpha_{\bf n}({\bf I})<<1$, the eigenfrequencies of the two branches are approximated by $\omega^2_+\approx\bar\omega^2_{\bf n}$ and $\omega^2_-\approx\omega^2_{\bf n}\omega^2_\mu/\bar\omega^2_{\bf n}$. The poles corresponding to $\omega_-$ are approximately canceled by the term $(\omega^2-\omega^2_{\bf n}\omega^2_\mu/\bar\omega^2_{\bf n})$in the numerators of Eq. (\ref{delrocd}).
Thus, the constant-$\Delta$ part of the present approach, which is based on the four equations of motion (\ref{coe1}--\ref{coe4}), gives results in good agreement with those of Ref. \cite{abdm} in which the fourth equation of motion was replaced by the constraint (\ref{pauli}).
\subsection{Relation with two-quasiparticle excitations}
\label{tqp}
The eigenfrequencies (\ref{ompm}) are related also to the energy of two-quasiparticle excitations.
In the quantum BCS theory , the simplest excitations of an even nucleus are two-quasiparticle states, with energy (cf., for example, Sect. 3.3 of \cite{bb})
\begin{equation}
{\cal E}(\epsilon_p,\epsilon_h)=E_p+E_h=\sqrt{(\epsilon_p-\mu)^2+\Delta^2}+\sqrt{(\mu-\epsilon_h)^2+\Delta^2}\,.
\end{equation}
If we change variables from $(\epsilon_p,\epsilon_h)$ to $(\epsilon_{ph},\bar\epsilon)$, with
\begin{eqnarray}
\epsilon_{ph}&=&\epsilon_p-\epsilon_h\,,\\
\label{ave}
\bar\epsilon&=&\frac{\epsilon_p+\epsilon_h}{2}\,,
\end{eqnarray}
then
\begin{eqnarray}
&&{\cal E}(\epsilon_{ph},\bar\epsilon)=
\sqrt{[{\frac{1}{2}}\epsilon_{ph}+(\bar\epsilon-\mu)]^2+\Delta^2}+\sqrt{[{\frac{1}{2}}\epsilon_{ph}-(\bar\epsilon-\mu)]^2+\Delta^2}\nonumber\,.
\end{eqnarray}
From the work on normal systems \cite{bdd}, we know that the eigenfrequencies $\omega_{\bf n}$ give a semiclassical approximation to the {\em particle-hole} energies\footnote{In order to simplify the notation, we will write also $\omega_{\bf n}(\epsilon)$, instead of $\omega_{\bf n}({\bf I}){\Big |}_\epsilon$.}
\begin{equation}
\label{norm}
\epsilon_{ph}\approx\omega_{\bf n}(\bar\epsilon)\,,
\end{equation}
then
\begin{eqnarray}
&&{\cal E}(\epsilon_{ph},\bar\epsilon)\approx\\
&&{\frac{1}{2}} {\Big (}\sqrt{\bar\omega_{\bf n}^2(\bar\epsilon)+2\omega_{\bf n}(\bar\epsilon)\omega_\mu(\bar\epsilon)}+\sqrt{\bar\omega_{\bf n}^2(\bar\epsilon)-2\omega_{\bf n}(\bar\epsilon)\omega_\mu(\bar\epsilon)} {\Big )}=\omega_+({\bf n},\bar\epsilon)\,.\nonumber
\end{eqnarray}
The last relation is exact, as can be easily checked by squaring both sides.
Thus, the eigenfrequencies $\omega_+({\bf n},\epsilon)$ correspond to two-quasiparticle excitations. There are two approximations involved: one is the relation (\ref{norm}), and the other is the replacement $\bar\epsilon\to\epsilon$ in the particle energy.
In the same way it is possible to show that the eigenfrequencies $\omega_-({\bf n},\epsilon)$ correspond to the combination $E_p-E_h$. Anderson \cite{pwa} called 'unphysical' the solutions corresponding to this combination of quasiparticle energies and Valatin \cite{jgv} pointed out that these solutions are eliminated by the supplementary condition required by the Pauli principle (see Appendix).
Here we find that the poles corresponding to these 'unphysical'
eigenfrequencies are practically canceled by corresponding zeros in the numerator of the density
fluctuations (\ref{delrocd}).
\section{The self-consistency relations}
The initial system of coupled differential equations has been reduced to a linear algebraic system, which is supplemented by the two integral relations
\begin{eqnarray}
\label{sc111}
\int d{\bf p} {\Big (}\delta\kappa^i({\bf r},{\bf p},\omega)+\frac{\delta\Delta^i({\bf r},\omega)}{2E(\epsilon)}{\Big)}&=&0\,,\\
\label{sc222}
\int d{\bf p} {\Big (}\delta\kappa^r({\bf r},{\bf p},\omega)+\frac{\delta\Delta^r({\bf r},\omega)}{2E(\epsilon)}{\Big)}&=&0\,,
\end{eqnarray}
which are the Fourier transform in time of Eqs. (\ref{sc11}, \ref{sc22}).
Multiplying both sides of these relations by $e^{-i{{\bf n}}\cdot\Phi}$ and integrating over ${\bf r}$, we get
\begin{eqnarray}
\label{pro1}
\int d{\bf r}\int d{\bf p} {\Big (}\delta\kappa^i({\bf r},{\bf p},\omega)+\frac{\delta\Delta^i({\bf r},\omega)}{2E(\epsilon)}{\Big)}e^{-i{{\bf n}}\cdot\Phi}&=&0\,,\\
\label{pro2}
\int d{\bf r}\int d{\bf p} {\Big (}\delta\kappa^r({\bf r},{\bf p},\omega)+\frac{\delta\Delta^r({\bf r},\omega)}{2E(\epsilon)}{\Big)}e^{-i{{\bf n}}\cdot\Phi}&=&0\,.
\end{eqnarray}
Now, changing variables from $({\bf r},{\bf p})$ to $({\bf I}, {\bf \Phi})$ and using the orthogonality of the functions $e^{i{\bf n}\cdot\Phi}$, we have
\begin{eqnarray}
\label{pro11}
\int d{\bf I} {\Big (}\delta\kappa_{{\bf n}}^i({\bf I},\omega)+\frac{\delta\Delta_{{\bf n}}^i({\bf I},\omega)}{2E(\epsilon)}{\Big)}&=&0\,,\\
\label{pro22}
\int d{\bf I} {\Big (}\delta\kappa_{{\bf n}}^r({\bf I},\omega)+\frac{\delta\Delta_{{\bf n}}^r({\bf I},\omega)}{2E(\epsilon)}{\Big)}&=&0\,.
\end{eqnarray}
Finally, by using the four equations (\ref{coe1}-\ref{coe4}),
the two integral relations (\ref{pro11}, \ref{pro22}) can be written in the form
\begin{eqnarray}
\label{sc13}
\int d{\bf I} [a_{11}({\bf n},{\bf I},\omega)\delta\Delta^r_{\bf n}({\bf I},\omega)+ a_{12}({\bf n},{\bf I},\omega)\delta\Delta^i_{\bf n}({\bf I},\omega)]&=&
\int d{\bf I} b_1({\bf n},{\bf I},\omega)\delta h_{\bf n}({\bf I})\,,\qquad\qquad\\
\label{sc23}
\int d{\bf I} [a_{21}({\bf n},{\bf I},\omega)\delta\Delta^r_{\bf n}({\bf I},\omega)+ a_{22}({\bf n},{\bf I},\omega)\delta\Delta^i_{\bf n}({\bf I},\omega)]&=&\int d{\bf I} b_2({\bf n},{\bf I},\omega)\delta h_{\bf n}({\bf I})\,,
\end{eqnarray}
with
\begin{eqnarray}
\label{a11}
a_{11}({\bf n},{\bf I},\omega)&=&-i\omega\frac{\omega_\mu}{\Omega^3}A_{\bf n}(\omega)=-i\omega\frac{\omega_\mu}{\Omega D_{\bf n}(\omega)} {\Big (}1+\frac{4\Delta^2}{\Omega^2}\frac{\omega^2_{\bf n}}{\omega^2-\omega^2_{\bf n}} {\Big )}\,,\\
a_{12}({\bf n},{\bf I},\omega)&=&\frac{1+A_{\bf n}(\omega)}{\Omega}=\frac{\omega^2}{\Omega D_{\bf n}(\omega)}\,,\\
a_{21}({\bf n},{\bf I},\omega)&=&\frac{\omega_\mu^2}{\Omega^3}A_{\bf n}(\omega)+\frac{1}{\Omega}\\
&=&
\frac{1}{\Omega D_{\bf n}(\omega)}{\Big [}\omega^2-4\Delta^2{\Big (}1+\frac{4\Delta^2}{\Omega^2}\frac{\omega^2_{\bf n}}{\omega^2-\omega^2_{\bf n}}{\Big)}{\Big ]}\nonumber\,,\\
a_{22}({\bf n},{\bf I},\omega)&=&i\omega\frac{\omega_\mu}{\omega^2}\frac{1+A_{\bf n}(\omega)}{\Omega}=i\omega\frac{\omega_\mu}{\Omega D_{\bf n}(\omega)}\,,\\
b_{1}({\bf n},{\bf I},\omega)&=&-\frac{2i\omega\Delta A_{\bf n}(\omega)}{\Omega^3}=-\frac{2i\omega\Delta}{\Omega D_{\bf n}(\omega)}{\Big (}1+\frac{4\Delta^2}{\Omega^2}\frac{\omega_{\bf n}^2}{\omega^2-\omega_{\bf n}^2}{\Big )}\,,\\
\label{b2}
b_{2}({\bf n},{\bf I},\omega)&=&\frac{2\Delta\omega_\mu A_{\bf n}(\omega)}{\Omega^3}
=\frac{2\Delta\omega_\mu}{\Omega D_{\bf n}(\omega)}{\Big (}1+\frac{4\Delta^2}{\Omega^2}\frac{\omega_{\bf n}^2}{\omega^2-\omega_{\bf n}^2}{\Big )}\,.
\end{eqnarray}
The external field fluctuations (\ref{eff}) depend on position and time, but not on the momentum ${\bf p}$. For a uniform system the self-consistency relations lead to pairing-field fluctuations $\delta\Delta({\bf r},\omega)$ which are also momentum-independent and Eqs. (\ref{sc13}, \ref{sc23}) become separable, giving an algebraic system. For finite systems, Eqs. (\ref{sc13}, \ref{sc23}) are not separable because the coefficients $\delta h_{\bf n}({\bf I})$ can depend on the action variables ${\bf I}$.
\section{'Ansatz' and solution}
The pairing-field fluctuations are given by the solution of the two coupled integral equations (\ref{sc13}, \ref{sc23}). In the spirit of a simplified approach, here we make an $Ansatz$ that simplifies the problem:
we assume that
\begin{eqnarray}
\delta\Delta^r_{\bf n}({\bf I},\omega)&=&F^r_{\bf n}(\omega)\delta h_{\bf n}({\bf I})\,,\\
\delta\Delta^i_{\bf n}({\bf I},\omega)&=&F^r_{\bf n}(\omega)\delta h_{\bf n}({\bf I})\,,
\end{eqnarray}
with the functions $F^{r,i}_{\bf n}(\omega)$ given by the solution of the algebraic system
\begin{eqnarray}
F^r_{\bf n}(\omega)\int d{\bf I} a_{11}({\bf n},{\bf I},\omega)+F^i_{\bf n}(\omega)\int d{\bf I} a_{12}({\bf n},{\bf I},\omega)&=&\int d{\bf I} b_1({\bf n},{\bf I},\omega)\,,\\
F^r_{\bf n}(\omega)\int d{\bf I} a_{21}({\bf n},{\bf I},\omega)+F^i_{\bf n}(\omega)\int d{\bf I} a_{22}({\bf n},{\bf I},\omega)&=&\int d{\bf I} b_2({\bf n},{\bf I},\omega)\,.
\end{eqnarray}
Then, defining
\begin{eqnarray}
\label{aij}
A_{ij}({\bf n},\omega) &=& \int d {\bf I} a_{ij}({\bf n},{\bf I},\omega)\,,\\
\label{bi}
B_i({\bf n},\omega) &=& \int d {\bf I} b_i ({\bf n},{\bf I},\omega))\,,
\end{eqnarray}
we have
\begin{eqnarray}
\label{ddr}
F^r_{\bf n}(\omega)&=&\frac{B_1A_{22}-B_2A_{12}}{A_{11}A_{22}-A_{21}A_{12}}\,,\\
\label{ddi}
F^i_{\bf n}(\omega)&=&\frac{B_2A_{11}-B_1 A_{21}} {A_{11}A_{22}-A_{21}A_{12}}\,.
\end{eqnarray}
The six integrals $A_{ij}$ and $B_i$ are conveniently expressed in terms of the four basic integrals
\begin{eqnarray}
\label{i1n}
I_1({\bf n},\omega)&=&\int d{\bf I} \frac{\omega_\mu(\epsilon)}{\Omega(\epsilon)}\frac{1}{D_{\bf n}(\omega)}\,,\\
I_2({\bf n},\omega)&=&\int d{\bf I} \frac{1}{\Omega(\epsilon)}\frac{1}{D_{\bf n}(\omega)}\,,\\
I_3({\bf n},\omega)&=&\int d{\bf I} \frac{\omega_\mu(\epsilon)}{\Omega^3(\epsilon)}\frac{1}{D_{\bf n}(\omega)}\frac{\omega_{\bf n}^2}{\omega^2-\omega_{\bf n}^2}\,,\\
\label{i4n}
I_4({\bf n},\omega)&=&\int d{\bf I} \frac{1}{\Omega^3(\epsilon)}\frac{1}{D_{\bf n}(\omega)}\frac{\omega_{\bf n}^2}{\omega^2-\omega_{\bf n}^2}\,
\end{eqnarray}
as
\begin{eqnarray}
A_{11}&=&-i\omega(I_1+4\Delta^2I_3)=-i\omega I_1'\,,\\
A_{12}&=&\omega^2 I_2\,,\\
A_{21}&=&(\omega^2-4\Delta^2)I_2-16\Delta^4I_4=\omega^2I_2-4\Delta^2I_2'\,\\
A_{22}&=&i\omega I_1\,,\\
B_1&=&-i\omega2\Delta (I_2+4\Delta^2I_4)=-i\omega2\Delta I_2'\,,\\
B_2&=&2\Delta(I_1+4\Delta^2I_3)=2\Delta I_1' \,,
\end{eqnarray}
with
\begin{eqnarray}
\label{ip1}
I_1'&=&I_1+4\Delta^2I_3\,,\\
\label{ip2}
I_2'&=&I_2+4\Delta^2I_4\,.
\end{eqnarray}
In terms of these integrals, the solution (\ref{ddr}, \ref{ddi}) reads
\begin{eqnarray}
\label{frn}
F^r_{\bf n}(\omega)&=&2\Delta\,\frac{4\Delta^2(I_1I_4-I_2I_3)}{I_1 I_1'- I_2(\omega^2I_2-4\Delta^2I_2')}\,,\qquad\\
\label{fin}
F^i_{\bf n}(\omega)&=&\frac{2\Delta}{i\omega}\,{\Big [}1+4\Delta^2\,\frac{I_3I_1'-I_4(\omega^2I_2-4\Delta^2I_2')}{I_1I_1'-I_2(\omega^2I_2-4\Delta^2I_2')}{\Big ]}\,.
\end{eqnarray}
With these expressions, the density fluctuations and the density-density response function can be evaluated explicitly. This is what we do in the following, by neglecting that part of the mean-field fluctuations that is not related to pairing, hence the expressions derived in the following correspond to the static mean-field approximation, called 'zero-order' in Ref. \cite{bdd}. Within this approximation,
the fluctuations of the hamiltonian are
\begin{equation}
\delta h_{\bf n} ({\bf I},\omega)=\eta Q_{\bf n}({\bf I})=\frac{\eta}{(2\pi)^3}\int d{\bf \Phi}e^{-i{\bf n}\cdot{\bf\Phi}}Q({\bf r})\,.
\end{equation}
\subsection{The mode ${\bf n}=0$}
The mode ${\bf n}=(0,0,0)$ is particularly interesting because, as shown in \cite{abdm}, this mode is the only one giving a spurious fluctuation of the number of particles in the system. When ${\bf n}=0$, the two integrals $I_3$ and $I_4$ vanish, $I_1'=I_1$, $I_2'=I_2$ and the solution is expressed in terms of the integrals $I_{1,2}$ alone, giving
\begin{eqnarray}
\label{fr0}
F^r_{{\bf n}=0}(\omega)&=&0\,,\\
\label{fi0}
F^i_{{\bf n}=0}(\omega)&=&\frac{2\Delta}{i\omega}\,,
\end{eqnarray}
with the possible exception of points where the denominators in Eqs. (\ref{frn},\ref{fin}) might happen to vanish.
By using the result (\ref{fr0}, \ref{fi0}), it can be easily shown that the spurious contribution to the mode ${\bf n}=0$ in the density fluctuations of the constant-$\Delta$ approximation, is exactly canceled by the fluctuations of the imaginary pairing field, since Eq. (\ref{delron}) gives
\begin{eqnarray}
\delta\rho^{ev}_{{\bf n}=0}&=&\frac{\Omega^2}{\omega^2-\Omega^2}\frac{4\Delta^2}{\Omega^3}\delta h_{\bf n}-\frac{1}{\omega^2-\Omega^2}\frac{4\Delta^2}{\Omega}\delta h_{\bf n}=0\,.
\end{eqnarray}
Another interesting feature of the mode ${\bf n}=0$ is that the condition
\begin{equation}
A_{11}({\bf n},\omega)A_{22}({\bf n},\omega)-A_{21}({\bf n},\omega)A_{12}({\bf n},\omega)=0\,,
\end{equation}
that determines the possible collective eigenfrequencies of the system, acquires a particularly simple form and does not depend on the equilibrium mean field:
\begin{equation}
\label{disp00}
\omega^2[I_1^2(\omega)-(\omega^2-4\Delta^2)I_2^2(\omega)]=0\,,
\end{equation}
with
\begin{eqnarray}
I_1(\omega)&=&\int_0^\infty d\epsilon \sqrt{\epsilon}\,\frac{\omega_\mu(\epsilon)}{\Omega(\epsilon)}\frac{1}{\omega^2-\Omega^2(\epsilon)}\,,\\
I_2(\omega)&=&\int_0^\infty d\epsilon \sqrt{\epsilon}\,\frac{1}{\Omega(\epsilon)}\frac{1}{\omega^2-\Omega^2(\epsilon)}\,.
\end{eqnarray}
By evaluating these integrals, we find that
the dispersion relation (\ref{disp00}) has only the solution $\omega^2=0$. This zero-frequency solution corresponds to the Anderson-Goldstone-Nambu mode, associated with rotations in gauge space (see e.g. Ch. 4 of \cite{bb}).
The other possibility
\begin{equation}
\label{pv}
[I_1^2(\omega)-(\omega^2-4\Delta^2)I_2^2(\omega)]=0
\end{equation}
instead, is reminiscent of the dispersion relation of pairing vibrations (cf., e.g. Eq. (5.23) of \cite{bb}, note the correspondence between our integrals $I_{1,2}(\omega)$ and the quantities $B$ and $A$ of Ref. \cite{bb}), however we find that the relation (\ref{pv}) has no solution in the interval $\omega=0-30$ MeV. The quantum counterpart of Eq. (\ref{pv}) instead, has many solutions, corresponding to single-particle levels $\epsilon_i$ close to the Fermi surface (cf. e.g. Eqs. (J.31) and (J.32) of \cite{bb}), but these solutions have a single-particle, rather than a collective character.
\section{Response function and dispersion relation}
\label{ewsr}
Equations (\ref{delron}, \ref{ddr}, \ref{ddi}) allow us to calculate the density response function, defined as
\begin{equation}
\label{denr}
{\cal R} (\omega)=\frac{1}{\eta}\int d{\bf r} Q^*({\bf r})\delta\varrho({\bf r},\omega)\,,
\end{equation}
where $\delta\varrho({\bf r},\omega)$ is the time Fourier transform of the density fluctuation (\ref{denf}). According to Eq. (\ref{delron}), the density response function is given by the sum of three terms:
\begin{equation}
\label{denresp}
{\cal R} (\omega)= {\cal R}^{cd}(\omega) + {\cal R}^r(\omega) +{\cal R}^i(\omega)\,,
\end{equation}
with the first term containing the contribution of the constant-$\Delta$ approximation, the second and third terms, the contributions of the real and imaginary parts of the pairing-filed fluctuations, respectively.
The three components of the density response function are given by
\begin{eqnarray}
\label{cdresp}
{\cal R}^{cd}(\omega)&=&4\sum_{\bf n} \int d{\bf I} C^{cd}_{\bf n}({\bf I},\omega) Q_{\bf n}({\bf I}) Q_{\bf n}^*({\bf I})\,,\\
\label{drresp}
{\cal R}^{r}(\omega)&=&4\sum_{\bf n} F^r_{\bf n}(\omega) \int d{\bf I} C^r_{\bf n}({\bf I} ,\omega) Q_{\bf n}({\bf I})Q^*_{\bf n}({\bf I} )\,,\\
\label{diresp}
{\cal R}^{i}(\omega)&=&4\sum_{\bf n} F^i_{\bf n}(\omega) \int d{\bf I} C^i_{\bf n}({\bf I} ,\omega) Q_{\bf n}({\bf I})Q^*_{\bf n}({\bf I})\,,
\end{eqnarray}
with
\begin{eqnarray}
\label{ccdn}
C^{cd}_{\bf n}({\bf I},\omega)&=&\frac{\omega^2\bar\omega_{\bf n}^2-\omega_{\bf n}^2\omega_\mu^2}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}\frac{4\Delta^2}{\Omega^3}\,\\
C^r_{\bf n}({\bf I},\omega)&=&-\frac{\omega^2\bar\omega_{\bf n}^2-\omega_{\bf n}^2\omega_\mu^2}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}\frac{2\Delta\omega_\mu}{\Omega^3}\,,\\
\label{cin}
C^i_{\bf n}({\bf I},\omega)&=&-i\omega\frac{2\Delta}{\Omega}\frac{\omega^2}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}\,.
\end{eqnarray}
The solutions (\ref{frn}, \ref{fin}) can be written as
\begin{eqnarray}
F^r_{\bf n}(\omega)&=&{2\Delta}\tilde F^r_{\bf n}(\omega)\,,\\
\label{fitiln}
F^i_{\bf n}(\omega)&=&\frac{2\Delta}{i\omega}+\frac{2\Delta}{i\omega}\tilde F^i_{\bf n}(\omega)\,,
\end{eqnarray}
with
\begin{eqnarray}
\tilde F^r_{\bf n}(\omega&=&4\Delta^2\frac{(I_1I_4-I_2I_3)}{I_1 I_1'- I_2(\omega^2I_2-4\Delta^2I_2')}\,,\qquad\\
\,,\\
\tilde F^i_{\bf n}(\omega)&=&4\Delta^2\,\frac{I_3I_1'-I_4(\omega^2I_2-4\Delta^2I_2')}{I_1I_1'-I_2(\omega^2I_2-4\Delta^2I_2')}\,.
\end{eqnarray}
Since the function $F^i_{\bf n}(\omega)$ is the sum of the two terms in Eq. (\ref{fitiln}), the response function (\ref{diresp}) also becomes the sum of two response functions
\begin{eqnarray}
{\cal R}^{i}(\omega)&=&4\sum_{\bf n}\frac{2\Delta}{i\omega} \int d{\bf I} C^i_{\bf n}({\bf I} ,\omega) Q_{\bf n}({\bf I})Q^*_{\bf n}({\bf I})+\tilde{\cal R}^{i}(\omega)\,,
\end{eqnarray}
with
\begin{eqnarray}
\tilde{\cal R}^{i}(\omega)&=&4\sum_{\bf n} \frac{2\Delta}{i\omega}\tilde F^i_{\bf n}(\omega) \int d{\bf I} C^i_{\bf n}({\bf I} ,\omega) Q_{\bf n}({\bf I})Q^*_{\bf n}({\bf I})\,.
\end{eqnarray}
The first part of ${\cal R}^{i}(\omega)$ cancels the spurious part of ${\cal R}^{cd}(\omega)$ (the term $\omega^2\Omega^2$ of the product $\omega^2\bar\omega_{\bf n}^2=\omega^2\omega_{\bf n}^2+\omega^2\Omega^2$ in the numerator of (\ref{ccdn})) and the density response function becomes
\begin{equation}
\label{corr}
{\cal R} (\omega)= \tilde{\cal R}^{cd}(\omega) + {\cal R}^r(\omega) +\tilde{\cal R}^i(\omega)\,,
\end{equation}
This response function can be written in a more compact form as
\begin{eqnarray}
\label{comp}
{\cal R}(\omega)&=&4\sum_{\bf n} \int d{\bf I} G_{\bf n}({\bf I},\omega) Q_{\bf n}({\bf I}) Q_{\bf n}^*({\bf I})\,,
\end{eqnarray}
with
\begin{eqnarray}
G_{\bf n}({\bf I},\omega)&=&\frac{4\Delta^2}{\Omega^3}{\Big (}\frac{\omega_{\bf n}^2(\omega^2-\omega_\mu^2)}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}
-\frac{\omega^2\bar\omega_{\bf n}^2-\omega_{\bf n}^2\omega_\mu^2}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}\omega_\mu\tilde F^r_{\bf n}(\omega)\\
&-&\frac{\omega^2\Omega^2}{(\omega^2-\omega_{\bf n}^2)D_{\bf n}(\omega)}\tilde F^i_{\bf n}(\omega){\Big )}\nonumber\,.
\end{eqnarray}
The response function (\ref{comp}) can be considered as a finite-system version of the response function derived by the authors of Ref. \cite{cks} for uniform systems (cf. Eq. (B22) of \cite{cks}).
The dispersion relation of the Bogoliubov-Anderson mode of \cite{cks} is here replaced by the dispersion relation
\begin{eqnarray}
\label{disprel}
I_1({\bf n},\omega)I_1'({\bf n},\omega)-I_2({\bf n},\omega)[\omega^2I_2({\bf n},\omega)-4\Delta^2I_2'({\bf n},\omega)]=0\,
\end{eqnarray}
that takes into account the finite size of the system. This dispersion relation is the main result of our paper and in the next section we will study the possibility of a collective mode, analogous to that studied in \cite {cks}, in 'small' systems like nuclei.
It can also be easily checked, simply by looking at the large-$\omega$ behavior of $G_{\bf n}$ \cite{ls}, that the EWSR is satisfied by our response function (\ref{corr}), because the spurious contribution of ${\cal R}^{cd}(\omega)$ is canceled by the pairing-field fluctuations (more specifically, by the first term on the right-hand side of Eq. (\ref{fitiln})).
\section{Spherical cavity}
The four three-dimensional integrals (\ref{i1n}--\ref{i4n}), which appear in the dispersion relation (\ref{disprel}), are not simple when ${\bf n}\neq 0$. In a spherical systems, however, these integrals can be reduced to two-dimensional. In such a system, the action variables are conveniently defined as (\cite{gol}, p. 476, see also \cite{dmb})
\begin{eqnarray}
I_1&=&\lambda_z\,,\\
I_2&=&\lambda\,,\\
I_3&=&\lambda+\frac{1}{2\pi}\oint dr p_r\,,
\end{eqnarray}
where $\lambda$ is the magnitude of the particle angular momentum, $\lambda_z$ its $z$-component and $p_r$ the radial component of the particle momentum. It is convenient to change variables from ${\bf I}$
to $(\epsilon,\lambda,\cos\beta)$, with $\cos\beta=\lambda_z/\lambda$, the Jacobian of the transformation is
\begin{equation}
J(\epsilon,\lambda)=\frac{\lambda T(\epsilon,\lambda)}{2\pi}\,,
\end{equation}
where $T(\epsilon,\lambda)$ is the period of radial motion.
Because of the spherical symmetry, the eigenfrequencies $\omega_{\bf n}({\bf I})$ do not depend on $\lambda_z$. Moreover, as shown in \cite{bdd}, the eigenfrequencies can be labeled with only two components of the integer vector ${\bf n}$. We put $n_3=n$, $n_2=N$ and $n_1$ becomes redundant. Thus, in spherical systems
\begin{equation}
\omega_{\bf n}({\bf I})\to\omega_{nN}(\epsilon,\lambda)\approx \epsilon_{r+n,l+N}-\epsilon_{r,l}\,
\end{equation}
(in a spherical mean field, the single-particle levels $\epsilon_{r,l}$ depend only on two quantum numbers).
Also, by expanding the external field $Q({\bf r})$ in multipoles as
\begin{equation}
Q({\bf r})=\sum_{LM}Q_L(r) Y_{LM}(\hat{\bf r})\,,
\end{equation}
one finds that the response function (\ref{comp}) becomes
\begin{equation}
\label{Lresp}
{\cal R}(\omega)=\sum_L\;{\cal R}_L(\omega)\,,
\end{equation}
with
\begin{eqnarray}
\label{respl}
{\cal R}_L(\omega)&=&\frac{8}{2L+1}\sum_{N=-L}^L\sum_{n=-\infty}^\infty |Y_{LN}(\frac{\pi}{2},0)|^2\int_0^\infty d\epsilon\int_0^{\bar\lambda(\epsilon)} d\lambda \frac{\lambda T(\epsilon,\lambda)}{2\pi}\nonumber\\
&\times&|Q^{(L)}_{nN}(\epsilon,\lambda)|^2G(N,n,\epsilon,\lambda,\omega)\,,
\end{eqnarray}
and
\begin{eqnarray}
\label{gfun}
G(N,n,\epsilon,\lambda,\omega)&=&\frac{4\Delta^2}{\Omega^3}{\Big (}\frac{\omega_{nN}^2(\omega^2-\omega_\mu^2)}{\omega^4-\omega^2\bar\omega^2_{nN}+\omega^2_\mu\omega^2_{nN}}\\
&-&\frac{\omega^2\bar\omega^2_{nN}-\omega^2_{nN}\omega_\mu^2}{\omega^4-\omega^2\bar\omega^2_{nN}+\omega^2_\mu\omega^2_{nN}}\omega_\mu\tilde F^r(N,n,\omega)\nonumber\\
&-&\frac{\omega^2\Omega^2}{\omega^4-\omega^2\bar\omega^2_{nN}+\omega^2_\mu\omega^2_{nN}}\tilde F^i(N,n,\omega){\Big )}\nonumber\,.
\end{eqnarray}
In Eq. (\ref{respl}), only terms in which the integer $N$ has the same parity as $L$ appear in the sum over $N$, otherwise the spherical harmonics $Y_{LN}(\frac{\pi}{2},0)$ vanish.
The Fourier coefficients $Q_{\bf n}({\bf I})$ in Eq. (\ref{comp}) are replaced by the radial coefficients $Q^{(L)}_{nN}(\epsilon,\lambda)$, defined as in \cite{bdd}, which correspond to the radial matrix elements of the quantum approach.
The quantity $\bar\lambda(\epsilon)$ is the maximum possible value of $\lambda$ for a particle with energy $\epsilon$.
For a spherical cavity of radius $R$, $\bar\lambda(\epsilon)=\sqrt{2m\epsilon}R$ and
the eigenfrequencies $\omega_{nN}(\epsilon,\lambda)$ are conveniently written as
\begin{equation}
\omega_{nN}(\epsilon,\lambda)=\omega_F(\epsilon) s_{nN}(x)\,,
\end{equation}
where $x$ is a dimensionless parameter, which is related to the particle angular momentum by
\begin{eqnarray}
x&=&\sin\alpha\,,\\
\cos\alpha&=&\frac{\lambda}{\bar\lambda(\epsilon)}\,.
\end{eqnarray}
The functions $s_{nN}(x)$ are
\begin{equation}
s_{nN}(x)=\frac{n\pi+N\arcsin(x)}{x}\,,
\end{equation}
while the frequency $\omega_F(\epsilon)$ is given by
\begin{equation}
\omega_F(\epsilon)=\frac{\sqrt{2\epsilon/m}}{R}\,.
\end{equation}
The multipole response function (\ref{respl}) becomes
\begin{eqnarray}
\label{resplsc}
{\cal R}_L(\omega)&=&\frac{9}{2}\frac{A}{\mu^{3/2}}\sum_{N=-L}^L\sum_{n=-\infty}^\infty \frac{|Y_{LN}(\frac{\pi}{2},0)|^2}{2L+1}\int_0^\infty d\epsilon\sqrt{\epsilon}\nonumber\\
&\times&\int_0^1 d x x^2|Q^{(L)}_{nN}(x)|^2G(N,n,\epsilon,x,\omega)\,.
\end{eqnarray}
The functions $G(N,n,x,\omega)$ are given by Eq.(\ref{gfun}), with
\begin{eqnarray}
I_1(N,n,\omega)&=&{\cal N}\int_0^\infty d\epsilon\sqrt{\epsilon}\,\frac{\omega_\mu(\epsilon)}{\Omega(\epsilon)}J_1(N,n,\epsilon,\omega)\,,\\
I_2(N,n,\omega)&=&{\cal N}\int_0^\infty d\epsilon\sqrt{\epsilon}\,\frac{1}{\Omega(\epsilon)}J_1(N,n,\epsilon,\omega)\,,\\
I_3(N,n,\omega)&=&{\cal N}
\int_0^\infty d\epsilon\sqrt{\epsilon}\,\frac{\omega_\mu(\epsilon)}{\Omega^3(\epsilon)}\omega^2_F(\epsilon)J_2(N,n,\epsilon,\omega)\,,\\
I_4(N,n,\omega)&=&{\cal N}\int_0^\infty d\epsilon\sqrt{\epsilon}\,\frac{1}{\Omega^3(\epsilon)}\omega^2_F(\epsilon)J_2(N,n,\epsilon,\omega)\,,\\
J_1(N,n,\epsilon,\omega)&=&\int_0^1 dxx^2 \frac{1}{D(N,n,\epsilon,x,\omega)}\,,\\
J_2(N,n,\epsilon,\omega)&=&\int_0^1 dx x^2\frac{1}{D(N,n,\epsilon,x,\omega)}\frac{s^2_{nN}(x)}{\omega^2-\omega_F^2(\epsilon)s^2_{nN}(x)}\,,\nonumber\\
\end{eqnarray}
and ${\cal N}$ is a normalization factor which cancels in the final expression.
The denominator
$D(N,n,\epsilon,x,\omega)$ is given explicitly by
\begin{equation}
D(N,n,\epsilon,\lambda,\omega)=\omega^2-\Omega^2(\epsilon)-4\Delta^2\frac{\omega^2_{F}(\epsilon)s_{nN}^2(x)}{\omega^2-\omega^2_{F}(\epsilon)s_{nN}^2(x)}\,.
\end{equation}
The dispersion relation (\ref{disprel}) takes the form
\begin{equation}
\label{scdisprel}
K(N,n,\omega)=0\,,
\end{equation}
with
\begin{eqnarray}
K(N,n,\omega)&=&I_1(N,n,\omega)I_1'(N,n,\omega)\\
&-&I_2(N,n,\omega)[\omega^2I_2(N,n,\omega)-4\Delta^2I_2'(N,n,\omega)]\,.\nonumber
\end{eqnarray}
\section{Results}
The first figure shows the absolute value of $K(N,n,\omega)$ for a few low-energy modes $(N,n)$ in the range
$\omega=0-30$ MeV for a spherical cavity containing $A=208$ nucleons within a radius $R=1.2\; A^{1/3}$ fm. The chemical potential $\mu$ is determined from Eq. (\ref{anorm}) to be $\mu\approx33.33$ MeV and a value $\Delta=1$ Mev has been assumed for the static pairing gap.
Except for the modes $(N=\pm1, n=0)$ (top right panel), all other modes display a smooth behavior suggesting that for these modes the dispersion relation (\ref{scdisprel}) has no solution in this range of $\omega$. The modes $N=\pm1,n=0$ instead, have a minimum at $\omega\approx 7.7$ MeV, implying that, for this value of $\omega$, the dispersion relation (\ref{scdisprel}) is approximately satisfied. Thus we can expect (mild) collective effects in the odd multipole channels, to which these modes contribute. The dipole response gets its main contribution from the mode $(N=\pm1,\;n=0)$, however this channel is spurious because of the breaking of translation invariance, so the first physically meaningful collective effect is expected in the low-energy octupole resonance (corresponding to $L=3, N=\pm 1,n=0$) while the high-energy octupole resonance (corresponding to $L=3, N=\pm3, n=0)$ is not affected by collective effects (cf. bottom right panel of Fig.1)
\begin{figure}[h]
\label{fig1}
\vspace{.2in}
\centerline {
\includegraphics[width=4in]{disparray.eps}
}
\vspace{.2in}
\caption{ Absolute value of $K(N,n,\omega)$ as a function of $\omega$ (in MeV) for $n=0$ and $N=0$ (top left), $N=\pm 1$ (top right), $N=\pm2$ (bottom left), $N=\pm3$ (bottom right). The curve corresponding to the modes $N=\pm1,\;n=0$ has a local minimum at $\omega\approx 7.7$ MeV, corresponding to an approximate solution of the dispersion relation $K(N,n,\omega)=0$.}
\end{figure}
The second figure shows an enlargement of the top right panel of Fig. 1, in the region of the minimum.
\begin{figure}[h]
\label{fig2}
\vspace{.2in}
\centerline {
\includegraphics[width=3in]{disp2.eps}
}
\vspace{.2in}
\caption{ Absolute value of the function $K(N=1,n=0,\omega)$ in the region of the local minimum.}
\end{figure}
The other figures (Figs. 3-6) show the strength function
\begin{equation}
S(\omega)=-\frac{1}{\pi}{\rm Im}\,{\cal R}_L(\omega)\,
\end{equation}
associated with the multipolarities $L=0,\;2,\;3$.
The radial Fourier coefficients needed for the monopole and quadrupole response functions are (for the quadrupole and octupole response, we take $Q_L(r)=r^L$, while for the monopole response, we take $Q_L(r)=r^{L+2}$)
\begin{eqnarray}
\label{q2n0}
Q^{(2)}_{n 0}(x)&=&(-)^nR^2\frac{2}{s^2_{n0}(x)}\,,\qquad\qquad(n\neq 0)\\
Q^{(2)}_{n,\pm2}(x)&=&(-)^n R^2\frac{2}{s^2_{n\pm2}(x)}[1\pm2\frac{\sqrt{1-x^2}}{s_{n,\pm2}(x)}]\,,\\
\end{eqnarray}
while those entering the octupole response are
\begin{eqnarray}
Q^{(3)}_{nN}(x)&=&(-)^nR^3\frac{3}{s^2_{nN}(x)}[1+\frac{4}{3}N\frac{\sqrt{1-x^2}}{s_{nN}(x)}\nonumber\\
&-&\frac{2}{s^2_{nN}(x)}+4(|N|-1)\frac{1-x^2}{s^2_{nN}(x)}]\qquad\qquad N=\pm1,\pm3\,.
\label{q33}
\end{eqnarray}
\begin{figure}[h]
\vspace{.2in}
\centerline {
\includegraphics[width=3in]{monop10.eps}
}
\vspace{.2in}
\caption{ Contribution of the modes $(N=0,n=\pm1)$ to the monopole response function of a spherical cavity containing $A=208$ nucleons. Parameters are the same as for Fig. 1, a small imaginary part $\varepsilon=0.1$ MeV has been added to $\omega$. The giant monopole resonance is centred at about $25$ MeV because the long-range part of the residual interaction has not been taken into account. }
\end{figure}
\newpage
\begin{figure}[h]
\vspace{.2in}
\centerline {
\includegraphics[width=3in]{quadrup10.eps}
}
\vspace{.2in}
\caption{ Contribution of the modes $(N=\pm2,n=0,)$ to the quadrupole strength function of the same system of Fig.3. }
\end{figure}
\newpage
In all the figures showing the multipole strength functions there are two curves. The dashed curve corresponds to the strength function associated with the response function $\tilde{\cal R}^{cd}_L(\omega)$, while the solid curve corresponds to ${\cal R}_L(\omega)$.
For the monopole response of Fig.3, the two curves are practically coincident.
For the quadrupole response of Fig.4, the two curves are barely distinguishable, while for the octupole response of Figs. 5 and 6, the two curves differ in the region of the low-energy octupole resonance, while they are very similar for the high-energy component of the giant octupole resonance.
\begin{figure}[h]
\label{fig2}
\vspace{.2in}
\centerline {
\includegraphics[width=3in]{octupfig1.eps}
}
\vspace{.2in}
\caption{ Contribution of the modes $(N=\pm1,\pm3,\,n=0)$ to the octupole strength function. The two peaks at $\omega\approx 8$ and $\omega\approx 25$ MeV correspond to the low and high-energy components of the giant octupole resonance. The energy of the two peaks is larger than experiment because the long-range part of the residual interaction has not been included in these calculations. The low-energy peak is due to the modes $N=\pm1$, while the high-energy peak to the modes $N=\pm 3$.}
\end{figure}
\newpage
\begin{figure}[h]
\vspace{.2in}
\centerline {
\includegraphics[width=2in]{octupfig2.eps}
}
\vspace{.2in}
\caption{Same as Fig.5, in the region of the low-energy octupole resonance.}
\end{figure}
\newpage
The pronounced minimum of the dispersion curve in the top-right panel of Fig. 1 at $\omega \approx 7.7$ MeV is at the origin of the increase of the strength function given by the solid curve in Figs. 5 and 6. This increase is a collective effect due to the pairing interaction.
\newpage
\newpage
\section{Summary and conclusions}
We have studied the solutions of an improved kinetic equation that includes dynamic pairing correlations, as well as finite-size effects, and compared it with a more simplified approach based on the constant-$\Delta$ approximation. The particle-number symmetry, which is broken in the constant-$\Delta$ approximation, is restored by the fluctuations of the imaginary pairing field and all spurious contributions are canceled, leading to a correct value of the energy-weighted sum rule. The restoration of particle-number conservation is not the only effect of the pairing-field fluctuatons since they
introduce also the possibility of new collective modes of the system, generated by the pairing interaction. The frequencies of these collective modes are determined by the solution of appropriate dispersion relations which are a generalization of the dispersion relation of the Bogoliubov--Anderson mode derived by the authors of Ref.\cite{cks} for a uniform system. In the present paper these modes have been studied in detail for a small system like a nucleus, where the pairing gap is small compared with the collective frequency. The nuclear mean field has been approximated with a spherical square-well potential and it has been found that the isoscalar octupole strength function shows an increase in the region of the low-energy octupole resonance, which is due to collective pairing effects.
We conclude that the semiclassical treatment of pairing introduced in \cite{abdm} and improved in the present paper, gives results in good agreement with previous quantum approaches and that the self-consistent pairing-field fluctuations considered here have two main effects:
\begin{itemize}
\item{the fluctuations of the imaginary part are essential for restoring the particle-number symmetry which is broken by the constant-$\Delta$ approximation (this is already well known from studies of infinite systems);}
\item{there is a mild collective effect in the octupole response function due to the pairing interaction and no similar effect is found in the monopole and quadrupole response function.}
\end{itemize}
\section*{Appendix}
\subsection*{Supplementary condition}
In this Appendix it is shown that only one branch of the solutions of the equations of motion (\ref{coe1}--\ref{coe4}) approximately satisfy the supplementary condition (\ref{pauli}) required by the Pauli principle for the modes ${\bf n}\neq 0$.
For this purpose we consider the homogeneous system of equations associated with Eqs. (\ref{coe1}--\ref{coe4}) in constant-$\Delta$ approximation:
\begin{eqnarray}
\label{eom1}
\omega\delta\rho^{ev}_{\bf n}(\omega)&=&\omega_{\bf n}\delta\rho^{od}_{\bf n}(\omega)-2i\Delta\delta \kappa^i_{\bf n}(\omega)\,,\\
\label{eom2}
\omega\delta\rho^{od}_{\bf n}(\omega)&=&\omega_{\bf n}\delta\rho^{ev}_{\bf n}(\omega)\,,\\
\label{eom3}
i\omega\delta\kappa^i_{\bf n}(\omega)&=&\omega_\mu\delta\kappa^r_{\bf n}(\omega)-2\Delta\delta\rho^{ev}_{\bf n}(\omega)\,,\\
\label{eom4}
-i\omega\delta\kappa^r_{\bf n}(\omega)&=&\omega_\mu\delta\kappa^i_{\bf n}(\omega)\,.
\end{eqnarray}
The following relation can be easily obtained from these equations, either by using all four of them or only the first three:
\begin{equation}
\label{mult}
(\omega^2-\omega_{\bf n}^2-4\Delta^2)\delta\rho^{ev}_{\bf n}(\omega)=-2\Delta \omega_\mu\delta\kappa^r_{\bf n}(\omega)\,.
\end{equation}
It is important to note that, if we put $\omega=\bar\omega_{\bf n}$ in this relation, it gives
\begin{equation}
\delta\kappa^r_{\bf n}(\bar\omega_{\bf n})=-\frac{\omega_\mu}{2\Delta}\delta\rho^{ev}_{\bf n}(\bar\omega_{\bf n})\,,
\end{equation}
in agreement with the Pauli constraint (\ref{pauli}).
The solutions of the homogeneous system (\ref{eom1}--\ref{eom4}) are non vanishing only at the eigenfrequencies $\omega_+({\bf n},{\bf I})$ and $\omega_-({\bf n},{\bf I})$, but we have seen that the frequencies
$\omega_+({\bf n},{\bf I})$ are well approximated by $\bar\omega_{\bf n}$ and the fact that the condition (\ref{pauli}) is satisfied when $\omega=\bar\omega_{\bf n}$, means that the branch of solutions corresponding to $\omega_+({\bf n},{\bf I})$ approximately satisfies the supplementary condition (\ref{pauli}):
\begin{equation}
\delta\kappa^r_{\bf n}(\omega_+)\approx-\frac{\omega_\mu}{2\Delta}\delta\rho^{ev}_{\bf n}(\omega_+)\,,
\end{equation}
while the same is not true for the solutions corresponding to the frequencies $\omega_-({\bf n},{\bf I})$. Thus, following Anderson \cite{pwa} and Valatin \cite{jgv}, we can say that only the solutions corresponding to the branch $\omega_+({\bf n},{\bf I})\approx\bar\omega_{\bf n}$ are acceptable. An important point is that the constraint (\ref{pauli}) need not be satisfied for all values of $\omega$, but only at the eigenfrequencies of the system.
|
\section{Introduction}
\label{intro}
Collisional ring galaxies represent a very specific case of galaxy interaction for which a companion drops through a much more massive spiral galaxy \citep[e.g.][]{lynds76,theys77,hern93}. First it requires very specific initial conditions to occur. The mass of the companion must be small compared to the main galaxy and the primary must be a disk galaxy. Second, the characteristic ring of star formation only lasts for a few 10$^8$\,yr, which is not a large temporal window compared to the Hubble timescale. Although they are rare objects in the local universe, they are thought to be more common at high redshifts \citep{lav96,elm06}.
NGC\,922 is a drop-through ring galaxy for which the perturber, responsible for the ring pattern, has recently been discovered \citep[][hereafter W06]{wong06}. Using simulations, the authors found that NGC\,922's companion (named S2) interacted gravitationally with NGC\,922 about 330\,Myr ago through an off-centered collision, creating the C-shape morphology seen in the disk of NGC\,922. The dwarf compact companion is about 20 times less massive than NGC\,922, the latter having an estimated mass of 2.8$\times$10$^{10}$\,M$_{\odot}$. NGC\,922 is one of the nearest collisional ring galaxies known, at a distance of 43\,Mpc \citep{meurer06}, about three times closer than the iconic Cartwheel Galaxy \citep[e.g.][]{fos77,dav82,str96}. It has a heliocentric velocity of 3077\,km\,s$^{-1}$, and an estimated star formation rate (SFR) of 7-8\,M$_{\odot}$\,yr$^{-1}$. A more detailed description of the physical properties of both NGC\,922 and S2 can be found in the work of W06.
\begin{figure*}[!ht]
\epsscale{1.15}
\plottwo{pellerin_fig1a_astroph.eps}{pellerin_fig1b_astroph.eps}
\caption{\label{cmyk} Left: Three-color HST/WFPC2 image of NGC\,922. Blue: F300W; Green: F547M; Red: F814W. The image is 2.5$\arcmin\times$2.0$\arcmin$ ($\sim$30.7$\times$24.6\,kpc$^2$). Right: H$\alpha$ image of NGC922 from SINGG. The resolution is $\sim$1.5$\arcsec$, lower than the HST image. Blue: R-band; Green: H$\alpha$+ continuum; Red: H$\alpha$ with continuum substracted. North is up and east is left.\\}
\end{figure*}
\section{Data processing and PSF Photometry}
\label{data}
NGC\,922 is a very interesting object to study. Since the companion responsible for the star formation activity is known, dynamical simulations can be better constrained and its recent star formation history can be modeled. Furthermore, because of its proximity, its stellar cluster population can be resolved and observed with the {\it Hubble Space Telescope} (HST). In this paper, we present recent HST optical images of NGC\,922. The overall goal is to study the recent star cluster formation history of NGC\,922 and to compare it to the dynamical models in order to constrain the models and the recent dynamical history of NGC\,922 as well as better understanding the star formation and cluster evolution processes.
In the next section, we present the HST data as well as the photometric analysis of the star clusters detected in NGC\,922. In section~\ref{syn} we present the synthesis modeling performed to estimate cluster ages and masses. In section~\ref{pop} we describe the star cluster properties. Section~\ref{test} discusses the effect of cluster dissolution on the observed star cluster population. We then present an N-body/SPH simulation of the collision and compare the results to the observations (\S\ref{model}). In section~\ref{large} we investigate the large-scale star formation of NGC\,922 and compare it to distant ring galaxies. Section~\ref{tail} presents diffuse features detected in NGC\,922. We then discuss the fate of NGC\,922 and its star clusters in section~\ref{fate}. We conclude in \S\ref{conclusion}.
Six sets of images of NGC\,922, two slightly different pointings with three exposures each, have been obtained on 25 and 26 June 2007 using the Wide Field Planetary Camera 2 (WFPC2) onboard HST in the four filters F300W, F439W, F547M, and F814W (hereafter U, B, V, and I for simplicity). The F439W and F547M filters were chosen over F450W and F555W to avoid contamination by the nebular emission lines of [O\,{\sc{iii}}] and H$\beta$ in the passbands. The WFPC2 field covers the whole ring galaxy, but excludes S2.
The data were first reduced by the STScI calibration pipeline, which includes treatment for the bias, dark, and flat fields. Then we corrected for warm pixels, geometric distortion and flux calibration, as well as charge transfer efficiency. We used the variable-pixel linear reconstruction method (or drizzling) and the IRAF task {\it {multidrizzle}} to combine the individual exposures, remove cosmic rays and to improve the spatial resolution of the final images. The chips were drizzled individually to an output pixel size of 0.075$\arcsec$ pixel$^{-1}$ (15.6\,pc~pixel$^{-1}$) for the WF chip.
A three-color image from our HST/WFPC2 data is shown in Figure~\ref{cmyk} together with a H$\alpha$ image from the Survey for Ionization in Neutral-Gas Galaxies \citep[SINGG;][]{meurer06}.
Figure~\ref{morph} presents the main morphological features of NGC\,922 over-plotted on the F814W WFPC2 image. The blue ring has a projected diameter of 15\,kpc and is particularly prominent on the eastern side, giving a ``C" like morphology to NGC\,922. While the eastern side of the ring displays sharp edges with star-forming regions well distributed along it, the western side shows a patchy distribution and weak diffuse H$\alpha$ emission. A 9\,kpc off-center bar, which appears yellow in Fig.~\ref{cmyk}a, displays spokes emanating from it. A relatively isolated star-forming region is also observed on the South-West portion of the ring but is disconnected from the ``C". Another star-forming structure, like an accent over the ``C", is observed north of the ring, and merges quite well with the ring near the bar end. On the H$\alpha$ image an inner arc can be seen internal to the ring, perhaps as an extension of the ``accent". However it is impossible to conclude here if this structure is physically real or just a random distribution of star-forming regions misleading the eye.
\begin{figure*}
\epsscale{0.75}
\plotone{pellerin_fig2_astroph.eps}
\caption{\label{morph} Morphological features of NGC\,922 over-plotted on the F814W image. The circle shows the expanding ring of star formation. The $+$ symbol identifies the nucleus. The white lines follow the bar. The dotted lines indicate the location of possible spokes. The polygon highlights the tidal plume. The solid line, North of the circle, shows a feature of star formation outside of the expanding ring. The physical scale is the same as in Fig.~\ref{cmyk}.\\}
\end{figure*}
To study the star cluster properties, we performed PSF photometry using the IRAF/DAOPHOT package. PSF photometry is required because of crowding. For each chip and each pointing, we used bright unsaturated sources to create a PSF model of 4.5 pixels radius and a FWHM$\sim$2.5 pixels, which slightly varies depending on the chip. In the case of the PC1 chip, there were too few sources to create a reliable PSF model. For this reason, and because the same region was observed on the WF2 chip of the other pointing, we are using only the photometry based on the WF chips for this work. A summary of the observational datasets is presented in Table~\ref{hstobs}. The absolute photometric calibration is in the Vegamag system and was obtained with the zero points calculated for each chip as described on the STScI/WFPC2 cookbook website\footnote{http://www.stsci.edu/hst/wfpc2/analysis/wfpc2\_cookbook.html}. Because the Multidrizzle software converts the WFPC2 images from DN to e- in order to properly reject cosmic rays, we adjusted the photometric zeropoint to account for this gain. We calculated and applied the aperture correction for each of the four filters and chips.
\begin{deluxetable*}{lccccc}
\tabletypesize{\scriptsize}
\tablecaption{\label{hstobs} HST/WFPC2 Observational Data Parameters of NGC\,922}
\tablehead{
\colhead{Filter} & \colhead{Pointing} & \colhead{Exp. Time} & \colhead{Detection Limit} & \colhead{Chip} & \colhead{PSF FWHM} \\
\colhead{} & & sec & mag & & pixels
}
\startdata
F300W & A &1100 & 23.8 & WF2 & 2.5 \\
& A & & & WF3 & 3.3 \\
& A & & & WF4 & 2.6 \\
& B & 1100 & & WF2 & 2.5 \\
& B & & & WF3 & 2.8 \\
F439W & A &600 & 24.3 & WF2 & 2.3 \\
& A & & & WF3 & 2.5 \\
& A & & & WF4 & 2.3 \\
& B & 600 & & WF2 & 2.3 \\
& B & & & WF3 & 2.5 \\
F547M & A &600 & 26.2 & WF2 & 2.3 \\
& A & & & WF3 & 2.5 \\
& A & & & WF4 & 2.5 \\
& B & 600 & & WF2 & 2.4 \\
& B & & & WF3 & 2.1 \\
F814W & A &600 & 27.1 & WF2 & 2.2 \\
& A & & & WF3 & 2.8 \\
& A & & & WF4 & 2.3 \\
& B &600 & & WF2 & 2.5 \\
& B & & & WF3 & 2.5 \\
\enddata
\end{deluxetable*}
To determine the photometric uncertainties and the completeness level of each cluster, we generated a list of artificial sources of various magnitudes and background levels for each combination of chip and filter. This is critical to take into account that more clusters will remain undetected in regions of high level background (often linked to crowding), and that the photometry will also be less accurate. We used this list as input for the DAOPHOT/addstar task to generate aritificial sources in the WFPC2 images. We then performed the PSF photometry again on the new images with added stars. We used the results to create functions of photometric uncertainties and completeness levels that takes into account the background level. Then we used these functions to associate the proper photometric uncertainty and completeness level values to each detected source and each filter. On average, the 50\% completeness limits are reached at 23.8, 24.3, 26.2, and 27.1 magnitudes for the U, B, V, and I filters respectively. We also applied a foreground extinction correction of E(B$-$V)=0.02 \citep{sch98}. We find that these detection limits may include individual bright stars among faint cluster samples (see discussion in \S\ref{distrib}). Finally we do not expect much contamination from background galaxies, i.e. about 5 to 15 galaxies over the entire field, to our exposure depth, based on the Hubble Deep Field \citep{sha98}.
We created four lists for all the clusters detected in each of the four filters. We then cross-correlated the four lists to find sources with matching coordinates. The matching objects had to be within 2.5~pixels from each other to be considered the same physical source. Overall we observed 2248 star clusters\footnote{Due to the very large number of clusters, the detailed photometry is not published here, but is available upon request.} detected in the four filters simultaneously through PSF photometry (hereafter UBVI sample). An additional 1569 clusters were also observed in the three BVI filters (hereafter BVI sample), but displayed no conterpart in the U-band image. Because of the given spatial resolution of 39\,pc (2.5~pixels), it is probable that some of the sources detected contain multiple clusters. A completeness correction was applied to each cluster found in each filter and based on the completeness level determined in that filter. For simultaneous detections in multiple filters, we combined the completeness corrections of the appropriate filters.
An examination of the residual images showed that three extended sources were significantly bigger than the PSF size, and were therefore badly fitted. They comprise two large clusters and the nucleus. For those sources, we performed a simple aperture photometry using polygons and the IRAF task {\it {polyphot}}. The photometric results for those three sources are given in Table~\ref{complex}.
\section{Stellar Population Modeling}
\label{syn}
We used the population synthesis model Starburst99\footnote{http://www.stsci.edu/science/starburst99/} \citep{lei99} to estimate cluster ages and masses. We used an instantaneous burst of star formation with a standard initial mass function \citep[IMF;][]{kroupa02} ranging from 0.1 to 100\,M$_{\odot}$, as well as the Padova stellar evolutionary tracks with stars from the asymptotic giant branch. A solar metallicity was chosen to match the result of W06 and based on the nuclear spectrum from the 6dF Galaxy Redshift Survey \citep{jones04}. For comparison, we also performed the synthesis using different model properties with 1) an IMF ranging from 0.1 to 30\,M$_{\odot}$\ and 2) with a metallicity of 0.4\,Z$_{\odot}$. However, considering the photometric uncertainties, these two additional models did not give significantly different results.
For an accurate comparison between the stellar population models and our photometric data, we applied the IRAF/SYNPHOT package on the Starburst99 spectra to calculate the theoretical magnitudes that would be seen in our four WFPC2 filters.
Then we performed a $\chi^2$ fitting on the photometric data points, basically the slope of the spectral energy distribution (SED), to obtain the best-fit age. Finally, we calculated the best-fit mass for each cluster by averaging the magnitude differences between the observation and the models in the four filters for the UBVI sample, and in the three filters for the BVI sample. The mass and age uncertainties were estimated by fitting the models within the photometric error bars using the bootstrap technique.
Most of the clusters found in this galaxy are very young, with a mean age of 16\,Myr and a median age of 6\,Myr for the UBVI sample, and a mean age of 34\,Myr and a median age of 6\,Myr for the BVI sample. Interestingly, the sources of the BVI sample are mainly faint sources and we obtained young ages for most of them.
\subsection{Photometric Modeling and Extinction}
\begin{deluxetable*}{lccccccccr}
\tabletypesize{\scriptsize}
\tablecaption{\label{complex}Large Sources with Aperture Photometry}
\tablehead{
\colhead{Source} & \colhead{RA} & \colhead{DEC} & \colhead{F300W} & \colhead{F439W} &
\colhead{F547M} & \colhead{F814W} & \colhead{Age} & \colhead{Log(M$_{Stellar}$)} & \colhead{Area} \\
\colhead{} & degrees & degrees & mag & mag & mag & mag & Myr & {M$_{\odot}$} & kpc$^2$
}
\startdata
Cluster \#1 & 36.27565 & -24.79476 & 18.4$\pm$0.4 & 18.6$\pm$0.4 & 19.8$\pm$0.4 & 19.7$\pm$0.5 & 6$^{+0}_{-2}$ & 5.5$\pm$0.2 & 41.1 \\
Cluster \#2 & 36.27287 & -24.79845 & 17.2$\pm$0.3 & 17.4$\pm$0.3 & 18.8$\pm$0.3 & 18.5$\pm$0.4 & 6$^{+0}_{-2}$ & 6.0$^{+0.1}_{-0.2}$ & 143.9 \\
Nucleus & 36.26851 & -24.78820 & 15.34$\pm$0.04 & 14.88$\pm$0.01 & 16.08$\pm$0.02 & 15.56$\pm$0.03 & 90$^{+0}_{-20}$ & 8.1$^{+0}_{-0.2}$ & 328.3 \\
Tidal Plume & \nodata & \nodata & \nodata & 17.0 & 15.06 & 15.45 & 1300$^{+3700}_{-300}$ & 8.9$\pm$0.2 & 30260.3 \\
\enddata
\tablecomments{The RA and DEC positions are in J2000 and approximations and are not necessarily related to the exact photometric center of the sources. \\}
\end{deluxetable*}
In an effort to estimate the extinction for each cluster, we used the Q-method applied to cluster colors as described by \citet{john53} and \citet{van68}. This method allows, in principal, an accurate age estimate of the clusters since the Q-Parameters are independent of reddening \citep[see a good example of this method applied to star clusters by][]{whit99}. However, this method requires very good photometry because the error bars get larger when color information is involved. For this reason, and despite the high quality of the WFPC2 images, we did not obtain extinction estimates that are precise enough to apply to each cluster. The ages obtained with the Q-Parameter method were overestimated by $\sim$100\,Myr compared to the more precise SED fitting described previously, which biased the extinction calculation. Note that the overall E(B-V) values we obtained were consistent with the dust features seen by eye in the WFPC2 images, with more obscured zones where larger E(B-V) values were found, but the cluster-to-cluster extinction variation was too large to reliably apply it to each cluster. Although most clusters found in this work are already very young and cannot be that much younger, we must keep in mind that we probably overestimated the age of some clusters, especially the reddest ones.
\section{The Stellar Cluster Population}
\label{pop}
\subsection{Cluster Luminosity Function}
Figure~\ref{clf} shows the cluster luminosity functions (CLF). For each filter (one per panel) we compare the CLF of four groups. The first group contains all the sources detected in each individual filter and is represented by the black line. The second group, in blue, contains only the clusters from the UBVI sample. We can see that many faint sources are lost when applying the four color detection criteria. The third group, in red, contains the clusters from the BVI sample, i.e., those undetected in the U-band. Several reasons can explain the non-detection of a cluster in the U-band: a) the cluster is old, b) the cluster is faint or c) the cluster is reddened by dust. The CLF in Figure~\ref{clf} clearly shows that most of the sources undetected in the U-band are simply faint, and therefore, probably simply less massive. This is consistent with the young mean and median ages found for this sample as noted in section~\ref{syn}. This is also confirmed by the fact that 59\% of the clusters from the BVI sample have masses smaller than 7$\times$10$^3$\,M$_{\odot}$. The fourth curve in cyan is the CLF for the UBVI and BVI samples combined together. We can see that this combined sample is in better agreement with the raw sample (black) at fainter luminosity compared to the UBVI sample alone (blue).
\begin{figure*}
\epsscale{0.8}
\plotone{pellerin_fig3_astroph.eps}
\caption{\label{clf} Cluster luminosity function for NGC\,922. Black: all sources detected in each individual filter. Blue: Sources detected simultaneously in the UBVI filters. Red: Sources detected simultaneously in the BVI filters only. Green: combination of the UBVI plus BVI samples (i.e. Blue$+$Red curves). The dotted lines represent the samples uncorrected for incompleteness. The arrows show the 50\% completeness limit.}
\end{figure*}
We calculated the slope $\alpha$ of the CLF as described by \citet{elm97}. For the raw samples, we obtained slope values of 2.15$\pm$0.08, 2.22$\pm$0.05, 2.05$\pm$0.05 and 2.10$\pm$0.05 for the U, B, V, and I bands respectively. The slopes tend to steepen to $\sim$2.3 for clusters with UBVI simultaneous detection, when we fit up to the 50\% completeness limit magnitude. Our values are consistent with those of young clusters found in nearby galaxies \citep[e.g.][]{whit99,lar02}.
\subsection{Age Distributions}
\label{distrib}
In Figure~\ref{mapage}, we present spatial maps of clusters for various age ranges. The age group 0-7\,Myr corresponds to clusters capable of ionizing their surrounding gas and therefore producing H$\alpha$ emission. They represent 69\% of the detected clusters, in number, and they are mostly located within the ring of star formation and along the bar, as well as in a more isolated region, South-West of the nucleus. Their location is consistent with the H$\alpha$ emission (Fig.~\ref{cmyk}). We also find that 19\% of the clusters have ages between 7 and 50\,Myr, and only 3\% are older than 100\,Myr.
The oldest clusters are mainly located in the nuclear region, but are also seen along the bar and in the disk. One must keep in mind that some of them might be reddened and therefore younger than estimated here.
\begin{figure*}
\epsscale{0.9}
\plotone{pellerin_fig4_astroph.eps}
\caption{\label{mapage} Maps of star clusters detected in NGC\,922 for both UBVI and BVI samples. a) Clusters younger than 7\,Myr and capable of producing H$\alpha$ emission. b) Young clusters with ages from 7 to 30\,Myr. c) Clusters with ages from 30 to 150\,Myr. d) Clusters older than 150\,Myr. The grey star symbols show the location of the two unresolved clusters and the nucleus (see Table~\ref{complex}).}
\end{figure*}
Figure~\ref{lumage} shows the magnitude-age distributions of star clusters in NGC\,922. It can be seen that at older ages ($>$100\,Myr), only the most massive clusters ($>$1$\times$10$^5$\,M$_{\odot}$) can be detected in this galaxy. They are not numerous and they mostly have just BVI photometry. Those sources must be analyzed with care since they are mostly related to the nuclear region, or other very dense regions where the field is very crowded and lumpy, and therefore may include underlying stellar populations. Also, some of them may be particularly affected by extinction and may have ages biased toward larger values, which would lead to overestimated masses.
\begin{figure*}
\epsscale{1.0}
\plotone{pellerin_fig5_astroph.eps}
\caption{\label{lumage} Magnitude-age distributions of star clusters observed in NGC\,922, including both UBVI and BVI samples shown in term of their a) U, b) B, c) V, and d) I magnitudes. Solid circles: clusters with combined completeness level greater or equal to 50\%. Open squares: clusters with combined completeness level smaller than 50\%. The solid lines represent stellar populations with masses of 1$\times$10$^4$, 1$\times$10$^5$, and 1$\times$10$^6$\,M$_{\odot}$, from bottom to top. The vertical dotted lines indicate the age at which the sample of clusters more massive than 1$\times$10$^5$\,M$_{\odot}$\ is complete in the specified filter. The horizontal dashed lines determine the brightness upper limit for a single star in each band. The right axis is displayed in absolute magnitude.}
\end{figure*}
W06 estimate the impact epoch to be $\sim$300\,Myr ago. Because of the small number of clusters at similar ages and their large age uncertainties, we do not believe it is possible to constrain the impact epoch from the cluster age distribution with the current dataset.
Around 10-30\,Myr, there is an effect from the photometric age-dating technique that results in an underestimation of the number of clusters in this age interval. It is commonly claimed as being an artifact from the photometric age-dating technique \citep[e.g.][]{gieles05,lee05} , but more recently \citet{maiz09} proposed that it is more specifically related to stochastic effects in the stellar initial mass function sampling.
In Figure~\ref{lumage} we also indicate for each band the magnitude of the brightest stars, and where confusion can occur between a bright star and a faint cluster. The limits are based on the work of \citet[][their Fig. 4 to 6]{mas06} that includes a large sample of stars in M\,31 and M\,33. In the V band, this is consistent with the work of \citet{hum83}. \citet{hill93} obtained a good stellar photometry in the B-band for massive stars in 30\,Dor and reported a limit value that is slightly fainter than \citet{mas06}. Therefore an averaged limit value was used in this work for the B band. Note that clusters more massive than 1$\times$10$^5$\,M$_{\odot}$\ do not fall below the star/cluster brightness limit for UBV bands.
Age uncertainties are displayed in Figure~\ref{errage} as a function of cluster brightness and age. It shows no particular trend other than larger uncertainties are observed for fainter sources. Square symbols indicates sources for which the completeness level is lower than 50\%. Note that the BVI sample includes only such sources. The bottom panel clearly shows the population of old clusters ($>$300\,Myr) has a low completeness level, and is therefore not reliable to draw scientific conclusions.
\begin{figure*}
\epsscale{0.7}
\plotone{pellerin_fig6_astroph.eps}
\caption{\label{errage} Top panel: Age uncertainties as a function of cluster brightness. Bottom panel: age uncertainties as a function of cluster age. Solid circles: clusters with completeness level greater or equal to 50\%. Open squares: clusters with completeness level smaller than 50\%. }
\end{figure*}
To determine the age at which the sample is complete, we use Figure~\ref{lumage} and find the age where the model lines (solid) cross the lower envelope of the data points with $>$50\%\ completeness and the horizontal (dashed) line marking the brightest stars. The lowest age of these two determines the age completeness for that filter and mass limit, and is marked by a dotted line in each panel of Figure~\ref{lumage}; the lowest of these for U, B, V, and I determines our final age limit, since clusters must be detected in at least these filters to have a photometric age. This results in an age limit of 125\,Myr for the 1$\times$10$^5$\,M$_{\odot}$\ sample which is effectively defined by both the brightest star luminosity and the 1$\times$10$^5$\,M$_{\odot}$\ model line in the U band.
Those values indicate that we do not have to worry about statistical fluctuations of the initial mass function that can be seen for clusters less massive than a few times 10$^4$\,M$_{\odot}$\, \citep{cer04}. Also, it means that we cannot extend our mass-limited sample analysis to ages greater than $\sim$125\,Myr.
The age distribution of stellar clusters in NGC\,922 is presented in Figure~\ref{agedist}. A detailed list of the completeness corrections applied to Fig.~\ref{agedist} is presented in Table~\ref{tabcomp}. The age interval from 1 to 330\,Myr covers essentially the time over which NGC\,922 should have experienced a relatively high star formation activity and produced a large amount of star clusters due to the collision with S2, and the subsequent propagation of the density wave through the disk (W06). The number of clusters was corrected for incompleteness (see \S\ref{data}), and the age distributions include only sources with combined completeness levels higher than 50\%.
\begin{figure*}[!t]
\epsscale{0.9}
\plotone{pellerin_fig7.ps}
\caption{\label{agedist} a) Observed age distribution of star clusters of at least 1$\times$10$^5$\,M$_{\odot}$\ using a bin size of 0.5 dex in log(age). b) Same as in panel a) but for a bin size of 0.3 dex. Linear fits for 6.5$\ge$Log($\tau$/yr)$\ge$8.0, where the sample is considered as complete (see \S\ref{distrib}), are over-plotted. c) Simple cluster age histograms for the same sample using a bin size of 0.5 dex. d) Same as in panel c) but using a bin size of 0.3 dex. Dotted lines: sample uncorrected for incompleteness. Solid lines: sample corrected for incompleteness. Clusters with completeness lower than 50\%\ are excluded.\\}
\end{figure*}
\begin{deluxetable*}{lccc}
\tabletypesize{\scriptsize}
\tablecaption{\label{tabcomp} Incompleteness Corrections of Observed Cluster Numbers and Age Distribution Values}
\tablehead{
\colhead{Age Bin} & \colhead{No. Clusters\tablenotemark{a}} & \colhead{No. Cluster\tablenotemark{b}} & \colhead{Log(dN/dt)} \\
\colhead{} & \colhead{Uncorrected} & \colhead{Corrected} & \colhead{Corrected}}
\startdata
0.3\,dex bins & & & \\
6.6-6.9 & 19.0$\pm$4.3 & 23.7$\pm$5.4 & -5.22 \\
6.9-7.2 & 8.0$\pm$2.8 & 9.8$\pm$3.5 & -5.91 \\
7.2-7.5 & 4.0$\pm$2.0 & 4.7$\pm$2.3 & -6.53 \\
7.5-7.8 & 26.0$\pm$5.1 & 30.9$\pm$6.1 & -6.01 \\
7.8-8.1 & 32.0$\pm$5.7 & 39.5$\pm$7.0 & -6.20 \\
\hline
0.5\,dex bins & & & \\
6.5-7.0 & 20.0$\pm$4.5 & 24.7$\pm$5.5 & -5.44 \\
7.0-7.5 & 11.0$\pm$3.3 & 13.5$\pm$4.1 & -6.21 \\
7.5-8.0 & 43.0$\pm$6.5 & 52.3$\pm$8.0 & -6.12 \\
\enddata
\tablenotetext{a}{Number of clusters observed and uncorrected for incompleteness. The values corresponds to the dotted lines in Fig.~\ref{agedist}c,d.}
\tablenotetext{b}{Number of clusters observed and corrected for incompleteness. The values corresponds to the solid lines in Fig.~\ref{agedist}c,d.}
\end{deluxetable*}
We find slopes of $-$0.7$\pm$0.2 for the age distributions in Figure~\ref{agedist}. The slopes were obtained using ordinary least-squares of Y$|$X linear regressions \citep{fei92} over the mass ranges that are not affected by incompleteness, i.e. 6.5\,$<log(\tau)<$\,8.0 if using a bin size of 0.5 dex and 6.6\,$<log(\tau)<$\,8.1 if using a bin size of 0.3 dex. Note that a bin size of 0.3 dex is smaller than the typical age uncertainties as shown in Fig~\ref{errage}. We excluded the first bins (0-5\,Myr) because clusters may still be buried in their native dust cocoon at this young age, adding to the incompleteness at visible wavelengths. We also excluded the bin around 30\,Myr in Figure~\ref{agedist}b due to the age bias discussed above. The 30\,Myr bin was included in the fit in Figure~\ref{agedist}a because of the very low number of bins available for the fit.
We also performed a Kolmogorov-Smirnov (KS) test as an independent test to check the slope of the observed age distribution. The test is presented in Fig.~\ref{kstest} and show results consistent with negative slopes. We find a probability of less than 1\% for a positive slope, while we obtain the highest probabilities ($>$50\%) for slopes between $-$0.4 and $-$0.8. Note that due to jumps in the cumulative distribution, we cannot find a single best fit.
The evolution of the SFR as traced by star clusters should, to first order, tell us something about the distribution of the dense interstellar medium (ISM), or more specifically the molecular gas, from which the clusters formed. In a collisional ring galaxy much of the star formation occurs in the expanding ring. Hence, one may expect the SFR to be proportional to the rate of the dense ISM being swept up by this ring. If the expansion velocity is nearly constant, and the ring radius is significantly larger than the radius at which the impact occurred, and if there is no cluster dissolution then a flat cluster age distribution would mean that the dense ISM must have a surface density $\Sigma_g$ that falls off with radius, $r$, as $r^{-1}$. While the neutral ISM in galaxies can fall off this shallowly \citep{bos81, hoek01,leroy08}, the molecular ISM, from which stars form, has a steeper fall-off more similar to the stellar distribution \citep{leroy08}. This will produce positive slopes in $\log(N)$ versus $\log({\rm age})$, as we indeed find in the simulations presented in Section~6. A negative slope suggests a shallower drop-off than $\Sigma_g(r) \propto r^{-1}$; while one can not rule this out in any particular galaxy, it would represent a peculiar molecular ISM distribution. Assuming a normal steep $\Sigma_g(r)$ profile, one way to reconcile the negative slope observed in NGC\,922 with what we understand of collisional ring galaxies is to introduce young star cluster disruption and dissolution. In the next section we explore this possibility.
\section{Star Cluster Disruption in NGC\,922}
\label{test}
\begin{figure*}
\epsscale{0.8}
\plotone{pellerin_fig8.ps}
\caption{\label{kstest} Cumulative distribution function of star clusters of at least 1$\times$10$^5$\,M$_{\odot}$\ (circles). For comparison, the dashed, dotted, plain, and dash-dotted lines show what would be the cumulative distribution function of an age distribution with slopes of $-$1, $-$0.6, 0, and $+$0.37, respectively. The positive slope represents a typical case of a ring galaxy the size of NGC\,922 as simulated in \S\ref{simresult}.\\}
\end{figure*}
The disruption of star clusters is a concept not trivial to test observationally, especially because the star formation history is often unknown. The recent work of \citet{bas09} expresses well this problem. NGC\,922 represents a good case-study for the star cluster disruption scenario since its cluster formation history can be relatively well recovered with dynamical and star formation simulations.
In their work on the Antennae galaxies, \citet{fall05} proposed that about 90\% of star clusters dissolve for each dex of age, up to 1\,Gyr, independently of cluster mass. We will call this empirically based relation the Mass Independent (MI) scenario, following the works of \citet{fall05} and \citet{whit07}. Concerns about the mass independent, and continuous nature of the cluster destruction, as well as the assumption of a constant SFR have been raised \citep[e.g.][]{bas09}.
There is an alternative scenario in which cluster dissolution rate varies over time. According to this scenario, infant mortality is responsible for cluster dissolution during the first 10$^7$\,yr or so, leading to a steep decreasing slope in the age-distribution. Infant mortality of star clusters is a phenomenon thought to happen to young clusters experiencing a serious loss of gas and dust during the supernovae explosion phase, leaving the clusters gravitationally unbound and easy to disrupt within 10 to 50\,Myr \citep[e.g.][]{lada03,fall05,good08}.
According to a recent work, infant mortality is hardly seen at visible wavelengths \citep{gieles08}.
Then internal dynamical evolution and stellar population fading occurs between 10$^7$ and 10$^8$\,yr.
There are relatively few cluster destroyed during that phase. This results in a flat age distribution as sketched by Lamers (2008), although the precise form may not be exactly flat, which he also notes.
Finally mass dependent cluster dissolution, due to galactic tidal effects and other external effects, produces a decreasing slope for the 10$^8$-10$^9$\,yr. We will call this scenario the Mass Dependent (MD) scenario, after the works of \citet{bou03}, \citet{gieles05}, \citet{lamers05}, \citet{lamers06}, and \citet{gieles07}.
A good comparison between the two scenarios is presented by \citet{lamers08}. As shown in Figure~\ref{agedist}, there might be some evidence of a flat distribution in NGC\,922 between 10 and 100\,Myr, as proposed by the MD scenario, but that could also be due to the artificial lack of clusters around 10-30\,Myr mentioned above.
For every cluster detected in NGC\,922, we applied a correction for the loss of stellar clusters due to dissolution.
We used the two cluster disruption functions discussed above. A starting age of 4\,Myr was chosen instead of 0\,Myr because it corresponds better to the age of the first supernova explosions and, presumably, the time when the clusters can start dissolving. The fraction of star clusters F$_{sc}$ left after dissolution, according to the MI model, is then given by
\begin{equation}
log(\rm{F}_{sc})=1,~for~log(\tau)\leq6.6
\end{equation}
\begin{equation}
log(\rm{F}_{sc})=6.6-log(\tau),~for~log(\tau)>6.6
\end{equation}
\noindent
where $\tau$ is the age of the cluster in years. Following the MD model, the fraction can be expressed by
\begin{equation}
log(\rm{F}_{sc})=1,~for~log(\tau)\leq6.6
\end{equation}
\begin{equation}
log(\rm{F}_{sc})=16.5-2.5log(\tau),~for~6.6<log(\tau)<7.0
\end{equation}
\begin{equation}
log(\rm{F}_{sc})=-1,~for~7.0<log(\tau)<8.0
\end{equation}
\begin{equation}
log(\rm{F}_{sc})=12.6-1.7log(\tau),~for~log(\tau)>8.0
\end{equation}
For the early ages of the MD model, we used a 90\% infant mortality rate per decade of time.
Note that because we do not know the star formation history of NGC\,922 prior to the collision, we cannot discuss the issue of whether or not the cluster disappearance in this age range is mass-dependent (see \S\ref{distrib}).
The age distributions of star clusters corrected for cluster dissolution and incompleteness is presented in Figure~\ref{agedistback}. We find positive slopes of $+$0.5$\pm$0.2 for clusters with masses of at least 1$\times$10$^5$\,M$_{\odot}$\ when we use the MI model. We find a relatively flat slopes (0.0$\pm$0.1) when we use the MD model. The slopes were calculated up to $log(\tau)=$ 8.1.
\begin{figure*}
\epsscale{0.9}
\plotone{pellerin_fig9.ps}
\caption{\label{agedistback} a) Observed age distribution of star clusters of at least 1$\times$10$^5$\,M$_{\odot}$\ and corrected for cluster disruption and incompleteness. b) Same as in panel a) but for a 0.3\,dex bin size. c) Simple cluster age histogram for the sample corrected for cluster disruption and incompleteness. d) Same as in panel c) but for a different bin size. The solid lines show the sample corrected using the MI model while the dashed lines show the sample corrected using the MD model. The error bars on the MD corrected histograms are omitted here so that the plots are not too cluttered. They are similar in size to those for the MI corrected histograms.\\ }
\end{figure*}
The positive slope value below 100\,Myr now suggests that the cluster formation rate was decreasing in NGC\,922 during that period, assuming that the dissolution correction is realistic and that there are no other major mechanisms affecting the number of clusters. The slope value for clusters older than 100\,Myr is still steep despite the correction for dissolution but this is attributed to incompleteness.
\section{Dynamical and Star-Cluster Simulations}
\label{model}
In order to better understand the age distribution of star clusters observed in NGC\,922, we performed N-body/SPH simulations on possible formation sites of star clusters in NGC\,922 colliding with its companion S2. We considered that the observed ring-like morphology of NGC\,922 is the result of a high-speed, off-centered collision between NGC\,922 and S2 (W06). We therefore describe the simulation results with unique orbital configurations and disk inclinations. Here we focus mainly on the age and spatial distributions of star clusters within NGC\,922 and not the chemical and morphological evolution.
\subsection{Model Parameters}
In the current simulation NGC\,922 is represented by a fully self-consistent disk galaxy model whereas S2 is represented by a point mass. NGC\,922 is modeled as a disk galaxy embedded in a massive dark matter halo with a total mass of $M_{\rm dm}$ and the universal ``NFW profile'' \citep{nav96}. The disk has an exponential radius and the total disk mass and size are $M_{\rm d}$ and $R_{\rm d}$, respectively. Our results were obtained using $M_{\rm d}=$2$\times$10$^{10}$\,M$_{\odot}$, $R_{\rm d}=13.4$\,kpc, and $M_{\rm dm}/M_{\rm d}=9$. The gas mass fraction $f_{\rm g}$ of the disk is assumed to be a free parameter.
More details about the numerical methods and techniques we employ for modeling chemodynamical evolution of interacting galaxies with star formation can by found in \citet{bekki02}.
The radial ($R$) and vertical ($Z$) density profiles of the disk are assumed to be proportional to $\exp (-R/R_{0}) $ with scale length $R_{0} = 0.2R_{\rm d}$ and to ${\rm sech}^2 (Z/Z_{0})$ with scale length $Z_{0} = 0.04R_{\rm d}$, respectively. In addition to the rotational velocity attributable to the gravitational field of the disk and halo components, the initial radial and azimuthal velocity dispersions are added to the disk component in accordance with the epicyclic theory, and with a Toomre Q parameter value of 1.5 \citep{bin87}. The interstellar medium in NGC 922 is represented by SPH particles. The mass resolution (i.e., the minimum mass of a particle) and the size resolution (the minimum gravitational softening length) are 2$\times$10$^5$\,M$_{\odot}$\ and $13.4$\,pc, respectively.
The mass ratio of the companion dwarf to the spiral ($m_{\rm 2}$) is set to be 0.2. The initial location and velocity of the companion S2, with respect to be the center of the spiral, are represented by ${\bf X}_{\rm g}$ and ${\bf V}_{\rm g}$, respectively. The spiral is assumed to be inclined by 80 degrees with respect to the $x$-$y$ plane. Although we have investigated models with different ${\bf X}_{\rm g}$ and ${\bf V}_{\rm g}$, we show the results of the model with ($x$,$y$,$z$) = ($-4R_{\rm d}$, 0.2$R_{\rm d}$, 0) and ($v_{\rm x}$,$v_{\rm y}$,$v_{\rm z}$) = ($6v$, 0, 0), where $v$ is the simulation unit ($=80.1$ km s$^{-1}$). Stars that are initially within the spiral disk are referred to as ``stars'' (or ``old stars'') whereas stars that are newly formed from gas are as ``new stars''.
\subsection{Possible formation sites of star clusters}
We adopted the following star formation laws to convert gaseous particles into either star clusters or field stars. For star cluster formation, we considered that interstellar gaseous pressure ($P_{\rm gas}$) in star forming regions of a galaxy drives the collapse of pressure-confined magnetized self-gravitating molecular clouds to form compact clusters, providing that $P_{\rm gas}$ is larger than the surface pressure ($P_{\rm s}$) of the clouds \citep[e.g.][]{elm97}:
\begin{equation}
P_{\rm gas} \ge P_{\rm s} \sim 2.0\times 10^5 k_{\rm B}. \;
\end{equation}
\noindent
where $k_B$ is the Boltzman constant. This adopted value for the threshold gaseous pressure $P_{\rm gas}$ would be appropriate for the formation of massive star clusters, such as the 1$\times$10$^5$\,M$_{\odot}$\ mass-limited sample. Our star formation prescription allows for the formation of low mass clusters, associations or field stars even if $P_{\rm gas}$ is smaller than the above threshold value. However, we do not track these stars in the plots presented here.
For field stars, we adopt the Schmidt law \citep{sch59} with an exponent of 1.5 \citep{ken89}. Although we cannot investigate the detailed physical processes of cluster formation in the present and global simulation (i.e. from $\sim$ 100 pc to 10 kpc scale), we expect that the adopted phenomenological approach enables us to identify the plausible formation sites of star clusters in NGC 922. For models with $f_{\rm g}=0.4$ (note that this is different from the 0.2 value adopted by W06), the present possible mass of young stars in NGC\,922 can be well reproduced. For the adopted orbital configurations, NGC\,922 can experience a moderately strong starburst when S2 passes through NGC\,922.
\subsection{Simulation Results}
\label{simresult}
The best-fit model produces a peak of star formation during the passage of S2 within the disk of NGC\,922 about 150\,Myr ago, then a relatively constant decrease of the cluster formation rate over time. Figure~\ref{simhist} shows the simulated age distribution of 4240 star clusters formed as described above \citep[see a detailed discussion about the star cluster or globular cluster formation models by][]{bekki02}. The overall shape of the simulated age distribution displays a slope of $+$0.37$\pm$0.02 between 1 and 100\,Myr.
\begin{figure*}
\epsscale{0.8}
\plotone{pellerin_fig10.ps}
\caption{\label{simhist} Top: simulated age distribution of clusters formed after NGC\,922's collision with S2. Bottom: Simulated cluster age histogram.}
\end{figure*}
The expected cumulative distribution function of clusters formed with this age distribution and no disruption is shown in Fig.~\ref{kstest} with a dot-dashed line. It shows that the age slope of $-0.7 \pm 0.2$ we find without any cluster dissolution correction is significantly steeper than our simulation result by $\sim5\sigma$. As discussed in Section~\ref{test} and illustrated in Fig.~\ref{agedistback}, after correction for MI we find an age slope of $0.5 \pm 0.2$ which is within $\sim 1\sigma$ of the simulated results. Correction using the MD scenario produces a flatter age slope of $0.0 \pm 0.1$, which differs from our simulation by $\sim4\sigma$. We therefore conclude that correction for cluster dissolution, using previously defined prescriptions, is capable of bringing the observed age distribution into better accord with the simulations, and that the MI model for the correction produces results somewhat more consistent with the simulations than the MD based correction. Because of the possibility that any given galaxy may have a peculiar cluster formation history that is not captured in a simple model, one should not see our results as a sweeping proof of cluster destruction following the MI scenario, the MD scenario still agrees within 4$\sigma$ of the simple model. However, it does suggest that further comparisons between simulations and well modelled galaxies may provide good constraints on which cluster disruption path works best.
One possible problem with the MI scenario is the vast number of clusters that would have had to be destroyed. To test if the implied SFR is consistent with the observations of W06, we integrated the MI cluster formation rate for all detected clusters uncorrected for incompleteness.
If we use the MI scenario, we obtain a total stellar mass post-collision of 7$\times$10$^8$\,M$_{\odot}$. This is 14\%\ of the total stellar mass reported by W06 (5$\times$10$^9$\,M$_{\odot}$) for the whole galaxy. This stellar mass is a lower limit if we consider that the collision was 300\,Myr ago as suggested by W06, and that our sample is not complete beyond 125\,Myr ago. However, we found an
error in the conversion of the 2MASS fluxes to ABmag presented by W06. We re-calculated the photometry and masses (see Table~\ref{newphot}) and obtain a total stellar mass of 2.8$\times$10$^{10}$\,M$_{\odot}$\ for NGC\,922 based on the K-band, which makes the corrected mass originally created in clusters to be 2.5\%\ of the total stellar mass. We concluded that the integrated cluster mass produced in this starburst is not an unreasonable fraction of the total mass.
\begin{deluxetable}{lcc}[!h]
\tabletypesize{\scriptsize}
\tablecaption{\label{newphot} Corrected 2MASS Photometry for NGC\,922 and S2}
\tablehead{
\colhead{} & \colhead{NGC\,922} & \colhead{S2}}
\startdata
NUV (AB mag) & 13.04 & 15.98 \\
R (AB mag) & 11.57 & 14.71 \\
J (ABmag) & 11.57 & 14.87 \\
H (AB mag) & 11.49 & 14.89 \\
K (AB mag) & 11.82 & 14.80 \\
M$_{\star}$ (M$_{\odot}$) & 2.8$\times$10$^{10}$ & 1.81$\times$10$^9$ \\
\enddata
\tablecomments{All values have been corrected for extinction as in W06.}
\end{deluxetable}
\begin{figure*}[!t]
\epsscale{0.72}
\plotone{pellerin_fig11_astroph.eps}
\caption{Large-scale star-forming complexes are labeled on the F547M image of NGC 922.}\label{figbig}
\end{figure*}
The simulated star clusters younger than 100\,Myr are located preferentially in the ring. This result confirms that the high-density, high-pressure ring formed after the NGC\,922-S2 collision is the ideal site for star cluster formation, and is consistent with the observations.
Also, according to the present simulation, NGC\,922 and the companion collided about 150\,Myr ago which triggered a starburst in the central region of NGC\,922. This burst population does not show a ring-like distribution and is thus responsible for a more dispersed distribution for star clusters with ages between 100 and 200 Myr. This population is not obvious in the HST images, but this is related to the detection limit of our sample or cluster dissolution, most of the older clusters being seen either within the nuclear region or the ring. However, in contradiction to the models, the observations show that star clusters also formed inside the ring within the last 100\,Myr. Those were possibly formed along some spoke features, as shown in other collisional ring simulations \citep{hern93,apple96,struck97}.
\section{Large-Scale Star Formation}
\label{large}
Large-scale star formation in galaxies provides insight into global dynamical processes. In normal spiral galaxies, the largest star-forming units are kpc-size complexes which incorporate many smaller clumps or clusters. The sizes of these largest complexes scale with the absolute magnitude and the diameter of the galaxy \citep[][hereafter EELS]{eels}. They are governed by a Jeans instability \citep{elm94a,elm94b}, which often leads to rather regular spacing along spiral arms \citep{elm83} and in circumnuclear rings \citep{buta93,maoz96,elmetal97}. The gas mass of H\,{\sc{i}} supercloud complexes is on the order of $10^7$\,M$_{\odot}$\ for spiral arms \citep{elm87} and the total stellar mass upper limit is less than $10^6$\,M$_{\odot}$\, consisting of several embedded clusters; in circumnuclear rings, the hotspot stellar masses are typically a few times $10^5$\,M$_{\odot}$\ \citep{elmetal97}. Complexes are also the smallest structures that can be resolved in high redshift galaxies, so these star-forming units are useful to examine in NGC 922 for comparison with ring galaxies in the early universe.
\subsection{Star-forming complexes}
\label{complexes}
Complexes in NGC 922 were identified by visual inspection of the WFPC2 images. Figure \ref{figbig} shows the F547M image with the star-forming complexes labeled; these encompass the smaller-scale clusters discussed in sections~\ref{pop} to \ref{model}. These complexes, which range from 750 pc to 2 kpc in diameter, are also prominent in the SINGG H$\alpha$ image of NGC 922 (see Fig.~\ref{cmyk}). In nearby quiescent spiral galaxies, a late-type galaxy with NGC 922's absolute magnitude of M$_B$=-21.24 \citep{meurer06} would be expected to have a maximum complex size of $\sim$1\,kpc (see EELS).
Photometry was done on the complexes in each of the WFPC2 images by defining boxes around the complexes and their brightest subunits and using the IRAF task {\it imstat}. The results were compared with evolutionary models of an instantaneous burst, as in section 3. The stellar population synthesis results are presented in Table~\ref{synclump}. The complexes have ages ranging from 4 to 86\,Myr and masses from 9$\times$10$^5$ to 3$\times$10$^7$\,M$_{\odot}$.
\begin{deluxetable*}{cccccccccc}
\tabletypesize{\scriptsize}
\tablecaption{\label{synclump} Physical Parameters of Star-Forming complexes}
\tablehead{
\colhead{Complex ID} & \colhead{RA} & \colhead{DEC} & \colhead{F300W} & \colhead{F439W} &
\colhead{F547M} & \colhead{F814W} & \colhead{Age} & \colhead{Log(M$_{Stellar}$)} & \colhead{Diameter} \\
\colhead{} & degrees & degrees & mag & mag & mag & mag & Myr & {M$_{\odot}$} & kpc
}
\startdata
1 & 36.272875 & -24.798485 & 15.49$\pm$0.07 & 16.05$\pm$0.07 & 16.6$\pm$0.1 & 17.4$\pm$0.2 & 4$^{+50}_{-4}$ & 6.7$^{+1.1}_{-0.3}$ & 0.76 \\
2 & 36.275468 & -24.795138 & 16.72$\pm$0.04 & 17.01$\pm$0.03 & 17.46$\pm$0.08 & 18.1$\pm$0.1 & 7$^{+92}_{-6}$ & 6.3$^{+1.2}_{-0.2}$ & 0.71 \\
3 & 36.275073 & -24.790066 & 16.00$\pm$0.01 & 15.97$\pm$0.01 & 16.42$\pm$0.03 & 17.02$\pm$0.04 & 54$^{+294}_{-50}$ & 7.6$^{+1.5}_{-1}$ & 1.04 \\
5 & 36.274900 & -24.786761 & 17.39$\pm$0.02 & 17.67$\pm$0.02 & 18.13$\pm$0.05 & 18.70$\pm$0.07 & 7$^{+291}_{-6}$ & 6.0$^{+2.3}_{-0.1}$ & 0.43 \\
6 & 36.274171 & -24.783779 & 17.17$\pm$0.01 & 17.25$\pm$0.02 & 17.65$\pm$0.04 & 18.19$\pm$0.07 & 54$^{+303}_{-50}$ & 7.1$^{+1.4}_{-1}$ & 0.6 \\
8-9 & 36.272113 & -24.784431 & 16.17$\pm$0.02 & 16.13$\pm$0.02 & 16.52$\pm$0.04 & 17.05$\pm$0.07 & 54$^{+50}_{-308}$ & 7.5$^{+1.4}_{-1.1}$ & 0.92 \\
11 & 36.266513 & -24.800236 & 16.63$\pm$0.02 & 16.96$\pm$0.02 & 17.45$\pm$0.04 & 18.07$\pm$0.06 & 7$^{+80}_{-6}$ & 6.3$^{+1.4}_{-0.4}$ & 0.78 \\
12-13 & 36.268845 & -24.801279 & 16.78$\pm$0.02 & 17.02$\pm$0.02 & 17.52$\pm$0.04 & 18.13$\pm$0.06 & 7$^{+279}_{-3}$ & 6.2$^{+2.3}_{-0.2}$ & 0.76 \\
14 & 36.270572 & -24.800231 & 16.77$\pm$0.01 & 17.02$\pm$0.01 & 17.53$\pm$0.03 & 18.14$\pm$0.05 & 7$^{+290}_{-6}$ & 6.2$^{+2.4}_{-0.3}$ & 0.71 \\
15 & 36.256402 & -24.797122 & 16.34$\pm$0.01 & 16.47$\pm$0.01 & 16.85$\pm$0.03 & 17.43$\pm$0.04 & 54$^{+268}_{-50}$ & 7.4$^{+1.4}_{-1.2}$ & 1.19 \\
16 & 36.264253 & -24.794724 & 16.26$\pm$0.005 & 16.05$\pm$0.008 & 16.47$\pm$0.02 & 17.05$\pm$0.02 & 54$^{+450}_{-50}$ & 7.5$^{+1.5}_{-1.3}$ & 1.19 \\
17 & 36.265672 & -24.793388 & 16.08$\pm$0.005 & 16.52$\pm$0.008 & 16.88$\pm$0.02 & 17.42$\pm$0.03 & 87$^{+456}_{-82}$ & 7.5$^{+1.4}_{-1.6}$ & 0.91 \\
18 & 36.262733 & -24.790521 & 17.51$\pm$0.005 & 17.36$\pm$0.007 & 17.72$\pm$0.02 & 18.28$\pm$0.02 & 54$^{+337}_{-50}$ & 7.0$^{+1.6}_{-1.4}$ & 0.67 \\
\enddata
\tablecomments{The RA and DEC positions are in J2000 and approximations and are not necessarily related to the exact photometric center of the sources.\\}
\end{deluxetable*}
\subsection{Comparison with High Redshift Collisional Ring Clumps}
\label {goods}
The largest complexes in NGC 922 are close to the limit of resolved star-forming regions in galaxies in the Hubble Ultra Deep Field \citep[UDF][]{beck06} out to redshifts of about 4-5; W06 noted that several clump cluster galaxies in the UDF \citep{elm05} resemble NGC 922 in having outer rings. Galaxies in the GEMS and GOODS fields at redshifts 0.6 to 1.4 also include possible collisional ring galaxies \citep{elm06}. Two UDF galaxies, with catalog numbers 5190 and 9159, have large-scale star-formation in their outer ring-like structures \citep[see Fig.~1 in][]{elm05}. Their photometric redshifts are 1.25 and 1.45, respectively. They each have a diameter of about 10 kpc, and the complexes range from 0.5 to 2 kpc. The largest complex is about 13\% of the galaxy size in UDF5190 and 20\% in UDF9159, which is 3 times larger than the average for nearby late-type spirals but about the same as the ratio for NGC922, 15\% based on its diameter of 13.4 kpc (W06).
The absolute complex sizes are about the same in the distant galaxies as in NGC 922, but the distant galaxies are physically smaller.
Photometric data for the clumps in these two galaxies were compared with evolutionary models that included bandshifting effects, interstellar reddening, and intergalactic H\,{\sc{i}} absorption \citep[see details in][]{elm05,elm06}. The clump ages are about $10^8$\,yr, and their masses range from 2$\times10^7$ to $2\times$10$^8$\,M$_{\odot}$. Thus, they are fairly similar to the largest complexes in NGC 922.
A GOODS galaxy with a well-defined outer ring of star formation and an off-center nucleus is shown in Figure \ref{figMAST} (left) in color from the MAST archives\footnote{http:/archive.stsci.edu}. It is number 21238 in the COMBO-17 catalog, with a redshift of 0.664 \citep{wolf03}. Star-forming complexes are identified in Figure~\ref{figMAST} (right). Number 1 is the nucleus; 2 is the brightest star-forming region on the ring, and other ring clumps are 3-13. Features 14-16 are ``feathers'' on the western side, reminiscent of the plume in NGC 922, and 17 is a blue clump. Photometry was done on the features in the BVIz images. Figure \ref{figcmd} shows their color-magnitude diagram of V$_{606}$ versus (V$_{606}$-z$_{850}$). Overlaid are evolutionary tracks for clusters with different masses, as labeled. The open circles along the lines are timesteps, with two times indicated. Most of the clumps are 3-7$\times$10$^7$\,M$_{\odot}$. The feathers are in the same range. The blue dot, Feature 17, is much less massive, closer to 1$\times$10$^7$\,M$_{\odot}$. The nucleus is about 5$\times$10$^8$\,M$_{\odot}$. The ages of the ring clumps are 80 to 150\,Myr. The clumps on the southeast side of the ring are systematically redder than those on the northwest, possibly due to some reddening.
\begin{figure*}
\epsscale{1.0}
\plotone{pellerin_fig12_astroph.eps}
\caption{Left: Collisional ring galaxy COMBO-17 21238 from the Hubble GOODS field number 12. Right: The complexes and features are labeled. See section~\ref{goods} for a detailed identification of the morphological features. The image is 3$\arcsec$ in length, or about 25\,kpc.\\}\label{figMAST}
\end{figure*}
Other GEMS and GOODS ring galaxies at redshifts z=0.7 to 1.4 have ring clumps that are about 0.1 times the galaxy diameter, with ages of a few$\times10^8$\,yr. Their masses range from several$\times10^6$\,M$_{\odot}$\ for the lower redshifts to a few$\times10^8$\,M$_{\odot}$ for the higher redshifts \citep{elm06}. These masses are slightly smaller than the giant clumps found in several non-ring clump cluster galaxies at redshift z=1.8 to 3.0, whose masses average 6$\times10^8$ M$_{\odot}$ \citep{elm05}.
\section{The Tidal Plume and Diffuse Emission of NGC\,922}
\label{tail}
The three-color image of NGC\,922 (Fig.~\ref{cmyk}) shows a faint yellow-red tail on the western side of the nucleus, pointing roughly toward the companion S2. This structure appears clearly in the WF2 chip. A smooth and unresolved stellar population comprises the tail; any clusters present are below the detection limit in U and B.
As was done for the nucleus and the two large clusters and nucleus (Section~\ref{data}), we made photometric measurements for the tail (see Table~\ref{complex}) in the area shown by a polygon in Figure~\ref{morph}. There was no detection in the F300W image. The photometric uncertainties depend mainly on the size of the polygon.
Our population synthesis of the plume indicates stellar ages of 1 to 5\,Gyr, and a total stellar mass of the order of 7$\times$10$^8$\,M$_{\odot}$. The age found for the tail strongly suggests either that it was part of the main disk of NGC\,922, and was pulled away by S2 after the collision or it was part of S2 and was ripped off during the encounter. Based on the I-band residual image, we can see the tail extending up to 7-8\,kpc from star forming regions in the disk. If the stars come from NGC\,922's disk, this implies that the most distant stars have a projected linear velocity of about 20-30\,km\,sec$^{-1}$. Even though the material in the tail was subject to intense mixing and compressions, it did not produce significant star formation. This suggests that the plume may have formed from an ISM-free portion of the galaxy (i.e. bulge or thick disk).
\begin{figure*}
\epsscale{0.7}
\plotone{pellerin_fig13.eps}
\caption{Color-magnitude diagram showing the complexes and feathers in COMBO-17 21238, with evolutionary tracks overlaid showing cluster ages and masses.}\label{figcmd}
\end{figure*}
In addition to the tidal plume, another diffuse emission component is observed in the whole disk of NGC\,922 after removing the clusters from the images. We calculated the fraction of diffuse light and found that 79.2, 89.5, 93.0, and 94.1\% of the light was produced by this diffuse component for the U, B, V, and I filters, respectively. This indicates the fraction of light in clusters is more important at shorter wavelengths. The fraction values are also consistent with the work of \citet{meurer95} where they found that about 80\% of the UV light is found outside of star clusters in nearby starburst galaxies. The diffuse fraction is known to be higher in normal galaxies \citep{larsen00}. This indicates that NGC\,922 is more a starburst type of galaxy, i.e. with enhanced cluster formation activity. Synthesis modeling suggests an age of 250$\pm$40\,Myr for the diffuse emission, indicating that the diffuse stellar component was mainly formed before the collision $\sim$150\,Myr ago. We obtained that log(M$_{Stellar})=10.0\pm0.1$\,M$_{\odot}$\ for the diffuse emission, although the use of an instantaneous burst model might be less appropriate in this case.
\section{The Fate of NGC\,922 and its Star Clusters}
\label{fate}
According to general dynamical models, the ring in NGC\,922 will eventually be fragmented into several massive clumps, which would resemble high-redshift complexes and substructures. Its galactic disk will become thicker and it will evolve in morphology to become an S0.
On a smaller scale point-of-view, \citet{bur05} proposed that collisional ring galaxies were excellent candidates for producing faint fuzzy clusters. They are known to be old, relatively large and non-compact clusters \citep[8-15\,Gyr, 7-15\,pc][]{lar00,bro02}. \citet{elm08} showed that they are actually the lucky survivors of star formation in regions where the gas density was barely enough to produce bound clusters but where the environment also presented low tidal forces so the clusters were not torn apart. The author explains that such environments can exist in the molecular clouds of outer galactic regions that are highly shocked such as in the expanding collisional rings of star formation. Consequently, the most massive older clusters (e.g. clusters $>$50\,Myr in Fig.~\ref{lumage}) observed in NGC\,922 are excellent candidates for progenitors of faint fuzzy clusters.
\section{Conclusions}
\label{conclusion}
In this work, we presented the four-band HST/WFPC2 images of NGC\,922, one of the nearest known collisional ring galaxies. We performed PSF photometry and detected 3817 stellar clusters in at least three bands, as well as two other clusters bigger than the PSF. We used the stellar population synthesis model Starburst99 to estimate ages and masses for each observed clusters. We also simulated the dynamical and star-forming events related to the collision between NGC\,922 and its companion S2.
We found that most observed clusters are 7\,Myr old, or younger. Those clusters are mainly located within the ring or the bar, consistent with H$\alpha$ emission and in agreement with the simulation. A large fraction of clusters older than 50\,Myr are found in the nuclear region, which is consistent with the model producing many older clusters (100-200\,Myr) in the bulge. Those also tend to be massive, according to single stellar population modeling, but are probably contaminated by an older underlying stellar population which would then bias the age and mass estimation.
We also observed a stellar tidal plume pointing in the direction of the companion. We did not detect clusters in the plume. We performed integrated photometry for this plume and found that it is consistent with stellar populations of 1 to 5\,Gyr old. This suggests that the material was stripped off from NGC\,922 by S2 and did not produce significant star formation related to the collision.
The cluster luminosity function observed for NGC\,922 has a slope typical of young cluster populations. The age distribution of star clusters shows a decreasing slope, consistent with cluster dissolution and disruption. By comparing the N-body/SPH star formation simulation to the observed age distribution corrected for cluster disruption (MI and MD models), we find that the MI model is over-correcting for dissolution while the MD model is under-correcting in the 10-100\,Myr age range. In any case, cluster dissolution and disruption is a plausible physical process to consider in order to reach good agreement between the simulation and observation.
We found a significant number of massive ($>$10$^5$\,M$_{\odot}$) star clusters in NGC\,922 that are old enough ($>$50\,Myr) to have survived the supernova explosion phase. Since some of them are likely to be born in highly shocked clouds and that they are now located at high galactic radii where the tidal disruption forces are presumably low, those clusters are excellent candidate progenitors of faint fuzzy clusters.
NGC\,922 has a morphology similar to several distant galaxies observed in the GEMS and GOODS fields. We detected about a dozen big star-forming complexes in the ring of NGC\,922. A direct comparison with the knots observed in a ring galaxy from the GOODS field (\#21238 in COMBO-17, z$=$0.7) and other high redshift ring galaxies shows that NGC\,922's complexes are of comparable sizes and masses even though the distant galaxies are physically smaller.
\acknowledgments
AP is grateful to Barry Rothberg for support with HST/WFPC2 data and to R. Chandar for helpful scientific comments. This work was supported by the HST grant HST-GO-11112. GRM was also supported by NASA LTSA grant NAG5-13083. KB acknowledges financial support from the Australian Research Council (ARC). The numerical simulations reported here were carried out on SGI supercomputers at Australian Partnership for Advanced Computing (APAC) in Australia.
|
\section{Introduction}
In the present paper we call \textit{ordered} spirals those that have
as building block a set of stable periodic orbits. The standard
example are the spirals in Contopoulos \& Grosb{\o}l (1986) models,
which end at their inner 4:1 resonance. In this type of models the
spiral structure terminates at a distance beyond which there are no
more quasiperiodic orbits trapped around stable, elliptical periodic
orbits of the x$_1$ family. On the other hand we will call
\textit{chaotic} spirals those that we believe are constituted from
stars in chaotic motion. Kaufmann \& Contopoulos (1996) argued for the
first time that part of the spirals in barred-spiral systems with a
single pattern speed are chaotic. The galactic morphologies we study
are grand design independently of whether their spirals are ordered or
chaotic. We use response models in trying to model these
morphologies. The potentials we use are either analytic forms that
describe the structures we study, or potentials that have been
directly determined from near-infrared observations of specific
galaxies.
\begin{figure*}[t!]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[scale=0.55]{patsis_f01.ps}}
\end{center}
\caption{\footnotesize
The flow of stars in a time-independent stellar response model of
Contopoulos \& Grosb{\o}l (1986) type. (a) The response
morphology. Corotation is at $\approx$ 24~kpc and at the end of the
strong spirals, at $\approx$ 12.5~kpc, is the inner 4:1
resonance. (b) The velocity field of the model indicated by arrows.}
\label{normal}
\end{figure*}
\section{Normal spiral galaxies}
Normal spiral galaxies have no or very weak bar components. Orbital
models (Patsis et al. 1991) and gaseous response models (Patsis et
al. 1994, 1997a) have shown that open, grand design, \textit{normal}
spiral galaxies are best modeled by slowly rotating patterns. Then
corotation is beyond the end of the spiral arms, or at least beyond
the end of a characteristically bisymmetric part of the arms. Since
the galaxies we refer to are grand design, their symmetric part
dominates their morphology. Nevertheless, especially in gaseous
response models, bifurcations of the arms appear at the end of the
symmetric part. Also off-phase extensions of the spirals with respect
to the imposed potential minima may appear beyond the point at which
symmetry breaks. The inclusion of an $m=1$ component in phase with the
main $m=2$ spiral and with the same pattern speed, is crucial for
modeling the observed structures (see Fig.~4 in Patsis et
al. 1997a). The point at which the symmetric part of the spirals ends
can be used for the estimation of their pattern speed, since it is
associated with the inner 4:1 resonance (Contopoulos \& Grosb{\o}l
1986). Very frequently we observe there a characteristic bifurcation
of the arms.
In the last 12 years, stellar and gaseous response models for normal
spirals, or for spirals that have their own pattern speeds in
barred-spiral systems, consistently put corotation beyond the end of
the symmetric part of the spirals and in some cases even beyond the
end of the spiral structure altogether (Mulder \& Combes 1996; Kranz
et al. 2001; Bissanz et al. 2003; Kranz et al. 2003; Pichardo et
al. 2003; Martos et al. 2004; Martos \& Yanez 2005; Vorobyov 2006;
Vorobyov \& Shchekinov 2006; Minchev \& Quillen 2008; Patsis et
al. 2009). All these models strongly indicate that the flow in normal
spirals follows `precessing ellipses'. Since the arms extend inside
corotation, they are located in a disk region, where order dominates
and thus structures can be built by quasi-periodic orbits trapped
around stable periodic orbits.
Fig.~\ref{normal} shows a typical stellar response model to an imposed
bisymmetric potential of the type used by Contopoulos \& Grosb{\o}l
(1986). Only an $m=2$ term is included. Corotation is at 24~kpc. The
main spirals end at their inner 4:1 resonance, inside corotation
(Fig.~\ref{normal}a) and the flow of the stars is along `precessing
ellipses' (Fig.~\ref{normal}b).
\begin{figure*}[t!]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[scale=0.7]{patsis_f02.ps}}
\end{center}
\caption{\footnotesize
(a) An $R$-band image of NGC~4314 showing the boxiness of the outer
isophotes of the bar. We observe the limited azimuthal extent of the
spirals. (b) A stellar response model of the Quillen et al. (1994)
potential. The pattern rotates with $\Omega_{\rm p}=38.23$~km~s$^{-1}$~kpc$^{-1}$. The
model reproduces the basic morphological features of the galaxy.}
\label{n4314}
\end{figure*}
\section{Barred spiral galaxies}
Bars end close, but before, corotation (Contopoulos 1980). In
barred-spiral systems with one pattern speed and the observed spiral
arms attached to the ends of the bar, the spirals have to cross the
corotation region where chaos dominates. We present the orbits that we
find to support such a spiral structure in two types of barred-spiral
potentials. Both of them result from near-infrared observations of real
galaxies. The values of the pattern speeds ($\Omega_{\rm p}$) we use
in the examples we present below does not necessarily give the best
fit of the galaxy morphology in all cases. We just use template models
to present the orbital behavior in barred-spiral systems with boxy
bars and with bars with ansae morphology.
\subsection{Boxy bars}
We call `boxy' the bars that have, on the plane of the galaxy, outer
isophotes with rectangular-like shape. A standard example is the
nearly face-on, early type barred-spiral galaxy NGC~4314. In
Fig.~\ref{n4314}a we give an R-image (Gadotti \& de Souza 2006) with
overplotted isophotes showing the boxiness in discussion. The spiral
structure appears as a continuation of the bar and is confined
azimuthally in an angle less than $\pi/2$. Patsis et al. (1997b), have
shown that the stellar orbits associated with the observed boxy
isophotes are chaotic. The potential used for this study has been
estimated by Quillen et al. (1994) from the $K$-image of
NGC~4314. Stellar responses are similar for a range of $\Omega_{\rm
p}$ values $38<\Omega_{\rm p}<45$~km~s$^{-1}$~kpc$^{-1}$. The pattern in the response
model of Fig.~\ref{n4314}b rotates with $\Omega_{\rm
p}=38.23$~km~s$^{-1}$~kpc$^{-1}$. Patsis (2006) has shown that the chaotic orbits that
reinforce the outer boxy isophotes of the bar are practically the
same, which sustain also the spirals that emerge from the ends of the
bar. They belong to a population that visits both the bar and the disk
area further out. During the time these chaotic orbits visit the bar
region, they have a morphological resemblance with quasi-periodic
orbits trapped around rectangular-like 4:1 type stable periodic orbits
we find in generic barred potentials (Contopoulos 1988). However, in
the potential we study, this family is stable over a tiny energy
interval, and the space their invariant curves occupy in the phase
space is tiny (Patsis et al. 1997b). In Fig.~\ref{orbits} we give
characteristic examples of orbits of particles that are located on the
spiral arms in Fig.~\ref{n4314}b.
\begin{figure*}[t!]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[scale=0.4]{patsis_f03.ps}}
\end{center}
\caption{\footnotesize
Chaotic orbits of particles that support the spiral structure we
observe in Fig.~\ref{n4314}. During the time they spend in the bar
region they have a rectangular-like morphology, typical of orbits at
the 4:1 region of bars. They are integrated for 10 bar periods.}
\label{orbits}
\end{figure*}
\subsection{Bars with ansae}
Another type of barred-spiral morphology we have studied has an ansae
type bar. The potential we used in this case is from an estimation of
the potential of NGC~1300, from a $K$-band image of this galaxy
(Kalapotharakos et al. 2008). For the figures we present below we have
used as inclination and position angles for the bar $42\fdg2$ and
$100\fdg6$ respectively. The $K$-band image of the galaxy deprojected
with these values is given in Fig.~\ref{n1300bar}.
\begin{figure}[t!]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[scale=0.7]{patsis_f04.ps}}
\end{center}
\caption{\footnotesize
The deprojected $K$-band image of the barred-spiral galaxy NGC~1300,
which we used for our calculations in potentials with ansae type
bars.}
\label{n1300bar}
\end{figure}
It is interesting that in these models we find qualitatively different
effective potentials for nearby values of $\Omega_{\rm p}$. In
Fig.~\ref{effpot} we give the isocontours of two effective potentials
we studied. In Fig.~\ref{effpot}a, $\Omega_{\rm p}$=21~km~s$^{-1}$~kpc$^{-1}$, while in
(b) $\Omega_{\rm p}$=26~km~s$^{-1}$~kpc$^{-1}$. In the case with $\Omega_{\rm
p}$=26~km~s$^{-1}$~kpc$^{-1}$\ the model has two sets of unstable Lagrangian points, and
the ansae morphology appears also in the isocontours. Nevertheless,
when $\Omega_{\rm p}$=21~km~s$^{-1}$~kpc$^{-1}$, we have a more conventional
configuration with two unstable Lagrangian points close to the ends of
the bar. The models are asymmetric, because in the Fourier analysis we
have taken into account also the odd terms (Kalapotharakos et
al. 2008). Despite the differences these two models give similar
stellar responses. Again here we consider the orbits that support the
spirals found. They are chaotic and their typical morphologies can be
seen in Fig.~\ref{13orbits}. As in the previous barred-spiral case we
have orbits that visit both the bar region as well as the disk beyond
its end. During the time they stay at the bar region they develop a
morphology which again has a 4:1 resonance character. However, this
morphology is not rectangular-like, but like another type of 4:1
orbits. These are elliptical-like with loops at their apocentra
(Contopoulos 1980; Patsis et al. 1997b). These chaotic orbits shape
also the ansae, performing a number of loops at the corresponding
regions.
\section{Other cases}
\subsection{Barred galaxies with extended spiral structure}
In both barred-spiral galaxies we have studied above, the spiral arms
are radially and/or azimuthally confined to short
distances. Especially in NGC~1300, the spiral structure is very
asymmetric. The right arm, as depicted in Fig.~\ref{n1300bar},
practically breaks in two parts. In NGC~4314, the extent of the
spirals azimuthally does not exceed $\pi/2$. This is a morphology that
differs from the logarithmic spirals of nearly normal grand design
spiral galaxies, where we do not observe major gaps (Grosb{\o}l \&
Patsis 1998). A case of a barred-spiral system with a radially
extended spiral structure without major gaps and discontinuities along
the arms, has been studied by Boonyasait et al. (2005) and Patsis et
al. (2009) using a potential estimated from near-infrared observations
of NGC~3359. For these `well organized' spiral arms the best matching
of the flow at their region has been obtained by a
`precessing-ellipses' flow. The spiral structure in the barred-spiral
galaxy NGC~3359 is supported by regular orbits. This is an indication
that there is a correspondence between morphological features of the
arms (extent, presence of gaps etc.) and the dynamical mechanism that
sustains them in each individual case.
\begin{figure*}[t!]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[scale=0.45]{patsis_f05.ps}}
\end{center}
\vspace*{-0.5cm}
\caption{\footnotesize
Equipotentials of effective potentials in models for NGC~1300. In
(a) $\Omega_{p}$=21~km~s$^{-1}$~kpc$^{-1}$\ and in (b) $\Omega_{p}$=26~km~s$^{-1}$~kpc$^{-1}$. Heavy dots
indicate the unstable Lagrangian points in the two models.}
\label{effpot}
\end{figure*}
\subsection{Normal double spirals}
As we have noted in Sect.~2, in both gaseous and stellar response
models of normal open spirals we find the density response maxima
along the imposed potential minima only up to the inner 4:1
resonance. Beyond that distance we find in some cases off-phase
extensions between the 4:1 resonance and corotation in gaseous
models. Nevertheless, if we fine tune our models, so that we have a
perturbing force of the order of 10\% of the axisymmetric background
at corotation, and populate the region beyond corotation with enough
particles, we observe another set of spirals \textit{beyond}
corotation. These spirals are weak and do not have the pitch angle of
the imposed potential. This is the case depicted in Fig.~\ref{dble}.
The strong spirals we observe in Fig.~\ref{dble}a end at the 4:1
resonance and are supported by regular orbits. The weak spirals beyond
corotation, emerge close to the unstable Lagrangian points of the
system and follow a chaotic flow, which is shown in Fig.~\ref{dble}b.
Double spirals in response models with a gap at the corotation region
are obtained also by Vorobyov (2006). These morphologies are not
frequently observed. However, grand design spiral galaxies like
NGC~1566 and NGC~5248 show this morphology and are candidates of
hosting both ordered and chaotic spirals.
\section{Conclusions and discussion}
The orbital response models that we have studied lead us to the
following conclusions:
\begin{figure*}[t!]
\begin{center}
\includegraphics[scale=0.5]{patsis_f06.ps}
\end{center}
\vspace*{-2.3cm}
\caption{\footnotesize
Typical orbits that support barred-spiral morphologies with
ansae. They support both the spirals as well as the ansae
features. They are integrated for 10 bar periods.}
\label{13orbits}
\end{figure*}
\begin{figure*}[t!]
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[clip=true]{patsis_f07.ps}}
\end{center}
\caption{\footnotesize
A model of a double normal spiral. In (a) we observe the inner
strong and the outer weak spiral. In (b) we give the velocity field
with arrows, so that we can see the flow along the spirals beyond
corotation.}
\label{dble}
\end{figure*}
\begin{enumerate}
\item Response models of (nearly) normal
grand design spiral galaxies indicate that the spiral arms in this
type of galaxies rotate slowly. Corotation is beyond the end of the
bisymmetric part of the spirals and the grand design is supported by
regular orbits. There are several published simulations in the
literature that reproduce this result. As soon as the perturbing
forces are larger than 5\% of those of the axisymmetric background,
the nonlinear phenomena described by Contopoulos \& Grosb{\o}l
(1986) are present in the models. The final response morphology is
obtained after a few pattern rotations. This means that these models
can predict galactic morphologies that last for a few Gyr.
\item Barred-spiral systems with a single pattern speed
may have spirals that are supported by chaotic orbits. In the cases
we have studied and we have found that the spirals are supported by
particles in chaotic motion (NGC~4314, NGC~1300), the spiral arms
are either confined azimuthally in their extent, or asymmetric with
gaps. Contrarily in a barred-spiral galaxy with a well described
spiral structure (NGC~3359) our results are in agreement with a flow
of regular orbits. The conditions under which barred-spiral systems
have ordered or chaotic spirals needs further investigation. We
mention the work of other groups based on $N$-body models and
orbital theory that present models of barred-spiral systems with a
well developed spiral structure consisting entirely by particles in
chaotic motion (see Efthymiopoulos et al. 2008 and references
therein, as well as Romero-Gomez et al. 2008 and references
therein).
\item The chaotic orbits that support the spirals
belong to a population that visits both the bar and the disk region
beyond corotation. In the two examples we studied we find that they
are the same orbits that shape the outer structure of the bars. They
support the boxy and the ansae type morphologies of our models. This
means that they belong to a very important orbital population for
understanding the shapes of the bars. On a surface of section we
find the initial conditions in a chaotic sea, that support the bar
if integrated for a certain number of bar revolutions (of the order
of 10). During the time these orbits stay at the bar region, they
follow trajectories with a morphology similar to those of 4:1
resonance regular orbits trapped around stable periodic orbits. This
gives to the bars shapes resembling periodic orbits at the 4:1
resonance region.
\end{enumerate}
\begin{acknowledgements}
PAP thanks ESO for an invitation to visit ESO Garching for two months
during the summer of 2008, where this work has been completed. We
thank Prof. G. Contopoulos for valuable comments.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
Many interesting problems in knot theory involve the question of when one knot complement can be mapped onto another, or at the algebraic level, when the fundamental group of one knot complement can be mapped onto the fundamental group of another. For example, Simon's Conjecture asserts that the fundamental group of any knot in $S^3$ can surject onto only finitely many knot groups. This conjecture has been established for 2-bridge knots by Boileau, Boyer, Reid and Wang \cite{BBRW:2009}. On the other hand, given any 2-bridge knot $K_{p/q}$, Ohtsuki, Riley, and Sakuma \cite{ORS} show how to construct infinitely many 2-bridge knots $K_{p'/q'}$ whose complements map onto the complement of $K_{p/q}$. Their construction induces an epimorphism $\phi: \pi_1(S^3-K_{p'/q'}) \to \pi_1(S^3-K_{p/q})$ of knot groups which {\it preserves peripheral structure}, that is, $\phi$ takes the subgroup generated by a longitude and meridian of $K_{p'/q'}$ into a conjugate of the subgroup generated by a longitude and
meridian of $K_{p/q}$.
For any pair of knots, not just 2-bridge knots, Silver and Whitten define a partial ordering $\ge$ on
knots in $S^3$ by declaring that $K_1\ge K_2$ if there exists
an epimorhism $\phi: \pi_1(S^3 - K_1) \to \pi_1(S^3 -
K_2)$ which preserves peripheral structure \cite{SW:2006}. Thus, for a given 2-bridge knot $K_{p/q}$, the Ohtsuki-Riley-Sakuma (ORS) construction gives infinitely many 2-bridge knots $K_{p'/q'}$ such that $K_{p'/q'} \ge K_{p/q}$. This leads naturally to the following question posed in \cite{ORS}.
\begin{question}\label{ORS question}(Ohtsuki-Riley-Sakuma) Is every pair of 2-bridge knots $(K_{p'/q'}, K_{p/q})$ with $K_{p'/q'} \ge K_{p/q}$ given by the ORS construction?
\end{question}
Gonz\'{a}lez-Acu\~{n}a and Ram\'{i}rez \cite{GR:2001} give an affirmative answer to Question~\ref{ORS question}
in the case where $K_{p/q}$ is a 2-bridge torus knot. In this paper we give additional evidence that the answer to Question~\ref{ORS question} is yes in general. Numerous computations with non-torus knot ORS pairs $(K_{p'/q'}, K_{p/q})$ with $K_{p'/q'} \ge K_{p/q}$ suggest that in any such pair the larger knot $K_{p'/q'}$ must be sufficiently complex with respect to a variety of knot invariants such as: crossing number, degree of the Alexander polynomial, degree of the A-polynomial, number of distinct boundary slopes, etc. Thus, if the answer to Question~\ref{ORS question} is yes, then knots with ``small'' complexity should be minimal. Here, a knot $K_1$ is said to be {\em minimal} with respect to the Silver--Whitten partial ordering if $K_1\ge K_2$ implies that either $K_2=K_1$, $K_2$ is the mirror image of $K_1$, or $K_2$ is the unknot. Throughout the remainder of this paper, we will consider two knots $K_1$ and $K_2$ to be equivalent, $K_1 \equiv K_2$, if either $K_2=K_1$ or $K_2$ is the mirror image of $K_1$.
One of the themes of this paper will be to look at the set of distinct boundary slopes of a knot. These are easily computed for 2-bridge knots due to the classification of Hatcher and Thurston \cite{HT:1985}.
Our first main result gives a lower bound on the number of distinct boundary slopes for the larger knot in an ORS pair.
\setcounter{section}{4}
\begin{theorem}
If $K_{p'/q'} \ge K_{p/q}$ is a nontrivial ORS pair, then either
\begin{enumerate}
\item[{\it i}.] $K_{p'/q'} $ and $K_{p/q}$ are both torus knots and $K_{p'/q'} $ has precisely two distinct boundary slopes, or
\item [{\it ii}.]$K_{p'/q'} $ has at least five distinct boundary slopes.
\end{enumerate}
\label{thm1}
\end{theorem}
\setcounter{section}{1}
Our second main result establishes the minimality of 2-bridge knots with exactly three boundary slopes.
\setcounter{section}{5}
\setcounter{theorem}{0}
\begin{theorem}
If $K_{p/q}$ is a 2-bridge knot with exactly 3 distinct boundary slopes, then $K_{p/q}$ is minimal with respect to the Silver--Whitten partial order.
\label{thm2}
\end{theorem}
\setcounter{section}{1}
Notice that if the answer to Question~\ref{ORS question} is yes, then Theorem~\ref{thm1} implies Theorem~\ref{thm2}. Thus, we may think of Theorem~\ref{thm2} as evidence in support of an affirmative answer to Question~\ref{ORS question}. It is also worth noting that Theorem~\ref{thm1} is sharp. That is, there exist non-minimal 2-bridge knots with 5 distinct slopes. We provide an example of this in Section~4.
This project began as joint work of the first author and two undergraduate students, T\'omasz Przytycki and Rebecca Nachison, in the summer of 2007 at the Claremont Colleges REU program. Thanks are due to T\'omasz and Rebecca for working out most of Theorem~\ref{list of knots with given number of slopes}. We thank the National Science Foundation, the Claremont Colleges, and Pitzer College in particular, for their support of the REU program. The second author would also like to thank the Claremont Colleges for their hospitality and Pitzer College for its support during the completion of this work. Finally, thanks are due to Alan Reid for helpful conversations regarding character varieties.
\section{Preliminaries}
We begin with some notation for 2-bridge knots. Recall that each 2-bridge knot corresponds to a relatively prime pair of integers $p$ and $q$ with $q$ odd. We denote the knot as $K_{p/q}$. Furthermore, $K_{p/q}$ and $K_{p'/q'}$ are ambient isotopic as unoriented knots if and only if $q'=q$ and $p' \equiv p^{\pm 1}\ (\mbox{mod } q)$ (see \cite{BZ:2003} for details). As mentioned earlier, we will not distinguish between a knot $K_{p/q}$ and its mirror image $K_{-p/q}$. Therefore, in this paper we consider two 2-bridge knots $K_{p/q}$ and $K_{p'/q'}$ to be equivalent if and only if $q'=q$ and either $p' \equiv p^{\pm 1}\ (\mbox{mod } q)$ or $p' \equiv -p^{\pm 1}\ (\mbox{mod } q)$. Because of this classification, it is sometimes convenient to assume that $0<p<q$.
The theory of 2-bridge knots is closely tied to continued fractions. We adopt the convention used in \cite{ORS} and define $p/q=r+[b_1, b_2, \dots, b_n]$ as the continued fraction
$$p/q=r+[b_1, b_2, \dots, b_n]=r+\frac{1}{b_1+\displaystyle \frac{1}
{
\begin{array}{ccc}
b_2+&&\\
&\ddots&\\
&&+\displaystyle \frac{1}{b_n}\\
\end{array}
}
}$$
It is well known that if $0<p<q$ then we may write $p/q$ uniquely as $p/q=0+[b_1, b_2, \dots , b_n]$ where each $b_i>0$ and $b_n>1$. Furthermore, we may also assume that $b_1>1$ for the following reason. If $b_1=1$, reversing the order of the partial quotients will give $p'/q=[b_n, \dots,b_3, b_2, 1]=[b_n, \dots, b_3, b_2+1]$ with $0<p'<q$ and $p p'\equiv (-1)^{n+1}\ \text{ (mod $q$)}$. Thus, for any 2-bridge knot $K$, there exists a continued fraction $0+[b_1, b_2, \dots, b_n]$ with $b_i>0, \ b_1>1$, and $b_n>1$ representing $K$ (or its mirror image). In what follows, we will refer to such a continued fraction as {\it strongly positive}. Moreover, we will also refer to a vector $(b_1, b_2, \dots, b_n)$ of positive integers with $b_1 > 1$ and $b_n >1$ as strongly positive. Finally, the negation of such a vector we call {\it strongly negative}.
\begin{figure}[t]
\psfrag{a}{\Huge$b_1$}
\psfrag{b}{\Huge$b_2$}
\psfrag{c}{\Huge$b_3$}
\psfrag{d}{\Huge$b_n$}
\begin{center}
\leavevmode
\scalebox{.4}{\includegraphics{Figures/4platt.eps}}
\end{center}
\caption{A 4-plat diagram for $K_{p/q}$ where $p/q = r+[b_1, b_2, \dots, b_n]$ ($n$ even).}
\label{4plat fig}
\end{figure}
If $p/q=r+[b_1, b_2, \dots, b_n]$, then $K_{p/q}$ can be represented by the standard 4-plat diagram shown in
Figure~\ref{4plat fig}. In each box there are $b_i$ crossings. If $i$ is odd, then the crossings are right-handed if $b_i >0$ and left-handed otherwise. For $i$ even the convention reverses, with $b_i > 0$ corresponding to left-handed crossings and $b_i < 0$ for right-handed. Thus, the 4-plat diagram is alternating precisely when all $b_i$ have the same sign. Notice also that there are two possible ways to ``close up'' the strands at the right end of the braid in Figure~\ref{4plat fig} depending on whether $n$ is even or odd. Figure~\ref{4plat fig} depicts a 4-plat with $n$ even. If $n$ is odd, strings 1 and 2, and strings 3 and 4, connect pairwise at the right end of the braid just as they do at the left end. With this notation, the fraction $1/3=[3]$ gives the right-hand trefoil knot while $2/5=[2,2]$ gives the figure eight knot.
Now suppose that $\phi$ is a homomorphism between knot groups that preserves peripheral structure. Since all nontrivial knots in $S^3$ have Property P \cite{KM:2004}, there are certain restrictions placed on the images of a meridian and longitude of $K_1$.
\begin{theorem} Let $K_1$ and $K_2$ be nontrivial knots with $K_1\ge
K_2$. Then there exists an epimorphism $\phi: \pi_1(
S^3-K_1) \to \pi_1(S^3-K_2)$ and an integer $d$ such that
\begin{eqnarray}
\phi(\mu_1)&=&\mu_2, \mbox{ and}\\
\phi(\lambda_1)&=&\lambda_2^d
\end{eqnarray}
for some choice $\{ \mu_i, \lambda_i \}$ of meridian-(preferred)\,longitude pairs
for $K_i$.
\label{abcdcriteria}
\end{theorem}
\noindent {\em Proof.} Since $\phi$ preserves peripheral structure we
may assume (after conjugation, if necessary) that
\begin{eqnarray*}
\phi(\mu_1)&=&\mu_2^a \lambda_2^b, \mbox{ and}\\
\phi(\lambda_1)&=&\mu_2^c \lambda_2^d
\end{eqnarray*}
for some integers $a, b, c$ and $d$. If we consider the
abelianization map $ab:\pi_1(S^3-K_2) \rightarrow \mathbb Z$,
then $ab(\phi (\lambda_1)) = 0$ since $\lambda_1$ is in the
commutator subgroup. However, $ab(\phi (\lambda_1)) = ab ( \mu_2^c
\lambda_2^d) = c$. This shows that $c=0$. Moreover, $ab(\phi (\mu_1))
= ab(\mu_2^a \lambda_2^b)=a$ must generate $\mathbb Z$ because $ab
\circ \phi$ is onto and $\pi_1(S^3-K_1)$ is generated by
conjugates of $\mu_1$. Therefore, $a = \pm 1$. By replacing $\mu_2$
with its inverse if necessary, we may assume that $a=1$.
Now consider $1/b$ Dehn filling on $K_2$. This filling kills the
image of $\mu_1$ and any of its conjugates, and so kills the entire
group, since $\phi$ is onto. Thus the manifold we obtain after filling is simply connected.
Since $K_2$ is a nontrivial knot with Property P, we must have done
trivial surgery. Therefore, $b=0$.
\hfill $\Box$
Because of Theorem~\ref{abcdcriteria}, it makes sense to write $$K_1\ge_dK_2$$ whenever $K_1\ge K_2$ and there exists an epimorphism $\phi$ as in Theorem~\ref{abcdcriteria} with $\phi(\lambda_1)=\lambda_2^d$. In general, $d$ can take on any integer value. In particular, given the knot $K_2$ and any integer $d$, there exists a knot $K_1$ and epimorphism $\phi$ realizing $d$. The reader is referred to \cite{JL:1989} for entr\'e into this subject. On the other hand, if $K_1$ and $K_2$ are 2-bridge knots, then $d$ must be odd.
\begin{theorem}
\label{2bridge means d is odd}
If $K_{p'/q'}\ge_d K_{p/q}$ where $K_{p'/q'}$ and $K_{p/q}$ are 2-bridge knots and $K_{p/q}$ is nontrivial, then $d$ is odd. In particular, the image $\phi(\lambda_1)$ under the epimorphism $\phi$ cannot be trivial.
\end{theorem}
\noindent {\em Proof.} Suppose $\phi:\pi_1(S^3-K_{p'/q'}) \to
\pi_1(S^3-K_{p/q})$ is an epimorphism taking $\lambda_1$ to $\lambda_2^d$ and let $\rho$ be an irreducible
parabolic representation of $\pi_1(S^3-K_{p/q})$ into $SL(2, \mathbb C)$ (such a representation exists by Riley \cite{R:1984}).
Composing $\rho$ with $\phi$ gives such a representation for $K_{p'/q'}$.
It follows from \cite{HS:2004} that any irreducible parabolic representation of a
2-bridge knot into $SL(2,\mathbb C)$ must take the longitude into the conjugacy class of an
upper triangular matrix with diagonal entries of $-1$. Hence $\phi$
cannot take $\lambda_1$ to an even power of $\lambda_2$.
\hfill $\Box$
The following theorem summarizes some additional properties of the partial ordering which we will refer to in this article. The first two parts are contained in \cite{SW:2006}, the third comes from \cite{BBRW:2009}, and the fourth appears in \cite{BB:2007}.
\begin{theorem}
\label{properties of partial order}
(Silver-Whitten, Boileau-Boyer, Boileau-Boyer-Reid-Wang) Suppose that $K_1 \ge_d K_2$.
\begin{enumerate}
\item[{\it i}.] If $\Delta^{(i)}_K(t)$ is the $i$-th Alexander polynomial of $K$, then $\Delta^{(i)}_{ K_2}(t)$ divides $\Delta^{(i)}_{K_1}(t)$.
\item[{\it ii}.] If $A_K(L,M)$ is the A-polynomial of $K$, then
$A_{K_2}(L,M)$ divides $(L^d-1)A_{K_1}(L^d,M)$.
\item[{\it iii}.] If $X(K)$ is the $SL(2, \mathbb C)$ character variety of $K$, then the induced map $\phi^* : X(K_2) \rightarrow X(K_1)$ is an injective, algebraic and closed (in the Zariski topology) mapping.
\item[{\it iv}.] If $K_1$ is a 2-bridge knot, then $K_2$ is a 2-bridge knot. Futhermore, if $K_1=K_{p'/q'}$ and $K_2=K_{p/q}$, then either $K_1 \equiv K_2$ or $q'=kq$ with $k>1$.
\end{enumerate}
\end{theorem}
Notice that the fact that $q$ divides $q'$ in Theorem~\ref{properties of partial order}({\em iv}) is easily derived from ({\em i}) and the fact that $q=| \Delta_{K_{p/q}}(-1) |$ is the determinant of $K_{p/q}$.
Because of the relationship between the A-polynomial and boundary slopes of a knot, Theorem~\ref{properties of partial order}({\it ii}) enables us to use boundary slopes as a tool to study whether one knot is greater than another. Recall that a {\it boundary slope} $r=\frac{a}{b}$ of a knot $K$ is an element of $\mathbb Q \cup \infty$ such that $K$ contains a properly embedded, incompressible, boundary incompressible surface $S$ in its exterior whose boundary is the curve $ a \mu + b \lambda$ (or multiple copies of this curve). Here $\{\mu, \lambda\}$ is a meridian-(preferred) longitude pair for $K$. The {\it Newton polygon} of the A-polynomial $A_K(L,M)=\sum a_{i\,j}L^i M^j$ is the convex hull of the set of points $\{(i,j) \in \mathbb R^2\, |\, a_{i\, j} \ne 0\}$. If $r$ is the slope of a side of the Newton polygon of $A_K(L,M)$, then it is proven in \cite{CCGLS:1994}, that $r$ is a boundary slope of $K$. A boundary slope of $K$ which appears as a slope of the Newton polygon is called {\it strongly detected}. We shall tacitly assume that the A-polynomial always includes the component $L-1$ corresponding to abelian representations so that $0$ is always a strongly detected slope.
\begin{corollary} Suppose $K_1\ge_d K_2$, $d\ne 0$, and
that $r$ is a strongly detected boundary slope of $K_2$. Then $dr$ is
a strongly detected boundary slope of $K_1$.
\end{corollary}
\noindent {\em Proof.} We establish this result using the following well-known property of Newton polygons. For any polynomial $P(x,y)$ let $S_P$ denote the set of slopes of the sides of the Newton polygon of $P$. Then for all polynomials $P(x,y)$ and $Q(x,y)$, $S_P \cup S_Q = S_{PQ}$ (a proof may be found in \cite{HS:2004}). Now suppose that $r$ is a strongly detected boundary slope of $K_2$. Then $r$ is the slope of a side of the Newton polygon of $A_{K_2}(L,M)$. Since $A_{K_2}(L,M)$ divides
$(L^d-1)A_{K_1}(L^d,M)$ it follows that $r=0$ or $r$ is a slope of a side of
the Newton polygon of $A_{K_1}(L^d,M)$. If $r=0$, then $dr=0$ is a strongly detected slope of $K_1$. Now assume that
$r$ is a slope of a side of the Newton polygon of $A_{K_1}(L^d,M)$. The Newton polygon of $A_{K_1}(L^d,M)$ is obtained from the Newton polygon of $A_{K_1}(L,M)$ by simply replacing every vertex $(i,j)$ with $(i/d,j)$.
Consequently, $dr$
is the slope of a side of the Newton polygon of $A_{K_1}(L,M)$, and therefore, is a strongly detected boundary slope of $K_1$. \hfill $\Box$
By work of Ohtsuki \cite{O:1994}, all boundary slopes of 2-bridge knots are strongly detected. Thus, for 2-bridge knots we have a stronger result.
\begin{corollary}
\label{boundary slopes are subsets}
Suppose $K_1\ge_d K_2$ and that
$K_1$ is a 2-bridge knot. If $\{r_1, r_2,
\dots, r_m\}$ is the set of boundary slopes of $K_2$, then $\{dr_1,
dr_2, \dots, dr_m\}$ is a subset of the set of boundary slopes of
$K_1$.
\end{corollary}
For any knot $K$, the {\it diameter} of its set of boundary slopes is the maximum difference between any two slopes. We denote the diameter by $diam(K)$. For 2-bridge knots, in fact for all alternating Montesinos knots, $cr(K)=diam(K)/2$, where $cr(K)$ is the crossing number of $K$. See \cite{HS:2007}, \cite{IM:2008} and \cite{MMR:2008}. This leads to the following results.
\begin{corollary}
\label{crossing number}
Suppose $K_1$ is a 2-bridge knot and $K_1\ge_d K_2$.
Then $cr(K_1)\ge |d|cr(K_2)\ge cr(K_2)$.
\end{corollary}
Using Theorem~\ref{properties of partial order}({\it iii}) and Corollary~\ref{crossing number} it is now easy to show the following. (This result was first proven by Boileau, Rubinstein, and Wang \cite{BRW:2005}.)
\begin{corollary} A 2-bridge knot can only be greater than or equal
to finitely many other knots.
\end{corollary}
\section{2-bridge knots with four or less boundary slopes}
In preparation for proving Theorem~\ref{thm1}, we determine in this section those 2-bridge knots with small numbers of distinct boundary slopes. In \cite{MMR:2008}, the authors determine necessary but not sufficient conditions on $p$ and $q$ so that $K_{p/q}$ has 4 or less distinct boundary slopes. In order to exploit Corollary~\ref{boundary slopes are subsets} we require a stronger result than what appears in \cite{MMR:2008}. In particular, we need both a complete classification of 2-bridge knots with 4 or less boundary slopes as well as an explicit description of the associated slope sets.
We assume the reader is familiar with Hatcher and Thurston's paper \cite{HT:1985} where a method for computing the boundary slopes of $K_{p/q}$ is given. We will also make use of a method equivalent to theirs which is described in \cite{HS:2007}. Rather than describing the methodology in complete detail, we provide a brief description illustrated by an example.
Recall that the Farey graph may be thought of as the edges of the ideal modular tessellation of $\mathbb H^2$ obtained by starting with the ideal triangle whose vertices are $1/0,\ 0/1$, and $1/1$ and then reflecting it in all possible ways across its edges. The ideal vertices of this tessellation are $\mathbb Q \cup \{\infty\}$. Two vertices $a/b$ and $c/d$ are joined by an edge precisely when $\det \begin{pmatrix}a&c\\b&d\end{pmatrix}=\pm 1$ and we call this value the {\it determinant} of the directed edge from $a/b$ to $c/d$. If $a/b$ and $c/d$ are joined by an edge, then so are $a/b$ and $(a+c)/(b+c)$ as well as $c/d$ and $(a+c)/(b+d)$. In this case the two fractions $a/b$ and $c/d$, together with their {\it mediant} $(a+c)/(b+c)$, form the vertices of an ideal triangle in the tessellation. Finally, given any continued fraction expansion of $p/q$, the associated sequence of convergents defines a path in the Farey graph from $1/0$ to $p/q$. The number of triangles through which the path ``turns'' at each vertex is equal to the corresponding partial quotient.
For example, consider $7/17=0+[2,2,3]$. We picture the relevant portion of the Farey graph in Figure~\ref{7/17}. The sequence of convergents is $\{0, 1/2, 2/5, 7/17\}$. The partial quotients are $2,2,3$ and the path turns through 2 triangles on the left at $0$, 2 triangles on the right at $1/2$ and 3 triangles on the left at $2/5$. If all the partial quotients are positive, as in this example, then the turning at each vertex alternates between triangles on the left and right. Negative partial quotients correspond to turning in the opposite direction. Again, the reader is urged to consult \cite{HT:1985} and \cite{HS:2007}.
\begin{figure}
\psfrag{a}{\huge $\frac{1}{0}$}
\psfrag{b}{\huge $\frac{1}{1}$}
\psfrag{c}{\huge $\frac{1}{2}$}
\psfrag{d}{\huge $\frac{3}{7}$}
\psfrag{e}{\huge $\frac{5}{12}$}
\psfrag{f}{\huge $\frac{7}{17}$}
\psfrag{g}{\huge $\frac{0}{1}$}
\psfrag{h}{\huge $\frac{1}{3}$}
\psfrag{i}{\huge $\frac{2}{5}$}
\begin{center}
\leavevmode
\scalebox{.5}{\includegraphics{Figures/Example1}}
\end{center}
\caption{The portion of the Farey graph between $1/0$ and $7/17$.}
\label{7/17}
\end{figure}
For any $p/q$, there are infinitely many edge paths in the Farey graph that begin at $1/0$ and end at $p/q$, but only finitely many that are {\it minimal}. A path is minimal if it never backtracks and if it never traverses two edges of the same triangle. The minimal paths for $K_{7/17}$ and their corresponding continued fractions are shown in the first two columns of Table~\ref{slope data 7/17}. In general, only one of the minimal paths is {\it even}, that is, all the partial quotients are even. In Table~\ref{slope data 7/17}, the first path is even. According to Hatcher and Thurston, each minimal path determines a boundary slope. The slope can be computed from the path $\gamma$ by the formula
$$-2(m(\gamma)-m(\gamma_{even}))$$
where $\gamma_{even}$ is the even path and $m(\gamma)$ is the sum of the determinants of each edge of the path, excluding the first edge. Thus, the even path always gives a slope of zero. For any path $\gamma$, we call $m(\gamma)$ the {\it unadjusted} slope. The last two columns of Table~\ref{slope data 7/17} give the unadjusted slopes and the boundary slopes of $K_{7/17}$.
\begin{table}
\begin{center}
\caption{Boundary slope data for $K_{7/17}$}
\vskip .1 in
\begin{tabular}{l|l|l|l}
minimal path $\gamma$ & fraction & $m(\gamma)$ & slope\\
\hline
$\{1/0, 1/1, 1/2, 3/7, 5/12, 7/17\}$&$1+[-2,4,-2,2]$& 4 & 0\\
$ \{1/0, 0/1, 1/2, 3/7, 5/12, 7/17\}$&$0+[2,3,-2]$& 2 & 4\\
$ \{1/0, 1/1, 1/2, 2/5, 7/17\}$&$1+[-2,3,3]$& 1 & 6\\
$\{1/0, 0/1, 1/2, 2/5, 7/17\}$&$ 0+[2,2,3]$& -1 & 10\\
$\{1/0, 1/3, 2/5, 7/17\}$&$0+[3,-2,4]$& -3 & 14
\end{tabular}
\label{slope data 7/17}
\end{center}
\end{table}
For any $p/q$, if $\gamma_1$ and $\gamma_2$ are two paths from $1/0$ to $p/q$, then we can move from one path to the other by a sequence of {\it triangle moves}. By a triangle move, we mean replacing two consecutive edges that lie in a single triangle with the third edge of the triangle, or vice-versa. If $\gamma_1$ is changed to $\gamma_2$ by a single triangle move, and furthermore, the triangle lies on the right of $\gamma_1$, then we call the move a {\it right} triangle move. Right triangle moves define a partial order on the set of paths: we say that $\gamma_1>\gamma_2$ if $\gamma_1$ can be changed to $\gamma_2$ by a sequence of right triangle moves. Furthermore, it is shown in \cite{HS:2007} that right triangle moves strictly decrease the unadjusted slope of the path, thus, $\gamma_1 > \gamma_2$ implies that $m(\gamma_1) > m(\gamma_2)$. The {\it upper} and {\it lower} minimal paths can now be defined as ones that are either maximal or minimal with respect to this partial order, respectively. These paths are well-defined because there is a unique minimal path with no triangles on its left and a unique minimal path with no triangles on its right. In Table~\ref{slope data 7/17}, the first path is the upper path and the last path is the lower path.
The following lemma gives a lower bound on the number of distinct boundary slopes.
\begin{lemma}
Suppose $p/q=r+[a_1, a_2, \dots, a_m]$ is a strongly positive continued fraction. Then the number of distinct boundary slopes of $K_{p/q}$ is at least $2+\lfloor m/2\rfloor$ and at most $f_{m+1}$ where $f_m$ denotes the $m$-th Fibonacci number (assuming $f_0=0$ and $f_1=1$).
\label{bounds on number of paths}
\end{lemma}
\noindent {\em Proof.} For the lower bound we will create a sequence of minimal paths $\gamma_\text{upper}$, $\gamma_1$, $\dots$, $\gamma_{\lfloor m/2 \rfloor}$, $\gamma_\text{lower}$ with strictly decreasing unadjusted slopes. Let $\gamma_\text{upper}$ be the upper path. Next, let $\gamma_1$ be the path that goes from $1/0$ to $r/1$ to $r+[a_1]$ and then continues with the rest of the upper path. The path $\gamma_1$ is obtained from $\gamma_\text{upper}$ by a sequence of $a_1$ right triangle moves. Now let $\gamma_2$ begin $\{1/0, r/1, r+[a_1, a_2], r+[a_1, a_2, a_2]\}$ and then continue with the rest of the upper path. Notice that $\gamma_2$ is obtained from $\gamma_1$ by a sequence of $a_2 + a_3$ right triangle moves. We continue in this manner to generate each path $\gamma_i$. The fact that both $a_1>1$ and $a_m>1$ guarantee that all of these paths are minimal. Finally, their unadjusted slopes are strictly decreasing since each path is obtained from the previous one by performing right triangle moves. This establishes the lower bound. It is shown in \cite{HT:1985} that $f_m$ is an upper bound on the number of minimal paths, each of which might give a distinct boundary slope.
\hfill $\Box$
As the number $m$ of partial quotients grows, it becomes tedious to characterize all knots with $m$ partial quotients and exactly $k$ distinct boundary slopes. Therefore, in the following theorem, we only classify 2-bridge knots with up to four distinct boundary slopes. Furthermore, we provide the set of slopes for only one knot from each chiral pair. (Remember that switching from a knot to its mirror image will negate the set of boundary slopes.)
\begin{theorem}
\label{list of knots with given number of slopes} Let $K_{p/q}$ be a 2-bridge knot with 2, 3, or 4, distinct boundary slopes. Then a strongly positive continued fraction representing $K_{p/q}$, or its mirror image, together with the associated slope set, is given in the following table.
\begin{center}
\begin{tabular}{|l|l|l|}
\hline
& continued fraction & slope set \\
& &*=multiplicity 2, **=multiplicity 3\\
\hline
$2$ & $[a_1], a_1$ odd& $\{0, 2a_1\}$\\
\hline
$3_i$ & $[a_1, a_2]$, $a_1$ even, $a_2$ even& $\{-2a_1, 0, 2 a_2\}$\\
$3_{ii}$ & $[a_1, a_2]$, $a_1$ even, $a_2$ odd& $\{0, 2a_1, 2a_1+2a_2\}$\\
$3_{iii}$& $[a_1, 1, a_1], a_1$ odd& $\{-4a_1-2,-2a_1-2^*,0\}$\\
\hline
$4_i$ & $[a_1, a_2, a_1]$, $a_1$ odd, $a_2$ odd, $a_2 > 1$ &$\{-4a_1-2a_2, -2a_1-2a_2^*, -2a_2, 0\}$\\
$4_{ii}$ & $[a_1, 1, a_3]$, $a_1\ne a_3$, $a_1$ odd, $a_3$ odd & $\{-2a_1-2a_3-2,-2a_3-2,-2a_1-2,0\}$\\
$4_{iii}$ & $[a_1, 1, a_3]$, $a_1 \ne a_3$, $a_1$ even, $a_3$ odd &$\{-2a_1,0,-2a_1+2a_3,2a_3+2\}$\\
$4_{iv}$& $[a_1,1,a_1,a_1+1], a_1$ even &$\{-2a_1, 0^*, 2a_1+2^*, 4a_1+4\}$\\
$4_{v}$& $[2,1,1,1,2]=8/21$&$\{-8, -4^*, 0^{**}, 6\}$\\ \hline
\end{tabular}
\end{center}
\end{theorem}
\begin{figure}[t]
\psfrag{a}{\Huge $a_1$}
\psfrag{b}{\Huge $a_2$}
\psfrag{c}{\Huge $a_3$}
\psfrag{d}{\Large $1/0$}
\psfrag{e}{\Large $0/1$}
\psfrag{f}{\Large $[a_1]$}
\psfrag{g}{\Large $[a_1,a_2]$}
\psfrag{h}{\Large $[a_1,a_2,a_3]$}
\psfrag{i}{\Large $\gamma_\text{upper}$}
\psfrag{j}{\Large $m=a_1+a_3-1$}
\psfrag{k}{\Large $\gamma_1$}
\psfrag{l}{\Large $m=a_3-1$}
\psfrag{m}{\Large $\gamma_2$}
\psfrag{n}{\Large $m=a_1-1$}
\psfrag{o}{\Large $\gamma_3$}
\psfrag{p}{\Large $m=1$}
\psfrag{q}{\Large $\gamma_\text{lower}$}
\psfrag{r}{\Large $m=-a_2-1$}
\begin{center}
\leavevmode
\scalebox{.4}{\includegraphics{Figures/pathgraph.eps}}
\end{center}
\caption{The partially ordered set of minimal paths when $m=4$.}
\label{graph of paths}
\end{figure}
{\em Proof.} Suppose that $K_{p/q}$ has 4 or less distinct boundary slopes and strongly positive continued fraction expansion $p/q = 0 + [a_1,a_2,..., a_m]$. By Lemma~\ref{bounds on number of paths}, it follows that $m\le 4$. We will examine only the case $m=3$ and leave the remaining cases, which are handled in a similar manner, to the interested reader.
Figure~\ref{graph of paths} shows all possible minimal paths for $p/q=[a_1, a_2, a_3]$ arranged according to the partial ordering on paths. Each of the five figures is made up of three main triangles which are subdivided into $a_1$, $a_2$, and $a_3$ smaller triangles, respectively, as we move from left to right. The five paths $\gamma_\text{upper}$, $\gamma_1$, $\gamma_2$, $\gamma_3$, and $\gamma_\text{lower}$ are depicted with heavy dark lines. Also listed is the unadjusted slope $m$ for each path under the path's label.
Since $a_1 >1$ and $a_3 > 1$ we have that $\gamma_\text{upper}$, $\gamma_1$, $\gamma_2$, and $\gamma_\text{lower}$ are minimal. If $\gamma_3$ is minimal and $a_1 \neq a_3$, then we would have 5 distinct boundary slopes. Therefore, we have two cases to consider.
Assume $\gamma_3$ is not minimal, which requires $a_2 =1$, and also that $a_1\ne a_3$. Thus, $p/q = 0 + [a_1,1,a_3]$ and we have four distinct boundary slopes. In order to compute the (adjusted) slopes we need to determine the unique even path. Notice that
$$\frac{p}{q}=[a_1,1,a_3] = \frac{1+a_3}{a_1 + a_3 +a_1 a_3}$$
and so in order for $q$ to be odd we need at least one of $a_1$ or $a_3$ to be odd. This gives three sub-cases which are listed below.
\begin{center}
\begin{tabular}{|l|l|l|l|}
\hline
$a_1$ & $a_3$ & even path & boundary slopes \\
\hline
even & odd & $\gamma_1$ & $\{-2a_1,0,-2a_1+2a_3,2a_3+2\}$ \\
odd & even & $\gamma_2$ & $\{-2a_3,2a_1-2a_3,0,2a_1+2\}$ \\
odd & odd & $\gamma_\text{lower}$ & $\{-2a_1-2a_3-2,-2a_3-2,-2a_1-2,0\}$ \\ \hline
\end{tabular}
\end{center}
These slopes appear in parts $4_{ii}$ and $4_{iii}$ of Theorem~\ref{list of knots with given number of slopes}. The first two rows correspond to equivalent knots and only appear in one row of the theorem.
Next assume that $\gamma_3$ is not minimal and $a_1 = a_3$. In this case, we have 3 distinct boundary slopes and
$$\frac{p}{q}=[a_1,1,a_1] = \frac{1+a_1}{a_1^2 + 2 a_1}.$$
Here $q$ is odd if and only if $a_1$ is odd. In this case, the even path is $\gamma_\text{lower}$ and the boundary slopes are
$$\{-4a_1-2,-2a_1-2,0\}$$
where the second slope $-2a_1-2$ has multiplicity two, that is, it corresponds to two paths, $\gamma_1$ and $\gamma_2$. This data appears in part $3_{iii}$ of Theorem~\ref{list of knots with given number of slopes}.
Finally, assume that $\gamma_3$ is minimal, which means $a_2>1$, and $a_1=a_3$. We now have four distinct boundary slopes and
$$\frac{p}{q}=[a_1,a_2,a_1] = \frac{1+a_1 a_2}{a_1^2 a_2 + 2 a_1}.$$
In order for $q = a_1^2 a_2 + 2 a_1$ to be odd we must have that both $a_1$ and $a_2$ are odd. This implies that $\gamma_\text{lower}$ is the even path and we obtain the following boundary slopes:
$$\{-4a_1-2a_2, -2a_1-2a_2, -2a_2, 0\}.$$
Notice that the second boundary slope $-2a_1-2a_2$ has multiplicity two.
This data appears in part $4_i$ of Theorem~\ref{list of knots with given number of slopes}. The remaining cases follow by a similar analysis with $m=1$, $m=2$, or $m=4$. \hfill $\Box$
By working out the fractions $p/q$ for each case of Theorem~\ref{list of knots with given number of slopes} we obtain the following corollary which strengthens Theorem~3 of \cite{MMR:2008}. As in Theorem~\ref{list of knots with given number of slopes}, we consider a knot and its mirror image to be equivalent.
\begin{corollary} Let $K$ be a 2-bridge knot.
\begin{itemize}
\item[i.] $K$ has exactly two distinct boundary slopes if and only if $K \equiv K_{1/q}$.
\item[ii.] $K$ has exactly three distinct boundary slopes if and only if $K$ is not equivalent to a knot from part (i) and $K \equiv K_{p/q}$ for which either
\begin{itemize}
\item[a.] $p \mid q-1$ or
\item [b.] $p^2 = q+1$.
\end{itemize}
\item[iii.] $K$ has exactly four distinct boundary slopes if and only if $K$ is not equivalent to a knot from parts (i) or (ii) and $K \equiv K_{p/q}$ for which either
\begin{itemize}
\item[a.] $p + 1 \mid q$ and $q \mid p^2-1$,
\item[b.] $p \mid q+1$,
\item[c.] $(p-1)^3 = q^2$, or
\item[d.] $p/q = 8/21$.
\end{itemize}
\end{itemize}
\end{corollary}
\section{Ohtsuki, Riley, Sakuma knot pairs}
In this section we briefly review the method of Ohtsuki, Riley, and Sakuma for constructing a pair of two bridge knots $K_{p'/q'}$ and $K_{p/q}$ with $K_{p'/q'} \ge K_{p/q}$. The reader is encouraged to consult \cite{ORS} for a more detailed description. One begins with any 2-bridge knot (or link) $K_{p/q}$ given by the four-plat shown in Figure~\ref{ORS construction}. The 4-string braid defining $K_{p/q}$ is denoted by $\beta$. There are three associated braids $\beta^{-1}, \beta_{-}$, and $\beta^{-1}_{-}$ which are obtained by rotating $\beta$ 180 degrees around an axis perpendicular to the plane of the diagram, reflecting $\beta$ through a plane perpendicular to the plane of the diagram, and the composition of these two motions, respectively. In order to construct a 2-bridge knot $K_{p'/q'}$ that is greater than $K_{p/q}$ we consider a 4-plat with an odd number of ``boxes'' each containing either $\beta, \beta^{-1}, \beta_{-}$, or $\beta^{-1}_{-}$. Starting from the left in Figure~\ref{ORS construction}, the first box contains $\beta$ and every other box after that contains $\beta$ or $\beta_{-}$. The remaining boxes each contain $\beta^{-1}$ or $\beta^{-1}_{-}$. Also between each braid box in $K_{p'/q'}$ we may insert an even number of half-twists in the middle two strands of the 4-plat. Finally, a {\it branched fold map} $f$ is constructed from the complement of $K_{p'/q'}$ onto the complement of $K_{p/q}$ as follows. The complement of $K_{p'/q'}$ is cut along a collection of parallel 2-spheres into two 3-balls and a number of $S^2 \times I$'s. The complement of $K_{p/q}$ is decomposed into two 3-balls and a single $S^2 \times I$. A continuous mapping is then defined in a piecewise manner. First each 3-ball ``upstairs'' is mapped by the identity onto a corresponding 3-ball ``downstairs.'' Next, each $S^2 \times I$ containing one braid box upstairs is mapped homeomorphically onto the $S^2 \times I$ downstairs in a way that depends on the presence of $\beta, \beta^{-1}, \beta_{-}$, or $\beta^{-1}_{-}$. Finally, the mapping is extended to the remaining $S^2 \times I$'s upstairs onto the 3-balls downstairs using 2-fold branched mappings that depend on the combinations of $\beta, \beta^{-1}, \beta_{-}$, or $\beta^{-1}_{-}$ in the adjacent components. From the way that $f$ is constructed, the two meridional generators associated with the bridges at either end of the 4-plat diagram upstairs are taken to the corresponding generators downstairs. Furthermore, the longitude upstairs is taken to a power of the longitude downstairs. Thus $f$ induces an epimorphism on fundamental groups which preserves peripheral structure.
\begin{figure}[t]
\psfrag{X}{$K_{p/q}$}
\psfrag{Y}{$\beta$}
\psfrag{Z}{$K_{p'/q'}$}
\psfrag{a}{$a_1$}
\psfrag{b}{$a_2$}
\psfrag{c}{$a_3$}
\psfrag{g}{$2c_1$}
\psfrag{h}{$2c_2$}
\psfrag{d}{$\beta$}
\psfrag{e}{$\beta^{-1}$}
\psfrag{f}{$\beta_{-}$}
\psfrag{j}{$f$}
\psfrag{k}{$S^2 \times I$}
\psfrag{i}{$B^3$}
\begin{center}
\leavevmode
\scalebox{.75}{\includegraphics{Figures/OhtsukiConstruction.eps}}
\end{center}
\caption{The Ohtsuki, Riley, Sakuma construction.}
\label{ORS construction}
\end{figure}
Given a 2-bridge knot $K_{p/q}$, Ohtsuki, Riley and Sakuma determine which knots $K_{p'/q'}$ are produced by their construction. In particular, they show that $p'/q'$ gives such a knot if and only if it lies in the orbit of $p/q$ under the action of $\Gamma_\infty * \Gamma_{p/q}$, where $\Gamma_{r/s}$ is the infinite dihedral group generated by all reflections in edges of the Farey graph which end at $r/s$. Let $A_1$ and $A_2$ be generators of $\Gamma_\infty$ defined by reflection in the edges $<1/0, 0/1>$ and $<1/0, 1/1>$, respectively. For $p/q=[a_1, a_2, \dots, a_m]$ let $B_1$ and $B_2$ be generators of $\Gamma_{p/q}$ defined by reflection in the edges $<p/q, [a_1, a_2, \dots, a_{m-1}]>$ and $B_2=<p/q, [a_1, a_2, \dots, a_m-1]>$. Thus $p'/q'=W(p/q)$ where $W$ is a word in $A_1, A_2, B_1$ and $B_2$. Since $B_1$ and $B_2$ fix $p/q$, and $A_1$ and $A_2$ take $p/q$ to a fraction representing the same knot, we may assume that we begin by applying a nontrivial word in $A_1$ and $A_2$ and end by applying a nontrivial word in $B_1$ and $B_2$. Hence we may write $W$ as
\begin{equation}\label{W word}
W=W_1 W_2\dots W_n, \mbox{ with $n$ even},
\end{equation}
where
\begin{equation}\label{W syllables}
W_i=\left\{
\begin{array}{lll}
(B_1B_2)^{c_i}B_1^\frac{1-\eta_i}{2}, &c_i \in \mathbb Z, \eta_i \in \{1, -1\}, &\mbox{ if $i$ is odd, and}\\
&&\\
(A_1A_2)^{c_i}A_1^\frac{1-\eta_i}{2}, &c_i \in \mathbb Z, \eta_i \in \{1, -1\},& \mbox{ if $i$ is even.}
\end{array}
\right.
\end{equation}
Ohtsuki, Riley and Sakuma then show (Lemma 5.3 in \cite{ORS}) that
\begin{equation}
\label{ORS formula}
p'/q'=[ \epsilon_1 {\bf a}, 2 \epsilon_1 c_1, \epsilon_2 {\bf a}^{-1}, 2 \epsilon_2 c_2, \epsilon_3 {\bf a}, 2 \epsilon_3 c_3, \epsilon_4 {\bf a}^{-1},\dots, \epsilon_{n+1} {\bf a}],
\end{equation}
where ${\bf a}$ is the vector ${\bf a}=(a_1, a_2, \dots, a_m),\ {\bf a}^{-1}$ means ${\bf a}$ written backwards, $\epsilon_1 = 1$, and $\epsilon_i=\prod_{j=1}^{i-1}-\eta_j$.
The following lemma will prove helpful in simplifying continued fractions.
\begin{lemma}
\label{cf moves} Let $\bf a$ and $\bf b$ be vectors of integers, possibly empty, and $m$ and $n$ be any integers. Then
\begin{enumerate}
\item[i.] $r+[{\bf a}, m, -n, {\bf b}] =r+ [{\bf a}, m-1,1,n-1, -{\bf b}]$ and
\item[ii.] $r+[{\bf a}, m, 0, n, {\bf b}] = r+[{\bf a}, m+n, {\bf b}]$
\end{enumerate}
\end{lemma}
\noindent {\em Proof.}
If $p/q=r_0+[r_1, r_2, \dots, r_k]$ then it is well known that
$$\pm \begin{pmatrix}p\\q\end{pmatrix}= M_{(r_0, r_1, \dots, r_k)}\begin{pmatrix}1\\0\end{pmatrix},$$
where
$$\ M_{(r_0, r_1, \dots, r_k)}=
\begin{pmatrix}r_0&1\\1&0\end{pmatrix}
\begin{pmatrix}r_1&1\\1&0\end{pmatrix}\cdots
\begin{pmatrix}r_k&1\\1&0\end{pmatrix}.$$
It is an easy exercise to show that
$$\begin{pmatrix}-1&0\\0&1\end{pmatrix} M_{\bf b}=(-I)^{|{\bf b}|}M_{-{\bf b}}\begin{pmatrix}-1&0\\0&1\end{pmatrix},$$
where $|{\bf b}|$ is the dimension of the vector ${\bf b}$ and $I$ is the identity matrix.
The proof of Lemma~\ref{cf moves}({\it i}) follows from
\begin{eqnarray*}
M_{{\bf a}, m, -n, {\bf b}} \begin{pmatrix}1\\0\end{pmatrix}&=&
M_{\bf a}\begin{pmatrix}m&1\\1&0\end{pmatrix} \begin{pmatrix}-n&1\\1&0\end{pmatrix} M_{\bf b}
\begin{pmatrix}1\\0\end{pmatrix}\\
&=&
M_{\bf a}\begin{pmatrix}m-1&1\\1&0\end{pmatrix}
\begin{pmatrix}1&1\\1&0\end{pmatrix}
\begin{pmatrix}n-1&1\\1&0\end{pmatrix}
\begin{pmatrix}-1&0\\0&1\end{pmatrix}
M_{\bf b}
\begin{pmatrix}1\\0\end{pmatrix}\\
&=&
(-I)^{|{\bf b}|} M_{{\bf a}, m-1, 1, n-1, -{\bf b}}
\begin{pmatrix}-1&0\\0&1\end{pmatrix}
\begin{pmatrix}1\\0\end{pmatrix}\\
&=&
(-I)^{|{\bf b}|} M_{{\bf a}, m-1, 1, n-1, -{\bf b}}
\begin{pmatrix}-1\\0\end{pmatrix}\\
&=&
(-I)^{|{\bf b}|+1} M_{{\bf a}, m-1, 1, n-1, -{\bf b}}
\begin{pmatrix}1\\0\end{pmatrix}
\end{eqnarray*}
The second part of the lemma follows in a similar way.
\hfill $\Box$
We are now prepared to show that in any ORS pair $(K_{p'/q'}, K_{p/q})$, if $K_{p'/q'}$ is not a torus knot, then it has at least five distinct boundary slopes.
\begin{theorem}
\label{5slopes}
If $K_{p'/q'} \ge K_{p/q}$ is a nontrivial ORS pair, then either
\begin{enumerate}
\item[{\it i}.] $K_{p'/q'}$ and $K_{p/q}$ are both torus knots and $K_{p'/q'}$ has precisely two distinct boundary slopes, or
\item [{\it ii}.]$K_{p'/q'}$ has at least five distinct boundary slopes.
\end{enumerate}
\end{theorem}
\noindent {\em Proof.} Let $p/q=[a_1, a_2, \dots, a_m]=[{\bf a}]$ be a strongly positive continued fraction. Suppose that $p'/q' = W(p/q)$ where $W$ is described in Equations (\ref{W word}) and (\ref{W syllables}). By Equation~(\ref{ORS formula}) we have that
\begin{equation}
\label{ORS formula 2}
p'/q'=[ \epsilon_1 {\bf a}, 2 \epsilon_1 c_1, \epsilon_2 {\bf a}^{-1}, 2 \epsilon_2 c_2, \epsilon_3 {\bf a}, 2 \epsilon_3 c_3, \epsilon_4 {\bf a}^{-1},\dots, \epsilon_{n+1} {\bf a}].
\end{equation}
Suppose $c_i=0$ for some $i$. Because $W_i$ is not the identity, it follows that $\eta_i=-1$ and hence $\epsilon_i=\epsilon_{i+1}$. Therefore, on either side of $2 \epsilon_i c_i=0$, the continued fraction appears as
$p'/q'=[\dots,\epsilon_i {\bf a}^{\pm 1}, 0, \epsilon_i {\bf a}^{\mp 1},\dots]$. Using Lemma~\ref{cf moves}, we may eliminate the zero to combine $\epsilon_i {\bf a}^{\pm 1}$ and $\epsilon_i {\bf a}^{\mp 1}$ into a single strongly positive or negative vector.
Let $1\le j_1<j_2<\dots <j_k\le n$ be the indices of the nonzero $c_i$'s.
Eliminating each of the $n-k$ zero entries as just described gives the following continued fraction expansion with $n(m-1)+m+2k$ partial quotients,
\begin{equation}
\label{nozerosfrac}
p'/q'=[ \epsilon_1 {\bf v}_1, 2 \epsilon_{j_1} c_{j_1}, \epsilon_{j_2} {\bf v}_2, 2 \epsilon_{j_2} c_{j_2},\epsilon_{j_3} {\bf v}_3,\dots, 2 \epsilon_{j_k} c_{j_k}, \epsilon_{n+1} {\bf v}_{k+1}]
\end{equation}
where $\epsilon_1=1$, each ${\bf v}_i$ is a strongly positive vector, and $c_{j_i} \neq 0$ for all $i$. Notice that if $m=1$, then ${\bf a}^{-1}={\bf a}=a_1$ and each vector ${\bf v}_i$ is a single integer given by ${\bf v}_i=(j_i-j_{i-1})a_1$ if $i\le k$ and ${\bf v}_{k+1}=(n+1-j_k)a_1$. Otherwise, if $m>1$, $|{\bf v}_i| \ge 2$ for all $i$.
We now claim that the continued fraction given in Equation~(\ref{nozerosfrac}) can be changed into a strongly positive one with at least as many partial quotients. Before proving this claim, consider its consequences. If $m=1$ then the number of partial quotients is at least $2k+1$ which can only be less than 6 if $k=0, 1$, or $2$. If $m>1$, the number of partial quotients can only be less than 6 if $n=m=2$ and $k=0$. Hence, except in the cases where $m=1$ and $0\le k\le 2$, or $n=m=2$ and $k=0$, we can express $p'/q'$ as a strongly positive continued fractions with at least 6 partial quotients. Lemma~\ref{bounds on number of paths} now implies that $K_{p'/q'}$ has at least 5 distinct boundary slopes. Thus, after proving this claim, we must analyze these four cases.
To prove the claim, we first assume that $m>1$ and proceed by induction on the number of sign changes in the sequence of partial quotients in Equation~(\ref{nozerosfrac}). If there are none, then it is easy to check that Equation~(\ref{nozerosfrac}) is already strongly positive. If there are sign changes, then consider the smallest value of $r$ so that $\epsilon_{j_r}=1$ and either $ c_{j_r}$ or $\epsilon_{j_{r+1}}$ is negative.
If $ c_{j_r}>0$ and $\epsilon_{j_{r+1}}=-1$, then using Lemma~\ref{cf moves}, we obtain
\begin{eqnarray*}
p'/q'&=&[\dots, {\bf v}_r, 2 c_{j_r}, -{\bf v}_{r+1}, -2 c_{j_{r+1}},\dots]\\
&=&[\dots, {\bf v}_r, 2 c_{j_r}, -v_{r+1}^1, -v_{r+1}^2, \dots, -v_{r+1}^{|{\bf v}_{r+1}|}, -2 c_{j_{r+1}},\dots]\\
&=&[\dots, {\bf v}_r, 2 c_{j_r}-1,1, v_{r+1}^1-1, v_{r+1}^2, \dots, v_{r+1}^{|{\bf v}_{r+1}|}, 2 c_{j_{r+1}},\dots].
\end{eqnarray*}
This continued fraction is now of the same form as the original, has fewer sign changes, and has more partial quotients. The result now follows by induction. It is important here to remember that since $m>1$, $|{\bf v}_i| \ge 2$ for all $i$.
If $ c_{j_r}<0$ and $\epsilon_{j_{r+1}}=-1$, then using Lemma~\ref{cf moves}, we obtain
\begin{eqnarray*}
p'/q'&=&[\dots, {\bf v}_r, 2 c_{j_r}, -{\bf v}_{r+1}, -2 c_{j_{r+1}},\dots]\\
&=&[\dots, v_r^1, v_r^2, \dots, v_r^{|{\bf v}_r|}, 2 c_{j_r}, -{\bf v}_{r+1}, -2 c_{j_{r+1}},\dots]\\
&=&[\dots, v_r^1, v_r^2, \dots, v_r^{|{\bf v}_r|}-1, 1, -2 c_{j_r}-1, {\bf v}_{r+1}, 2 c_{j_{r+1}},\dots].
\end{eqnarray*}
Again, this continued fraction is now of the same form as the original, has fewer sign changes, and has more partial quotients. The result now follows by induction.
Finally, suppose that $ c_{j_r}<0$ and $\epsilon_{j_{r+1}}=1$. Applying Lemma~\ref{cf moves} twice, we obtain
\begin{eqnarray*}
p'/q'&=&[\dots, {\bf v}_r, 2 c_{j_r}, {\bf v}_{r+1}, \dots]\\
&=&[\dots, v_1^1, \dots, v_r^{|{\bf v}_r|}-1, 1, -2c_{j_r}-2, 1, v_{r+1}^1-1, v_{r+1}^2, \dots, v_{r+1}^{|{\bf v}_{r+1}|}, \dots].
\end{eqnarray*}
If $c_{j_r}\ne -1$, then once again we have arrived a continued fraction of the same form as the original, but with fewer sign changes and more partial quotients. If instead, $c_{j_r}= -1$, then we may use Lemma~\ref{cf moves} to eliminate the zero entry and arrive at a continued fraction with the same form as the original, fewer sign changes, but now an equal number of partial quotients. In either case, the result follows by induction.
We now consider the case where $m=1$. Recall that if $m=1$, then ${\bf a}^{-1}={\bf a}=a_1$, ${\bf v}_i=(j_i-j_{i-1})a_1$ for $i\le k$, and ${\bf v}_{k+1}=(n+1-j_k)a_1$. Equation~(\ref{nozerosfrac}) is now of the form
\begin{equation}
\label{m=1 form}
p'/q'=[\epsilon_1 j_1 a_1, 2\epsilon_{j_1}c_{j_1}, \epsilon_{j_2}(j_2-j_1)a_1, 2\epsilon_{j_2}c_{j_2},
\dots, 2 \epsilon_{j_k}c_{j_k}, \epsilon_{n+1}(n+1-j_k)a_1].
\end{equation}
Because $a_1\ge 3$, this continued fraction is a member of a family $\cal F$ of continued fractions all having the form $[{\bf v}, {\bf w}]$ where
\begin{itemize}
\item ${\bf v}$ is a strongly positive vector (possibly of length 1),
\item the first entry of the vector ${\bf w}$ is negative,
\item the entries of $\bf w$ alternate between being even and having magnitudes greater than 2
(the first entry of $\bf w$ can have either property), and
\item the last entry of $\bf w$ has magnitude greater than 2.
\end{itemize}
We call $\bf v$ the {\it strongly positive part} and ${\bf w}$ the {\it tail} of such a continued fraction. We shall prove, by induction on the length of the tail, that any such continued fraction can be changed to a strongly positive one with at least as many partial quotients. Since ${\bf v}$ is strongly positive, if ${\bf w}$ has length zero then we are done. Now consider such a continued fraction where ${\bf w}$ is not empty. We have
\begin{eqnarray*}
[\bf v, w]&=&[v_1, v_2, \dots, v_r, w_1, w_2, \dots, w_s]\\
&=&[v_1, v_2, \dots, v_r-1, 1, -w_1-1, -w_2, \dots, -w_s]
\end{eqnarray*}
If $-w_2>0$, or if $-w_2<0$ and $-w_1-1>1$, then we have arrived at a member of $\cal F$ with more partial quotients and a shorter tail. Hence the result follows by induction. Otherwise, we have $-w_1-1=1$ and $-w_2<0$. Thus
\begin{eqnarray*}
[\bf v, w]&=&[v_1, v_2, \dots, v_r, -2, w_2, \dots, w_s]\\
&=&[v_1, v_2, \dots, v_r-1, 1, 1, -w_2, \dots, -w_s]\\
&=&[v_1, v_2, \dots, v_r-1, 1, 0, 1, w_2-1, w_3, \dots, w_s]\\
&=&[v_1, v_2, \dots, v_r-1, 2, w_2-1, w_3, \dots, w_s].
\end{eqnarray*}
Because the entries of $\bf w$ alternate between being even and having magnitudes greater than 2, it must be the case that $w_2>2$ and hence $w_2-1>1$. Hence we have arrived at a member of $\cal F$ with the same number of partial quotients but with a shorter tail. Once again, the result now follows by induction.
We have now completed the proof of the claim, and it remains to analyze the four cases already mentioned above. If $m=n=2$ and $k=0$, we have $p/q=[a_1, a_2]$ and $p'/q'=[a_1, a_2, 0, a_2, a_1, 0, a_1, a_2]=[a_1, 2a_2, 2a_1, a_2]$. Comparing this strongly positive continued fraction to the table in Theorem~\ref{list of knots with given number of slopes} we see that $K_{p'/q'}$ cannot have less than 5 distinct boundary slopes.
If $m=1$ and $k=0$ the continued fraction for $p'/q'$ collapses to $p'/q'=[(n+1)a_1]$. Thus $K_{p'/q'}$ is a torus knot with exactly 2 slopes.
Next, suppose that $m=1$ and $k=1$. Then exactly one of the $c_i$'s is nonzero, say $c_{j_1}$. Equation~\ref{nozerosfrac} now gives $p'/q'=[j_1 a_1, 2 \epsilon_{j_1}c_{j_1}, \epsilon_{n+1}(n+1-j_1)a_1]$. If all of these partial quotients are positive, then comparing this strongly positive continued fraction to the table in Theorem~\ref{list of knots with given number of slopes}, we see that $K_{p'/q'}$ cannot have less than 5 distinct boundary slopes.
If not, there are either one or two sign changes in the sequence of partial quotients. These considerations lead to the following cases which are handled in a fashion similar to that when $m>1$:
\begin{itemize}
\item If $\epsilon_{j_1}c_{j_1}>0$ and $\epsilon_{n+1}=-1$, then
$p'/q'=[j_1a_1, 2\epsilon_{j_1}c_{j_1}-1, 1, (n+1-j_1)a_1-1]$.
\item If $\epsilon_{j_1}c_{j_1}<0$ and $\epsilon_{n+1}=-1$, then
$p'/q'=[j_1a_1-1,1, -2\epsilon_{j_1}c_{j_1}-1, (n+1-j_1)a_1]$.
\item If $\epsilon_{j_1}c_{j_1}=-1$ and $\epsilon_{n+1}=1$, then
$p'/q'=[j_1a_1-1, 2, (n+1-j_1)a_1-1]$.
\item If $\epsilon_{j_1}c_{j_1}<-1$ and $\epsilon_{n+1}=1$, then
$p'/q'=[j_1a_1-1,1, -2\epsilon_{j_1}c_{j_1}-2, 1, (n+1-j_1)a_1-1]$.
\end{itemize}
In all four of these cases, comparison to the table in Theorem~\ref{list of knots with given number of slopes} gives that $K_{p'/q'}$ cannot have less than 5 distinct boundary slopes.
Our final case to consider is when $m=1$ and $k=2$. Now two of the $c_i$'s are nonzero, say $c_{j_1}$ and $c_{j_2}$. Eliminating zeroes leads to
$$p'/q'=[j_1 a_1, 2\epsilon_{j_1}c_{j_1}, \epsilon_{j_2}(j_2-j_1)a_1, 2\epsilon_{j_2}c_{j_2}, \epsilon_{n+1}(n+1-j_2)a_1].$$
If all of these partial quotients are positive, then comparing this strongly positive continued fraction to the table in Theorem~\ref{list of knots with given number of slopes}, we once again see that $K_{p'/q'}$ cannot have less than 5 distinct boundary slopes. If not, then using Lemma~\ref{cf moves} to move to a strongly positive expansion will increase the number of partial quotients to 6 or more, again resulting in 5 or more boundary slopes, unless the first entry of the tail is $-2$. In this case, that entry will remain even, and comparison with the table in Theorem~\ref{list of knots with given number of slopes} again shows that this knot cannot have less than 5 distinct boundary slopes.
\hfill $\Box$
The lower bound of five boundary slopes in part {\it ii} of Theorem~\ref{5slopes} is sharp. For example, with $p/q=1/3=[3]$ and $p'/q'=7/45=[3,0,3,2,3]=[6,2,3]$ we have $K_{7/45} \ge K_{1/3}$. Furthermore, using Figure~\ref{graph of paths}, it is easy to show that $K_{7/45}$ has exactly 5 distinct boundary slopes.
Recall that if $K_{p'/q'}$ is a 2-bridge knot and $K_{p'/q'} \ge K_2$, then $K_2$ is also a 2-bridge knot. Combining this fact with the result in Theorem~\ref{5slopes} gives the following corollary.
\begin{corollary}
If the answer to Question~\ref{ORS question} is yes, then 2-bridge knots with exactly 3 or 4 distinct boundary slopes are minimal with respect to the Silver-Whitten partial ordering.
\end{corollary}
Notice that if $K$ is a 2-bridge knot with exactly 2 distinct boundary slopes, then $K$ is the torus knot $T_{2, q}$ and is minimal if and only if $q$ is prime.
\section{Two-bridge knots with three distinct boundary slopes}
In this section we provide evidence for an affirmative answer to Question~\ref{ORS question} by proving that 2-bridge knots with exactly three distinct boundary slopes are in fact minimal.
\begin{theorem}
If $K$ is a 2-bridge knot with exactly 3 distinct boundary slopes, then $K$ is minimal with respect to the Silver--Whitten partial order.
\label{minknots}
\end{theorem}
{\em Proof.} Suppose $K_1$ is a 2-bridge knot with exactly three distinct boundary slopes and that $K_1\ge_d K_2$ with $K_2$ nontrivial. Then, by Theorem~\ref{properties of partial order} and Corollary~\ref{boundary slopes are subsets}, $K_2$ is a 2-bridge knot with exactly two or three distinct boundary slopes. If $K_2$ has two distinct slopes then it is a torus knot and so, by Gonz\'alez-Acu\~na and Ram\'irez \cite{GR:2001}, $K_1$ is given by the ORS construction. Now Theorem~\ref{5slopes} contradicts that $K_1$ has exactly three boundary slopes. Therefore, from Theorem~\ref{list of knots with given number of slopes} there are three possibilities for $K_1$ and three for $K_2$, giving a total of 9 different cases to consider. In what follows, the symbol $(3_{ii}, 3_{i})$, for example, will be used to denote the case where $K_1$ is the second type of knot with three slopes and $K_2$ is the third type of knot with three slopes, as listed in Theorem~\ref{list of knots with given number of slopes}.
Since boundary slopes of 2-bridge knots are always even, we first consider them mod 4. If $K_2$ has $r$ distinct slopes equal to 0 mod 4 and $s$ distinct slopes equal to 2 mod 4, then $K_1$ has the same number of slopes of each type because of Corollary~\ref{boundary slopes are subsets} and the fact that $d$ is odd. This observation rules out four cases: $(3_i, 3_{ii}), (3_i, 3_{iii}), (3_{ii}, 3_i), (3_{iii}, 3_i)$.
Before considering any other cases we note that the Alexander polynomial of of a 2-bridge knot with three boundary slopes is given in the following table. Note that in each case, no nontrivial constant can be factored from each polynomial.
\begin{center}
\begin{tabular}{|ll|l|}
\hline
& knot type & Alexander polynomial\\
\hline
& & \\
$3_i$ & $0+[a_1, a_2],\ a_1$ even, $a_2$ even & $\displaystyle -\frac{a_1 a_2}{4}+ \left(1-\frac{a_1 a_2}{2} \right)t-\frac{a_1 a_2}{4}t^2$\\
& & \\
$3_{ii}$ & $0+[a_1, a_2],\ a_1$ even, $a_2$ odd & $\displaystyle \frac{a_2+1}{2} + a_2 \sum_{i=1}^{a_1-1} (-1)^i t^i + \frac{a_2+1}{2} t^{a_1}$\\
& & \\
$3_{iii}$ & $0+[a_1, -1, a_1],\ a_1$ odd&$\displaystyle \frac{(a_1+1)^2}{4}+ \left(1-\frac{(a_1+1)^2}{2}\right)t+\frac{(a_1+1)^2}{4}t^2$\\
& & \\
\hline
\end{tabular}
\end{center}
\noindent{\bf Case $(3_i, 3_i)$} Suppose $K_1$ corresponds to $0+[a_1,a_2]$ and $K_2$ corresponds to $0+[b_1, b_2]$. Because of Corollary~\ref{boundary slopes are subsets}, it follows that
$$\{ -2db_1,0,2db_2\} = \{-2a_1,0,2a_2\}$$
and so $a_1 a_2=d^2 b_1 b_2$. Furthermore, $\Delta_{K_2}$ divides $\Delta_{K_1}$ which implies that $d=\pm1$. Hence, $\{a_1,a_2\} = \{b_1,b_2\}$ and we have $K_2 \equiv K_1$.
\noindent {\bf Case $(3_{ii}, 3_{iii})$} Suppose $K_1$ corresponds to $0+[a_1,a_2]$ and $K_2$ corresponds to $0+[b_1,1, b_1]$. From Corollary~\ref{boundary slopes are subsets} we obtain that $d<0$, $a_1=-(b_1+1)d$, and $a_1+a_2=-(2 b_1+1)d$. From these it follows that $a_2=-b_1 d$.
Since the determinants of these knots are $a_1 a_2+1$ and $b_1^2+2 b_1$, respectively, it follows from Theorem~\ref{properties of partial order} that $k(b_1^2+2 b_1)=a_1a_2+1=b_1^2 d^2+b_1 d^2+1$ for some integer $k$. Hence $b_1$ divides 1 which is a contradiction.
\noindent {\bf Case $(3_{iii}, 3_{ii})$} Suppose $K_1$ corresponds to $0+[a_1,1,a_1]$ and $K_2$ corresponds to $0+[b_1,b_2]$. From Corollary~\ref{boundary slopes are subsets} we obtain that $d<0,\ a_1+1=-b_1 d$, and $2a_1+1=-(b_1+b_2)d$. From these we obtain that $d=-1,\ b_1=a_1+1$, and $b_2=a_1$. However, the division of the Alexander polynomials implies that $b_1=2$ and this contradicts that $a_1>1$.
\noindent {\bf Case $(3_{iii}, 3_{iii})$} Suppose $K_1$ corresponds to $0+[a_1,1,a_1]$ and $K_2$ corresponds to $0+[b_1, 1, b_2]$. The division of the Alexander polynomials implies that $b_1=a_1$ and hence $K_2=K_1$.
\noindent {\bf Case $(3_{ii}, 3_{ii})$} Suppose $K_1$ corresponds to $0+[a_1,a_2]$ and $K_2$ corresponds to $0+[b_1, b_2]$. Using Corollary~\ref{boundary slopes are subsets} we immediately obtain that $a_1=b_1d$ and $a_2=b_2 d$. All evidence suggests that if $d \neq 1$, then the Alexander polynomials do not divide (in fact, they are almost certainly both irreducible), however, we were unable to prove this in general. Instead we turn to character varieties to settle this last case.
By Theorem~\ref{properties of partial order}, we have that $\phi^* : X(K_2) \rightarrow X(K_1)$ is an injective, algebraic and closed (in the Zariski topology) mapping. In particular, this implies that the image $\phi^*(X(K_2))$ is a subvariety of $X(K_1)$ that is birationally equivalent to $X(K_2)$. By Theorem~6.5 of Macasieb, Petersen, and Van Luijk \cite{MPV:2009}, both $X(K_1)$ and $X(K_2)$ are irreducible curves and so we conclude that $X(K_1)$ and $X(K_2)$ are birationally equivalent, and therefore have the same genus. Moreover, by Theorem~6.5 of \cite{MPV:2009}, the genus of $X(K_2)$ is
\begin{equation}
\label{genus2}
3 \left( \frac{b_2+1}{2} \right) \left( \frac{b_1}{2} \right) - \frac{b_2+1}{2} - 4 \frac{b_1}{2} +2.
\end{equation}
(To interpret Theorem~6.5 in our setting, note that $k=-b_2$ and $l = b_1$.)
Similarly, the genus of $X(K_1)$ is
\begin{equation}
\label{genus1}
3 \left( \frac{d b_2+1}{2} \right) \left(\frac{d b_1}{2} \right)- \frac{d b_2+1}{2} - 4 \frac{d b_1}{2} +2.
\end{equation}
Equating (\ref{genus2}) and (\ref{genus1}) gives
$$d \left(3 d b_1 b_2 - 5b_1 - 2b_2 \right) = 3 b_1 b_2 - 5b_1 - 2b_2.$$
Since $d$, $b_1$, and $b_2$ are all positive, this can only be true if $d=1$, in which case $K_2 = K_1$.
\hfill $\Box$
Since Theorem~\ref{list of knots with given number of slopes} also classifies those 2-bridge knots with exactly four distinct boundary slopes, we could hope to apply arguments similar to those above to prove that 2-bridge knots with four slopes are minimal. In particular, looking at boundary slopes will eliminate most of the 40 cases involved. However, without additional information about the relevant Alexander polynomials or character varieties, we were unable to settle resolve several cases. We close with the following conjecture and question.
\begin{conjecture}
A 2-bridge knot with exactly four distinct boundary slopes is minimal with respect to the Silver-Whitten partial ordering.
\end{conjecture}
\begin{question}
Does there exist a non 2-bridge, non-minimal knot with exactly 3 distinct boundary slopes?
\end{question}
\bibliographystyle{gtart}
|
\section{Tuning weights (Phase~\ref{phase:ComputeWeights}) }\label{tuning}
First, let us address the question of finding a vector $\Weights$ such that the
multidimensional rejection scheme (Phase~\ref{phase:RejectComposition}) is as efficient as possible. We propose and explore two
alternatives, both computing a weights vector that make the expected and targeted compositions coincide.
The first one uses a numerical Newton iteration. The second one uses an asymptotic approximation for
the value of $z$ which greatly simplifies the weights/frequencies relationship.
\paragraph{Tuning by expectation.} Newton's methods are based on successive linear (or higher order) approximations in order to obtain numerical estimates of a root of a system of equations. It is generally an efficient algorithm assuming that the initial values are close enough to a root. Here, we are interested in finding the unique root $(z_0,\Weights_{\TargComp})$ of the system $\Freqs(z_0,\Weights)=n\TargComp$. Algorithm~\ref{algo1} is a slightly revisited version of Newton's method which tests at each step if Newton's approximation has improved the estimate of the root. This test fails if and only if the current parameters are too far from the solution. In this case,
we search using dichotomy an intermediate target that is closer to the solution than the current parameters.
\begin{proposition}\label{prop0}
Let $\TargComp$ and $n$ be the targeted composition and size respectively.
Assume that the Jacobian matrix $J(\mathbb{E}_{z_0}(\Weights_{\TargComp}))$ is not singular\footnote{I.e. there is no linear dependency between the expected numbers of different letters.}, then Algorithm~\ref{algo1} returns $(z_0,\Weights_1)$ such that the expected composition $\Freqs(z_0,\Weights_1)$ satisfies $||\Freqs(z_0,\Weights_1)-n\TargComp||<\epsilon$.\\
Moreover, there exists a neighborhood $B$ of $(z_0,\Weights_{\TargComp})$ such that, for any $\Weights_0\in B$, Algorithm~\ref{algo1} with
initial condition $\Weights_0$ quadratically converges to $\Weights_{\TargComp}$ (i.e. $\exists C>1$ such that $\forall k\geq 0, ||\Weights_k-\Weights_{\TargComp}||\leq C^{-2k}$ where $\Weights_{k+1} := J(\mathbb{E}_{z_0})^{-1}(\Weights_k)\cdot(n\TargComp-\mathbb{E}_{z_0}(\Weights_k))+\Weights_k$).
\end{proposition}
\begin{algorithm}[t]\label{algo1}\begin{framed}
\caption{Tracking the weights.}
\KwIn{Initial parameters $z_0$ and $\Weights_0$, a composition $\TargComp$, a size $n$ and $\epsilon$ a numerical precision}
\KwOut{The valid weights}
Let $\mathbb{E}_{z_0}$ be the map from the space of the weights into $\mathbb{R}^k_+$ such that $\mathbb{E}_{z_0}(\Weights)=\Freqs(z_0,\Weights)$\;
Let $J(\mathbb{E}_{z_0}(\Weights))$ be the Jacobian matrix of \YCorr{$\mathbb{E}_{z_0}$ evaluated at $\Weights$}\;
$\Weights:=\Weights_0$\;
\Repeat{end=true}{
end:=true; $\boldsymbol{c}:=n\TargComp$; $N:=||\boldsymbol{c}-\mathbb{E}_{z_0}(\Weights)||$\;
\While{$N>\epsilon$}{
$\Weights_{aux} := \Weights$\;
$\Weights := J(\mathbb{E}_{z_0})^{-1}(\Weights)\cdot(n\TargComp-\mathbb{E}_{z_0}(\Weights))+\Weights$\;
\If{$N<||\boldsymbol{c}-\mathbb{E}_{z_0}(\Weights)||$}{$\Weights:=\Weights_{aux}$; $\boldsymbol{c}:=(\boldsymbol{c}+\mathbb{E}_{z_0}(\Weights))/2$; end:=false;}
}}
\Return $\Weights$
\end{framed}\end{algorithm}
\paragraph{Asymptotic tuning.}
Since one generally attempts to generate large objects, a natural option consists in solving the simpler asymptotic system.
\begin{proposition}\label{prop0a}
Let us consider a language whose grammar is irreducible and aperiodic and whose generating function $\WGF{C}$ admits $\rho(\Weights)$ as dominant
singularity. \Olive{Then\YCorr{, for any letter $\At$ and as $z$ tends to $\rho(\Weights)$,} it holds that:}
{\centering\YCorr{\begin{tabular}{cl}
$\mathbb{E}_{z,\Weights}(N_\At)\sim \frac1{2}\WF_\At n\frac{\frac{\partial}{\partial \WF_\At}\rho(\Weights)}{\rho}$ & if $\rho(\Weights)$ is a rational singularity,\\
$\mathbb{E}_{z,\Weights}(N_\At)\sim -\WF_\At n\frac{\frac{\partial}{\partial \WF_\At}\rho(\Weights)}{\rho}$ & if $\rho(\Weights)$ is an algebraic singularity.
\end{tabular}}\\}
\end{proposition}
\begin{remark} Considering the expectation $\mathbb{E}_n(N_\At)$ of the number of letters $\At$ in a word of \Def{fixed} size $n$. Then, from~\cite{Denise2000}, similar asymptotic estimates
holds for $\mathbb{E}_n(N_\At)$ and the weights computed by our methods can therefore be used by the recursive approach.
\end{remark}
\section{Complexity of the multidimensional rejection (Phase \ref{phase:RejectComposition})}
Our approach relies on a rejection scheme that generalizes that of the classic -- univariate -- Boltzmann sampling.
Words are drawn from a weighted distribution -- \Olive{rejecting those whose frequencies
are too distant from the targeted one -- until an acceptable one is found and returned.
This gives the following rejection sampler $\Gamma_3\mathcal{A}{(x,\Weights;n,\boldsymbol{m},\varepsilon,\sigma)}$ for a language $\mathcal{A}$ where $x$ is real,
$\Weights$ a real $k$-vector, $\boldsymbol{m}$ a map from $\mathbb{N}$ to $\mathbb{R}^k$, and $\varepsilon$ the tolerance:\\
\begin{algorithm}[H]
\begin{framed}
\caption{$\Gamma_3\mathcal{A}{(x,\Weights;n,\boldsymbol{m},\varepsilon,\sigma)}$ }
\KwIn{The parameters $x,\Weights ,n,\boldsymbol{m},\varepsilon,\sigma$}
\KwOut{An object of $\mathcal{A}$ of size $s$ in $I(n, \varepsilon) $} and for every parameter $\WF_i$, the number of occurrences of $Z_i$ is in $I(m_i(s), \varepsilon,\sigma) := [m_i(s)-m_i(s)^\sigma\varepsilon, m_i(s)+m_i(s)^\sigma\varepsilon]$\\
\Repeat{ $\forall i, |\gamma|_i \in I(m_i(s), \varepsilon,\sigma)$ }{$\gamma := \Gamma_2\mathcal{A}{(x,\Weights;n,\varepsilon)}$}
\Return($\gamma$)
\end{framed}
\end{algorithm}
In many important classes of combinatorial structures, the composition of a random object is concentrated around its mean. It follows that
a rejection-based generation can succeed after few attempts, provided that the expected composition matches the targeted one.
Our main result is that, for any irreducible and simple context-free language, a suitably parameterized multidimensional rejection sampler
generates a word of targeted composition after $\BigO{n^{k/2}}$ attempts.
Moreover, allowing a $n^\beta$ ($\beta>1/2$) tolerance on the number of occurrences of each letters yields a sampler that succeeds in expected number
of attempts asymptotically constant.
Now, let us denote by $U_n(\Weights_0)$ the $k$-multivariate random variable which follows the probability $$\mathbb{P}(U_n(\Weights_0)=\boldsymbol{a})=\frac{\Weights_0^{\boldsymbol{a}}\cdot[z^n \Weights^{\boldsymbol{a}}]C_{\Weights}(z)}{[z^n]C_{\Weights_0}(z)},$$ i.e. the distribution of the parameters for objects of size $n$.
Moreover, let us denote by $\boldsymbol{\mu}(n,\Weights_0)$ the mean-vector of $U_n(\Weights_0)$ and by $\boldsymbol{V}(n,\Weights_0)$ its variance-covariance matrix.
If we do not have any strict correlation between the parameters, the matrix $\boldsymbol{V}(n,\Weights_0)$ is positive definite (and so, invertible). We can then \YCorr{associate a norm to each composition vector $\boldsymbol{u}$ through} $||\boldsymbol{u}||_{\boldsymbol{V}^{-1}}:=\sqrt{\boldsymbol{u}^T\boldsymbol{V}(n,\Weights_0)^{-1}\boldsymbol{u}}.$ Now, let $\boldsymbol{V}$ be a positive definite matrix, we denote by $\kappa\left(\boldsymbol{V}\right):=\inf\limits_{||\boldsymbol{u}||_\infty=1}\{||\boldsymbol{u}||_{\boldsymbol{V}} \}$, the infinum distance\footnote{Recall that the infinity norm is defined as $||\boldsymbol{u}||_\infty=\max{(|u_1|,\cdots,|u_k|)}$} from the unit sphere to the center of the Banach space.
\begin{definition}
The \emph{$\sigma$-concentrated condition} is defined as :\\
$$\limsup_{n\to \infty} \left(||\boldsymbol{\mu}(n,\Weights)||_\infty\right)^\sigma\cdot\kappa\left(\boldsymbol{V}(n,\Weights)^{-1}\right)=c>\sqrt{k}/\varepsilon.$$
\end{definition}
\begin{theorem}[Approximate composition]\label{theo3}
Let $x_n$ and $\Weights_{\boldsymbol{a}}$ be the solution of $\mathbb{E}_{x,\Weights}(N)=n$ and $\mathbb{E}_{x,\Weights}(N_i)=a_i$. The map $\boldsymbol{m}$ is defined as the \Olive{$\boldsymbol{m}: s\mapsto \mathbb{E}_{s,\Weights_{\boldsymbol{a}}}(N_i)$} and assume that
the $\sigma$-concentrated condition holds for some $\sigma\leq 1$.
Then the expected number of trials (of $\Gamma_2\mathcal{A}{(x_n,\Weights;n,\varepsilon)}$) of the rejection sampler $\Gamma_3\mathcal{C}{(x_n,\Weights_{\boldsymbol{a}};n,\boldsymbol{m},\varepsilon,\sigma)}$ is upper-bounded by $$\sup_{s \in I(n, \varepsilon) }\frac{\left(\varepsilon \cdot\kappa\left(\boldsymbol{V}(s,\Weights_{\boldsymbol{a}})^{-1}\right)\cdot||\boldsymbol{\mu}(s,\Weights_{\boldsymbol{a}})||^\sigma_\infty\right)^2}{\left(\varepsilon\cdot \kappa\left(\boldsymbol{V}(s,\Weights_{\boldsymbol{a}})^{-1}\right)\cdot||\boldsymbol{\mu}(s,\Weights_{\boldsymbol{a}})||^\sigma_\infty\right)^2-k}$$ which tends to a constant as $n\rightarrow \infty$.\\
\end{theorem}
\begin{theorem}[Exact composition]\label{theo4}
Assume that $(U_n(\Weights_{\boldsymbol{a}}))$
admits a multidimensional Gaussian law with mean $\boldsymbol{\mu}$ and variance-covariance matrix $\boldsymbol{V}$ proportional to $f(n)$ as limiting distribution when $n$ tends to the infinity,
then the exact-composition rejection sampler $\Gamma_3\mathcal{C}{(x_n,\Weights_{\boldsymbol{a}};n,\boldsymbol{m},0,1)}$ succeeds after an expected number of trials equal to $(2\pi)^{k/2}(\det(\boldsymbol{V}))^{1/2}=\BBigO{f(n)^{k/2}}$.
\end{theorem}
\begin{proof} Just notice that the probability to draw an exact composition corresponds to take $\boldsymbol{u}=\boldsymbol{\mu}$ in the asymptotic estimate $$p(\boldsymbol{u}) = \frac{1}{ (2\pi)^{k/2}\left(\det(\boldsymbol{V})\right)^{1/2} }
\exp\Big( {-\tfrac{1}{2}}(\boldsymbol{u}-\boldsymbol{\mu})^t \; \boldsymbol{V}^{-1}(\boldsymbol{u}-\boldsymbol{\mu})+o(1) \Big).$$
Consequently the expected number of attempts is $(2\pi)^{k/2}\det(\boldsymbol{V})^{1/2}=\BigO{f(n)^{k/2}}.$
\end{proof}
\subsection{Rational languages: Bender-Richmond-Williamson theorem}
The Bender-Richmond-Williamson theorem~\cite[Theorem~1]{Bender1983b} defines sufficient conditions such that the limiting
distribution of a rational language $\CS{R}$ is a multidimensional Normal distribution.
Let us remind that a rational language is \Def{irreducible} if its minimal automaton $\mathcal{A}$ is strongly-connected, and
\Def{aperiodic} -- if the cycle lengths in $\mathcal{A}$ have greatest common divisor equal to $1$.
Additionally the \Def{periodicity parameter lattice} $\Lambda$, defined in~\cite{Bender1983b} (Definition 2) is required to be full dimensional
to avoid trivial correlations in the occurrences of letters.
\begin{theorem}\label{th:rational}
Let $\CS{R}_{\Weights}$ be a weighted rational language whose minimal automaton is irreducible and aperiodic, and
$x_n$ be the root in $[0, \rho_{\Weights})$ of $\mathbb{E}_{x,\Weights}(N)=n$.
Assume that the periodicity parameter lattice $\Lambda$ is full dimensional; Then:
\begin{itemize}
\item $\forall \sigma>1/2$, the approximate-composition sampler $\Gamma_3 \CS{R}(x_n,\Weights; n,\varepsilon,\sigma)$ succeeds after $\BigO{1}$ trials
\item For $\sigma=1/2$, $\exists \varepsilon_0$ such that $\forall \varepsilon>\varepsilon_0$ $\Gamma_3 \CS{R}(x_n,\Weights; n,\varepsilon,\sigma)$ succeeds after $\BigO{1}$ trials
\item The exact-composition rejection sampler $\Gamma_3 \CS{R}(x_n,\Weights; n,0,1)$ succeeds after $\BigO{n^{k/2}}$ trials.
\end{itemize}
\end{theorem}
\begin{proof}
From the system of language equations $\boldsymbol {\mathcal{L}}={\boldsymbol M}\cdot \boldsymbol{\mathcal{L}}+\boldsymbol{\mathcal{E}}$,
we directly obtain the system $\boldsymbol {{L}}=z{\boldsymbol M}\cdot \boldsymbol{{L}}+\boldsymbol{{E}}$ for the generating function.
In this case the Perron-Frobenius theorem ensures that the dominating pole of every $L_i$ in $\boldsymbol{L}$ is \YCorr{the smallest positive real root}
of $\det(\mathbb{I}-z\cdot{\boldsymbol M})=0$ and that this pole is simple. Now, assume that the \emph{periodicity parameter lattice}
$\Lambda$ defined in~\cite{Bender1983b} (Definition 2) is full dimensional.
Assume also that we have a compact set $\Pi_1$ for the parameters in which the singular exponent is constant and equal to $1$. Then from the Bender-Richmond-Williamson theorem (see~\cite{Bender1983b},
Theorem 1 and~\cite{Bender1983a}), it follows that for any fixed parameter in the compact set $\Pi_1$, the limiting distribution of the parameters is a multidimensional Gaussian distribution with
mean and variance-covariance matrix proportional to $n$. Consequently, Theorem~\ref{theo3} applies for $\sigma>1/2$, Theorem~\ref{theo4} applies with $f(n) = n$, and the result follows.
\end{proof}
Let us discuss the prerequisites of Theorem~\ref{th:rational}. If the matrix $\boldsymbol{M}$ is not aperiodic, there exists a power $d$ such that $\boldsymbol{M}^d$ is aperiodic. So, we can always reduce the problem to a list of $d$ aperiodic ones, and Theorem~\ref{th:rational} applies under the same assumptions (full dimensional periodicity parameter lattice and compact set with constant singular exponents).
The \emph{irreducibility} requirement may be lifted when one of the strongly connected components dominates asymptotically, i.e. when the associated schema only involves subcritical and supercritical compositions \cite[Theorem~IX.2]{Flajolet2009}. However the case of a competition between different components in a non irreducible automaton is much more challenging and requires serious developments that cannot be included in this short paper. Finally we point out that, with minor modifications, similar results could be obtained for more general transfer matrix models.
\subsection{Context-free languages: Drmota theorem}
A theorem of~\cite{Drmota1997} gives very similar sufficient conditions for the limiting multivariate distribution to satisfy the conditions of Theorem~\ref{theo4}.
Namely, the irreducibility condition needs \YCorr{being} fulfilled by the \Def{dependency graph} of the grammar -- the directed graph on non-terminals whose
edges connect left hand sides of rules to their associated right-hand sides. The lattice and aperiodicity properties are replaced by the very similar concept of
\Def{simple type} grammar, requiring the existence of a \emph{positive} $k+1$ dimensional cone centered on $\boldsymbol{0}$ in the space of coefficients.
\begin{theorem}\label{th:contextfree}
Let $\CS{C}_{\Weights}$ be a weighted context-free language \YCorr{generated from a grammar $\mathcal{G}$
of simple-type \cite[Theorem~1]{Drmota1997} and whose dependency graph is strongly connected.}
Then the complexities summarized in Theorem~\ref{th:rational} also hold for $\CS{C}_{\Weights}$.
\end{theorem}
Again, the strong-connectedness requirement could be relaxed for disconnected grammars whose behavior is dominated by that of a single connected component. A formal characterization of such grammars \Olive{can be interpreted in} the theory of (sub/super)-critical compositions \cite[Theorem~IX.2]{Flajolet2009}.
\section{Introduction}
Random generation is the core of the simulation of complex data. It
appears in real applicative domains such as complex networks (biology, Internet or social relationship), or software testing (validation, benchmarking).
It helps us to predict the behavior of algorithms (complexities and statistical significance of results), to visualize limit properties (such as transition
phases in statistical physics), to model real contexts (random graphs for web simulation).
\begin{comment}
This simulation of the real discrete world first implies purely combinatorial
modelizations, relying on a deep understanding of the properties of the
real data, together with a good expertise in combinatorial methods for
decomposing and transforming real data into combinatorial structures
providing a good representation.
\end{comment}
Following the pioneering work of Flajolet \emph{et al}~\cite{Flajolet1994}, decomposable combinatorial classes can be specified
using standard specifications. Two major techniques can then be applied to draw $m$ objects of size $n$ at random from such a class.
On one hand, the recursive approach~\cite{Wilf1977} precomputes the cardinalities of sub-classes for sizes up to $n$ and uses these numbers to
perform local choices that are consistent with the targeted uniformity. The best known optimization of this technique~\cite{Denise2000} uses
certified floating point arithmetics and works in $\BigO{m\cdot n^{1+o(1)}}$ but its implementation remains highly non-trivial due to its sophisticated precomputations.
On the other hand, the Boltzmann sampling techniques, recently introduced by Duchon \emph{et al}~\cite{Duchon2004}, achieves a random generation
for most unlabelled~\cite{Flajolet2007} and labelled specifications in $\BigO{m\cdot n^2}$ operations at an optimally low $\BigO{m\cdot n}$ memory cost.
Instead of enforcing a strict -- and costly -- control on the size of generated objects, this general technique rather induces an appropriate
distribution on the size of sampled objects, and performs rejection until a suitable object is found.
In the present work, we investigate a natural multivariate extension of Boltzmann sampling, aiming at drawing objects, uniformly at random,
having a prescribed composition in the different terminal letters. From a combinatorial perspective, such a generation allows the so-called
symbolic method to reclaim combinatorial classes and languages that \emph{fall slightly off} of its natural expressivity. For instance, restrictions
of rational languages may not admit a rational (or even context-free) specification under the additional hypothesis that some letters co-occur strictly
(One may consider the triple-copy language).
For context-free languages on $k$ letters, this problem was previously addressed within the recursive framework~\cite{Wilf1977} by Denise \emph{et al}~\cite{Denise2000}, deriving algorithms in
$\Theta(n^k)$ and $\Theta(n^{2k})$ arithmetic operations, respectively for rational and context-free languages.
Using properties of holonomic series, Bertoni\emph{ et al}~\cite{Bertoni2003} revisited the problem and proposed a method for the uniform sampling from rational languages on two letters in $\Theta(n)$.
Unfortunately a direct generalization of the technique yields an algorithm in $\Theta(n^{k-1})$ for $k$ letters, as pointed out in Radicioni's thesis~\cite{Radicioni2006}.
Following the general philosophy of Boltzmann sampling, our algorithm will first relax the compositional constraint, using non-uniform samplers to draw objects whose average
composition is fine-tuned to match the targeted one, and perform rejection until an acceptable object is found. By acceptable, one understands that
generated objects must feature prescribed size and composition, while tolerances may be allowed on both requirements.
Our programme can then be summarized in the three following phases:
\begin{enumerate}[{Phase }I.]
\item\label{phase:ComputeWeights} Figure out a set of weights such that the expected composition matches the targeted one.
\item\label{phase:GenerateWeighted} Draw structures from a weighted distribution, using either the recursive approach (See~\cite{Denise2000})
or a weighted Boltzmann sampler (See Section~\ref{sec:WeightedBolztmannSamplers}).
\item\label{phase:RejectComposition} Reject structures of unsuitable compositions, until an adequate object is generated and returned.
\end{enumerate}
Although phases \ref{phase:GenerateWeighted} and \ref{phase:RejectComposition} are independently addressed in our analyses,
one can (and will) combine them into a single rejection step when a weighted Boltzmann sampler is used for Phase~\ref{phase:GenerateWeighted}.
The algorithmic aspects of our programme will essentially build on and extend previous works addressing the uniform version, but a general analysis
of its overall performance is more challenging. Indeed, the complexity of the rejection Phase~\ref{phase:RejectComposition} is heavily related to a general
analysis of the limiting distribution of the associated multivariate -- parameter-induced -- generating functions.
For each phase, we attempt to give mathematical characterizations of classes having proper behaviors. In particular, for context free languages whose grammars are \Def{strongly connected} and \Def{aperiodic},
we obtain for each combination of tolerances, the complexities summarized in Table~\ref{tab:compSummary}.
\begin{table}[t]
\centering
{\setlength{\tabcolsep}{8pt}
\renewcommand{\arraystretch}{1.5}
\YCorr{\begin{tabular}{| c | c | c | c |}\hline
\multicolumn{2}{|c|}{\multirow{2}{*}{{Tolerance}}} & \multicolumn{2}{|c|}{Composition}\\ \cline{3-4}
\multicolumn{2}{|c|}{} & None & $\Omega(\sqrt{n})$\\ \hline
\multirow{2}{*}{{Size}} & None & $\BigO{n^{2+k/2}}$ & $\BigO{n^{2}}$ \\
& $\Theta({n})$ & $\BigO{n^{1+k/2}}$ & $\BigO{n}$\\ \hline
\end{tabular}}\\[1em]
}
\caption{\YCorr{Average-case complexities of our samplers for a word of length $n$ over $k$ letters in strongly connected context-free languages under different tolerances.}}\label{tab:compSummary}
\end{table}
The plan of this paper follows the different phases : Section 2 defines the concepts and notations used throughout the
paper. Section~3 explains how to tune efficiently the parameters such that the targeted composition matches the average behavior (Phase~\ref{phase:ComputeWeights}).
In Section~4, we discuss the complexity of Phase~\ref{phase:GenerateWeighted}, the number of rejections needed to reach a word of suitable size
(or suitable approximate size). The complexity of the multidimensional rejection (Phase~\ref{phase:RejectComposition}) is addressed in Section~5.
We illustrate our method in Section~6 by sampling perfect Tetris tessellations -- tessellations of a $w\times h$ rectangles using balanced lists of tetraminoes.
Finally we conclude with a short overview of future works.
\section{Notations and definitions}
Following traditional mathematical notations, we will use bold symbols for multi-dimensional variables/functions (i.e. ${\boldsymbol x}$), and use subscripts to access a specific dimension (i.e. $x_i$).
Throughout the rest of the document, we will denote by $\Atoms$ the \Def{alphabet} of $k$ letters, by $\CS{C}$ a \Def{context-free language} over $\Atoms$, and by $n$
the length of generated words.
\subsubsection*{Composition and tolerance.}
Define the \Def{composition} of sampled words as the \Def{frequency} of occurrences of each letter $\At_i$ in a
word $\Struct\in\CS{C}$, denoted by $\Comp{\Struct} := \left(|\Struct|_{\At_i}/n\right)_{i\in [1,k]}.$
Our main goal is to generate -- uniformly at random -- some word $\Struct\in\CS{C}$ having a composition that is \emph{close to} a \Def{targeted
composition} $\TargComp\in[0,1]^k$ such that $\sum_{i\in[1,k]} \TargComp_i = 1$.
We make this notion of proximity explicit, and formalize the notion of acceptability for a sampled word.
Namely let $\ErrFactor$ be a $k$-tuple of positive real numbers and $\ErrExp\in \mathbb{Q}^+$ a rational exponent, an object $\Struct\in\CS{C}$
qualifies as \Def{$(\ErrFactor, \ErrExp)$-acceptable} if and only if
$$ \CompE{\Struct}{i} \in I(\Target_i, \ErrFactorEntry_i,\ErrExp) ,\;\text{for all } i\in[1,k]$$
where \Olive{$I(f,e,a) := [f-f^{a}n^{a-1} e, f+f^{a}n^{a-1} e]$.}
This definition captures the case of fixed (exact) compositions by setting $\ErrExp=1$ and $\ErrFactorEntry_i = 1/n, \text{for all } i\in[1,k]$.
\subsubsection*{Weighted distributions.}
\begin{figure}[t]
\begin{equation*}
\renewcommand{\arraystretch}{1.8}
\setlength{\tabcolsep}{8pt}
\YCorr{\begin{array}{| l l l l|}\hline \label{tab1}
\text{ Epsilon} & \CS{C} = 1 & \WGF{C} = 1 & \Gen{C} := \varepsilon\\
\text{ Letters} & \CS{C} = \At_i & \WGF{C} = \AtW{\At_i}z & \Gen{C} := \At_i\\
\text{ Union} & \CS{C} = \CS{A} + \CS{B} & \WGF{C} = \WGF{A} +\WGF{B} & \Gen{C} := \Bern\left(\displaystyle\frac{\WGFb{A}{x}}{\WGFb{C}{x}},\frac{\WGFb{B}{x}}{\WGFb{C}{x}}\right) \longrightarrow \Gen{A}\;|\;\Gen{B}\\
\text{ Product} & \CS{C} = \CS{A} \times \CS{B} & \WGF{C} = \WGF{A}\times \WGF{B} & \Gen{C} := \Gen{A}.\Gen{B}\\ \hline
\end{array}}
\end{equation*}
\caption{ \YCorr{Weighted generating functions and associated Boltzmann sampler $\Gen{C}$ for context-free languages.}\label{fig:constructions}}
\end{figure}
The following notions and definitions, recalled here for the sake of self-containment, can be found in Denise~\emph{et al}~\cite{Denise2000}.
A positive \Def{weight} vector $\Weights$ assigns positive weights $\WF_i\in\mathbb{R}^+$ to each letter $\At_i\in\Atoms$.
The weight is then extended multiplicatively on any object $\Struct$ by $\Weight{\Struct} = \prod_{x \in \Struct} \AtW{x}.$
This gives rise to the notion of \Def{weighted generating function} $\WGF{C}$ for a context-free language $\CS{C}$, a natural
generalization of the size (enumerative) generating function where each structure is counted with multiplicity equal to its weight
$$ \WGF{C} = \sum_{\Struct\in\CS{C}} \Weight{\Struct}z^{|\Struct|}
\YCorr{= \sum_{\Struct\in\CS{C}} \WF_{\At_{1}}^{|\Struct|_{\At_{1}}}\cdots \WF_{\At_k}^{|\Struct|_{\At_k}} z^{|\Struct|}}
= \sum_{n\ge 0} \PF{n} z^n$$
where $\PF{n}$ is the total weight\footnote{This quantity is essentially similar to the partition function in statistical mechanics, introduced by L. Boltzmann.}
of objects of size $n$. Notice that this generating function can be re-interpreted as a multivariate generating function in $\Weights$ and $z$
This weighting scheme implicitly defines a \Def{weighted distribution} on the set $\CS{C}_n$ of words of size $n$, such that
$$ \WProb{\Struct\;|\;\YCorr{n=|\Struct|}} = \frac{\Weight{\Struct}}{\sum_{\Struct'\in \CS{C}_n} \Weight{\Struct'}} = \frac{\Weight{\Struct}}{\PF{n}}.$$
Finally, the weighted distribution \YCorr{generalizes to} a {\bf Boltzmann weighted distribution} on the whole language such that
\begin{equation}\Prob{\Struct\;|\;\YCorr{n=|\Struct|}} = \frac{\Weight{\Struct}x^n }{\sum_{\Struct'\in \CS{C}} \Weight{\Struct'}x^{|\Struct'|} } = \frac{\Weight{\Struct}x^n}{\WGFun{C}(x)}.\label{eq:WeightedBoltzmannDistribution}\end{equation}
\begin{property}
Let $N$ (resp. $N_i$) be the random variable \YCorr{associated with} the size (resp. number of occurrences of a letter $\At_i$) of a word in
a $(x,\Weights)-$ Boltzmann weighted distribution over a class $\CS{C}$.
Then the \YCorr{expectations of $N$ and $N_i$
are \YCorr{related to} the partial derivatives of the multivariate generating function $\WGF{C}$ through
\begin{equation}
\mathbb{E}_{x,\Weights}(N) = x\dfrac{\frac{d\WGFun{C}(x)}{dx}}{\WGFun{C}(x)} \text{\YCorr{\quad \quad and \quad \quad }} \mathbb{E}_{x,\Weights}(N_i) = \dfrac{\WF_i\frac{\partial}{\partial \WF_i}\WGFun{C}(x)}{\WGFun{C}(x)}\label{eq:expectations}
\end{equation}
\end{property}
In the sequel we will denote by $\Freqs(x,\Weights)$
the vector of expectations $(\mathbb{E}_{x,\Weights}(N_1),\cdots,\mathbb{E}_{x,\Weights}(N_k))$.
\begin{comment}
\subsubsection*{Rational generating functions of rational languages}
Let us consider the weighted generating function $\RatGF$ of a rational language $\RatLang$.
It is a classic result of language theory that any rational language $\RatLang$ is recognized by a deterministic minimal automaton $\mathcal{A}$. Such an
automaton can in turn be transformed into a left-linear grammar generating exactly $\RatLang$.
From the deterministic nature of the automaton, the resulting grammar is unambiguous and only uses the Cartesian product and disjoint union operators,
respectively denoted by $\times$ and $+$ in the so-called symbolic method. One can therefore systematically translate such a grammar, using
rules summarized in Table~\ref{fig:constructions}, into a system of functional equations involving the weighted generating
function associated with each state. The system can be expressed using the following classical matrix representation
$\boldsymbol {\mathcal{L}}={\boldsymbol M}\cdot \boldsymbol{\mathcal{L}}+\boldsymbol{\mathcal{E}}$, where $\boldsymbol{\mathcal{E}}$ is
the final state $\{0,\epsilon\}$-vector and the matrix ${\boldsymbol M}$ is the \Def{transition matrix} of $\mathcal{A}$.
The resulting system is linear in each of the generating function associated with non-terminals, and can therefore always be solved
through Gaussian elimination. By considering the non-terminal/generating function associated with the initial state, one obtains the
weighted generating function $\RatGF$ of $\RatLang$. Since this function is the solution of a linear system, it can be written as a
fraction of polynomials in $z$ and $\Weights$, i. e. a {\bf rational generating function}.
\end{comment}
\input{ComputingWeights-Rational}
\input{WeightedGeneration-Rational}
\input{EfficiencyRejection-Rational}
\input{tetris}
\section{Conclusion}
In this paper, we adapted and applied a general methodology for the multivariate random generation of combinatorial objects.
Under explicit and natural conditions, random generators having complexity in $\mathcal{O}(n^{2+k/2})$ were derived for the exact size
and composition generation, outperforming best known algorithms (in $\mathcal{O}(n^k)$ and $\mathcal{O}(n^{2k})$ respectively for rational
and context-free languages) for this problem. Furthermore, provided a small (linear) tolerance is allowed on the size of generated objects,
and a $\Omega(\sqrt{n})$ one is allowed in the other dimensions, our generators generate objects in linear expected time. We applied these
principles to the generation of perfect Tetris tessellations with uniform statistic in tetraminoes and discussed the generation of Tetris
games from this model.
This paper is the first step toward a general analysis of the multi-parameters Boltzmann sampling. Compared to its alternative using the recursive
method, the resulting method is not only theoretically faster, but also only requires $\BigO{n}$ storage and its time complexity seems less
affected by larger specifications.
Nevertheless, many questions are left open, for instance with respect to the nature of the dependency between the
weights and \emph{reasonable} frequencies, which would allow us to address the complexities of Phase 2 in a much more general setting.
Furthermore the success of our programme critically depends on the existence of suitable weights, which is not guaranteed, e.g. when the targeted
distribution is incompatible with some dependencies induce by the grammar. A future direction of this work might investigate non-trivial, sufficient -- yet tight --
conditions such that the targeted composition can be achieved on the average.
Since multivariate Boltzmann samplers can be obtained in any situation where the distribution is well-concentrated, one may envision extensions to other classes, including
constrained trees, permutations with a fixed number of cycles, functional graphs with a controlled number of components\ldots
A first extension may consider simple Polya operators and extend some of the multivariate theorems established in the present work.
The requirement of strong-connectedness (or irreducibility) could be questioned or categorized using (sub/super)-critical
compositions. Another direction is the use of Hwang's Quasi-powers theorem, giving rise to low variance distributions, for a general treatment
of the bivariate case.
\section*{Acknowledgements}
\YCorr{The authors wish to express their gratitude toward Pierre Nicodeme for his thorough inspection of a preliminary version of the manuscript.
This work was supported by the ANR-GAMMA 07-2$\_$195422 grant of the French \emph{Agence Nationale de la Recherche}.}
\bibliographystyle{amsplain}
\section{Efficiency of the size rejection scheme (Phase~\ref{phase:GenerateWeighted})}
At this point, we assume that a $k$-tuple of weights $\Weights$ has been found such that the average composition in the
weighted distribution matches the targeted one. We now need to perform a random generation of $m$ words
from the context-free language with respect to the $\Weights$-weighted distribution.
This problem was previously addressed in \cite{Denise2000} within the framework of
the recursive method, and an algorithm in $\BigO{m\cdot n}$ arithmetic operations was proposed.
Despite its apparent low complexity, the exponential growth of the numbers processed by the algorithm
increases the practical complexity to $\Theta(m\cdot n^2)$ in time and $\Theta(n^2)$ in memory.
\label{sec:WeightedBolztmannSamplers}
Let us investigate a weighted generalization of Boltzmann sampling.
First let us remind that Boltzmann sampling first relaxes the size constraint and draws objects in a Boltzmann distribution of parameter $x$.
To that purpose a fixed stochastic process, coupled with an (anticipated) rejection procedure,
is used (See \Olive{Algorithm~\ref{algo:rejection}}). The probabilities of the different alternatives are precomputed by an external procedure
called \Def{oracle} (Symbolic algebra, or numerical method in \cite{Pivoteau2008}).
A judicious choice of value for $x$ ensures a low probability of rejection and this approach yields, for large classes of structures (trees, sequences, runs,
mappings, fountains\ldots), generic algorithms in $\BigO{n^2}$ for objects of \Def{exact-size} $n$,
and in $\BigO{n}$ for objects of \Def{approximate-sizes} in $[n(1-\varepsilon),n(1+\varepsilon)]$, for some $\varepsilon>0$.
Through a minor modification of the oracle, one can easily turn unlabelled Boltzmann samplers, introduced in \cite{Flajolet2007}, into generators
for the weighted Boltzmann distribution (See Equation~\ref{eq:WeightedBoltzmannDistribution}).
Namely, one only needs to replace any occurrence of the generating function $\GF{C}(z)$ \YCorr{by} its weighted counterpart $\WGF{C}$, obtaining generic samplers
summarized in Figure~\Olive{1}, and use the classic size rejection process (Algorithm~\ref{algo:rejection}).
\begin{algorithm}[t]
\begin{framed}
\caption{Rejection algorithm $\Gamma_2\mathcal{A}{(x,\Weights;n,\varepsilon)}$ \label{algo:rejection}}
\KwIn{Parameters $x,\Weights$}
\KwOut{Object of $\mathcal{A}$ of size in $I(n, \varepsilon) := [n(1-\varepsilon), n(1+\varepsilon)]$}
\Repeat{$|\gamma| \in I(n, \varepsilon)$ }{$\gamma := \Gen{A}$}
\Return($\gamma$)
\end{framed}
\end{algorithm}
\begin{proposition}\label{prop1}
Let $\Weights$ be a $k$-tuple of weights, $x$ be a Boltzmann parameter, $C$ be a context-free specification
and $\WGF{C}$ its weighted generating function.\\
Then \YCorr{the samplers $\Gen{C}$ summarized in Figure~\Olive{1} generate} any word $w\in\CS{C}$ with probability
$$\Prob{\Struct\;|\;n} = \frac{\Weight{\Struct}x^n}{\WGFun{C}(x)}.$$
\end{proposition}
The (renormalized) restriction of a $\Weights$-weighted Boltzmann distribution to objects of size $n$ is clearly a
$\Weights$-weighted distribution, and this fact ensures the correctness of a rejection-based approach.
Let us qualify a context-free language as \Def{well-conditioned} iff the singular exponent $\alpha_{\Weights}$ of its dominant singularity is non negative.
\YCorr{Following \cite{Duchon2004}, we observe that any grammar can be pointed repeatedly until the exponent of its generating function becomes non-negative. Moreover the pointing operator leaves a weighted distribution unaffected within the subset of words of a given length. Therefore we can restrict our analysis to grammars associated with \emph{flat} Boltzmann \Olive{distributions}, generate words from the pointed grammars and \emph{erase} the point(s) afterward.}
\begin{theorem}[Essentially proven in \cite{Duchon2004}]\label{theophase2}
Let $\CS{C}_{\Weights}$ be a weighted well-conditioned context-free language and $x_n$ be the root in $[0, \rho_{\Weights})$ of $\mathbb{E}_{x,\Weights}(N)=n$.
Then the complexity $X_{\varepsilon}[n]$ of the sampler $\Gamma_2 \CS{C}(x_n,\Weights; n,\varepsilon)$ described in Algorithm~\ref{algo:rejection} is such that
\begin{itemize}
\item If $\varepsilon=0$ (exact size): $X_{\varepsilon}[n] \in \BBigO{\dfrac{\kappa\Gamma(\alpha_{\Weights})n^2}{\alpha_{\Weights}^{\alpha_{\Weights}}}+c(\Weights)n}$, and
\item If $\varepsilon>0$ (approximate-size): $X_{\varepsilon}[n] \in \BBigO{\dfrac{\kappa n}{\zeta_{\alpha_{\Weights}}(\varepsilon)}+c(\Weights)}$
\end{itemize}
where $\kappa$ is the cost-per-letter induced by the canonical Boltzmann samplers, $\alpha_{\Weights}$ is the singular exponent of the dominant singularity of $\WGF{C}$, $\zeta_{\alpha_{\Weights}}(\varepsilon):=\dfrac{\alpha_{\Weights}^{\alpha_{\Weights}}}{\Gamma(\alpha_{\Weights})}\displaystyle\int_{-\varepsilon}^{\varepsilon}(1+s)^{\alpha_{\Weights}-1}e^{-\alpha_{\Weights}(1+s)}ds$,
$\Gamma(x)$ is the gamma function, and $c(\Weights)$ does not depend on $n$.
\end{theorem}
In particular, for any fixed weight vector $\Weights$, Theorem~\ref{theophase2} implies a $\BigO{n}$ (resp. $\BigO{n^2}$) complexity for the approximate-size (resp. exact size) weighted
samplers. By contrast, using weights to enforce compositions that are unnatural (e.g. enforcing $\BigO{\sqrt{n}}$ occurrences of a letter occurring $\BigO{n}$ times in the uniform distribution) may lead
to a -- somewhat hidden --
dependency of $\Weights$ in $n$. Although we were unable to characterize these dependencies and their impact $c(\Weights)$ on both complexities, we expect the latter to remain limited, and conjecture similar complexities when \emph{meaningful} compositions are targeted. \YCorr{For instance, assuming at least one occurrence of each letter (a realistic assumption, since prohibition of a letter is simply achieved through a grammar modification), and the frequencies and the weights can therefore be assumed to be bounded by some function of $n$.}
In the case of rational languages, the following theorem provides a computable evaluation for the efficiency of the size-rejection process. It relies on the partial fraction expansion of rational functions, which can be obtained for any weighted generating function $C_{\Weights}(z)$, and is denoted by
\begin{equation}\label{eq1} C_{\Weights}(z)=\sum\limits_{i=1}^{r}\sum\limits_{k=1}^{m_i}(1-z/\rho_i)^{-\alpha_{i,k}}h_{i,k}+P(z)
\end{equation}
where \YCorr{$P(z)$ is a polynomial of degree bounded by a constant, $r$ the number of distinct singularities} and $m_i$ the multiplicity of $\rho_i$ which are sorted by increasing module.
In weighted generating functions \YCorr{the values of} $\rho_i$, $P(z)$, $h_{i,k}$, $k$ and $r$ depend on the actual values of the weights.
\begin{theorem}
Let $\CS{C}_{\Weights}$ be a weighted rational language, $x_n$ be the root in $[0, \rho_{\Weights})$ of $\mathbb{E}_{x,\Weights}(N)=n$ \YCorr{and $\varepsilon>0$ be a tolerance then the approximate-size sampler $\Gamma_2 \CS{C}(x_n,\Weights; n,\varepsilon)$ succeeds after an expected number of trials of $\Gamma C_{\Weights}(x,b)$ in}
$$\dfrac{C_{\Weights}(x_n)}{\left(\sum\limits_{i=1}^{r}\sum\limits_{k=1}^{m_i}\binom{n+k-1}{k-1}(\rho_i)^{-n}h_{i,k}+[z^n]P(z)\right)(x_n)^n}.$$
\end{theorem}
\section{Sampling perfect Tetris tesselations}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathcal{B}}{\mathcal{B}}
\newcommand{0.2}{0.2}
\newcommand{\includegraphics[scale=\Scale]{pics/SBlock}}{\includegraphics[scale=0.2]{pics/SBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/ZBlock}}{\includegraphics[scale=0.2]{pics/ZBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/LBlock}}{\includegraphics[scale=0.2]{pics/LBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/JBlock}}{\includegraphics[scale=0.2]{pics/JBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/OBlock}}{\includegraphics[scale=0.2]{pics/OBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/IBlock}}{\includegraphics[scale=0.2]{pics/IBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/TBlock}}{\includegraphics[scale=0.2]{pics/TBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/SSBlock}}{\includegraphics[scale=0.2]{pics/SSBlock}}
\newcommand{\includegraphics[scale=\Scale]{pics/ZZBlock}}{\includegraphics[scale=0.2]{pics/ZZBlock}}
\newcommand{\mathcal{B}}{\mathcal{B}}
\newcommand{\mathcal{P}}{\mathcal{P}}
\newcommand{p}{p}
In this short illustration, we address the generation of Tetris tessellations, i.e. tessellations using tetraminoes of a board having prescribed width $w$.
The Tetris game consists in placing falling tetraminoes (or {\bf pieces}) $\mathcal{P}$ in a $w\times h$ board.
The goal of the player is to create hole-free horizontal lines which are then eliminated, and the game goes on until \YCorr{some piece stacks
past the board ceiling}.
Most implementations of Tetris use the so-called \emph{bag strategy}, which consists in giving the player sequences of permutations of the 7 types of tetraminoes,
therefore inducing a \Olive{uniform} composition in each tetramino type. A rational specification (Built by Algorithm~\ref{algo:statespace}) exists for Tetris
tessellations of any fixed width, but the additional constraint on composition provably throws the associated language out of the context-free class.
Therefore, we choose to model the generation of uniformly distributed Tetris tessellations as a multivariate generation \YCorr{within} a rational language.
Such tessellations could in turn be used as a basic construct to build hard instances for the offline version of the algorithmic
Tetris problems studied in~\cite{Breukelaar2004} and \cite{Hoogeboom2004}.
\subsection{Building the automaton of Tetris tesselations}
\begin{algorithm}[t]
\begin{framed}
\begin{tabular}{cc}
\begin{minipage}{.46\textwidth}
\KwIn{Board width $w$ and flat boundary $\mathcal{B}_w$}
\KwOut{$Q$ the states set and $\sigma$ the transition \\ \quad\quad function of $\mathcal{A}_w=\left(\mathcal{P},Q,\mathcal{B}_w,\{\mathcal{B}_w\},\sigma \right)$}
\Begin
{
$(Q,\sigma)\leftarrow(\mathcal{B}_w,\varnothing)$\\
$S \leftarrow \{\mathcal{B}_w\}$\\
\While{$S \neq \varnothing$}
{
$S \Rightarrow_{\text{pop}} \mathcal{B}$\;
\For{$p \in \mathcal{P}_{\mathcal{B}}$}
{
$\mathcal{B}' \leftarrow \mathcal{B}- p$\;
\If{$\mathcal{B}'\notin Q$}
{
$Q \leftarrow Q \cup \{\mathcal{B}'\}$\;
$S \Leftarrow_{\text{push}} \mathcal{B}'$\;
}
$\sigma \leftarrow \sigma \cup \{(\mathcal{B},p,\mathcal{B}')\} $\;
}
}
\Return{$(Q,\sigma)$}
}
\end{minipage} & \begin{tabular}{c|c|c}
Width $w$ & \#States in $\mathcal{A}_w$ & \#States minimal\\ \hline
2 & 4 & \YCorr{4}\\
3 & 55 & \YCorr{55}\\
4 & 80 & 78\\
5 & 1686 & 1646\\
6 & 4247 & 4130\\
7 & 41389 & 40099\\
8 & 49206 & 47564\\
9 & 919832 & --
\end{tabular}
\end{tabular}
\caption{Constructing the automaton $\mathcal{A}_w$ for tessellations of width $w$. Right: Growth of the number of states for increasing values of $w$.}\label{algo:statespace}
\end{framed}
\end{algorithm}
First let us find an unambiguous decomposition of Tetris tessellations.
The idea is to focus on the state of the upper band of the tessellation \YCorr{of height $4$}, or \Def{boundary} of a partial tessellation.
In particular for (complete) Tetris tessellations the upper band is completely filled and the associated boundary is \Def{flat}.
One can investigate the different ways to get to a given boundary $\mathcal{B}$ by simulating the \Def{removal} from $\mathcal{B}$ of a
piece $p$\YCorr{, completing the boundary after each removal so that the highest non-empty position stays on the top row.}
Without further restriction on the position of removal, such a decomposition would be \emph{ambiguous} and give rise to an infinite number
of different boundaries. Consequently, we enforce a canonical order on the removal of pieces
by restricting it to a set of (possibly rotated) pieces $\mathcal{P}_{\mathcal{B}}$ positioned such that: \YCorr{a) the upper-rightmost position of the piece
matches that of the boundary and b) the piece is entirely contained in the boundary.
We refer to the induced decomposition as the \Def{disassembly decomposition}}.
\begin{proposition}
The disassembly decomposition generates sequences of removals from and to flat boundaries that are in bijection with Tetris tesselations.
\end{proposition}
\begin{proof}[Sketch]
Let us discuss briefly the correctness of this decomposition, or equivalent that the sequences of $k$ removals leading from a \Def{flat boundary}
$\mathcal{B}_w$ to itself are in bijection with the tessellations of width $w$. First let us notice that the decomposition is unambiguous,
since all the local removals share at least one position (the upper-rightmost of the boundary)
and are therefore strongly ordered. Furthermore \Olive{the \YCorr{decomposition} is also provably complete by induction on the number $n$ of piece}, since any
tessellation has a upper-rightmost position which, upon removal, gives another tessellation of smaller size, and completeness of
the decomposition propagates from tessellations of size $n$ to size $n+1$. Finally, it gives rise to a finite number of states since
the difference between the highest and lowest point in any reached boundary does not exceed the maximal height of a piece.
\end{proof}
The finiteness
of the state space suggests Algorithm~\ref{algo:statespace} that builds the automaton $\mathcal{A}_w$, generating tessellations of width $w$. Notice that the resulting automaton in not necessarily co-accessible, since the removal of some piece can create boundaries that cannot be completed
into a flat one through any sequence of removal.
Consequently, we added in our implementation a test of connectedness that discards any boundary having a (dis)connected component involving a number
of blocks that is not a multiple of 4, as such boundaries clearly cannot reach a flat state again.
Running a minimization algorithm of the resulting automata confirms the expected explosion in the number of states (See Algorithm~\ref{algo:statespace}) required for increasing values of $w$.
\subsection{Random generation}
First we point out that the automaton has matching initial and final states, so the strong connectedness is obviously ensured
and our theorems regarding the complexity of \YCorr{our generators} apply. One can then translate \YCorr{the automaton transitions} into a system of functional equations
involving the (rational) generating functions associated with each states. Solving the system gives the generating functions, from which one can extract many
informations.
For instance, fixing the width $w=6$ and a number $n=105$ of pieces, one obtains a number $h_{6,105} = 3.10^{71}$ of potential tessellations, and extracting
coefficients of suitable derivatives yields:
{\centering
\begin{tabular}{c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c}
Piece & \includegraphics[scale=\Scale]{pics/ZBlock}& \includegraphics[scale=\Scale]{pics/OBlock}& \includegraphics[scale=\Scale]{pics/LBlock}& \includegraphics[scale=\Scale]{pics/JBlock}& \includegraphics[scale=\Scale]{pics/IBlock}& \includegraphics[scale=\Scale]{pics/SBlock} & \includegraphics[scale=\Scale]{pics/TBlock}\\
Frequency ($\%$) & 7.90& 10.55 & 20.42 & 20.42& 17.00 &7.90& 15.81\\
\end{tabular}\\} \medskip
Consequently, the average composition of a Tetris tessellation is incompatible with the \emph{bag strategy}, which \YCorr{induces
uniformly distributed pieces.} \YCorr{One can then use the results of Section~\ref{tuning} to compute a set of weights
that ensures $1/7$-th proportions in each type of pieces.}
{\centering
\YCorr{\begin{tabular}{c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c}
Piece & \includegraphics[scale=\Scale]{pics/ZBlock}& \includegraphics[scale=\Scale]{pics/OBlock}& \includegraphics[scale=\Scale]{pics/LBlock}& \includegraphics[scale=\Scale]{pics/JBlock}& \includegraphics[scale=\Scale]{pics/IBlock}& \includegraphics[scale=\Scale]{pics/SBlock} & \includegraphics[scale=\Scale]{pics/TBlock}\\
Weight & 0.93&0.84 &0.38 &0.38 &0.46&0.93&0.42\\
Frequency ($\%$) & 14.3& 14.1& 14.2& 14.2& 14.2& 14.3& 14.5
\end{tabular}\\}}
\medskip
A weight random generation for the $w=6$ and $n=105$, coupled with a rejection that allows the numbers of any piece to
be equal to $15\pm 1$, gives the instances drawn in Figure~\ref{fig:example}.
\subsection{From random Tetris tessellations to Tetris instances}
\begin{figure}[t]
\centering \includegraphics[width=32em]{pics/singleInstance}
\caption{{\bf Left:} Tetris tessellations associated with a unique instance.
Only the most relevant dependency points are displayed here (arrows) and pieces are labelled with their rank in the only compatible instance. Duplicating the gadget preserves the uniqueness of the associated instance while allowing for the generation of tessellations of arbitrarily large dimensions. {\bf Right:} Tesselation realized by $\binom{h}{h/2}\in \Theta(2^n/\sqrt{n})$ different instances.}\label{fig:unique}
\end{figure}
\begin{proposition}
For any Tetris tesselation $\mathcal{T}$, there exists an instance (sequence of pieces) such that $\mathcal{T}$ can be obtained.
\end{proposition}
\begin{proof}
Let us assume that $\mathcal{T}$ is a tessellation of a $w\times n$ rectangle using tetraminoes,
and let us call \emph{dependency point} any contact between the southward face of a piece $\mathcal{B}_1$ and the northward
face of a piece $\mathcal{B}_2$. \YCorr{Such points induce dependencies $\mathcal{B}_1 \to \mathcal{B}_2$, which are the arcs of a \emph{dependency graph} $D=(\mathcal{T},E)$.}
Additionally, each edge is labelled with the coordinate of its associated dependency point.
It can be shown that $D$ is acyclic, by pointing out that any path along $D$ is labelled with coordinates that are either
increasing on the $y$-axis or monotonic on the $x$-axis. Let us start by noticing that, aside from the $\includegraphics[scale=\Scale]{pics/ZBlock}$ and
$\includegraphics[scale=\Scale]{pics/SBlock}$ pieces, all types of pieces exhibit northward faces that are strictly higher than their southward ones.
Furthermore, \YCorr{any assembly of distinct pieces exposes northward faces that are at greater $y$-coordinates than their
dependency point}, inducing an increase of $y$-coordinate in the path.
Consequently, there only exists two configurations of dependent pieces $A\to B$, namely $\includegraphics[scale=\Scale]{pics/SSBlock}$ and $\includegraphics[scale=\Scale]{pics/ZZBlock}$, such that
$B$ exposes a southward face at the same height as their dependency point. The only way for a path in $D$ not to increase
in $y$-coordinate is then to feature a sequence of $\includegraphics[scale=\Scale]{pics/SBlock}$ (resp. $\includegraphics[scale=\Scale]{pics/ZBlock}$) pieces, inducing a monotonic behavior which
proves our claim, and the acyclic nature of $D$ follows.
\YCorr{Finally, the acyclicity of $D$ implies the existence of a sequence of pieces realizing $\mathcal{T}$, since it is always possible to
removing a piece.}
\end{proof}
\begin{figure}[t]\newcommand{2.5em}{2.5em}
\centering \includegraphics[width=2.5em]{pics/Tetris-1}
\includegraphics[width=2.5em]{pics/Tetris-2}
\includegraphics[width=2.5em]{pics/Tetris-3}
\includegraphics[width=2.5em]{pics/Tetris-4}
\includegraphics[width=2.5em]{pics/Tetris-5}
\includegraphics[width=2.5em]{pics/Tetris-6}
\includegraphics[width=2.5em]{pics/Tetris-7}
\includegraphics[width=2.5em]{pics/Tetris-8}
\includegraphics[width=2.5em]{pics/Tetris-9}
\includegraphics[width=2.5em]{pics/Tetris-10}
\includegraphics[width=2.5em]{pics/Tetris-11}
\includegraphics[width=2.5em]{pics/Tetris-12}
\includegraphics[width=2.5em]{pics/Tetris-13}
\includegraphics[width=2.5em]{pics/Tetris-14}
\includegraphics[width=2.5em]{pics/Tetris-15}
\caption{Fifteen Tetris tesselations of width 6 having uniform composition (+/- 1) in the different pieces.}\label{fig:example}
\end{figure}
\YCorr{Let us discuss the limitations induced by Tetris tesselation as a model for Tetris instances. First it can be remarked that Tetris tesselations
do not capture every possible Tetris game ending with an empty board, as one may temporarily leave \emph{holes} which amount
to disconnecting pieces in the tesselation representation.
Secondly there generally exists different free pieces to choose from while rebuilding a tesselation, and therefore different instances can lead to a given tesselation.
Furthermore the number of instances highly depends on the actual tesselation (from one to an exponential in $n$, as illustrated in Figure~\ref{fig:unique}),
Consequently, using the DAGs associated with Tetris histories to draw instances for the offline version of Tetris algorithmic problems, studied in~\cite{Breukelaar2004},
would favor exponentially certain instances over others, and the uniform random generation of instances ensuring feasibility of a perfect Tetris game
remains a challenging problem.}
|
\section{Introduction}
Centaurus~A (hereafter Cen~A) is the posterchild of an active galaxy. Also known as NGC~5128 it is the closest elliptical galaxy, the closest recent merger, and hosts the closest active galactic nucleus (AGN). The distance to Cen~A has long been under debate and seems to converge to the value 3.8~Mpc \citep{rejkuba04,karachentsev07,harris09}.
Cen~A is one of the most studied galaxies in the sky and its proximity makes it an ideal candidate to test AGN models, the connection of merging and star formation, and the influence of the jet.
Its proximity also makes it one of the prime targets to reliably measure the mass of the supermassive black hole, suspected to hide behind the obscuring dust lane. This prominent dust lane has hindered the study of CenA's black hole with optical HST spectroscopy as was common practice for other galaxies \citep[e.g. M87,][]{macchetto97}, and requires to move to the near-infrared wavelength for a detailed study of its properties.\\
Interestingly, the first measurements of Cen~A's black hole mass \citep{marconi01,silge05} brought it almost a factor of ten above the so called M-$\sigma$ relation \citep{ferrarese00,gebhardt00}, and made Cen~A one of the largest outlier to this relation. It was not clear, whether this offset to the M-$\sigma$ relation is intrinsic to Cen~A, maybe due to the presence of an AGN or its merger history, or whether it can be accounted for by the fact that seeing limited studies do not fully resolve the sphere of influence of the black hole, which would be at $\sim 0\farcs4$, taking the predicted mass from the M-$\sigma$ relation \citep{ferrarese00,gebhardt00,tremaine02} for granted.\\
There are several methods to detect black holes in galaxy centres and to measure their masses. For nearby galaxies the most direct methods take gas and stars as kinematics traces of the nuclear region. Dynamical models are then constructed to explain the kinematics. These models rely on a couple of assumptions (circular gravitational motion of the gas, superposition of orbits for the stars). To test the reliability of both approaches it is important to compare their model results.
In general, the gas kinematical method has the advantage of using a relatively simple modelling approach and dealing with relatively short exposure times. However, the gas kinematics might be influenced by non-gravitational motions that bias or even falsify the method. Moreover, not every galaxy nucleus has detectable gas emission lines. The stellar dynamical approach has the clear advantage that stars are always present in galaxy nuclei and that their motion is always gravitational. However, the derivation of stellar kinematics requires long exposure times and gets very complicated in the close vicinity of an AGN, as the stellar absorption lines are diluted by the AGN continuum radiation. In addition, stellar dynamical models are very complex leading to potential indeterminacy \citep[e.g.][]{cretton04,valluri04}.
Centaurus~A is a prime example for testing both modelling approaches on our doorstep. It allows both gas and stellar kinematical studies at a spatial resolution of $\sim 2$pc (corresponding to $0\farcs1$) using either
the HST or ground based adaptive optics instrumentation.\\
This paper reviews the status of the measurement of the black hole mass in Centaurus~A and shows how the advance in instrumentation is driving the field of black hole mass modelling. Section \ref{studies} presents the different studies dedicated to the black hole mass measurement in Cen~A, while Section \ref{lessons} states what has been learned from these studies. Finally, Section \ref{conc} summarises the observational facts for Cen~A's black hole. All studies on Cen~A's black hole mass assumed a distance to Cen~A of D=3.5~Mpc. The derived black hole mass directly relates to the adopted distance ($\mathrm{M_{BH}\ }$$\propto$ D) and so all results need to be increased by 8.5\% given the distance measurement of 3.8~Mpc recommended by \cite{harris09}.
\section{Dedicated Studies\label{studies}}
This section briefly lists and describes the studies that were performed to get the most accurate black hole mass determination to date of the supermassive black hole in Centaurus~A. Each of these papers sets the state-of-the-art observations and modelling techniques available at the time the study was pursued. Gas and stellar kinematical studies are divided into two groups for comparative purposes. For details concerning the individual studies the reader is referred to the original papers.
\subsection{Gas kinematical studies}
\subsubsection*{Marconi et al. 2001\\ Peering through the dust}
An extensive imaging study with WFPC2 and NICMOS on board the {\it Hubble Space Telescope (HST)} (PI: E. Schreier) has revealed the nucleus of Cen~A in near-IR (K and H) and, for the first time, in the optical (I and V) at unprecedented spatial resolution \citep{marconi00}. \cite{schreier98} detected an elongated structure in Pa$\alpha$ along the jet position angle and interpreted it as a thin gas disc. \cite{marconi01} used the near-infrared spectrograph and imager ISAAC \citep{moorwood99} on the ESO Very Large Telescope to retrieve spectra in J, H, and K$_{\mathrm s}$ Band under excellent seeing conditions of 0\farcs5 - 0\farcs6. They used the HST narrow band images as a reference to position the slit along three position angles and derived the kinematics for both ionised and molecular gas from their emission lines (Pa$\beta$, [FeII], Br$\gamma$, and H$_2$). The ionised gas kinematics are modelled by a thin disc model in the gravitational potential of a point mass, representing the black hole. Free parameters in the model are the inclination angle of the modelled disc, the black hole mass and the systemic velocity of the system. The inclination angle and the central mass are strongly coupled, since the amplitude of the rotation curve is proportional to $\sqrt{\mathrm{M_{BH}}} \times \sin(i)$. The ISAAC data do reject disc inclinations $>60^\circ$, but do not constrain the disc inclination further. This gets reflected in the large uncertainties of the black home mass determination. The best fit value is $\mathrm{M_{BH}\ }$$= 2^{+3.0}_{-1.4} \times 10^8$$\mathrm{M_{\odot}}$.
\subsubsection*{H\"aring-Neumayer et al. 2006\\ Adaptive Optics works!}
In the year 2004 the resolving power of ground based instrumentation has taken a big leap ahead when the first adaptive optics systems went online. Using NAOS-CONICA (NACO) \citep{rousset03,lenzen98} at the ESO VLT, \citet{hn06} (hereafter HN+06)
followed the approach of \cite{marconi01} and used long-slit spectra at the same position angles to probe the kinematics of the ionised gas. Due to the correction of the atmospheric distortion, the spatial resolution of the spectra increased by almost a factor of five compared to the study of \cite{marconi01} and provided images and spectra at (or almost at) the diffraction limit of the VLT ($0\farcs06$). The data points are shown in Figure~\ref{compare_isaac_naco} along with the [FeII] kinematics derived by \cite{marconi01} from ISAAC seeing limited data. HN+06 note that the [FeII] velocity dispersion is very high along all slit positions and cannot be fully accounted for by unresolved rotation. They have thus included a pressure term in their gas kinematical model. Moreover, they take into account the underlying potential of the stars (that was not accounted for by \citet{marconi01}) and model the [FeII] kinematics in the joint potential of the stars and an additional point mass, assumed to be the black hole. Their data along four long-slit position angles does also not constrain the inclination angle of the modelled gas disc well. Their best fitting model gives $i=45^\circ$ and $\mathrm{M_{BH}\ }$$= 6.1^{+0.6}_{-0.8} \times 10^7$$\mathrm{M_{\odot}}$. The decrease in black hole mass compared to \cite{marconi01} is due to (i) accounting for the stellar potential, (ii) the higher resolution of the data, and (iii) the higher inclination angle of the gas disc.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.4, angle=0]{f1.eps}
\caption{Velocity curves extracted from the [FeII] emission line at $\lambda1.644 \, \mu$m from NACO (diamonds) and ISAAC data (triangles), along three slit positions centred on the nucleus of Cen~A (P.A.= $-44.5^\circ$, $32.5^\circ$, and $82.5^\circ$ from left to right). The NACO data have almost factor of five better resolution than the ISAAC data and are overall consistent.}\label{compare_isaac_naco}
\end{center}
\end{figure}
\subsubsection*{Marconi et al. 2006\\ Pushing STIS to the reddest end}
Knowing exactly where to look, \cite{marconi06} have pointed the HST spectrograph STIS on the dust enshrouded nucleus, that is not visible in the optical wavelengths. Pushing STIS to the reddest limit, they used the [SIII] $\lambda9533 \, {\mathrm \AA}$ line to study the
kinematics of the ionized gas in the nuclear region with a $0\farcs1$ spatial resolution.
The STIS data were analysed in conjunction with the ground-based near-infrared VLT ISAAC spectra used by \cite{marconi01}, and have a gain in spatial resolution of almost a factor of five (see Figure~\ref{compare_isaac_stis}). The gas kinematical analysis provides a mass of $\mathrm{M_{BH}\ }$\, $ = (1.1\pm 0.1) \times 10^8$$\mathrm{M_{\odot}}$\, for an assumed disc
inclination of $ i=25^\circ$ and $\mathrm{M_{BH}\ }$$=(6.5 \pm 0.7) \times 10^7 $$\mathrm{M_{\odot}}$\, for $i=35^\circ$, the largest inclination value allowed by their data.
In this study, \cite{marconi06} did extensive tests on the influence of the choice of the parametrisation of the surface brightness profile of the assumed gas disc, that is used to luminosity-weight the modelled emission lines and to resemble the observations. They found that the associated systematic errors are no larger than 0.08 in log$\mathrm{M_{BH}\ }$ comparable
with statistical errors and indicating that the method is robust. However, the intrinsic surface brightness distribution has a large
impact on the value of the gas velocity dispersion. They conclude that a mismatch between the observed and modelled velocity dispersion is not
necessarily an indication of non-circular motions or kinematically hot gas, but is as easily due to an inaccurate computation
arising from too course a model grid, or the adoption of an intrinsic brightness distribution which is too smooth.
Therefore, they did not include pressure support in their model. The stellar potential is accounted for by the model.
The spatial resolution of the data is $\sim 0\farcs1$ the same as in HN+06 and both studies use ionised gas as the kinematic tracer ([SIII] and [FeII], respectively). The largest discrepancy between the two studies is the modelled disc inclination angle. \cite{marconi06} find that the inclination angle is $i\le35^\circ$, while HN+06 find their best model at $i=45^\circ$ and exclude smaller inclination angles on the basis of the jet inclination angle (assuming a orthogonal jet-disc geometry and a jet inclination angle of $\sim 50^\circ$ \citep{tingay98,hardcastle03}). The inclination angle remains the strongest source of uncertainty in the model. The results of \cite{marconi06} and HN+06 are in very good agreement, and also agree with \cite{marconi01} inside the error-bars.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.9, angle=0]{f2.eps}
\caption{Comparison of average velocities from STIS derived from [SIII] emission and
ISAAC derived from [FeII] and Pa$\alpha$, showing the increase in spatial resolution of almost a factor of five. This is a reproduction of Fig. 8 from \cite{marconi06}}\label{compare_isaac_stis}
\end{center}
\end{figure}
\subsubsection*{Krajnovi\'c, Sharp \& Thatte 2007\\ Integral Field Spectroscopy to the rescue!}
All previous approaches were restricted to a limited number of slit position angles (3 for \cite{marconi01}, 4 for HN+06, and 2 for \cite{marconi06}) and did not allow to trace the kinematics of the gas outside these stripes. This changed when \cite{krajnovic07} analysed the data from the Cambridge Infrared Panoramic
Survey Spectrograph \citep[CIRPASS,][]{parry04}, an integral field spectrograph mounted on the Gemini South telescope. They detect two dimensional distributions of the following emission lines: [PII], [FeII], and Pa$\beta$, and extracted spatially resolved 2D kinematics from Pa$\beta$ and [FeII]. All emission-line regions are part of the same kinematic structure which shows a twist in the zero-velocity curve beyond $\sim 1^{\prime\prime}$ (for both Pa$\beta$ and [Fe II]). The kinematics of the two emission lines are similar, but the Pa$\beta$ velocity
gradient is steeper in the centre while the velocity dispersion is low everywhere. The velocity
dispersion of the [FeII] emission is relatively high featuring a plateau, approximately oriented
in the same way as the central part of the warped disc. \cite{krajnovic07} use the 2D kinematic information to test
the hypothesis that the ionized gas is distributed in a circularly rotating disc. Assuming simple
disc geometry they estimate the mass of the central black hole using Pa$\beta$ kinematics, which
is consistent with being distributed in a circularly rotating disc. Their best fit model gives $\mathrm{M_{BH}\ }$ $=8.25^{+2.25}_{-4.25} \times 10^7$$\mathrm{M_{\odot}}$, for PA $=-3^\circ$ and $i=25^\circ$, and a stellar mass-to-light ratio of 0.0 $\mathrm{M/L_{\odot}}$, indicating that the Pa$\beta$ gas disc is fully under the gravitational influence of the black hole.
To test their model against previous gas dynamical models, \cite{krajnovic07} constructed models with different geometrical orientations taken from previous studies. They find that these give similar results, and are often statistically indistinguishable. However, the mass predicted by the M-$\sigma$ relation \citep{ferrarese00,gebhardt00} is excluded by all their models.
\subsubsection*{Neumayer et al. 2007\\ Integral Field Spectroscopy combined with Adaptive Optics}
Another advance in observing techinque allowed a big jump ahead in the studies of black hole mass modelling. Combining the resolving capabilities of adaptive optics with the power of integral field spectroscopy, makes SINFONI on the VLT \citep{eisenhauer03,bonnet04} the perfect instrument for this case. SINFONI works in the near-infrared and allows to peer through the dust lane hiding Cen~A's nucleus. The adaptive optics assisted data presented by \cite{neumayer07} have a spatial resolution of $0\farcs12$ and cover the central $3^{\prime\prime} \times 3^{\prime\prime}$. The data cubes in H- and K-band exhibit a wealth of ionised gas emission lines: [FeII], [SiVI], HeI, Br$\gamma$, [CaVIII], and several transitions of H$_2$. In addition the data enable the extraction of stellar kinematics \citep[which will be presented in detail in Section \ref{cap} ;][]{cappellari09}.
The high spatial resolution two-dimensional maps of the ionised gas show clearly two distinct kinematic features: rotation and transverse motion (see Figure~\ref{sinfoni_cirpass}). The transverse motion is oriented along the jet direction (PA=55$^\circ$) and is strongest for the high-ionisation lines ([SiVI] and [CaVIII]). However, also the tracers used in previous studies like [FeII] and Br$\gamma$ are influenced by this non-gravitational motion that has not been accounted for by the dynamical models. Long-slit spectroscopy was not able to spot this additional kinematic feature and therefore leads to an over-simplified picture, given the assumption that the gas moves in a circularly rotating disc, which underlies the gas dynamical models.\\
The picture is surprisingly different when looking at the kinematics of molecular hydrogen rather than the ionised gas. This kinematic tracer shows no kinematic component along the jet direction and seems to be dominated by gravitational motion only. Therefore, \cite{neumayer07} have based their dynamical model on these kinematics.
The two dimensional kinematic maps constrain the geometry of the disc much better than the few long-slit position angles have done before. The disc is not flat, but appears warped. The disc parameters, i.e. the inclination angle and the major axis position angle, were fitted using the kinemetry software by \cite{krajnovic06}. The fitted mean inclination angle of the H$_2$ gas disc is $45^\circ\pm7^\circ$ and the mean position angle of the major axis is 155$^\circ$. The inclination angle is in very good agreement with the value derived by \cite{hardcastle03} from VLA data ($20^\circ < i < 50^\circ$), it is, however, somewhat smaller than the value from VLBI data derived by \cite{tingay98} ($50^\circ < i < 80^\circ$). This well constrained disc parameters enable a significant reduction of the uncertainty of the modelled black hole mass. The assumed mass-to-light ratio of the underlying stellar body, M/L$_K = (0.72 \pm 0.04) M_{\odot}/L_{\odot}$, is taken from the stellar kinematical model by \cite{silge05} and confirmed by \cite{cappellari09}. \cite{neumayer07} modelled the H$_2$ kinematics on tilted rings following the kinemetry model, keeping the relative inclination angle between the rings fixed but varying the overall inclination. The overall best-fit black hole mass is $\mathrm{M_{BH}\ }$ $=4.5^{+1.7}_{-1.0} \times 10^7$$\mathrm{M_{\odot}}$ and the best fit mean inclination angle is $34^\circ\pm4^\circ$ (3 $\sigma$ error bars).
\begin{figure}[h]
\begin{center}
\includegraphics[scale=1., angle=0]{f3.eps}
\caption{The [FeII] velocity maps compared from SINFONI adaptive optics and CIRPASS seeing limited data. The Figures are taken from \cite{neumayer07} and \cite{krajnovic07} respectively. The overall kinematic structure of the two maps is in good agreement. However, the adaptive optics assisted SINFONI map shows detailed kinematic substructures and reveals the influence of the jet on the kinematics.}\label{sinfoni_cirpass}
\end{center}
\end{figure}
\subsection{Stellar kinematical studies}
\subsubsection*{Silge et al. 2005\\ Near-Infrared long-slit spectra}
Using the Gemini Near-Infrared Spectrograph (GNIRS) on Gemini South, \cite{silge05} measured the stellar kinematics under very good natural seeing conditions of $0\farcs45$ and $0\farcs6$. They put the slit along two position angles across the nucleus of Cen~A, centred on the AGN. The first observation was made with the slit oriented perpendicular to the inner dust ring (along the major axis at large radii), while the second observation was made with the slit oriented parallel to the dust lane but offset from the centre by $0\farcs85$.
The stellar kinematics are extracted using the region around the CO band heads at $2.3\, \mu$m.
The central spatial bins of the first observations are dominated by emission from the AGN and the CO band heads are diluted by this emission. To recover the stellar signatures inside the central $0\farcs3$, \cite{silge05} have removed the AGN emission by measuring the AGN contribution by its dilution of the equivalent width of the $(2-1)^{12}$CO band head, assuming that the equivalent width is largely constant with radius in this galaxy for regions outside the AGN.
After subtracting the AGN contribution, the stellar kinematics were extracted from the spectra by fitting a stellar template spectrum (convolved with a velocity profile) to the observations. The choice of template star is important in order to avoid mismatch \citep{silge03} and so a variety of template stellar spectra is simultaneously fitted for the velocity profile and the stellar template weights. As a result the fitting procedure provides the line-of-sight velocity distribution (in a non-parametric form) as well as the stellar population information.
The luminosity weighted stellar velocity dispersion along the slit parallel to the dust disc is $138\pm10$ $\mbox{km s}^{-1}$.\\
The black hole mass is determined using axisymmetric orbit-based models. The full line-of-sight velocity distribution is fitted to the data and the best fit parameters, stellar M/L and $\mathrm{M_{BH}\ }$, are derived via a $\chi^2$ minimisation (the interested reader is referred to \cite{silge05} for details). From the long slit data along two slit positions alone it is not possible to identify the rotation axis and so the modelling procedure was performed twice: first matching the rotation axis to the dusk disc and second matching the rotation axis to the photometric major axis and non-rotating kinematics along the dust disc. It turns out that the two assumptions give very similar black hole masses but the second approach results in significantly lower $\chi^2$ values.\\
Although NGC~5128 appears very round, its actual shape is unknown and it could be intrinsically quite flattened. Therefore, \cite{silge05} repeated their models for three different intrinsic shapes (i.e intrinsic axis ratios) of Cen~A:
intrinsically round (1:1:1), at an inclination of $i=20^\circ$ (1:1:0.5), and $i=45^\circ$ (1:1:0.9). They find that this drastic change in flattening does not change the black hole mass much. The $\chi^2$ of the intrinsically round model is significantly less than either flattened model, and so this is adopted as the best-fit model with $\mathrm{M_{BH}\ }$$=2.4^{+0.3}_{-0.2}\times 10^8$$\mathrm{M_{\odot}}$ and stellar M/L$_{K}=0.68^{+0.01}_{-0.02}$ in solar units.
\subsubsection*{Cappellari et al. 2009\\ Integral field spectroscopy combined with adaptive optics\label{cap}}
Using the adaptive optics assisted integral field spectrograph SINFONI at the ESO Very Large Telescope, \cite{cappellari09} extracted the stellar kinematics from the nuclear region of NGC~5128 with a spatial resolution of $0\farcs17$. This means an increase in spatial resolution by a factor of three to four compared to the seeing limited data of \cite{silge05}. The high S/N data cover the central $8^{\prime\prime} \times 8^{\prime\prime}$ and allow the extraction of the stellar line-of-sight velocity distribution, which is given in parametric form of a Gauss-Hermite expansion. At radii where the nucleus dominates, the stellar absorptions are diluted, resulting in a strong decrease in the observed line-strength $\gamma$. An inaccurate modelling of the non-thermal dilution can cause large errors in the measured velocity dispersion $\sigma$. Similarly, the rise in the non-thermal continuum could be incorrectly interpreted as a variation of the stellar population, requiring a change in the stellar template mix and also producing an error in $\sigma$. As there is no evidence for a sudden change in the population in the nucleus of Cen~A, the authors argue the safest choice is to assume that the stellar template is fixed and to model the non-thermal continuum via additive polynomials. This approach allows the $\sigma$ to be reliably extracted in the high-resolution observations down to $\ge 0\farcs2$.
For the first time, a low-level clear sense of stellar rotation is detected for Cen~A which is counter-rotating with respect to the gas. The global kinematical major axis is at PA$_{\mathrm kin} = 167\pm8^\circ$ which is in between the dust disc and the galaxy's photometric major axis, and therefore between the two position angles that \cite{silge05} assumed for their kinematic models. The derived systemic velocity is v$_{\mathrm syst}= 531 \pm 5$ $\mbox{km s}^{-1}$.\\
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.8, angle=0]{f4.eps}
\caption{Spherical anisotropic Jeans model to check the $\mathrm{M_{BH}\ }$ determinations. The root mean square stellar velocity, $\mathrm{v_{rms} = \sqrt{v^2+\sigma^2}}$, is plotted as a function of radius from the centre of Cen~A, derived from SINFONI observations (100~mas, red open circles; 250~mas scale, blue open squares) and from GNIRS observations of \cite{silge05} along the major axis (filled downward
triangle) and the minor axis (filled upward triangle). The discrepancy between the SINFONI and GNIRS data inside $1^{\prime\prime}$ is due to a different subtraction of the AGN continuum. The solid
line shows the prediction of an anisotropic Jeans model, having the best-fitting $\mathrm{M_{BH}\ }$ and
M/L derived from the Schwarzschild model, the dashed lines represent the upper and lower 3$\sigma$ errors in the Schwarzschild model. The dotted lines has $\mathrm{M_{BH}\ }$$=2\times 10^8$$\mathrm{M_{\odot}}$, corresponding to the best-fitting value of \cite{silge05}. This figure is reproduced from \cite{cappellari09}. }\label{silge_cappellari}
\end{center}
\end{figure}
Having in mind the complex dynamical and morphological structure that Cen~A possesses,
\cite{cappellari09} constructed two types of models to test the
sensitivity of the $\mathrm{M_{BH}\ }$ estimate to the assumed geometry and dynamics; (i) an axisymmetric orbit-based model, to reproduce
in detail the kinematic observations and (ii) a simple anisotropic spherical Jeans model to qualitatively check the $\mathrm{M_{BH}\ }$ determination. This test gave very consistent results and is shown in Figure~\ref{silge_cappellari} along with a comparison to the data presented by \cite{silge05}. The best-fitting
value for the black hole mass from the Schwarzschild model is $\mathrm{M_{BH}\ }$$= (5.5\pm 3.0)\times 10^7$$\mathrm{M_{\odot}}$ (3$\sigma$ errors) and the corresponding best fitting stellar M/L$_{K}$ is $(0.65\pm0.15) \mathrm{M/L_{\odot}}$.
\section{Lessons learned \label{lessons}}
The ever increasing accuracy of the black hole mass measurement in Centaurus~A over the past eight years reflects the advancement in observing techniques, especially in the near infrared. High spatial resolution observations are crucial to determine the influence of the central black hole on the gas and stellar kinematics. Even in an object that is as nearby as Cen~A, where the sphere of influence of the black hole is comfortably resolved ($\mathrm{R_{BH}}\sim 0\farcs7$, using the mass measurement of \citet{cappellari09}) using HST or adaptive optics observations.\\
Adaptive optics actually works! With the advent of near-infrared adaptive optics assisted spectrographs, such as NACO at the VLT, even dust enshrouded galaxy nuclei became accessible at a spatial resolution of $0\farcs1$. This gets even more powerful when combined with integral-field spectroscopy. Mapping the gas and the stars in 3D allows to directly and simultaneously compare the morphology and kinematics of different gas species plus the stars. Having this powerful tool in hand, the influence of the inner jet on the kinematics of the ionised gas in Cen~A could be revealed, and moreover, molecular hydrogen could be identified as the ideal gas tracer for the central gravitational potential. The physical state of the gas is therefore very important when using it as a tracer for the dynamical models. The decrease of the best-fit value from \cite{marconi01} to \cite{marconi06} and HN+06 was due to the increase in spatial resolution plus taking into account the stellar contribution to the potential, while the decrease in $\mathrm{M_{BH}\ }$ from \cite{marconi06} and HN+06 to \cite{neumayer07} was due to the fact that the kinematic tracer changed from ionised gas to molecular hydrogen.\\
The presence of an AGN can definitely influence the kinematics of the gas, while the stellar kinematics should be unchanged by this. However, the extraction of the stellar kinematics from the spectral absorption features gets increasingly difficult in the close vicinity of the AGN, as the non-thermal continuum dilutes the stellar absorption lines and needs to be accounted for in the extraction process. This is the main difference in the analysis of \cite{silge05} and \cite{cappellari09}. While \cite{silge05} first subtract the AGN contribution and then fit the stellar line-of-sight-velocity distribution, \cite{cappellari09} include the fit of the AGN continuum in the the extraction of the stellar kinematics. To get reliable kinematics high signal-to-noise data are required. This is a very interesting lesson that we learned from Cen~A, and we should be cautious when extracting kinematics from other, more distant objects. Cen~A is indeed the closest AGN and at the same time it is very complex. Every leap in instrumentation development is likely to reveal more complex substructures. This warrants our continuous attention, in order to reveal intrinsic properties in the data and understand shortcomings in the models that aim to predict the observations.\\
\begin{figure*}[h]
\begin{center}
\includegraphics[scale=0.8, angle=0]{f5.eps}
\caption{Cen~A's black hole mass measurements are plotted on top of the black hole mass galaxy scaling relations. The left panel shows the $\mathrm{M_{BH}\ }$-$\sigma$ relation reproduced after \cite{gultekin09}, with the overplotted line being the best fit relation to the elliptical galaxy subsample of all plotted galaxies. The right panel shows the $\mathrm{M_{BH}\ }$-$\mathrm{M_{Bulge}\ }$ relation as presented in \cite{haering04} with Cen~A values over-plotted. The Cen~A plot symbols as well as their time sequence are indicated in the upper left corner of the left and right panels, respectively. Values published in the same year are separated horizontally for clarity. In the left panel (and for Cen~A measurements) triangles refer to gas kinematical measurements, while circles refer to dynamical models using stellar kinematics.}\label{relations}
\end{center}
\end{figure*}
\section{Summary \label{conc}}
In principle all presented $\mathrm{M_{BH}\ }$ determinations for Cen~A using gas kinematical models agree within the error bars. The only significant disagreement in $\mathrm{M_{BH}\ }$ is with the previous stellar dynamical $\mathrm{M_{BH}\ }$ determination by \cite{silge05} and subsequent measurements \citep[HN+06, ][]{marconi06,krajnovic07,neumayer07,cappellari09}. The disagreement is likely caused by a difference in the data quality and in the treatment of the contribution from the central non-thermal continuum in the kinematic extraction \citep[compared to][]{cappellari09}. It is not due to differences in the details of the modelling methods. \\
The latest mass determination for the black hole at the nucleus of Centaurus~A, using high resolution integral field stellar kinematics, gives $\mathrm{M_{BH}\ }$$=(5.5\pm3.0)\times 10^7$$\mathrm{M_{\odot}}$ \citep{cappellari09} and agrees very well with the value $\mathrm{M_{BH}\ }$$=4.5^{+1.7}_{-1.0} \times 10^7$$\mathrm{M_{\odot}}$ determined using $H_2$ gas kinematics from the same SINFONI data \citep{neumayer07} .
This gas versus stars $\mathrm{M_{BH}\ }$ com-parison constitutes one of the most robust and accurate ones, due to a very well resolved black hole sphere of influence ($R_{\mathrm BH}\sim 0\farcs7$ compared to a PSF FWHM$\sim 0\farcs12$) and thanks to the use of high-resolution integral-field data for both kinematical tracers. \\
Comparable and equally successful comparisons were done by \cite{shapiro06} and \cite{siopis09}, while a less good agreement was found in \cite{cappellari1459}, likely due to the disturbed gas kinematics.
There are certainly weaknesses in both the stellar and gas dynamical models, as e.g. the assumption of circular gravitational motions for gas kinematics and the complexity for the modelling of stellar kinematics. However, the agreement between the two modelling approaches for a growing number of cases strengthens the reliability of both methods.\\
With the latest gas and stars $\mathrm{M_{BH}\ }$ determination Cen~A lies within the errors on the $\mathrm{M_{BH}\ }$$-\sigma$ relation as given by either \cite{tremaine02}, \cite{ferrarese05}, or \cite{gultekin09}, and is in agreement with the relation of black hole mass to bulge mass as given by \cite{haering04} (see Figure~\ref{relations}). As the value of Cen~A's black hole mass has changed since 2001, so has the $\mathrm{M_{BH}\ }$$-\sigma$ relation. The scatter in the relation seems to depend on the Hubble type of the chosen galaxies and is smallest for elliptical galaxies \citep{hu08,graham08,gultekin09}. The latest SINFONI $\mathrm{M_{BH}\ }$ measurement is in agreement with the relation for ellipticals (Figure~5).
\section*{Acknowledgments}
I thank the DFG cluster of excellence 'Origin and Structure of the Universe' for financial support, and Davor Krajnovi\'c for a critical reading of the manuscript. Moreover, I would like to thank the science organizing committee for the invitation to the Many Faces of Centaurus~A Conference in Sydney as well as for financial support. Finally, I thank the anonymous referee, who helped to improve the manuscript.
|
\section{Introduction.}
\subsection{Background.}Let $p$ be a prime number and $K$ be a $p$-adic field, i.e. finite
extension of $\mathbb{Q}_p$. Trianguline representations, which form a class of $p$-adic representations
of $G_K:=\mathrm{Gal}(\bar{K}/K)$, play some crucial roles in the study of $p$-adic family of global or local
Galois representations. The aim of this article is, for any $p$-adic field $K$, to develop the deformation
theory of trianguline representations by using $B$-pairs.
In the study of $p$-adic local Langlands correspondence for $\mathrm{GL}_2(\mathbb{Q}_p)$, in [Co08], Colmez
defined the notion of trianguline representations, which is a class of $p$-adic representations of $G_K:=\mathrm{Gal}(\bar{K}/K)$
and is defined by using Fontaine, Kedlaya's theory of ($\varphi,\Gamma$)-modules over the Robba ring.
Since his work, when $K=\mathbb{Q}_p$, trianguline representations and their deformation theoretic properties have turned out to be
very important in number theory, especially in the study of $p$-adic local Langlands correspondence by Colmez ([Co09]), or in the study of $p$-adic
families of $p$-adic modular forms (more precisely, in the study of eigenvarieties), for example, by Kisin ([Ki03]), by Bella\"iche-Chenevier ([Bel-Ch]) and by Chenevier ([Ch08],[Ch09]).
For any $p$-adic field $K\not= \mathbb{Q}_p$ case, in [Na09], the author of this article generalized
many results in [Co08] for any $p$-adic fields. Namely, he studied some fundamental properties of trianguline representations and then classified two dimensional trianguline representations. In [Na09], we studied trianguline representations by using
$B$-pairs, which was defined by Berger in [Be08], instead of using ($\varphi,\Gamma$)-modules over the Robba ring.
The aim of this article is, by using the results in [Na09] and by using the theory of $B$-pairs, to generalize many results in [Bel-Ch], [Ch08], [Ch09] concerning to
deformations of trianguline representations to any $p$-adic fields. We develop the deformation theory of
trianguline representations (more generally, trianguline $B$-pairs) and generalize Chenevier's theorem in [Ch09]
concerning tangent spaces of these deformation rings to any $p$-adic fields.
When $K=\mathbb{Q}_p$, these results are crucial
for $p$-adic Hodge theoretic study of eigenvarieties, especially for problems about Zariski density of modular Galois representations
(or crystalline representations) in deformation spaces of global (or local) $p$-adic representations ([Ch08].[Ch09],[Co08],[Ki08b]).
In the next article, by using the results of this article (and by generalizing some results in [Ki03] to any $p$-adic fields),
the author will prove Zariski density of crystalline representations in deformation spaces of two dimensional
$p$-adic representations for any $p$-adic fields.
\subsection{Overview.}
We explain details of this article.
In $\S$ 2, we recall the definition of $B$-pairs
and study some fundamental properties of trianguline
$B$-pairs.
Let $E$ be a suitable finite extension of $\mathbb{Q}_p$ as in Notation below. In $\S$ 2.1, we recall the definition of $E$-$B$-pairs of $G_K$, which is the $E$-coefficient version of $B$-pairs. We write $B_e:=B_{\mathrm{cris}}^{\varphi=1}$.
An $E$-$B$-pair is a pair $W=(W_e,W^+_{\mathrm{dR}})$ where $W_e$ is a finite free $B_e\otimes_{\mathbb{Q}_p}E$-module
with a continuous semi-linear $G_K$-action such that $W^+_{\mathrm{dR}}\subseteq W_{\mathrm{dR}}:=B_{\mathrm{dR}}\otimes_{B_e}W_e$
is a $G_K$-stable $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E$-lattice of $W_{\mathrm{dR}}$. The category of $E$-representations
of $G_K$ is embedded in the category of $E$-$B$-pairs by $V\mapsto W(V):=(B_e\otimes_{\mathbb{Q}_p}V, B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}V)$.
We say that an $E$-$B$-pair is split trianguline if $W$ is a successive extension of rank one $E$-$B$-pairs, i.e. $W$ has a filtration
$0\subseteq W_1\subseteq W_2\subseteq \cdots \subseteq W_{n-1}\subseteq W_n=W$ such that $W_i$ is a saturated sub $E$-$B$-pair
of $W$ and $W_i/W_{i-1}$ is a rank one $E$-$B$-pair for any $1\leqq i\leqq n$. We say that $W$ is trianguline
if $W\otimes_{E}E'$ is a split trianguline $E'$-$B$-pair for a finite extension $E'$ of $E$. We say that an $E$-representation $V$ is split trianguline
(resp. trianguline) if $W(V)$ is split trianguline (resp. trianguline). By these definitions, to study trianguline $E$-$B$-pairs, we first need to classify
rank one $E$-$B$-pairs and then we need to calculate extension class group of them, which were studied in [Co08] for $K=\mathbb{Q}_p$ and in
[Na09] for general $K$. In $\S$ 2.1, we recall these results which we need to study deformations of trianguline $E$-$B$-pairs.
In $\S$ 2.2, we define the Artin ring coefficient version of $B$-pairs. Let $\mathcal{C}_E$ be the category of Artin local $E$-algebras with residue field $E$. For any $A\in \mathcal{C}_E$, we say that $W:=(W_e,W^+_{\mathrm{dR}})$ is an $A$-$B$-pair if $W_e$ is a finite free $B_e\otimes_{\mathbb{Q}_p}A$-module with a continuous semi-linear $G_K$-action and
$W^+_{\mathrm{dR}}\subseteq W_{\mathrm{dR}}:=B_{\mathrm{dR}}\otimes_{B_e}W_e$ is a $G_K$-stable $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-lattice. We generalize some results of [Na09] to $A$-$B$-pairs.
In $\S$ 3, we study two types of deformations for split trianguline $E$-$B$-pairs. In $\S$ 3.1, first we study usual
deformations for any $E$-$B$-pairs which are generalizations of homomorphismistic zero deformations
of $p$-adic Galois representations.
Let $W$ be an $E$-$B$-pair and $A\in \mathcal{C}_E$.
Then we say that $(W_A,\iota)$ is a deformation of $W$ over $A$ if $W_A$ is an $A$-$B$-pair and $\iota:W_A\otimes_A E\isom W$ is an
isomorphism. Then we define the deformation functor of $W$, $D_W:\mathcal{C}_E\rightarrow (\mathrm{Sets})$ by
$D_W(A):=\{$equivalent classes of deformations $(W_A,\iota)$ of $W$ over $A$ $\}$. We prove pro-representability and formally smoothness and dimension formula
of $D_W$ (Corollary $\ref{13}$). In $\S$ 3.2, we study the other type of deformations, i.e. trianguline deformations. Let $W$ be a split trianguline $E$-$B$-pair of rank $n$ and
$\mathcal{T}:0\subseteq W_1\subseteq W_2\subseteq \cdots \subseteq W_{n-1}\subseteq W_n=W$ be a fixed triangulation of $W$. For any $A\in \mathcal{C}_E$,
we say that $(W_A, \iota, \mathcal{T}_A)$ is a trianguline deformation of $(W,\mathcal{T})$ over $A$ if $(W_A,\iota)$ is a deformation of $W$
over $A$ and $\mathcal{T}_A:0\subseteq W_{1,A}\subseteq \cdots \subseteq W_{n-1,A}\subseteq W_{n,A}=W_A$ is a triangulation of $W_A$
( i.e. $W_{i,A}$ is a saturated sub $A$-$B$-pair of $W_A$ such that $W_{i,A}/W_{i-1,A}$ is a rank one $A$-$B$-pair for any $i$) such that
$\iota(W_{i,A}\otimes_A E)=W_i$ for any $i$. Then we define the trianguline deformation functor of $(W,\mathcal{T})$, $D_{W,\mathcal{T}}:
\mathcal{C}_E\rightarrow (\mathrm{Sets})$ by $D_{W,\mathcal{T}}(A):=\{ $equivalent classes of trianguline deformations $(W_A,\iota,\mathcal{T}_A)$
of $(W,\mathcal{T})$ over $A$$\}$. We prove pro-representability and formally smoothness and dimension formula of this functor (Proposition $\ref{19}$).
In $\S$ 4, we define the notion of benign $E$-$B$-pairs, which forms a special good class of split trianguline $E$-$B$-pairs, and prove
the main theorem of this article concerning to tangent spaces of the deformation rings of this class.
In $\S$ 4.1, we define the notion of benign $E$-$B$-pairs. Let $W$ be a potentially crystalline $E$-$B$-pair of rank $n$ such that
$W|_{G_L}$ is crystalline for a finite totally ramified abel extension $L$ of $K$ (we call such a representation a crystabelline representation).
We assume that $D_{\mathrm{cris}}^{L}(W):=(B_{\mathrm{cris}}\otimes_{B_e}W_e)^{G_L}=K_0\otimes_{\mathbb{Q_p}}Ee_1\oplus
\cdots \oplus K_0\otimes_{\mathbb{Q}_p}Ee_n$ such that $K_0\otimes_{\mathbb{Q}_p}Ee_i$ are preserved by $(\varphi,\mathrm{Gal}(L/K))$-action
and that $\varphi^f(e_i)=\alpha_ie_i$ for some $\alpha_i\in E^{\times}$, here $f:=[K_0:\mathbb{Q}_p]$ and $K_0$ is the maximal unramified extension of $\mathbb{Q}_p$ in $K$.
We write $\{k_{1,\sigma}.k_{2,\sigma},\cdots,k_{n,\sigma}\}_{\sigma:K\hookrightarrow \bar{K}}$ the Hodge-Tate weight of $W$ such that
$k_{1,\sigma}\geqq k_{2,\sigma}\geqq \cdots \geqq k_{n,\sigma}$ for any $\sigma:K\hookrightarrow \bar{K}$.
Let $\mathfrak{S}_n$ be the $n$-th permutation group. Then, for any $\tau\in\mathfrak{S}_n$, we can define a filtration on $D^L_{\mathrm{cris}}(W)$
by $0\subseteq K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(1)}\subseteq K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(1)}\oplus K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(2)}\subseteq
\cdots \subseteq K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(1)}\oplus \cdots \oplus K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(n-1)}\subseteq D^L_{\mathrm{cris}}(W)$ by
sub $E$-filtered $(\varphi, G_K)$-modules, where the filtration on $K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(1)}\oplus \cdots \oplus K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(i)}$ is
the one induced from $L\otimes_{L_0}D^L_{\mathrm{cris}}(W)$. Then, by the $B$-pair version of ``weakly admissible = admissible" theorem, for any $\tau\in \mathfrak{S}_n$
we have a triangulation $\mathcal{T}_{\tau}:0\subseteq W_{\tau,1}\subseteq W_{\tau,2}\subseteq \cdots \subseteq W_{\tau,n}=W$ such that
$W_{\tau,i}$ is potentially crystalline and $D^L_{\mathrm{cris}}(W_{\tau,i})\isom K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(1)}\oplus \cdots \oplus K_0\otimes_{\mathbb{Q}_p}E
e_{\tau(i)}$
Under this condition, we define the notion of twisted benign $E$-$B$-pairs as follows.
\begin{defn}
Let $W$ be an $E$-$B$-pair of rank $n$ as above. Then we say that $W$ is twisted benign if $W$ satisfies the following;
\begin{itemize}
\item[(1)] For any $i\not= j$, we have $\alpha_i/\alpha_j\not= 1, p^f, p^{-f}$,
\item[(2)] For any $\sigma:K\hookrightarrow \bar{K}$, we have $k_{1,\sigma}>k_{2,\sigma}>\cdots >k_{n-1,\sigma}> k_{n,\sigma}$,
\item[(3)] For any $\tau\in \mathfrak{S}_n$ and $i$, Hodge-Tate weight of $W_{\tau,i}$ is $\{k_{1,\sigma},k_{2,\sigma},\cdots,k_{i,\sigma}\}_{\sigma :K\hookrightarrow \bar{K}}$.
\end{itemize}
\end{defn}
In $\S$ 4.2, we prove the main theorem of this article. Let $W$ be a twisted benign $E$-$B$-pair of rank $n$ as above. Then, for any $\tau\in \mathfrak{S}_n$, we can define
the trianguline deformation functor $D_{W,\mathcal{T}_{\tau}}$. Let $R_W$ be the universal deformation ring of $D_W$ and
$R_{W,\mathcal{T}_{\tau}}$ be the universal deformation ring of $D_{W,\mathcal{T}_{\tau}}$ for any $\tau\in \mathcal{S}_{\tau}$.
Then $R_{W,\mathcal{T}_{\tau}}$ is a quotient ring of $R_W$. Let $t(R_{W})$ and $t(R_{W,\mathcal{T}_\tau})$ be the tangent spaces
of $R_W$ and $R_{W,\mathcal{T}_{\tau}}$. Then, for any $\tau\in \mathfrak{S}_n$, $t(R_{W,\mathcal{T}_{\tau}})$ is a sub $E$-vector space of $t(R_W)$.
The main theorem of this article is the following (Theorem $\ref{33}$), which is a generalization of Chenevier's theorem in [Ch09] in the
case of $K=\mathbb{Q}_p$.
\begin{thm}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$. Then we
have $$\sum_{\tau\in \mathfrak{S}_n}t(R_{W,\mathcal{T}_{\tau}})=t(R_W).$$
\end{thm}
This theorem is a crucial local results for applications to some Zariski density theorems of
local or global $p$-adic representations. In fact, by using
this theorem in the case of $K=\mathbb{Q}_p$, Chenevier, in [Ch09], proved Zariski density of unitary automorphic Galois representations
in deformation spaces of three dimensional self dual $p$-adic representations of $G_{F}$ for any CM field $F$
in which $p$ splits completely. In the next article, by using this main theorem, the author will prove Zariski density of crystalline
representations in deformation spaces of two dimensional $p$-adic representations for any $p$-adic fields, which is a generalization
of Colmez's and Kisin's theorem ([Co08], [Ki08b]) in the case of $K=\mathbb{Q}_p$.
The author of this article hopes that, by using this theorem, we can prove many kinds of Zariski density theorems for
general $p$-adic field cases or general global (CM or totally real) field cases.
\subsection*{Notation.}
Let $p$ be a prime number. $K$ be a finite extension of $\mathbb{Q}_p$, $\bar{K}$ a
fixed algebraic closure of $K$, $K_0$ the maximal unramified extension of $\mathbb{Q}_p$ in $K$,
$K^{\mathrm{nor}}$ the Galois closure of $K$ in $\bar{K}$. Let $G_K:=\mathrm{Gal}(\bar{K}/K)$ be the
absolute Galois group of $K$ equipped with profinite topology. $\mathcal{O}_K$ is the integer ring of $K$, $\pi_K\in\mathcal{O}_K$ is
a uniformizer of $K$, $k:=\mathcal{O}_K/\pi_K\mathcal{O}_K$ is the residue field
of $K$, $q=p^f:=\sharp k$ is the order of $k$, $\chi:G_K\rightarrow \mathbb{Z}_p^{\times}$ is the p-adic cyclotomic homomorphism
(i.e. $g(\zeta_{p^n})=\zeta_{p^n}^{\chi(g)}$ for any $p^n$-th roots of unity and for any $g\in G_K$). Let $\mathbb{C}_p:=\hat{\bar{K}}$
be the $p$-adic completion of $\bar{K}$, which is an algebraically closed $p$-adically complete field, and $\mathcal{O}_{\mathbb{C}_p}$ is its integer ring. We denote $v_p$ the normalized valuation on $\mathbb{C}_p^{\times}$ such that
$v_p(p)=1$. Let $E$ be a finite extension of $\mathbb{Q}_p$ such that
$K^{\mathrm{nor}}\subseteq E$. In this paper, we write $E$ as a coefficient of representations.
We put $\mathcal{P}:=\{\sigma:K\hookrightarrow \bar{K}\}=\{\sigma:K\hookrightarrow E\}$
Let $\chi_{\mathrm{LT}}:G_K\rightarrow \mathcal{O}_K^{\times}\hookrightarrow \mathcal{O}_E^{\times}$ be
the Lubin-Tate homomorphism associated with the fixed uniformizer $\pi_K$. Let $\mathrm{rec}_K:K^{\times}
\rightarrow G_K^{\mathrm{ab}}$ be the reciprocity map of local class field theory such that
$\mathrm{rec}_K(\pi_K)$ is a lifting of the inverse of $q$-th power Frobenius on $k$, then
$\chi_{\mathrm{LT}}\circ \mathrm{rec}_K:K^{\times}\rightarrow \mathcal{O}_K^{\times}$ satisfies
$\chi_{\mathrm{LT}}\circ\mathrm{rec}_K(\pi_K)=1$ and $\chi_{\mathrm{LT}}\circ\mathrm{rec}_K|_{\mathcal{O}_K^{\times}}
=\mathrm{id}_{\mathcal{O}_K^{\times}}$.
\subsection*{Acknowledgement.}
The author would like to thank Kenichi Bannai for reading the first version of this article and for helpful comments.
The author is supported by JSPS Core-to-Core program 18005 whose representative is Makoto Matsumoto.
This work is also supported in part by KAKENHI (21674001).
\section{$B$-pairs.}
\subsection{$B$-pairs.}
In this subsection, first we recall the definition of $E$-$B$-pairs ([Be08], [Na09]) and then recall fundamental
properties of them established in [Na09].
First, we recall some $p$-adic period rings ([Fo94]) which we need for defining $B$-pairs.
Let $\widetilde{\mathbb{E}}^+:=\varprojlim_n\mathcal{O}_{\mathbb{C}_p}\isom
\varprojlim_n\mathcal{O}_{\mathbb{C}_p}/p\mathcal{O}_{\mathbb{C}_p}$, where the limits are taken with respect to $p$-th power maps. It is known that $\widetilde{\mathbb{E}}^+$ is a complete valuation ring of homomorphismistic $p$ whose valuation
is defined by $\mathrm{val}(x):=v_{p}(x^{(0)})$ (here $x=(x^{(n)})\in \varprojlim_n\mathcal{O}_{\mathbb{C}_p}$).
We fix a system of $p^n$-th roots of unity $\{\varepsilon^{(n)}\}_{n{\geqq} 0}$
such that $\varepsilon^{(0)}=1$, $(\varepsilon^{(n+1) })^p=\varepsilon^{(n)}$,
$\varepsilon^{(1)}\not= 1$. Then $\varepsilon:=(\varepsilon^{(n)})$ is an element of $\widetilde{\mathbb{E}}^+$ such that $\mathrm{val}(\varepsilon -1)=p/(p-1)$. $G_K$ acts on this ring in natural way. We put $\widetilde{\mathbb{A}}^+:=W(\widetilde{\mathbb{E}}^+)$, where, for a perfect ring $R$,
$W(R)$ is the Witt ring of $R$. We put $\widetilde{\mathbb{B}}^+:=\widetilde{\mathbb{A}}^+[\frac{1}{p}]$. This ring also has natural $G_K$-action and Frobenius
action $\varphi$. Then we have a $G_K$-equivariant surjection $\theta:\widetilde{\mathbb{A}}^+
\rightarrow \mathcal{O}_{\mathbb{C}_p}:\sum_{k=0}^{\infty}p^k[x_k]\mapsto \sum_{k=0}^{\infty} p^k x_k^{(0)}$, where $[\,\, ]:\widetilde{\mathbb{E}}^+\rightarrow \widetilde{\mathbb{A}}^+$ is the
Teichm$\ddot{\mathrm{u}}$ller homomorphism.
By inverting $p$, we get a surjection $\widetilde{\mathbb{B}}^+\rightarrow \mathbb{C}_p$.
We put $B_{\mathrm{dR}}^+:=\varprojlim_n\widetilde{\mathbb{B}}^+/(\mathrm{Ker}(\theta))^n$, which is
a complete discrete valuation ring with residue field $\mathbb{C}_p$ and
is equipped with the projective limit topology of the $p$-adic topology
on $\widetilde{\mathbb{B}}^+/(\mathrm{Ker}(\theta))^n$ whose lattice is the image of $\tilde{\mathbb{A}}^+\rightarrow
\tilde{\mathbb{B}}^+/(\mathrm{Ker}(\theta))^n$.
Let $A_{\mathrm{cris}}$ be the $p$-adic completion of $\widetilde{\mathbb{A}}^+[\frac{([\tilde{p}]-p)^n]}{n!}]_{n\geqq 1}$, where $\tilde{p}:=(p^{(n)})$ is an element in $\widetilde{\mathbb{E}}^+$ such that $p^{(0)}=p, (p^{(n+1) })^p=p^{(n)}$. We put $B_{\mathrm{cris}}^+:=A_{\mathrm{cris}}[\frac{1}{p}]$ equipped with $p$-adic topology whose lattice is $A_{\mathrm{cris}}$. $A_{\mathrm{cris}}$ and
$B^+_{\mathrm{cris}}$ have $G_K$-actions and Frobenius actions $\varphi$. We have a natural
$G_K$-equivariant embedding $K\otimes_{K_0}B^+_{\mathrm{cris}}\hookrightarrow B^+_{\mathrm{dR}}$.
If we put $t:=\mathrm{log}([\varepsilon])=\sum_{n=1}^{\infty}(-1)^{n-1}\frac{([\varepsilon]-1)^n}{n}$, then we can see that $t\in A_{\mathrm{cris}}$,
$\varphi(t)=pt, g(t)=\chi(g)t$ for any $g\in G_K$ and $\mathrm{Ker}(\theta)=tB_{\mathrm{dR}}^+
\subset B_{\mathrm{dR}}^+$ is the maximal ideal of $B^+_{\mathrm{dR}}$. If we put $B_{\mathrm{cris}}:=B_{\mathrm{cris}}^+
[\frac{1}{t}], B_{\mathrm{dR}}:=B_{\mathrm{dR}}^+[\frac{1}{t}]$, we have a natural embedding
$K\otimes_{K_0}B_{\mathrm{cris}}\hookrightarrow B_{\mathrm{dR}}$.
We put $B_e:=B_{\mathrm{cris}}^{\varphi =1}$ on which we equipped the inductive limit topology of
$B_e=\cup_{n}(\frac{1}{t^n}B^{+}_{\mathrm{cris}})^{\varphi=1}$, where each $(\frac{1}{t^n}B^{+}_{\mathrm{cris}})^{\varphi=1}
=\frac{1}{t^n}B^{+ \varphi=p^n}_{\mathrm{cris}}$ is
equipped with $p$-adic topology. We put $\mathrm{Fil}^i B_{\mathrm{dR}}:=t^i B_{\mathrm{dR}}^+$ for any $i\in\mathbb{Z}$.
In this paper, we fix a coefficient field of $p$-adic representations or $B$-pairs of $G_K$.
So we start in this subsection by recalling an $E$-coefficient version of $p$-adic representations and $B$-pairs.
\begin{defn}\label{a}
An $E$-representation of $G_K$ is a finite dimensional $E$-vector space $V$ with a continuous $E$-linear action of $G_K$. We call $E$-representation for simplicity when there is no risk
of confusion about $K$.
\end{defn}
\begin{defn} \label{b}
A pair $W:=(W_e,W^+_{\mathrm{dR}})$ is an $E$-$B$-pair if
\begin{itemize}
\item[(1)] $W_e$ is a finite $B_e\otimes_{\mathbb{Q}_p}E$-module which is free over $B_e$ with
a continuous semi-linear $G_K$-action.
\item[(2)] $W^+_{\mathrm{dR}}$ is a $G_K$-stable finite sub $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E$-module
of $B_{\mathrm{dR}}\otimes_{B_e}W_e$ which generates $B_{\mathrm{dR}}\otimes_{B_e}W_e$ as a $B_{\mathrm{dR}}$-module.
\end{itemize}
\end{defn}
Then, by Lemma 1.7, 1.8 of [Na09], $W_e$ is a free $B_e\otimes_{\mathbb{Q}_p}E$-module and
$W^+_{\mathrm{dR}}$ is a free $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E$-module. We define the rank of $W$ by $\mathrm{rank}(W):=\mathrm{rank}_{B_e\otimes_{\mathbb{Q}_p}E}(W_e)$.
For $E$-$B$-pairs $W_1:=(W_{1,e},W^+_{1,\mathrm{dR}})$ and $W_2:=(W_{2,e}, W^+_{2,\mathrm{dR}})$,
we define the tensor product of $W_1$ and $W_2$ by $W_1\otimes W_2:=(W_{1,e}\otimes_{B_e\otimes_{\mathbb{Q}_p}E}W_{2,e}, W^+_{1,\mathrm{dR}}\otimes_{B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}W^+_{2,\mathrm{dR}})$ and define the dual of $W_1$
by $W_1^{\vee}:=(\mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}E}(W_{1,e}, B_{e}\otimes_{\mathbb{Q}_p}E), \{f\in \mathrm{Hom}_{B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}(B_{\mathrm{dR}}
\otimes_{B_e}W_{1,e}, B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E)| f(W^+_{1,\mathrm{dR}})\subseteq B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E\})$ (remark: there is a mistake in the
Definition 1.9 of [Na09]).
The category of $E$-$B$-pairs is not an abelian category, i.e. an inclusion $W_1\hookrightarrow W_2$ in general
does not have a quotient in the category of $E$-$B$-pairs. But we can always take the saturation $W^{\mathrm{sat}}_{1}:=(W^{\mathrm{sat}}_{1.e}, W^{+,\mathrm{sat}}_{1,\mathrm{dR}})$ such that $W^{\mathrm{sat}}_1$ sits in
$W_1\hookrightarrow W^{\mathrm{sat}}_1\hookrightarrow W_2$ and $W_{1,e}=W^{\mathrm{sat}}_{1,e}$ and $W_2/W^{\mathrm{sat}}_{1}$ is an $E$-$B$-pair (Lemma 1.14 of [Na09]). We say that an inclusion $W_1\hookrightarrow W_2$ is saturated if $W_2/W_1$ is an $E$-$B$-pair.
Next, we recall $p$-adic Hodge theory for $B$-pairs. Let $W=(W_e,W^+_{\mathrm{dR}})$ be an $E$-$B$-pair.
We define $D_{\mathrm{cris}}(W):=(B_{\mathrm{cris}}\otimes_{B_e}W_e)^{G_K}$, $D^L_{\mathrm{cris}}(W):=
(B_{\mathrm{cris}}\otimes_{B_e}W_e)^{G_L}$ for any finite extension $L$ of $K$, $D_{\mathrm{dR}}(W):=(B_{\mathrm{dR}}\otimes_{B_e}
W_e)^{G_K}$, $D_{\mathrm{HT}}(W):=(B_{\mathrm{HT}}\otimes_{\mathbb{C}_p}(W^+_{\mathrm{dR}}/tW^+_{\mathrm{dR}}))^{G_K}$, here
$B_{\mathrm{HT}}:=\mathbb{C}_p[T, T^{-1}]$ where $G_K$ acts by $g(aT^i):=\chi(g)^ig(a)T^i$ for any $g\in G_K, a\in \mathbb{C}_p, i\in \mathbb{Z}$.
\begin{defn}\label{c}
We say that $W$ is crystalline (resp. de Rham, resp. Hodge-Tate) if
$\mathrm{dim}_{K_{\ast}}D_{\mathrm{\ast}}(W)=[E:\mathbb{Q}_p]\mathrm{rank}(W)$ for $\ast= \mathrm{cris}$ (resp. $\ast=\mathrm{dR}$, resp.
$\ast=\mathrm{HT}$), where $K_{\ast}=K_0$ when $\ast=\mathrm{cris}$ and $K_{\ast}=K$ when $\ast=\mathrm{dR},\mathrm{HT}$.
We say that $W$ is potentially crystalline if $\mathrm{dim}_{L_0}(D^L_{\mathrm{cris}}(W))=[E:\mathbb{Q}_p]\mathrm{rank}(W)$ for
a finite extension $L$ of $K$, where $L_0$ is the maximal unramified extension of $\mathbb{Q}_p$ in $L$.
\end{defn}
\begin{defn}\label{d}
Let $L$ be a finite Galois extension of $K$ and $G_{L/K}:=\mathrm{Gal}(L/K)$.
We say that $D$ is an $E$-filtered ($\varphi,G_{L/K}$)-module over $K$ if
\begin{itemize}
\item[(1)] $D$ is a finite free $L_0\otimes_{\mathbb{Q}_p} E$-module with a $\varphi$-semi-linear
action $\varphi_D$ and a semi-linear action of $G_{L/K}$ such that
$\varphi_D:D\isom D$ is an isomorphism and that $\varphi_D$ and $G_{L/K}$-action commute, where ($\varphi$-)semi-linear means that
$\varphi_D(a\otimes b\cdot x)=\varphi(a)\otimes b\cdot \varphi_D(x)$, $g(a\otimes b\cdot x)=g(a)\otimes b\cdot g(x)$ for
any $a\in L_0, b\in E, x\in D, g\in G_{L/K}$.
\item[(2)] $D_L:=L\otimes_{L_0}D$ has a separated and exhausted decreasing filtration $\mathrm{Fil}^iD_L$ by
sub $L\otimes_{\mathbb{Q}_p}E$-modules such that $G_{L/K}$-action on $D_L$ preserves this filtration.
\end{itemize}
\end{defn}
Let $W$ be a potentially crystalline $E$-$B$-pair such that $W|_{G_L}$ is crystalline
for a finite Galois extension $L$ of $K$. Then we define an $E$-filtered ($\varphi, G_{L/K}$)-module
structure on $D^L_{\mathrm{cris}}(W)$ as follows. First, $D^L_{\mathrm{cris}}(W)$ has a Frobenius action
induced from that on $B_{\mathrm{cris}}$ and has a $G_{L/K}$-action induced from those on $B_{\mathrm{cris}}$ and
$W_e$.
We define a filtration on $L\otimes_{L_0}D^L_{\mathrm{cris}}(W)=L\otimes_{K}D_{\mathrm{dR}}(W)$ by
$\mathrm{Fil}^i(L\otimes_{L_0}D^{L}_{\mathrm{cris}}(W)):=(L\otimes_{K}D_{\mathrm{dR}}(W))\cap t^iW^+_{\mathrm{dR}}$.
Let $D:=L_0e$ be a rank one $\mathbb{Q}_p$-filtered ($\varphi, G_{L/K}$)-module with a base $e$. then
we define $t_N(D):=v_p(\alpha)$ where $\varphi_D(e)=\alpha\cdot e$ and define $t_H(D):=i$ such that
$\mathrm{Fil}^iD_L/\mathrm{Fil}^{i+1}D_L\not= 0$. For general $D$ of rank $d$, we define $t_N(D):=t_N(\wedge^dD)$,
$t_H(D):=t_H(\wedge^dD)$. Then we say that $D$ is weakly admissible if $t_N(D)=t_H(D)$ and $t_N(D')\geqq t_H(D')$ for any
sub $\mathbb{Q}_p$-filtered ($\varphi,G_{L/K}$)-module $D'$ of $D$.
\begin{thm}\label{e}
Let $L$ be a finite Galois extension of $K$. Then we have the following results.
\begin{itemize}
\item[(1)] The functor $W\mapsto D^L_{\mathrm{cris}}(W)$ gives an equivalence of categories between
the category of potentially crystalline $E$-$B$-pairs which are crystalline if restricted to $G_L$ and the
category of $E$-filtered $(\varphi,G_{L/K})$-modules over $K$.
\item[(2)] By restricting the above functor to $E$-representations, the functor $V\mapsto D^{L}_{\mathrm{cris}}(V)$
gives an equivalence of categories between the category of potentially crystalline $E$-representations which are
crystalline if restricted to $G_L$ and the category of weakly admissible $E$-filtered $(\varphi,G_{L/K})$-modules over $K$.
\end{itemize}
\end{thm}
\begin{proof}
See [Be08, Proposition 2.3.4 and Theorem 2.3.5] or [Na09 Theorem 1.18].
\end{proof}
Next, we define trianguline $E$-$B$-pairs, whose deformation theory we study in detail in this article.
\begin{defn}\label{f}
Let $W$ be an $E$-$B$-pair of rank $n$. Then we say that $W$ is split trianguline if
there exists a filtration $\mathcal{T}:0\subseteq W_1\subseteq W_2\subseteq \cdots \subseteq W_n=W$ by
sub $E$-$B$-pairs such that $W_{i}$ is saturated in $W_{i+1}$ and $W_{i+1}/W_{i}$ is a rank one
$E$-$B$-pair for any $i$. We say that $W$ is trianguline if $W\otimes_{E}E'$, base change of $W$ to $E'$,
is a split trianguline $E'$-$B$-pair for a finite extension $E'$ of $E$.
\end{defn}
By this definition, for studying split trianguline $E$-$B$-pairs, it is important to classify rank one $E$-$B$-pairs
and calculate extension classes by rank one objects, which were studied in the previous article by the author ([Na09]).
Now, we recall some results concerning these.
\begin{thm}\label{g}
There exists canonical one to one correspondence $\delta\mapsto W(\delta)$ between the set of
continuous homomorphisms $\delta:K^{\times}\rightarrow E^{\times}$ and the set of isomorphism classes
of rank one $E$-$B$-pairs.
\end{thm}
\begin{proof}
See Proposition 3.1 of [Co08] for $K=\mathbb{Q}_p$ and Theorem 1.45 of [Na09] for general $K$.
For construction of $W(\delta)$, see $\S$1.4 of [Na09].
\end{proof}
This correspondence is compatible with local class field theory, i.e. for a unitary homomorphism
$\delta:K^{\times}\rightarrow \mathcal{O}_E^{\times}$ if we take the $\tilde{\delta}:G^{\mathrm{ab}}_K\rightarrow \mathcal{O}_E^{\times}$
satisfying $\tilde{\delta}\circ \mathrm{rec}_K=\delta$, then we have $W(\delta)\isom W(E(\tilde{\delta}))$.
This correspondence is also compatible with tensor products and with duals, i.e for homomorphisms $\delta_1, \delta_2:K^{\times}
\rightarrow E^{\times}$, we have $W(\delta_1)\otimes W(\delta_2)\isom W(\delta_1\delta_2)$ and $W(\delta_1)^{\vee}
\isom W(\delta_1^{-1})$. There are important examples of rank one $E$-$B$-pairs which we recall now.
For any $\{k_{\sigma}\}_{\sigma\in\mathcal{P}}$ such that $k_{\sigma}\in\mathbb{Z}$, we define a homomorphism
$\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}}:K^{\times}\rightarrow E^{\times}:y\mapsto \prod_{\sigma\in \mathcal{P}}
\sigma(y)^{k_{\sigma}}$. Then we have $W(\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}})=(B_e\otimes_{\mathbb{Q}_p}E,
\oplus_{\sigma\in\mathcal{P}}t^{k_{\sigma}}B^+_{\mathrm{dR}}\otimes_{K,\sigma}E)$ ( Lemma 2.12 of [Na09]).
Let $N_{K/\mathbb{Q}_p}:K^{\times}\rightarrow \mathbb{Q}_p^{\times}$ be the norm map and
$|-|:\mathbb{Q}_p^{\times}\rightarrow \mathbb{Q}^{\times}\hookrightarrow E^{\times}$ be the $p$-adic absolute value
such that $|p|=\frac{1}{p}$, and we define $|N_{K/\mathbb{Q}_p}(x)|:K^{\times}\rightarrow E^{\times}$ the composite
of $N_{K/\mathbb{Q}_p}$ and $|-|$. Then we have $W(|N_{K/\mathbb{Q}_p}(x)|\prod_{\sigma\in \mathcal{P}}\sigma(x))=W(E(\chi))$, which is the $E$-$B$-pair
associated to $p$-adic cyclotomic homomorphism. These homomorphisms play special roles in the calculation of Galois cohomology of
$E$-$B$-pairs.
Next, we recall the definition and some properties of Galois cohomology of $E$-$B$-pairs.
For an $E$-$B$-pair $W:=(W_e, W^+_{\mathrm{dR}})$, we put $W_{\mathrm{dR}}:=B_{\mathrm{dR}}\otimes_{B_e}W_e$.
We have natural inclusions, $W_e\hookrightarrow W_{\mathrm{dR}}$, $W^+_{\mathrm{dR}}
\hookrightarrow W_{\mathrm{dR}}$. Then we define the Galois cohomology $\mathrm{H}^i(G_K, W)$ of $W$ to be
the Galois cohomology of the complex $W_e\oplus W^+_{\mathrm{dR}}\rightarrow W_{\mathrm{dR}}:(x,y)\mapsto
x+y$ (see $\S$2.1 of [Na09] for the precise definition). Then as in the usual $p$-adic representation case, we have $\mathrm{H}^0(G_K, W)
=\mathrm{Hom}_{G_K}(B_E, W)$ and $\mathrm{H}^1(G_K, W)=\mathrm{Ext}^1(B_E, W)$, where $B_E:=(B_e\otimes_{\mathbb{Q}_p}E,
B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E)$ is the trivial $E$-$B$-pair and $\mathrm{Hom}_{G_K}(-,-)$ is the group of homomorphisms
of $E$-$B$-pairs and $\mathrm{Ext}^1(-,-)$ is the extension class group in the category of $E$-$B$-pairs. If
$V$ is an $E$-representation of $G_K$, then we have $\mathrm{H}^i(G_K, V)\isom \mathrm{H}^i(G_K, W(V))$, which follows from
Bloch-Kato's fundamental short exact sequence $0\rightarrow \mathbb{Q}_p\rightarrow B_e\oplus B^+_{\mathrm{dR}}
\rightarrow B_{\mathrm{dR}}\rightarrow 0$.
Moreover, we have the following theorem, Euler-Poincar$\mathrm{\acute{e}}$ homomorphismistic formula and Tate duality theorem for
$B$-pairs.
\begin{thm}\label{h}
Let $W$ be an $E$-$B$-pair. Then$:$
\begin{itemize}
\item[(1)] for $i=0,1,2$, $\mathrm{H}^i(G_K, W)$ is finite dimensional over $E$ and $\mathrm{H}^i(G_K, W)=0$ for
$i\not= 0,1,2$.
\item[(2)] $\sum_{i=0}^2(-1)^i \mathrm{dim}_E\mathrm{H}^i(G_K, W)=[K:\mathbb{Q}_p] \mathrm{rank}(W)$.
\item[(3)] let $W$ be a $\mathbb{Q}_p$-$B$-pair, Then, for any $i=0,1,2$, there is a natural perfect pairing defined by cup
product,
\item[] \[
\begin{array}{ll}
\mathrm{H}^i(G_K, W)\times\mathrm{H}^{2-i}(G_K, W^{\vee}(\chi))&\rightarrow \mathrm{H}^2(G_K, W\otimes W^{\vee}(
\chi)) \\
&\rightarrow \mathrm{H}^2(G_K, W(\mathbb{Q}_p(\chi))\allowbreak \isom \mathbb{Q}_p,
\end{array}
\]
\item[] where the last isomorphism is the trace map.
\end{itemize}
\end{thm}
\begin{proof}
See Theorem 4.3, 4.7 of [Liu08] and Theorem 2.8 of [Na09].
\end{proof}
By using these formulae, we have the following dimension formulae for rank one $E$-$B$-pairs.
\begin{prop}\label{i}
Let $\delta:K^{\times}\rightarrow E^{\times}$ be a continuous homomorphism. Then we have$:$
\begin{itemize}
\item[(1)] $\mathrm{H}^0(G_K, W(\delta))\isom E $ if $\delta=\prod_{\sigma\in \mathcal{P}}\sigma(x)^{k_{\sigma}}$
such that $k_{\sigma}\in \mathbb{Z}_{\leqq 0}$ for any $\sigma\in \mathcal{P}$, and $\mathrm{H}^0(G_K, W(\delta))=0$ otherwise.
\item[(2)] $\mathrm{H}^2(G_K, W(\delta))\isom E$ if $\delta=|N_{K/\mathbb{Q}_p}|\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}}$
such that $k_{\sigma}\in \mathbb{Z}_{\geqq 1}$ for any $\sigma\in \mathcal{P}$, and $\mathrm{H}^2(G_K, W(\delta))=0$ otherwise.
\item[(3)] $\mathrm{dim}_E\mathrm{H}^1(G_K, W(\delta))=[K:\mathbb{Q}_p] +1$ if $\delta=\prod_{\sigma\in \mathcal{P}}\sigma(x)^{k_{\sigma}}$ such that $k_{\sigma}\in \mathbb{Z}_{\leqq 0}$ for any $\sigma\in \mathcal{P}$ or $\delta=
|N_{K/\mathbb{Q}_p}|\prod_{\sigma\in \mathcal{P}}\sigma(x)^{k_{\sigma}}$ such that $k_{\sigma}\in \mathbb{Z}_{\geqq 1}$ for any $\sigma\in \mathcal{P}$, and $\mathrm{dim}_E\mathrm{H}^1(G_K, W(\delta))=[K:\mathbb{Q}_p]$ otherwise.
\end{itemize}
\end{prop}
\begin{proof}
See Theorem 2.9 and Theorem 2.22 of [Co08] for $K=\mathbb{Q}_p$. For general $K$, the results may be proved by using
Proposition 2.14 and Proposition 2.15 of [Na09] and Tate duality for $B$-pairs.
\end{proof}
\subsection{$B$-pairs over Artin local rings}
Let $\mathcal{C}_E$
be the category of Artin local $E$-algebra $A$ with residue field $E$. The morphisms of $\mathcal{C}_E$ are given by
local $E$-algebra homomorphisms. For any $A\in \mathcal{C}_E$, let $m_A$ be the maximal ideal of $A$. We define the $A$-coefficient versions of $B$-pairs as follows.
\begin{defn}\label{1}
We call a pair $W:=(W_e, W^+_{\mathrm{dR}})$ an $A$-$B$-pair of $G_K$ if
\begin{itemize}
\item[(1)] $W_e$ is a finite $B_e\otimes_{\mathbb{Q}_p}A$-module which is flat over $A$
and is free over $B_e$, with a continuous semi-linear $G_K$-action.
\item[(2)] $W^+_{\mathrm{dR}}$ is a finite generated sub $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-module of $B_{\mathrm{dR}}\otimes_{B_e}W_e$ which is preserved by $G_K$-action and which generates
$B_{\mathrm{dR}}\otimes_{B_e}W_e$ as a $B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-module such that
$W^+_{\mathrm{dR}}/tW^+_{\mathrm{dR}}$ is flat over $A$.
\end{itemize}
For an $A$-$B$-pair $W:=(W_e,W^+_{\mathrm{dR}})$, we put $W_{\mathrm{dR}}:=B_{\mathrm{dR}}
\otimes_{B_e}W_e$.
\end{defn}
We simply call an $A$-$B$-pair if there is no risk of confusing about $K$.
\begin{lemma}\label{2}
Let $W:=(W_e,W^+_{\mathrm{dR}})$ be an $A$-$B$-pair. Then
$W_e$ is a finite free $B_e\otimes_{\mathbb{Q}_p}A$-module and
$W^+_{\mathrm{dR}}$ is a finite free $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-module
and $W^+_{\mathrm{dR}}/tW^+_{\mathrm{dR}}$ is a finite free $\mathbb{C}_p\otimes_{\mathbb{Q}_p}A$-module.
\end{lemma}
\begin{proof}
First, we prove for $W_e$. Because the sub module $m_A W_e\subseteq W_e$ is a $G_K$-stable finite generated
torsion free $B_e$-module, $m_AW_e$ is a finite free $B_e$-module by Lemma 2.4 of [Ke04].
Then, by Lemma 2.1.4 of [Be08], the cokernel $W_e\otimes_A E$ is also a finite free $B_e$-module
(with an $E$-action). Then, by Lemma 1.7 of [Na09], $W_e\otimes_AE$ is a finite free $B_e\otimes_{\mathbb{Q}_p}E$-module of
some rank $n$.
We take a $B_e\otimes_{\mathbb{Q}_p}A$-linear morphism $f:(B_e\otimes_{\mathbb{Q}_p}A)^n\rightarrow W_e$ which
is a lift of a $B_e\otimes_{\mathbb{Q}_p}E$-linear isomorphism $(B_e\otimes_{\mathbb{Q}_p}E)^n\isom W_e\otimes_A E$.
Then, by Nakayama's lemma, $f$ is surjective. Because $W_e$ is $A$-flat, we have $\mathrm{Ker}(f)\otimes_AE=0$,
so $\mathrm{Ker}(f)=0$. So $W_e$ is a free $B_e\otimes_{\mathbb{Q}_p}A$-module.
Next, we prove that $W^+_{\mathrm{dR}}$ is a free $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-module.
Because $W_e$ is a free $B_e\otimes_{\mathbb{Q}_p}A$-module, $W_{\mathrm{dR}}$ is a free $B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-module, in particular this is flat over $A$. Because $W^+_{\mathrm{dR}}/tW^+_{\mathrm{dR}}$ is a flat $A$-module, $W_{\mathrm{dR}}/W^+_{\mathrm{dR}}$ is also a flat $A$-module. So $W^+_{\mathrm{dR}}$ is also flat over $A$.
By the $A$-flatness of $W_{\mathrm{dR}}/W^+_{\mathrm{dR}}$, we have inclusion
$W^+_{\mathrm{dR}}\otimes_A E\hookrightarrow W_{\mathrm{dR}}\otimes_{A} E$, so
$W^{+}_{\mathrm{dR}}\otimes_A E$ is a finite generated torsion free $B^+_{\mathrm{dR}}$-module, so
free $B^+_{\mathrm{dR}}$-module. By using these and Lemma 1.8 of [Na09], in the same way as in the case of $W_e$, we can show that
$W^+_{\mathrm{dR}}$ is a free $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-module.
$\mathbb{C}_p\otimes_{\mathbb{Q}_p}A$-freeness of $W^+_{\mathrm{dR}}/tW^+_{\mathrm{dR}}$ follows from
$B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-freeness of $W^+_{\mathrm{dR}}$.
\end{proof}
By using this lemma, we can define the rank of $A$-$B$-pair.
\begin{defn}
Let $W=(W_e, W^+_{\mathrm{dR}})$ be an $A$-$B$-pair. Then we define the rank of $W$ as
$\mathrm{rank}(W):=\mathrm{rank}_{B_e\otimes_{\mathbb{Q}_p}A}(W_e)$.
\end{defn}
\begin{defn}\label{3}
Let $f:A\rightarrow A'$ be a morphism in $\mathcal{C}_E$ and $W=(W_e,W_{\mathrm{dR}})$ be an $A$-$B$-pair.
Then we define the base change of $W$ to $A'$ by $W\otimes_A A':=(W_e\otimes_A A',W^+_{\mathrm{dR}}\otimes_A A')$.
By using lemma $\ref{2}$, we can easily see that this is an $A'$-$B$-pair.
\end{defn}
\begin{defn}\label{4}
Let $W_1=(W_{e,1},W^+_{\mathrm{dR},1})$, $W_2=(W_{e,2}, W^+_{\mathrm{dR},2})$ be $A$-$B$-pairs.
Then we define the tensor product of $W_1$ and $W_2$ by $W_1\otimes W_2:=(W_{e,1}\otimes_{B_e\otimes_{\mathbb{Q}_p}A}W_{e,2},
W^+_{\mathrm{dR},1}\otimes_{B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A}W^+_{\mathrm{dR},2})$ ,
and define the dual of $W_1$ by $W_1^{\vee}:=(\mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}A}(W_{e,1},\allowbreak B_e\otimes_{\mathbb{Q}_p}A),
W^{+,\vee}_{\mathrm{dR},1})$. Here, $W^{+,\vee}_{\mathrm{dR},1}:=\{f\in\mathrm{Hom}_{B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A}(
W_{\mathrm{dR},1}, B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A)|f(W^{+}_{\mathrm{dR},1})\subseteq B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A\}$. By using lemma $\ref{2}$, we can easily see that these are $A$-$B$-pairs.
\end{defn}
Next, we classify rank one $A$-$B$-pairs.
Let $\delta:K^{\times}\rightarrow A^{\times}$ be a continuous homomorphism.
Then we define the rank one $A$-$B$-pair $W(\delta)$ as follows.
Let $\bar{u}\in E^{\times}$ be the reduction of $u:=\delta(\pi_K)$.
We define a homomorphism $\delta_0:K^{\times}\rightarrow A^{\times}$ such that $\delta_0|_{\mathcal{O}_K^{\times}}
=\delta|_{\mathcal{O}_K^{\times}}$, $\delta_0(\pi_K)=u/\bar{u}$. Then, because $u/\bar{u}\in 1+m_A$ , $(u/\bar{u})^{p^n}$
converges to $1\in A^{\times}$. So if we fix an isomorphism $K^{\times}\isom \mathcal{O}_K^{\times}\times \mathbb{Z}
:v\pi_K^n\mapsto (v,n)$ (here $v\in\mathcal{O}_K^{\times}$), then $\delta_0$ extends to $\delta_0':\mathcal{O}_K^{\times}\times
\hat{\mathbb{Z}}\rightarrow \mathcal{O}_K^{\times}\times \mathbb{Z}_p\rightarrow A^{\times}$. Hence, by local class field theory, there exists unique homomorphism $\tilde{\delta}_0:G_K^{\mathrm{ab}}\rightarrow A^{\times}$ such
that $\delta_0=\tilde{\delta}_0\circ\mathrm{rec}_K$, here $\mathrm{rec}_K:K^{\times}\rightarrow G_K^{\mathrm{ab}}$ is the reciprocity map as in Notation. We define the
$\mathrm{\acute{e}}$tale rank one $A$-$B$-pair $W(A(\tilde{\delta}_0))$, which is the $A$-$B$-pair associated to the rank one $A$-representation $A(\tilde{\delta}_0)$.
Next, we define a non-$\mathrm{\acute{e}}$tale rank one $A$-$B$-pair by using $\bar{u}\in E^{\times}$.
For this, first we define a rank one $E$-filtered $\varphi$-module $D_{\bar{u}}:=K_0\otimes_{\mathbb{Q}_p}Ee_{\bar{u}}$ such that
$\varphi^f(e_{\bar{u}}):=\bar{u}e_{\bar{u}}$ and $\mathrm{Fil}^0(K\otimes_{K_0}D_{\bar{u}}):=K\otimes_{K_0}D_{\bar{u}}$, $\mathrm{Fil}^1(K\otimes_{K_0}D_{\bar{u}}):=0$.
From this, we get the rank one crystalline $E$-$B$-pair $W(D_{\bar{u}})$ such that $D_{\mathrm{cris}}(W(D_{\bar{u}}))\isom D_{\bar{u}}$ which is
pure of slope $\frac{v_p(\bar{u})}{f}$.
By tensoring these, we define the rank one $A$-$B$-pair $W(\delta)$ by $W(\delta):=(W(D_{\bar{u}})\otimes_EA)\otimes W(A(\tilde{\delta}_0))$,
which is pure of slope $\frac{v_p(\bar{u})}{f}$.
The following proposition is the $A$-coefficient version of Theorem 1.45 of [Na09].
\begin{prop}\label{5}
This construction $\delta\mapsto W(\delta)$ does not depend on the
choice of uniformizer $\pi_K$ and gives a bijection between
the set of continuous homomorphisms $\delta:K^{\times}\rightarrow A^{\times}$ and
the set of isomorphism classes of rank one $A$-$B$-pairs.
\end{prop}
\begin{proof}
The independence of the choice of uniformizer and the injection is proved in the same
way as in the proof of Theorem 1.45 of [Na09]. We prove the surjection. Let $W$ be a rank one $A$-$B$-pair. Then
as an $E$-$B$-pair, $W$ is a successive extension of rank one $E$-$B$-pair $W\otimes_A E$.
$W\otimes_AE$ is pure of slope $\frac{n}{fe_E}$ for some $n\in \mathbb{Z}$ by Lemma 1.42 of [Na09].
So, by Theorem 1.6.6 of [Ke08], $W$ is also pure of slope $\frac{n}{fe_E}$.
We define the rank one $E$-filtered $\varphi$-module $D_{\pi_K^n}:=K_0\otimes_{\mathbb{Q}_p}Ee_{\pi_K^n}$
in the same way as in $D_{\bar{u}}$. Then $W\otimes (W(D_{\pi_K^n}) \otimes_E A)^{\vee}$ is pure of slope zero by Lemma
1.34 of [Na09]. So there exists $\tilde{\delta}':G_K^{\mathrm{ab}}\rightarrow A^{\times}$ such that
$W\otimes (W(D_{\pi^n})\otimes_E A)^{\vee}\isom W(A(\tilde{\delta}'))$. If we put $\delta':=\tilde{\delta}'\circ\mathrm{rec}_K:
K^{\times}\rightarrow A^{\times}$ and put $\delta:K^{\times}\rightarrow A^{\times}$ such that $\delta|_{\mathcal{O}_K^{\times}}
:=\delta'|_{\mathcal{O}_K^{\times}}$ and $\delta(\pi_K):=\delta'(\pi_K)\pi_K^n$, then we have an isomorphism
$W\isom W(\delta)$ which can be easily seen from the construction of $W(\delta)$.
\end{proof}
By local class field theory, we have a canonical bijection $\delta\mapsto A(\tilde{\delta})$ from the set of unitary continuous homomorphisms
from $K^{\times}$ to $A^{\times}$ (where unitary means that the reduction of the image of $\delta$ is contained in $\mathcal{O}_E^{\times}$)
to the set of isomorphism class of rank one $A$-representations of $G_K$, where $\tilde{\delta}:G_K^{ab}\rightarrow
A^{\times}$ is the continuous homomorphism such that $\delta=\tilde{\delta}\circ\mathrm{rec}_K$.
By the definition of $W(\delta)$ and by the above proof, it is easy to see that there exists an isomorphism
$W(\delta)\isom W(A(\tilde{\delta}))$ for any unitary homomorphism $\delta:K^{\times}\rightarrow A^{\times}$.
Moreover, it is easy to see that for any continuous homomorphisms $\delta_1,\delta_2: K^{\times}
\rightarrow A^{\times}$ we have isomorphisms $W(\delta_1)\otimes W(\delta_2)\isom W(\delta_1\delta_2)$ and $W(\delta_1)^{\vee}
\isom W(\delta_1^{-1})$.
\section{Deformations of trianguline $B$-pairs}
In this section, we develop the deformation theory of $B$-pairs, in particular
we study trianguline deformation functors for trianguline $B$-pairs, which
are the generalization of the deformation theory of usual $p$-adic Galois
representations and the generalization of the theory of trianguline deformation of
[Bel-Ch] in the case of $K=\mathbb{Q}_p$ by using
$(\varphi,\Gamma)$-modules. We prove the pro-representabilities and formally smoothnesses
of these deformation functors and calculate their dimensions by using the calculations of
Galois cohomology of $B$-pairs in the last section.
\subsection{Deformations of $E$-$B$-pairs}
In this subsection, we study the usual deformation theory for $E$-$B$-pairs,, which
are the generalization of the usual (homomorphismistic zero) deformation theory for
$E$-representations.
\begin{defn}\label{6}
Let $A$ be an object in $\mathcal{C}_E$, $W$ be an $E$-$B$-pair.
Then we say that a pair $(W_A,\iota)$ is a deformation of $W$ over $A$
if $W_A$ is an $A$-$B$-pair and $\iota: W_A\otimes_AE\isom W$
is an isomorphism of $E$-$B$-pairs. Let $(W_A,\iota)$, $(W'_A,\iota')$ be
two deformations of $W$ over $A$. Then we say that $(W_A,\iota)$ and $(W'_A,\iota')$ are
equivalent if there exists an isomorphism $f:W_A\isom W'_A$ of $A$-$B$-pairs which
satisfies $\iota=\iota'\circ\bar{f}$, where $\bar{f}:W_A\otimes_A E\isom W'_A\otimes_AE$
is the reduction of $f$.
\end{defn}
\begin{defn} \label{7}
Let $W$ be an $E$-$B$-pair. Then we define the deformation functor
$D_W$ from the category $\mathcal{C}_E$ to the category of sets by defining, for any
$A\in\mathcal{C}_E$,
$D_W(A):=\{$ equivalent classes $(W_A,\iota)$ of deformations of $W$ over $A$ $\}$.
\end{defn}
We simply write $W_A$ if there is no risk of confusing about $\iota$.
Next, we prove the pro-representability
of the functor $D_W$ under suitable conditions.
For this, we recall Schlessinger's criterion for pro-representability of functors from
$\mathcal{C}_E$ to the category of sets.
We call a morphism $f:A'\rightarrow A$ in $\mathcal{C}_E$ a small extension if it is surjective and
the kernel $\mathrm{Ker}(f)=(t)$ is generated by a nonzero single element $t\in A'$
and $\mathrm{Ker}(f)\cdot m_{A'}=0$. $E[\varepsilon]$ is the ring of dual number over $E$,
i.e. defined by $E[\varepsilon]:=E[X]/(X^2)$.
\begin{thm}\label{8}
Let $F$ be a functor from $\mathcal{C}_E$ to the category of sets such that
$F(E)$ is a single point. For morphisms $A'\rightarrow A$, $A"\rightarrow A$ in
$\mathcal{C}_A$, consider the natural map
\begin{itemize}
\item[(1)] $F(A'\times_{A}A")\rightarrow F(A')\times_{F(A)}F(A")$.
\end{itemize}
Then $F$ is pro-representable if and only if $F$ satisfies properties $(H_1)$,
$(H_2)$, $(H_3)$, $(H_4)$ below:
\begin{itemize}
\item[$(H_1)$] $(1)$ is surjective if $A"\rightarrow A$ is surjective.
\item[$(H_2)$] $(1)$ is bijective when $A=E$ and $A"=E[\varepsilon]$.
\item[$(H_3)$] $\mathrm{dim}_E(t_F)<\infty$ $($ where $t_F:=F(E[\varepsilon])$ and, under
the condition $(H_2)$, it is known that $t_F$ has a natural $E$-vector space structure$)$.
\item[$(H_4)$] $(1)$ is bijective if $A'=A"$ and $A'\rightarrow A$ is a small extension.
\end{itemize}
\end{thm}
\begin{proof}
See [Sch68] or $\S 18$ of [Ma97].
\end{proof}
By using this criterion, we prove the pro-representability of
$D_W$.
\begin{prop}\label{9}
Let $W$ be an $E$-$B$-pair.
If $\mathrm{End}_{G_K}(W)=E$, i.e. $W$ has only trivial morphism
of $E$-$B$-pairs, then $D_W$ is pro-representable by a complete
noetherian local $E$-algebra $R_W$ with residue field $E$.
\end{prop}
For proving this proposition, we first prove some lemmas.
\begin{lemma}\label{10}
Let $\mathrm{ad}(W):=\mathrm{Hom}(W,W)$ be the internal endomorphism of $W$.
Then there exists an isomorphism of $E$-vector spaces
$D_W(E[\varepsilon])\isom \mathrm{H}^1(G_K,\mathrm{ad}(W))$.
\end{lemma}
\begin{proof}
Let $W_{E[\varepsilon]}:=(W_{E[\varepsilon],e}, W^+_{E[\varepsilon],\mathrm{dR}})$
be a deformation of $W$ over $E[\varepsilon]$. From this, we define an element
in $\mathrm{H}^1(G_K,\mathrm{ad}(W))$ as follows. Because $\varepsilon W_{E[\varepsilon],e}\isom
W_e$ and $W_{E[\varepsilon],e}/\varepsilon W_{E[\varepsilon],e} \isom W_e$ (where we put $W:=(W_e,W^+_{\mathrm{dR}})$),
we have a natural exact sequence of $B_e\otimes_{\mathbb{Q}_p}E[G_K]$-modules
\begin{equation*}
0\rightarrow W_e\rightarrow W_{E[\varepsilon],e}\rightarrow W_e\rightarrow 0
\end{equation*}
We fix an isomorphism of $B_e\otimes_{\mathbb{Q}_p}E$-modules
$W_{E[\varepsilon],e}\isom W_e e_1\oplus W_ee_2$ such that first factor $W_e e_1$is equal to $\varepsilon W_{E[\varepsilon]}$
as $B_e\otimes_{\mathbb{Q}_p}E[G_K]$-module and that the above natural projection maps the second factor $W_e e_2$ to $W_e$ by $xe_2\mapsto x$ for any $x\in W_e$. Then we define a continuous one cocycle $c_e: G_K\rightarrow \mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}E}(W_e,W_e)$
by $g(ye_2):=c_e(g)(gy)e_1+gye_2$ for any $g\in G_K$ and $y\in W_e$.
For $W^+_{\mathrm{dR}}$, we fix an isomorphism $W^+_{E[\varepsilon],\mathrm{dR}}\isom W^+_{\mathrm{dR}}e_1\oplus
W^+_{\mathrm{dR}}e'_2$ as in the case of $W_e$, then we define a one cocycle $c_{\mathrm{dR}}:G_K\rightarrow \mathrm{Hom}_{B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}(W^+_{\mathrm{dR}},W^+_{\mathrm{dR}})$ by $g(ye'_2):=c_{\mathrm{dR}}(g)(gy)e_1+gye'_2$ for
any $g\in G_K$ and $y\in W^+_{\mathrm{dR}}$. Next, we define an element $c\in\mathrm{Hom}_{B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}
(W_{\mathrm{dR}},W_{\mathrm{dR}})$ as follows. By tensoring $W_{E[\varepsilon],e}$ and $W^+_{E[\varepsilon],\mathrm{dR}}$ with
$B_{\mathrm{dR}}$ over $B_e$ or $B^+_{\mathrm{dR}}$, we have
an isomorphism $f:W_{\mathrm{dR}}e_1\oplus W_{\mathrm{dR}}e_2\isom W_{E[\varepsilon],\mathrm{dR}}\isom
W_{\mathrm{dR}}e_1\oplus W_{\mathrm{dR}}e'_2$ of $B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E$-modules.
We define $c:W_{\mathrm{dR}}\rightarrow W_{\mathrm{dR}}$ by $f(ye_2):=c(y)e_1+ye'_2$ for any $y\in W_{\mathrm{dR}}$.
Then, by definition, the triple ($c_e$, $c_{\mathrm{dR}}$, $c$) satisfies $c_e(g)-c_{\mathrm{dR}}(g)=gc-c$ in $\mathrm{Hom}_{B_{\mathrm{dR}}
\otimes_{\mathbb{Q}_p}E}(W_{\mathrm{dR}},W_{\mathrm{dR}})$ for any $g\in G_K$, i.e. the triple $(c_e,c_{\mathrm{dR}},c)\in
\mathrm{H}^1(G_K,\mathrm{ad}(W))$ by the definition of Galois cohomology of $B$-pairs ($\S$ 2.1 of [Na09]). Then it is standard to check that this definition is independent of
the choice of fixed isomorphism $W_{E[\varepsilon],e}\isom W_ee_1\oplus W_ee_2$, etc, and it is easy to check that
this map defines an isomorphism $D_W(E[\varepsilon])\isom \mathrm{H}^1(G_K,\mathrm{ad}(W))$.
\end{proof}
\begin{lemma}\label{11}
Let $W_A$ be a deformation of $W$ over $A$.
If $\mathrm{End}_{G_K}(W)=E$, then $\mathrm{End}_{G_K}(W_A)=A$.
\end{lemma}
\begin{proof}
We prove this lemma by induction on the length of $A$.
When $A=E$, this is trivial. We assume that this lemma is proved for length $n$
rings and assume $A$ is length $n+1$. We take a small extension
$f:A\rightarrow A'$. Because $\mathrm{End}_{G_K}(W)=\mathrm{H}^0(G_K,
W^{\vee}\otimes W)$, then we have the following short exact sequence,
\begin{equation*}
0\rightarrow \mathrm{Ker(f)}\otimes_E\mathrm{End}_{G_K}(W)
\rightarrow \mathrm{End}_{G_K}(W_A)\rightarrow
\mathrm{End}_{G_K}(W_A\otimes_A A').
\end{equation*}
From this and induction hypothesis, we have
$\mathrm{length}(\mathrm{End}_{G_K}(W_A))\leqq
\mathrm{length}(\mathrm{End}_{G_K}\allowbreak (W_A\otimes_A A'))+
\mathrm{length}(\mathrm{Ker}(f)\otimes_E \mathrm{End}_{G_K}(W))
=\mathrm{length}(A')+1=\mathrm{length}(A)$. On the other hand,
we have a natural inclusion $A\subseteq \mathrm{End}_{G_K}(W_A)$.
So, by comparing length, we have $A=\mathrm{End}_{G_K}(W_A)$.
\end{proof}
\begin{proof} (of proposition)
Let $W$ be a rank $n$ $E$-$B$-pair satisfying $\mathrm{End}_{G_K}(W)=E$.
For this $W$, we check the conditions $(H_i)$ of Schlessinger's criterion.
First, by Lemma $\ref{10}$, we have $\mathrm{dim}_E(D_{W}(E[\varepsilon]))=
\mathrm{dim}_E(\mathrm{H}^1(G_K,\mathrm{ad}(W)))<\infty$, so $(H_3)$ is satisfied.
Next we check $(H_1)$. Let $f:A'\rightarrow A$, $g:A"\rightarrow A$ be morphisms in
$\mathcal{C}_E$ such that $g$ is surjective. Let $([W_{A'}],[W_{A"}])$ be an element in
$D_W(A')\times_{D_W(A)}D_W(A")$. We take deformations $W_{A'}:=(W_{A' e},W^{ +}_{A' \mathrm{dR}})
$, $W_{A"}:=(W_{A" e},W^{+}_{A" \mathrm{dR}})$ over $A'$ and $A"$ contained in equivalent classes
$[W_{A'}]$ and $[W_{A"}]$ respectively. Then we have an isomorphism $h:W_{A'}\otimes_{A'}A\isom W_{A"}\otimes_{A"}A=:W_A
:=(W_{A e}, W^+_{A\mathrm{dR}})$
which defines an equivalent class in $D_W(A)$. We fix a basis $e_1,\cdots,e_n$ of $W_{A' e}$ as a $B_e\otimes_{\mathbb{Q}_p}A'$
-module and write $\bar{e}_1,\cdots,\bar{e}_n$ the basis of $W_{A' e}\otimes_{A'}A$ induced from $e_1,\cdots,e_n$.
Then, by the surjection of $g:A"\rightarrow A$ and by $A"$-flatness of $W_{A" e}$, we can take a basis
$\tilde{e}_1,\cdots,\tilde{e}_n$ of $W_{A" e}$ such that the basis
$\bar{\tilde{e}}_1,\cdots,\bar{\tilde{e}}_n$ of $W_{A" e}\otimes_{A"}A$ induced from $\tilde{e}_1,\cdots,\tilde{e}_n$
satisfies $h(\bar{e}_i)=\bar{\tilde{e}}_i$ for any $i$. Then $W^{'"}_e:=W_{A' e}\times_{W_{A e}}W_{A" e}
:=\{(x,y)\in W_{A' e}\times W_{A" e}| h(\bar{x})=\bar{y}\}$ is a free $B_e\otimes_{\mathbb{Q}_p}(A'\times_A A")$-module
with basis $(e_1,\tilde{e}_1), \cdots, (e_n,\tilde{e}_n)$. In the same way, we can define
$W^{'"+}_{\mathrm{dR}}:=W^{+}_{A' \mathrm{dR}}\times_{W^+_{A \mathrm{dR}}} W^{+}_{A" \mathrm{dR}}$,
which is a free $B^{+}_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}(A'\times_A A")$-module. Then
$W_{A'}\times_{W_A} W_{A"}:=(W^{'"}_e,W^{'" +}_{\mathrm{dR}})$ is a $(A'\times_A A")$-$B$-pair which is a deformation
of $W$ over $A'\times_A A"$ such that the equivalent class $[W_{A'}\times_{W_A}W_{A"}]\in D_W(A'\times_A A")$ maps
$([W_{A'}], [W_{A"}])\in D_W(A')\times_{D_W(A)}D_W(A")$. So we have proved $(H_1)$.
Finally, we prove the following. If $g:A"\rightarrow A$ is surjective, then the natural map
$D_W(A'\times_AA")\rightarrow D_W(A')\times_{D_W(A)}D_W(A")$ is bijective, which
proves the condition $(H_2)$ and $(H_4)$, so we can prove the pro-representability of $D_W$.
Let $W^{'"}_1$, $W^{'"}_2$ be deformations of $W$ over $A'\times_A A"$ such
that $[W^{'"}_1\otimes_{A'\times_AA"}A']=[W^{'"}_2\otimes_{A'\times_{A}A"}A']$ in $D_W(A')$ and
$[W^{'"}_1\otimes_{A'\times_AA"}A"]=[W^{'"}_2\otimes_{A'\times_{A}A"}A"]$ in $D_W(A")$.
Then we want to show $[W^{'"}_1]=[W^{'"}_2]$ in $D_W(A'\times_A A")$.
We put $W_{1 A'}:=W^{'"}_1\otimes_{A'\times_A A"}A'$, $W_{1 A"}:=W_1^{'"}\otimes_{A'\times_A A"}A"$
, $W_{1 A}:=W_1^{'"}\otimes_{A'\times_A A"} A$, and same for $W_{2 A'}$, $W_{2 A"}$, $W_{2 A}$.
Then we have natural isomorphisms $W^{'"}_1\isom W_{1 A'}\times_{W_{1 A}} W_{1 A"}$ and $W_2^{'"}
\isom W_{2 A'}\times_{W_{2 A}} W_{2 A"}$ defined as in the last paragraph. Because
$[W_{1 A'}]=[W_{2 A'}]$ and $[W_{1 A"}]=[W_{2 A"}]$, we have isomorphisms
$h':W_{1 A'}\isom W_{2 A'}$ and $h":W_{1 A"}\isom W_{2 A"}$. By reduction of these isomorphisms,
we get an automorphism $\bar{h}'\circ \bar{h}^{" -1}:W_{2 A}\isom W_{1 A}\isom W_{2 A}$.
Then, by Lemma $\ref{11}$ and the surjection of $g:A^{" \times}\rightarrow A^{\times}$, we can find an automorphism
$\tilde{h}:W_{2 A"}\isom W_{2 A"}$ such that $\bar{\tilde{h}}=\bar{h}'\circ \bar{h}^{" -1}$.
Then, ifwe define a morphism $h^{'"}:W_{1 A'}\times_{W_{1 A}} W_{1 A"}\rightarrow
W_{2 A"}\times_{W_{2 A}} W_{2 A'}: (x,y)\mapsto (h_1(x), \tilde{h}\circ h_2(y))$, we can see that this is
well-defined and is isomorphism. So we finish to prove this proposition.
\end{proof}
\begin{prop}\label{12}
Let $W:=(W_e, W^+_{\mathrm{dR}})$ be an $E$-$B$-pair of rank $n$.
If $\mathrm{H}^2(G_K, \mathrm{ad}(W))\allowbreak =0$, then the functor
$D_W$ is formally smooth.
\end{prop}
\begin{proof}
Let $A'\rightarrow A$ be a small extension in $\mathcal{C}_E$,
we denote the kernel by $I\subseteq A'$.
Let $W_A:=(W_{e,A}, W^+_{\mathrm{dR},A})$ be a deformation of $W$
over $A$. Then it suffices to show that there exists an $A'$-$B$-pair $W_{A'}$
such that $W_{A'}\otimes_{A'}A\isom W_A$.
We fix a basis of $W_{e,A}$ as a $B_e\otimes_{\mathbb{Q}_p}A$-module.
By this choice, from the $G_K$-action on $W_{e,A}$,
we get a continuous one cocycle $\rho_e:G_K\rightarrow \mathrm{GL}_n(B_e\otimes_{\mathbb{Q}_p}A)$.
In the same way, if we fix a basis of $W^+_{\mathrm{dR},A}$ as a $B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A$-module,
we get a continuous cocycle $\rho_{\mathrm{dR}}:G_K\rightarrow \mathrm{GL}_n(B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A)$.
From the canonical isomorphism $W_{e,A}\otimes_{B_e}B_{\mathrm{dR}}\isom W^+_{\mathrm{dR},A}
\otimes_{B^+_{\mathrm{dR}}}
B_{\mathrm{dR}}$, we get a matrix $P\in\mathrm{GL}_n(B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A)$ such that
$P\rho_e(g)g(P)^{-1}=\rho_{\mathrm{dR}}(g)$ for any $g\in G_K$.
We fix an $E$-linear section $s:A\rightarrow A'$ of $A'\rightarrow A$ and fix a lifting
$\tilde{P}\in \mathrm{GL}_n(B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A')$ of $P$. From this,
we get continuous liftings $\tilde{\rho}_e:=s\circ \rho_e:G_K\rightarrow \mathrm{GL}_n(B_e\otimes_{\mathbb{Q}_p}A')$ of $\rho_e$
and $\tilde{\rho}_{\mathrm{dR}}:=s\circ \rho_{\mathrm{dR}}:G_K\rightarrow \mathrm{GL}_n(B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A')$ of
$\rho_{\mathrm{dR}}$.
By using these liftings, we define $c_e:G_K\times G_K\rightarrow I\otimes_E \mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}E}(W_e, W_e)$
by $\tilde{\rho}_e(g_1g_2)g_1(\tilde{\rho}_e(g_2))^{-1}\tilde{\rho}_e(g_1)^{-1}=1+c_{e}(g_1,g_2)\in
1+I \otimes_{A'}\mathrm{M}_n(B_e\otimes_{\mathbb{Q}_p}A')=1+I\otimes_{E}\mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}E}
(W_e,W_e)$ for any $g_1,g_2\in G_K$. In the same way, define $c_{\mathrm{dR}}:G_K\times G_K \rightarrow I\otimes_E\mathrm{Hom}_{B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}
(W^+_{\mathrm{dR}}, W^+_{\mathrm{dR}})$ by $\tilde{\rho}_{\mathrm{dR}}(g_1g_2)g_1(\tilde{\rho}_{\mathrm{dR}}(g_2))^{-1}\tilde{\rho}_{\mathrm{dR}}(g_1)^{-1}
=1+c_{\mathrm{dR}}(g_1,g_2)$. We define $c:G_K\rightarrow I\otimes_E\mathrm{Hom}_{B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}(W_{\mathrm{dR}}, W_{\mathrm{dR}})$
such that $\tilde{P}\tilde{\rho}_e(g)g(\tilde{P})^{-1}\tilde{\rho}_{\mathrm{dR}}(g)^{-1}=1 + c(g)\in 1+I\otimes_E\mathrm{Hom}_{B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}(W_{\mathrm{dR}}, W_{\mathrm{dR}})$. Then $c_e$ and $c_{\mathrm{dR}}$ are continuous two cocycles, i.e. satisfy $g_1c_{\ast}(g_2, g_3)-c_{\ast}(g_1g_2,g_3)+c_{\ast}(g_1, g_2g_3)
-c_{\ast}(g_1,g_2)=0$ for any $g_1,g_2,g_3\in G_K$ ($\ast=e, \mathrm{dR}$). Moreover, we can check that $c_e$ and $c_{\mathrm{dR}}$ and $c$ satisfy
$c_e(g_1,g_2)-c_{\mathrm{dR}}(g_1,g_2)=g_1(c(g_2))-c(g_1g_2)+c(g_1)$ for any
$g_1,g_2,g_3\in G_K$ (note that the isomorphism $\mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}E}(W_e, W_e)
\otimes_{B_e}B_{\mathrm{dR}}\isom \mathrm{Hom}_{B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}(W^+_{\mathrm{dR}},W^+_{\mathrm{dR}})\otimes_{B^+_{\mathrm{dR}}}B_{\mathrm{dR}}$ is given by $c\mapsto \bar{P}^{-1}c\bar{P}$, where $\bar{P}\in \mathrm{GL}_n(B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E)$
is the reduction of $P\in \mathrm{GL}_n(B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A)$). By definition of Galois cohomology
of $B$-pairs, these mean that $(c_e,c_{\mathrm{dR}}, c)$ defines an element $[(c_e,c_{\mathrm{dR}},c)]$in $I\otimes_E\mathrm{H}^2(G_K, \mathrm{ad}(W))$ . (Until here, we don't use the condition $\mathrm{H}^2(G_K,\mathrm{ad}(W))=0$ and we can show, in the standard way, that $[(c_e,c_{\mathrm{dR}},c)]$ doesn't depend on the choice of $s$ or $\tilde{P}$, i.e. depends only on $W_A$.)
Now, by the assumption, we have $\mathrm{H}^2(G_K,\mathrm{ad}(W))=0$, So, by definition, there exists $(f_e,f_{\mathrm{dR}}, f)
$ such that $f_e:G_K\rightarrow I\otimes_{E}\mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}E}(W_e, W_e)$ and
$f_{\mathrm{dR}}:G_K\rightarrow I\otimes_E\mathrm{Hom}_{B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}(W^+_{\mathrm{dR}}, W^+_{\mathrm{dR}})$ are continuous map and $f\in I\otimes_E\mathrm{Hom}_{B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}E}(W_{\mathrm{dR}}, W_{\mathrm{dR}})$
and these satisfy $c_e(g_1,g_2)=g_1f_e(g_2)-f_e(g_1g_2)+f_e(g_1)$ and $c_{\mathrm{dR}}(g_1,g_2)=g_1f_{\mathrm{dR}}(g_2)
-f_{\mathrm{dR}}(g_1g_2)+f_{\mathrm{dR}}(g_1)$ and $c(g_1)=f_{\mathrm{dR}}(g_1)-\bar{P}^{-1}f_{e}(g_1)\bar{P}+(g_1f-f)$
for any $g_1,g_2\in G_K$.
Then we define new liftings $\rho'_{e}:G_K\rightarrow \mathrm{GL}_n(B_e\otimes_{\mathbb{Q}_p}A')$ by
$\rho'_e(g):=(1+f_e(g))\tilde{\rho}_e(g)$, $\rho'_{\mathrm{dR}}(g):G_K\rightarrow \mathrm{GL}_n(B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A')$ by $\rho'_{\mathrm{dR}}(g):=(1+f_{\mathrm{dR}}(g))\tilde{\rho}_{\mathrm{dR}}(g)$ and
$P':=(1+f)\tilde{P}\in \mathrm{GL}_n(B_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A')$. Then we can check
that these satisfy $\rho'_e(g_1g_2)=\rho'_e(g_1)\rho'_{e}(g_2)$ and $\rho'_{\mathrm{dR}}(g_1g_2)
=\rho'_{\mathrm{dR}}(g_1)\rho'_{\mathrm{dR}}(g_2)$ and $P'\rho'_e(g_1)g(P')^{-1}=\rho'_{\mathrm{dR}}(g_1)$ for
any $g_1,g_2\in G_K$. By definition of $A'$-$B$-pair, this means that $(\rho'_e,\rho'_{\mathrm{dR}},P')$ defines
an $A'$-$B$-pair and ,by definition, this is a lift of $W_A$. We finish the proof of this proposition.
\end{proof}
\begin{corollary}\label{13}
Let $W$ be an $E$-$B$-pair of rank $n$.
If $\mathrm{End}_{G_K}(W)=E$ and $\mathrm{H}^2(G_K, \mathrm{ad}(W))=0$
and if we denote $d:=[K:\mathbb{Q}_p]n^2 +1$,
then the functor $D_W$ is pro-representable by $R_W$ such that $R_W\isom E[[T_1,\cdots,T_d]]$.
\end{corollary}
\begin{proof}
The existence and formally smoothness of $R_W$ follows from Proposition $\ref{9}$ and Proposition $\ref{12}$.
For dimension, by Theorem $\ref{h}$ and Lemma $\ref{10}$, we have
$\mathrm{dim}_ED_W(E[\varepsilon])=\mathrm{dim}_E\mathrm{H}^1(G_K, \mathrm{ad}(W))=[K:\mathbb{Q}_p]n^2
+\mathrm{dim}_E\mathrm{H}^0(G_K,\mathrm{ad}(W))+\mathrm{dim}_E\mathrm{H}^2(G_K,\mathrm{ad}(W))=[K:\mathbb{Q}_p]n^2
+1$.
\end{proof}
\subsection{Trianguline deformations of trianguline $B$-pairs}
In this subsection, we define the trianguline deformation functor for
split trianguline $E$-$B$-pairs and prove the pro-representability and formally smoothness
under some conditions and calculate the dimension of the universal
deformation ring of this functor. In $\S$ 2 of [Bel-Ch 09], Bella\"iche-Chenevier
proved all these in the case
of $K=\mathbb{Q}_p$ by using $(\varphi,\Gamma)$-modules over the Robba ring and by using Colmez's
theory of trianguline representations for $K=\mathbb{Q}_p$ ([Co08]).
We generalize their theory by using
$B$-pairs and the theory of trianguline representations for any $p$-adic fields ([Na09] or $\S$ 2).
First, we define the notion of split trianguline $A$-$B$-pairs as follows.
\begin{defn}
Let $W$ be a rank $n$ $A$-$B$-pairs. Then we say that $W$ is a split trianguline
$A$-$B$-pair if there exists a sequence of sub $A$-$B$-pairs
$\mathcal{T}: 0=W_0\subseteq W_1\subseteq W_2\subseteq\cdots \subseteq W_{n-1}
\subseteq W_n=W$ such that $W_i$ is saturated in $W_{i+1}$ and the
quotient $W_{i+1}/W_i$ is a rank one $A$-$B$-pair for any $0\leqq i\leqq
n-1$.
By Proposition $\ref{5}$, there exist continuous homomorphisms $\delta_i:K^{\times}\rightarrow A^{\times}$
such that $W_i/W_{i-1}\isom W(\delta_i)$ for any $1\leqq i\leqq n$.
We say that $\{\delta_i\}_{i=1}^n$ is the parameter of triangulation
$\mathcal{T}$.
\end{defn}
Next, we define the trianguline deformation functor.
Let $W$ be a rank $n$ split trianguline $E$-$B$-pair. We fix a triangulation
$\mathcal{T}: 0\subseteq W_1\subseteq\cdots\subseteq W_{n-1}\subseteq
W_n=W$ of $W$. Under this condition, we define the trianguline deformation
as follows.
\begin{defn}
Let $A$ be an object in $\mathcal{C}_E$.
We say that $(W_A,\iota, \mathcal{T}_A)$ is a trianguline deformation of $(W,\mathcal{T})$ over $A$
if $(W_A,\iota)$ is a deformation of $W$ over $A$ and $W_A$ is a split trianguline $A$-$B$-pair with
a triangulation $\mathcal{T}_A:0\subseteq W_{1,A}\subseteq \cdots \subseteq W_{n,A}=W_A$
such that $\iota(W_{i,A}\otimes_A E)=W_i$ for any $1\leqq i\leqq n$.
Let $(W_A,\iota,\mathcal{T}_A)$ and $(W'_A,\iota',\mathcal{T}'_A)$ be two trianguline
deformations of $(W,\mathcal{T})$ over $A$. Then we say that $(W_A,\iota,\mathcal{T}_A)$ and
$(W'_A,\iota',\mathcal{T}'_A)$ are equivalent if there exists an isomorphism of $A$-$B$-pairs
$f:W_A\isom W'_A$ satisfying $\iota=\iota'\circ \bar{f}$ and $f(W_{i,A})=W'_{i,A}$ for
any $1\leqq i\leqq n$.
\end{defn}
\begin{defn}
Let $W$ be a split trianguline $E$-$B$-pair with a triangulation $\mathcal{T}$.
Then we define the trianguline deformation functor $D_{W,\mathcal{T}}$
from the category $\mathcal{C}_E$ to the category of sets by defining, for any $A\in\mathcal{C}_E$,
$D_{W,\mathcal{T}}(A):=\{$ equivalent classes $(W_A,\iota,\mathcal{T}_A)$ of trianguline
deformation of $(W,\mathcal{T})$ over $A$ $\}$.
\end{defn}
By definition, we have a natural map of functors from $D_{W,\mathcal{T}}$ to $D_{W}$ by forgetting
triangulation, i.e. by defining $D_{W,\mathcal{T}}(A)\rightarrow D_{W}(A): [(W_A,\iota,\mathcal{T}_A)]
\mapsto [(W_A,\iota)]$. In general, by this map, $D_{W,\mathcal{T}}$ is not a sub functor of
$D_W$, i.e. a deformation $W_A$ can have many liftings of triangulation of $\mathcal{T}$.
So we give a sufficient condition for $D_{W,\mathcal{T}}$ to be a subfunctor of $D_W$.
Let $\{\delta_i\}_{i=1}^n$ be the parameter of triangulation $\mathcal{T}$.
\begin{lemma}\label{14}
If, for any $1\leqq i<j\leqq n$, we have $\delta_j/\delta_i\not= \prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}}$ for
any $\{k_{\sigma}\}_{\sigma\in\mathcal{P}}\in \prod_{\sigma\in\mathcal{P}} \mathbb{Z}_{\leqq 0}$, then
the functor $D_{W,\mathcal{T}}$ is a sub functor of $D_W$.
\end{lemma}
\begin{proof}
Let $W_A$ be a deformation of $W$ over $A$, let $0\subseteq W_{A ,1}\subseteq \cdots \subseteq W_{A ,n-1}\subseteq W_{A}$
and $0\subseteq W'_{A ,1}\subseteq \cdots \subseteq W'_{A, n-1}\subseteq W_{A}$ be two triangulations which are
lifts of $\mathcal{T}$, then it suffices to show that $W_{A, i}=W'_{A, i}$ for any $i$. Moreover, by induction, it suffices
to show $W_{A ,1}= W'_{A ,1}$. For proving this, first we consider $\mathrm{Hom}_{G_K}(W_{1,A}, W_A)$.
This is equal to $\mathrm{H}^0(G_K, W^{\vee}_{1, A}\otimes W_A)$. Because $\mathrm{H}^0(G_K, -)$ is left exact
and $\mathrm{H}^0(G_K, W(\delta))=0$ for any $\delta:K^{\times}\rightarrow E^{\times}$ such that
$\delta\not= \prod_{\sigma\in \mathcal{P}}\sigma(x)^{k_{\sigma}}$ for any $\{k_{\sigma}\}_{\sigma\in\mathcal{P}}
\in \prod_{\sigma\in\mathcal{P}}\mathbb{Z}_{\leqq 0}$ by Proposition $\ref{i}$, we have $\mathrm{H}^0(G_K, W_{1, A}^{\vee}\otimes
(W_{i+1, A}/W_{i, A}))\allowbreak =\mathrm{H}^0(G_K, W_{1, A}^{\vee}\otimes( W'_{i+1, A}/W'_{i, A}))=0$ for any $i\geqq 2$.
So we get $\mathrm{Hom}_{G_K}(W_{1, A}, W_{1,A})\allowbreak =\mathrm{Hom}_{G_K}(W_{1,A}, \allowbreak W_A)=\mathrm{Hom}_{G_K}(W_{1, A},
W'_{1,A})$. This means that the given inclusion $W_{1, A}\hookrightarrow W_A$ factors through $W'_{1, A}\hookrightarrow W_A$.
By symmetry, the inclusion $W'_{1,A}\hookrightarrow W_A$ factors through $W_{1, A}\hookrightarrow W_{A}$.
So we get $W_{1,A}=W'_{1, A'}$.
\end{proof}
\begin{prop}\label{15}
Let $W$ be a trianguline representations with a triangulation $\mathcal{T}$
such that the parameter $\{\delta_i\}_{\i=1}^n$ satisfies, for any $1\leqq i<j\leqq n$, $\delta_j/\delta_i\not=\prod_{\sigma\in \mathcal{P}}
\sigma(x)^{k_{\sigma}}$ for any $\{k_{\sigma}\}_{\sigma\in\mathcal{P}}\in \prod_{\sigma\in\mathcal{P}}\mathbb{Z}_{\leqq 0}$.
Then the natural map of functors $D_{W,\mathcal{T}}\rightarrow D_{W}$ is relatively representable.
\end{prop}
\begin{proof}
By $\S 23$ of [Ma97], it suffices to check that the map $D_{W,\mathcal{T}}\rightarrow D_W$ satisfies the fallowing three conditions (1), (2), (3).
(1) :For any map $A\rightarrow A'$ in $\mathcal{C}_E$ and $W_{A}\in D_{W,\mathcal{T}}(A)$, then
$W_A\otimes_A A'\in D_{W,\mathcal{T}}(A')$. (2): For any maps $A'\rightarrow A$ and $A"\rightarrow A$ in $\mathcal{C}_E$
and $W^{'"}\in D_W(A'\times_A A")$, if $W^{'"}\otimes_{A'\times_A A"} A'\in D_{W,\mathcal{T}}(A')$ and
$W^{'"}\otimes_{A'\times_A A"}A"\in D_{W,\mathcal{T}}(A")$, then $W^{'"}\in D_{W,\mathcal{T}}(A'\times_A A")$.
(3): For any inclusion $A\hookrightarrow A'$ in $\mathcal{C}_E$ and $W_{A}\in D_W(A)$, if $W_A\otimes_A A'\in
D_{W,\mathcal{T}}(A')$, then $W_A\in D_{W,\mathcal{T}}(A)$.
The condition (1) is trivial. For (2), let $W^{'"}\in D_{W}(A'\times_A A")$ be such that $W_{A'}:=W^{'"}\otimes_{A'\times_A A"}A'
\in D_{W,\mathcal{T}}(A')$ and $W_{A"}:=W^{'"}\otimes_{A'\times_A A"}A"\in D_{W,\mathcal{T}}(A")$. We put $W_A:=W^{'"}\otimes_{A'\times_A A"}A$. Then as in proof of Proposition $\ref{9}$, we have $W^{'"}\isom W_{A'}\times_{W_A} W_{A"}$.
And, by Lemma $\ref{14}$, the triangulations of $W_A$ induced from $W_{A'}$ and $W_{A"}$ are same, so
these triangulations induce a triangulation of $W^{'"}\isom W_{A'}\times_{W_A} W_{A"}$, i.e. $W^{'"}\in D_{W,\mathcal{T}}(A'\times_A A")$.
Finally, we prove the condition (3). Let $W\in D_W(A)$ and $A\hookrightarrow A'$ be an inclusion such that
$W_A\otimes_A A'\in D_{W,\mathcal{T}}(A')$. Let $0\subseteq W_{1,A'}\subseteq\cdots\subseteq W_{n-1,A'}\subseteq W_{A}
\otimes_A A'$ be a triangulation lifting of $\mathcal{T}$. By induction on rank of $W$, it suffices to show that there exists
a rank one sub $A$-$B$-pair $W_{1,A}\subseteq W_A$ such that $W_{1,A}\otimes_{A}A'=W_{1,A'}$ and that $W_A/W_{1,A}$ is an $A$-$B$-pair.
By Proposition $\ref{5}$, by twisting,
we may assume that $W_{1,A'}\isom (B_e\otimes_{\mathbb{Q}_p}A', B^+_{\mathrm{dR}}\otimes_{\mathbb{Q}_p}A')$ is
trivial $A'$-$B$-pair (so $\delta_1:K^{\times}\rightarrow E^{\times}$ is trivial homomorphism).
Then, by the same way as in the proof of Lemma $\ref{14}$, we have $A'\isom \mathrm{H}^0(G_K, W_{1,A'})=
\mathrm{H}^0(G_K, W_{A}\otimes_A A')$ and $E=\mathrm{H}^0(G_K, W_{1})=\mathrm{H}^0(G_K, W)$.
Then, by using these and the left exactness of $\mathrm{H}^0(G_K,-)$, we have
$\mathrm{length}(A')=\mathrm{length}(\mathrm{H}^0(G_K, W_{A'}))\leqq \mathrm{length}(\mathrm{H}^0(G_K, W_A))+
\mathrm{length}(\mathrm{H}^0(G_K, W_A\otimes(A'/A))\leqq \mathrm{length}(A)+\mathrm{length}(A'/A)
=\mathrm{length}(A')$. So we have $\mathrm{length}(\mathrm{H}^0(G_K, W_A))=\mathrm{length}(A)$.
From this, by similar argument, we can show the natural sequence
$0\rightarrow \mathrm{H}^0(G_K, W_A\otimes_A m_A)\rightarrow \mathrm{H}^0(G_K, W_A)
\rightarrow \mathrm{H}^0(G_K, W)\rightarrow 0$ is exact.
We take a lift $v\in \mathrm{H}^0(G_K, W_A)$ which maps to $1\in E\isom \mathrm{H}^0 (G_K, W)$
by the surjection $\mathrm{H}^0(G_K, W_A)\rightarrow \mathrm{H}^0(G_K, W)$.
Then we claim that $A\isom Av=\mathrm{H}^0(G_K, W_A)$. Because the surjection $\mathrm{H}^0(G_K, W_A)\rightarrow \mathrm{H}^0(G_K, W)$ factors through
$\mathrm{H}^0(G_K, W_A)\hookrightarrow \mathrm{H}^0(G_K, W_A\otimes_A A')\rightarrow
\mathrm{H}^0(G_K, W)$ and $\mathrm{H}^0(G_K, W_{A}\otimes_A A')\isom A'$, the image of $v$ in $\mathrm{H}^0(G_K,
W_A\otimes_A A')$ must be a generator of $\mathrm{H}^0(G_K, W_A\otimes_A A')$. So we have $A\isom Av\subseteq
\mathrm{H}^0(G_K, W_A)$. By comparing length, we get the claim. By using this $v\in \mathrm{H}^0(G_K, W_A)
=\mathrm{Hom}_{G_K}(B_A, W_A)$ (here $B_A$ is the trivial $A$-$B$-pair), we define a map $v:B_A\rightarrow W_A$.
Then $v\otimes \mathrm{id}_{A'}:B_A'\rightarrow W_A'$ induces an isomorphism $B_A'\isom W_{1,A'}$ because
$\mathrm{H}^0(G_K, W_{1,A'})=\mathrm{H}^0(G_K, W_A\otimes_A A')$. From this, $v$ is injective and if we put $W_{1,A}
\subseteq W_A$ the image of $v$, then $W_{1,A}\otimes_A A'=W_{1,A'}$. Finally, we need to prove
that the quotient $W_A/W_{1,A}$ is an $A$-$B$-pair. First, $W_A/W_{1,A}$ is an $E$-$B$-pair because
$W_{1,A}$ is saturated in $W_{A}\otimes_A A'$ as an $E$-$B$-pair, so also saturated in $W_A$. For proving $A$-flatness
of $W_{A}/W_{1,A}:=(W_{0, e}, W^+_{0,\mathrm{dR}})$, by definition of $v$,
we have that the natural map $W_{1,A}\otimes_A E\hookrightarrow W_{A}\otimes_A E$ is injective. From this, we get
$A$-flatness of $W_{0,e}$ and $W^+_{0,\mathrm{dR}}/tW^+_{0,\mathrm{dR}}$, so
$W_A/W_{1,A}$ is an $A$-$B$-pair. We finish to prove this proposition.
\end{proof}
\begin{corollary}\label{16}
Let $W$ be a trianguline $E$-$B$-pair with a triangulation $\mathcal{T}$ such that $\mathrm{End}_{G_K}(W)=E$
and the parameter $\{\delta_i\}_{i=1}^n$ of $\mathcal{T}$ satisfies, for any $1\leqq i<j\leqq n$,
$\delta_j/\delta_i\not= \prod_{\sigma\in \mathcal{P}}\sigma(x)^{k_{\sigma}}$ for any $\{k_{\sigma}\}_{\sigma\in \mathcal{P}}
\in \prod_{\sigma\in\mathcal{P}}\mathbb{Z}_{\leqq 0}$.
Then the functor $D_{W,\mathcal{T}}$ is pro-representable by a quotient $R_{W,\mathcal{T}}$ of $R_W$.
\end{corollary}
\begin{proof} This follows from Proposition $\ref{9}$ and Proposition $\ref{15}$.
\end{proof}
Next, we prove the formally smoothness of the functor $D_{W,\mathcal{T}}$.
\begin{prop}\label{17}
Let $W$ be a trianguline $E$-$B$-pair of rank $n$ with triangulation $\mathcal{T}$ such that
whose parameter $\{\delta_i\}_{i=1}^{n}$ satisfies, for any $1\leqq i< j\leqq n$, $\delta_i/\delta_j\not=
|\mathrm{N}_{K/\mathbb{Q}_p}(x)|\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}} $ for any
$\{k_{\sigma}\}_{\sigma\in\mathcal{P}}\in \prod_{\sigma\in\mathcal{P}}\mathbb{Z}_{\geqq 1}$.
Then the functor $D_{W,\mathcal{T}}$ is formally smooth.
\end{prop}
\begin{proof}
We prove this proposition by induction of rank of $W$.
When $W$ is rank one, then we have $D_{W,\mathcal{T}}=D_W$ and
$\mathrm{ad}(W)=B_E$ is the trivial $E$-$B$-pair.
So $\mathrm{H}^2(G_K,\mathrm{ad}(W))=0$ by Proposition $\ref{i}$. So,
by Proposition $\ref{12}$, $D_{W,\mathcal{T}}$ is formally smooth in this case.
Let's assume that the proposition is proved for rank $n-1$ $E$-$B$-pairs,
let $W$ be a rank $n$ $E$-$B$-pair with triangulation
$\mathcal{T}:0\subseteq W_1\subseteq \cdots \subseteq W_{n-1}\subseteq W_n=W$
whose parameter $\{\delta_i\}_{i=1}^n$ satisfying the condition as above.
Let $A'\rightarrow A$ be a small extension in $\mathcal{C}_E$ and let
$W_A$ be a trianguline deformation of $(W,\mathcal{T})$ with triangulation
$\mathcal{T}_A:0\subseteq W_{1,A}\subseteq \cdots \subseteq W_{n-1,A}\subseteq W_{n,A}=W_A$
which is a lift of $\mathcal{T}$. It suffices to prove that there exists a split trianguline $A'$-$B$-pair
$W_{A'}$ with a triangulation $0\subseteq W_{1,A'}\subseteq \cdots \subseteq W_{n-1.A'}\subseteq
W_{n,A'}=W_{A'}$ lifting of $W_A$ and $\mathcal{T}_A$. We take a lifting as follows.
First, because $W_{n-1}$ is a trianguline $E$-$B$-pair of rank $n-1$ satisfying the conditions as above,
so by induction hypothesis, there exists a rank $n-1$ trianguline $A'$-$B$-pair $W_{n-1,A'}$
with a triangulation $0\subseteq W_{1,A'}\subseteq \cdots \subseteq W_{n-2,A'}\subseteq W_{n-1,A'}$
which is a lift of $W_{n-1,A}$ and $0\subseteq W_{1,A}\subseteq\cdots\subseteq W_{n-1,A}$.
If we put $\mathrm{gr}_nW_A:=W_{A}/W_{n-1,A}$, then, by rank one case and by Proposition $\ref{5}$, there exists
a continuous homomorphism $\delta_{n,A'}:K^{\times}\rightarrow A^{' \times}$ such that the rank one $A'$-$B$-pair
$W(\delta_{n,A'})$ satisfies $W(\delta_{n,A'})\otimes_{A'}A=W(\delta_{n,A})\isom \mathrm{gr}_nW_A$.
(where $\delta_{n,A}:K^{\times}\rightarrow A^{\times}$ is the reduction of $\delta_{n,A'}$.)
Then $[W_A]$ is an element in $\mathrm{Ext}^1(W(\delta_{n,A}), W_{n-1,A})\isom \mathrm{H}^1(G_K,
W_{n-1,A}(\delta_{n,A}^{-1}))$. If we take the long exact sequence associated
to
\begin{equation*}
0\rightarrow I\otimes_E W_{n-1}(\delta_n^{-1})\rightarrow W_{n-1,A'}(\delta^{-1}_{n,A'})\rightarrow
W_{n-1,A}(\delta^{-1}_{n,A})\rightarrow 0
\end{equation*}
(where $I\subseteq A'$ is the kernel of $A'\rightarrow A$), then we get a long exact sequence
\[
\begin{array}{ll}
\cdots \rightarrow \mathrm{H}^1(G_K,W_{n-1,A'}(\delta^{-1}_{n,A'}))& \rightarrow \mathrm{H}^1(G_K, W_{n-1,A}(\delta^{-1}_{n,A}))\\
&\rightarrow I\otimes_E \mathrm{H}^2(G_K, W_{n-1}(\delta^{-1}_n))\rightarrow\cdots
\end{array}
\]
By assumption on $\{\delta_i\}_{i=1}^n$ and by Proposition $\ref{i}$, we have $\mathrm{H}^2(G_K, W_{n-1}(\delta^{-1}_n))=0$.
So we can take a $[W_{A'}]\in\mathrm{Ext}^1(W(\delta_{n,A'}),W_{n-1,A'})\isom\mathrm{H}^1(G_K, W_{n-1,A'}(\delta^{-1}_{n,A'}))$
which is a trianguline lift of $[W_A]$. This proves the proposition.
\end{proof}
Next, we calculate the dimension of $D_{W,\mathcal{T}}$. For this, we interpret
$D_{W,\mathcal{T}}(E[\varepsilon])$ in terms of Galois cohomology of $B$-pair as in Lemma $\ref{10}$.
Let $W$ be a trianguline $E$-$B$-pair with triangulation $\mathcal{T}:0\subseteq W_1\subseteq \cdots\subseteq
W_{n-1}\subseteq W_n=W$ whose parameter is $\{\delta_i\}_{i=1}^n$.
Then we define an $E$-$B$-pair $\mathrm{ad}_{\mathcal{T}}(W)$ by
$\mathrm{ad}_{\mathcal{T}}(W):=\{f\in\mathrm{ad}(W)| f(W_i)\subseteq W_i $ for any $1\leqq i\leqq n \}$.
\begin{lemma}\label{18}
Let $W$ be a trianguline $E$-$B$-pair. Then there exists
a canonical bijection of sets $D_{W,\mathcal{T}}(E[\varepsilon])
\isom \mathrm{H}^1(G_K, \mathrm{ad}_{\mathcal{T}}(W))$.
In particular, if $D_{W,\mathcal{T}}$ has a canonical structure of $E$-vector apace
$($c.f. the condition $(2)$ in Shlessinger's criterion $\ref{8}$$)$, then this bijection is an $E$-linear isomorphism.
\end{lemma}
\begin{proof}
The construction of the map $D_{W,\mathcal{T}}(E[\varepsilon])\rightarrow
\mathrm{H}^1(G_K,\mathrm{ad}_{\mathcal{T}}(W))$ is same as in the proof of
Lemma $\ref{10}$. We put $\mathrm{ad}_{\mathcal{T}}(W):=(\mathrm{ad}_{\mathcal{T}}(W_e), \mathrm{ad}_{\mathcal{T}}(W^+_{\mathrm{dR}}))$. Let $W_{E[\varepsilon]}:=(W_{e,E[\varepsilon]},W^+_{\mathrm{dR},E[\varepsilon]})$ be a trianguline deformation
of $(W,\mathcal{T})$ over $E[\varepsilon]$ whose triangulation $\mathcal{T}_{E[\epsilon]}$ is $0\subseteq W_{1,E[\varepsilon]},\subseteq\cdots\subseteq
W_{n-1,E[\varepsilon]}\subseteq W_{n,E[\varepsilon]}=W_{E[\varepsilon]}$. Then we can take a splitting
$W_{e,E[\varepsilon]}=W_e e_1\oplus W_ee_2$ as a filtered $B_e\otimes_{\mathbb{Q}_p}E$-module such that
$W_e e_1=\varepsilon W_{e,E[\varepsilon]}$ and the natural map $W_ee_2\hookrightarrow W_{e,E[\varepsilon]}
\rightarrow W_{e,E[\varepsilon]}/\varepsilon W_{e,E[\varepsilon]}\isom W_e$ is such that $ye_2\mapsto y$ for any $y\in W_e$.
Then, if we define $c_e:G_K\rightarrow \mathrm{Hom}_{B_e\otimes_{\mathbb{Q}_p}E}(W_e, W_e)$ as in the proof
of Lemma $\ref{10}$, we can check that the image of $c_e$ is contained in $\mathrm{ad}_{\mathcal{T}}(W_e)$. In the same way,
we can define $c_{\mathrm{dR}}:G_K\rightarrow \mathrm{ad}_{\mathcal{T}}(W^+_{\mathrm{dR}})$ from a
filtered splitting $W^+_{\mathrm{dR},E[\varepsilon]}=W^+_{\mathrm{dR}}e_1\oplus W^+_{\mathrm{dR}}e'_2$.
Moreover, we can define $c\in \mathrm{ad}_{\mathcal{T}}(W_{\mathrm{dR}})$ by $ye_2=c(y)e_1+ye'_2$ for any $y\in W_{\mathrm{dR}}$.
Then the map $D_{W,\mathcal{T}}(E[\varepsilon])\rightarrow \mathrm{H}^1(G_K, \mathrm{ad}_{\mathcal{T}}(W)):[(W_{E[\varepsilon]},
\mathcal{T}_{E[\varepsilon]})]\mapsto [(c_e,c_{\mathrm{dR}},c)]$ defines an bijection and, when $D_{W,\mathcal{T}}(E[\varepsilon])$
has a canonical $E$-vector space structure, this is an $E$-linear isomorphism.
\end{proof}
We calculate the dimension of $R_{W,\mathcal{T}}$.
\begin{prop}\label{19}
Let $W$ be a trianguline $E$-$B$-pair of rank $n$ such that $\mathrm{End}_{G_K}(W)\allowbreak =E$ whose triangulation
$\mathcal{T}:0\subseteq W_1\subseteq \cdots\subseteq W_{n-1}\subseteq W_n=W$ satisfies
the following. The parameter $\{\delta_i\}_{i=1}^n$ satisfies, for any $1\leqq i < j\leqq n$, $\delta_j/\delta_i\not= \prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}}$
for any $\{k_{\sigma}\}_{\sigma\in\mathcal{P}}\in \prod_{\sigma\in\mathcal{P}}\mathbb{Z}_{\leqq 0}$ and
$\delta_i/\delta_j\not= |\mathrm{N}_{K/\mathbb{Q}_p}(x)|\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}}$
for any $\{k_{\sigma}\}_{\sigma\in\mathcal{P}}\in \prod_{\sigma\in\mathcal{P}}\mathbb{Z}_{\geqq 1}$.
Then the universal trianguline deformation ring $R_{W,\mathcal{T}}$ is a quotient ring of $R_W$ such
that $R_{W,\mathcal{T}}\isom E[[T_1,\cdots, T_{d_n}]]$ where $d_n:=\frac{n(n+1)}{2}[K:\mathbb{Q}_p]+1$.
\end{prop}
\begin{proof}
From Proposition $\ref{15}$ and Proposition $\ref{17}$ and Lemma $\ref{18}$, it suffices to show that $\mathrm{dim}_E\mathrm{H}^1(G_K, \mathrm{ad}_{\mathcal{T}}(W))=d_n$.
We prove this by induction on rank $n$ of $W$.
When $n=1$, then $\mathrm{ad}_{\mathcal{T}}(W)=\mathrm{ad}(W)=B_E$, so the proposition
follows from Proposition $\ref{i}$. Let $W$ be a rank $n$ trianguline $E$-$B$-pair as above, let $\mathcal{T}_{n-1}:0\subseteq
W_{1}\subseteq \cdots \subseteq W_{n-2}\subseteq W_{n-1}$ be the triangulation of $W_{n-1}(\subseteq W)$.
Then , by definition of $\mathrm{ad}_{\mathcal{T}}(W)$, for any $f\in \mathrm{ad}_{\mathcal{T}}(W)$, the restriction of $f$
to $W_{n-1}$ is an element of $\mathrm{ad}_{\mathcal{T}_{n-1}}(W_{n-1})$ and this defines a short exact sequence
of $E$-$B$-pair,
\begin{equation*}
0\rightarrow \mathrm{Hom}(W(\delta_n), W)\rightarrow \mathrm{ad}_{\mathcal{T}}(W)\rightarrow
\mathrm{ad}_{\mathcal{T}_{n-1}}(W_{n-1})\rightarrow 0
\end{equation*}
From this, we have $\mathrm{rank}(\mathrm{ad}_{\mathcal{T}}(W))=\mathrm{rank}(\mathrm{ad}_{\mathcal{T}_{n-1}}(W_{n-1}))
+n=1+2+\cdots +n=\frac{n(n+1)}{2}$. From this and Theorem $\ref{h}$,
it suffices to show that $\mathrm{H}^0(G_K,\mathrm{ad}_{\mathcal{T}}(W))=E$ and $\mathrm{H}^2(G_K,\mathrm{ad}_{\mathcal{T}}(W))=0$
. For $H^0$, this follows from $E\subseteq \mathrm{H}^0(G_K, \mathrm{ad}_{\mathcal{T}}(W))\subseteq \mathrm{H}^0(G_K,
\mathrm{ad}(W))=E$. We prove $\mathrm{H}^2(G_K, \mathrm{ad}_{\mathcal{T}}(W))=0$ by induction of rank of $W$.
When $n=1$, this follows from Proposition $\ref{i}$. When $W$ is rank $n$, then from the above short exact sequence, we have the following
long exact sequence ,
\begin{equation*}
\cdots \rightarrow \mathrm{H}^2(G_K, \mathrm{Hom}(W(\delta_n), W))\rightarrow \mathrm{H}^2(G_K, \mathrm{ad}_{\mathcal{T}}(W))
\rightarrow \mathrm{H}^2(G_K, \mathrm{ad}_{\mathcal{T}_{n-1}}(W_{n-1}))\rightarrow 0
\end{equation*}
Then, by Proposition $\ref{i}$ and by the assumption on $\{\delta_i\}_{i=1}^n$, we have $\mathrm{H}^2(G_K, \mathrm{Hom}\allowbreak (W(\delta_n), W))=0$.
So, by induction hypothesis, we have $\mathrm{H}^2(G_K, \mathrm{ad}_{\mathcal{T}}(W))=0$. We finish the proof of this
proposition.
\end{proof}
\section{Deformations of benign $B$-pairs}
In this final section, we study a very important class, called benign $B$-pairs,
of trianguline $E$-$B$-pairs, which satisfies very good properties for trianguline
deformations and plays a crucial role in Zariski density problem of modular Galois (or crystalline)
representations in some deformation spaces of global (or local) $p$-adic representations.
This class was defined by Kisin in the case of $K=\mathbb{Q}_p$ and rank $W$ is 2 in [Ki03] and [Ki08b].
He studied some deformation theoretic properties of this class in [Ki03] and used these in a crucial way in his
proof of Zariski density of two dimensional crystalline representations of $G_{\mathbb{Q}_p}$.
Bella\"iche-Chenevier ([Bel-Ch09]) and Chenevier ([Ch08],[Ch09]) generalized this class for
higher dimensional case and $K=\mathbb{Q}_p$ case, they defined and studied in detail
crystalline representations with non-critical refinements. In particular, Chenevier ([Ch08],[Ch09])
discovered and proved a cruicially important preoperty concerning to tangent spaces of these deformation
rings. (This was discovered by Kisin ([Ki08b]) implicitly when $K=\mathbb{Q}_p$ and rank 2 case.)
In fact, by using this property, Chenevier ([Ch08],[Ch09]) proved many types of theorems concerning
Zariski density of modular Galois
representations in some deformation spaces of global $p$-adic representations.
The aim of this section is to generalize this Chenevier's theorem to any $K$-case.
\subsection{Benign $B$-pairs}
Let $P(X)\in \mathcal{O}_K[X]$ be a polynomial such that $P(X)\equiv \pi_K X$ (mod deg $2$)
and that $P(X)\equiv X^q$ (mod $\pi_K$), where $q:=p^f$ and $f:=[K_0:\mathbb{Q}_p]$. Then we take the Lubin-Tate's formal group
$\mathcal{F}$ of $K$ such that $[\pi_K]=P(X)$, where $[-]:\mathcal{O}_K\isom \mathrm{End}(\mathcal{F})$. We put $K_n$ is the Galois extension of $K$
generated by $[\pi_K^n]$-torsion points of $\mathcal{F}$ for any $n$, then we have a canonical isomorphism $\chi_{\mathrm{LT},n}:\mathrm{Gal}(K_n/K)\isom( \mathcal{O}_K^{\times}/
\pi^n\mathcal{O}_K)^{\times}$. We put $K_{\mathrm{LT}}:=\cup_{n=1}^{\infty}K_n$ and we put $G_n:=\mathrm{Gal}(K_n/K)$.
In [Ki08b] or [Ch09] etc, benign representation is defined as a special class of crystalline representations.
But, as we show in the sequel, we can easily generalize the main theorem to some potentially crystalline representations.
So, before defining benign representations, we first define the following class of potentially crystalline representations.
\begin{defn}
Let $W$ be an $E$-$B$pair. Then we say that $W$ is crystabelline if $W|_{G_{L}}$ is a crystalline $E$-$B$-pair of
$G_{L}$ for a finite abel extension $L$ of $K$.
\end{defn}
\begin{rem}
Because a finite abel extension $L$ of $K$ is contained in $K_mL'$, where $L'$ is a finite unramified extension of $K$, by using Hilbert 90, we can easily show that $W$ is crystabelline if and only if
$W|_{G_{K_m}}$ is crystalline for some $m\geqq 0$.
\end{rem}
Let $W$ be a crystabelline $E$-$B$-pair of rank $n$ such that $W|_{G_{K_m}}$ is crystalline for some $m$.
Let $D^{K_m}_{\mathrm{cris}}(W):=(B_{\mathrm{cris}}\otimes_{\mathbb{Q}_p}W)^{G_{K_m}}$
be the rank $n$ $E$-filtered ($\varphi,G_m$)-module over $K$. Then, because $K_m$ is totally ramified over $K$, this is
a free $K_0\otimes_{\mathbb{Q}_p}E$-module of rank $n$.
We take an embedding $\sigma:K_0\hookrightarrow \bar{E}$, this defines a map
$\sigma:K_0\otimes_{\mathbb{Q}_p}E\rightarrow \bar{E}:x\otimes y\rightarrow \sigma(x)y$.
From this, we can define the $\sigma$-component $D^{K_m}_{\mathrm{cris}}(W)_{\sigma}:=D^{K_m}_{\mathrm{cris}}(W)
\otimes_{K_0\otimes_{\mathbb{Q}_p}E, \sigma}\bar{E}$, this has an $\bar{E}$-linear
$\varphi^f$-action and an $\bar{E}$-linear $G_m$-action. Let $\{\alpha_1,\cdots,\alpha_n\}$ be a solution in $\bar{E}$
(with multiplicities) of $\mathrm{det}_{\bar{E}}(T\cdot \mathrm{id}-\varphi ^f |_{D^{K_m}_{\mathrm{cris}}
(W)_{\sigma}})\in \bar{E}[T]$. Then, because $\varphi^f$ and $G_m$-action commute, so any generalized $\varphi^f$-eigenvector spaces
of $D^{K_m}_{\mathrm{cris}}(W)_{\sigma}$ are preserved by the action of $G_m$.
So we can take an $\bar{E}$-basis $e_{1,\sigma},\cdots,e_{n,\sigma}$ of $D^{K_m}_{\mathrm{cris}}(W)_{\sigma}$ such that, for any $i$, $e_{i,\sigma}$ is a generalized eigenvector of
$\varphi^f$ with generalized eigenvalue $\alpha_i\in \bar{E}^{\times}$ and $G_m$ acts on $e_{i,\sigma}$ by a homomorphism $\tilde{\delta}_i:G_m\rightarrow \bar{E}^{\times}$. We change numbering of $\{\alpha_1,\cdots,\alpha_n\}$ so that $e_{1,\sigma}, e_{2,\sigma},\cdots,e_{n,\sigma}$ gives a $\varphi^f$-Jordan decomposition of $D^{K_m}_{\mathrm{cris}}(W)_{\sigma}$ by this order.
Because $\{\sigma, \varphi^{-1}\sigma,\cdots, \varphi^{-(f-1)}\sigma\}=\mathrm{Hom}_{\mathbb{Q}_p}(K_0, \bar{E})$ and
$\varphi^i:D^{K_m}_{\mathrm{cris}}(W)_{\sigma}\isom D^{K_m}_{\mathrm{cris}}(W)_{\varphi^{-i}\sigma}:x\otimes y\mapsto \varphi^i(x)\otimes y$ for any $x\in D^{K_m}_{\mathrm{cris}}(W)$ and $y\in \bar{E}$, is an $\bar{E}[\varphi^f,G_m]$-isomorphism,
the set $\{\alpha_1,\cdots,\alpha_n\}$ doesn't depend on the choice of $\sigma:K_0\hookrightarrow \bar{E}$.
Then, if we put $e_i:=e_{i,\sigma}+\varphi(e_{i,\sigma})+\cdots +\varphi^{f-1}(e_{i,\sigma})\in D^{K_m}_{\mathrm{cris}}(W)\otimes_{E}\bar{E}$, we have $D^{K_m}_{\mathrm{cris}}(W)\otimes_E\bar{E}=K_0\otimes_{\mathbb{Q}_p}\bar{E}e_1\oplus\cdots \oplus K_0\otimes_E\bar{E}e_n$ such that the subspace $K_0\otimes_{\mathbb{Q}_p}\bar{E}e_1\oplus\cdots\oplus K_0\otimes_{\mathbb{Q}_p}\bar{E}e_i$ is preserved by
$\varphi$ and $G_m$-action for any $i$. Moreover, if we take a sufficiently large finite extension $E'$ of $E$, then we can
take $e_i\in D^{K_m}_{\mathrm{cris}}(W)\otimes_E E'$ such that $D^{K_m}_{\mathrm{cris}}(W)\otimes_E E'=
K_0\otimes_{\mathbb{Q}_p}E'e_1\oplus \cdots\oplus K_0\otimes_{\mathbb{Q}_p}E'e_n$ and $\alpha_i\in E'$ for any $i$ and $\tilde{\delta}_i:G_m\rightarrow E^{'\times}$.
By using these arguments, first we study a relation between crystabelline $E$-$B$-pairs and trianguline $E$-$B$-pairs.
\begin{lemma}\label{20.5}
Let $W$ be an $E$-$B$-pair of rank $n$. Then the following conditions are equivalent.
\begin{itemize}
\item[(1)]$W$ is crystabelline.
\item[(2)]$W$ is trianguline $($ i.e, there exists a finite extension $E'$ of $E$ such that $W\otimes_E E'$ is a
split trianguline $E'$-$B$-pair$)$ and potentially crystalline.
\end{itemize}
\end{lemma}
\begin{proof}
First we assume that $W$ is crystabelline. Then, by the above argument, for a sufficiently large finite extension
$E'$ of $E$, we have $D^{K_m}_{\mathrm{cris}}(W)\otimes_E E'=K_0\otimes_{\mathbb{Q}_p}E'e_1\oplus
\cdots \oplus K_0\otimes_{\mathbb{Q}_p}E'e_n$ as above. Then, for any $i$, $K_0\otimes_{\mathbb{Q}_p}E'e_1
\oplus \cdots\oplus K_0\otimes_{\mathbb{Q}_p}E'e_i$ is a sub $E'$-filtered ($\varphi,G_m$)-module
of $D^{K_m}_{\mathrm{cris}}(W\otimes_EE')$. So, by Theorem $\ref{e}$, $W\otimes_E E'$ is
split trianguline and potentially crystalline.
Next we assume that $W$ is trianguline and potentially crystalline. By extending coefficient, we may assume that $W$ is split trianguline.
We take a triangulation $0\subseteq W_1\subseteq \cdots \subseteq W_n=W$ of $W$.
Then, because sub objects and quotients of crystalline $B$-pairs are again crystalline, $W_i$ and $W_i/W_{i-1}$ are
all potentially crystalline. Then, by Lemma 4.1 of [Na09], for sufficiently large $m$, $W_i/W_{i-1}|_{G_m}$ is crystalline. Then we claim that $W|_{G_m}$ is also crystalline.
We prove this claim by induction on rank $n$ of $W$. When $n=1$, this is trivial.
We assume that the claim is proved for $n-1$ case. Then $W_{n-1}|_{G_m}$ is crystalline.
If we put $W/W_{n-1}\isom W(\delta_n)$, we have $[W]\in \mathrm{H}^1(G_K, W_{n-1}(\delta_n^{-1}))$. By assumption,
there exists a finite Galois extension $L$ of $K_m$ such that $[W]$ is contained in
$\mathrm{Ker}(\mathrm{H}^1(G_K, W_{n-1}(\delta_n^{-1}))\rightarrow \mathrm{H}^1(G_L, B_{\mathrm{cris}}\otimes_{B_e}
(W_{n-1}(\delta_n^{-1}))_e))$. Then it suffices to prove that the natural map $\mathrm{H}^1(G_{K_m}, B_{\mathrm{cris}}\otimes_{B_e}(W_{n-1}(\delta_n^{-1}))_e)\rightarrow \mathrm{H}^1(G_L, B_{\mathrm{cris}}\otimes_{B_e}(W_{n-1}(\delta_n^{-1}))_e)$ is injective. By inflation
restriction sequence, the kernel of this map is $\mathrm{H}^1(\mathrm{Gal}(L/K_m), D^L_{\mathrm{cris}}(W_{n-1}(\delta_n^{-1})))=0$.
So $W|_{G_m}$ is crystalline, i.e. $W$ is crystabelline.
\end{proof}
From here, we consider a crystabelline $E$-$B$-pair $W$ such that
$D^{K_m}_{\mathrm{cris}}(W)\isom K_0\otimes_{\mathbb{Q}_p}Ee_1\oplus \cdots \oplus
K_0\otimes_{\mathbb{Q}_p}Ee_n$ such that
$K_0\otimes_{\mathbb{Q}_p}Ee_i$ is preserved by ($\varphi,G_m$) and $\varphi^f(e_i)=\alpha_ie_i$ for
some $\alpha_i\in E^{\times}$ such that $\alpha_i\not= \alpha_j$ for any $i\not= j$.
Let $\mathfrak{S}_n$ be the $n$-th permutation group.
Then, for any $\tau\in \mathfrak{S}_n$, there is a filtration by $E$-filtered ($\varphi,G_m$)-modules on $D^{K_m}_{\mathrm{cris}}(W)$ ,
$\mathcal{F}_{\tau}:0\subseteq F_{\tau,1}\subseteq \cdots\subseteq F_{\tau, n-1}
\subseteq F_{\tau,n}=D^{K_m}_{\mathrm{cris}}(W)$ such that
$F_{\tau,i}:=K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(1)}\oplus \cdots\oplus K_0\otimes_{\mathbb{Q}_p}Ee_{\tau(i)}$ for any $1\leqq i \leqq n$
( the filtration on $F_{\tau,i}$ is induced from $D^{K_m}_{\mathrm{cris}}(W)$).
We put $\mathrm{gr}_{\tau,i}D^{K_m}_{\mathrm{cris}}(W):=F_{\tau,i}/F_{\tau,i-1}$ for any
$1\leqq i\leqq n$. Then, by Theorem $\ref{e}$, there exists a
filtration $\mathcal{T}_{\tau}:0\subseteq W_{\tau,1}\subseteq\cdots \subseteq W_{\tau,n-1}\subseteq W_{\tau,n}=W$
such that $W_{\tau,i}|_{G_m}$ is crystalline and $D^{K_m}_{\mathrm{cris}}(W_{\tau,i})=F_{\tau,i}$.
Then $W_{\tau,i}/W_{\tau,i-1}$ is a rank one crystabelline $E$-$B$-pair such that
$D^{K_m}_{\mathrm{cris}}(W_{\tau,i}/W_{\tau,i-1})\isom \mathrm{gr}_{\tau,i}D^{K_m}_{\mathrm{cris}}(W)$. So, by Lemma 4.1 of [Na09] and by its proof,
there exists $\{k_{(\tau,i),\sigma}\}_{\sigma\in\mathcal{P}}\in\prod_{\sigma\in \mathcal{P}}\mathbb{Z}$ and
homomorphisms $\tilde{\delta}_{i}:K^{\times}\rightarrow E^{\times}$ satisfying $\tilde{\delta}_{i}|_{1+\pi_K^m\mathcal{O}_K}=1$ and $\tilde{\delta}_{i}(\pi_K)=1$,
such that
$W_{\tau,i}/W_{\tau,i-1}\isom W(\delta_{\alpha_{\tau(i)}}\tilde{\delta}_{\tau(i)}\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{(\tau,i),\sigma}})$
for any $1\leqq i\leqq n$, where $\delta_{\alpha_i}:K^{\times}\rightarrow E^{\times}$ is the homomorphism such that
$\delta_{\alpha_i}|_{\mathcal{O}_K^{\times}}=1$ and $\delta_{\alpha_i}(\pi_K)=\alpha_i$. Then, for any $\sigma\in\mathcal{P}$, the set $\{k_{(\tau,1),\sigma},k_{(\tau,2),\sigma},\cdots,k_{(\tau,n),\sigma}\}$ is independent of $\tau\in \mathfrak{S}_n$ because these numbers are
$\sigma$-part of Hodge-Tate weights of $W$. We write this set (with multiplicities) by $\{k_{1,\sigma},k_{2,\sigma},\cdots, k_{n,\sigma}\}$
such that $k_{1,\sigma}\geqq k_{2,\sigma}\geqq\cdots\geqq k_{n,\sigma}$ for any $\sigma\in\mathcal{P}$. Under these notations, we define the notion of benign $E$-$B$-pair as follows.
\begin{defn}\label{26}
Let $W$ be a rank $n$ crystabelline $E$-$B$-pair as above. Then
we say that $W$ is a benign $E$-$B$-pair if the following conditions hold:
\begin{itemize}
\item[(1)] For any $i\not= j$, we have $\alpha_i/\alpha_j\not=1, p^f, p^{-f}$.
\item[(2)] For any $\sigma\in\mathcal{P}$, we have $0=k_{1,\sigma}>k_{2,\sigma}>\cdots >k_{n-1,\sigma}>k_{n,\sigma}$.
\item[(3)] For any $\tau\in \mathfrak{S}_n$ and $\sigma\in \mathcal{P}$, we have
$k_{(\tau,i),\sigma}=k_{i,\sigma}$ for any $1\leqq i\leqq n$.
\end{itemize}
We say that $W$ is twisted benign if there exists a benign $E$-$B$-pair $W'$ and
$\{k_{\sigma}\}_{\sigma\in\mathcal{P}}\in \prod_{\sigma\in \mathcal{P}}\mathbb{Z}$
such that $W\isom W'\otimes W(\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{\sigma}})$.
(i.e. we drop the condition $k_{1,\sigma}=0$ in (2)).
\end{defn}
\begin{rem}
By definition, if $W$ is a twisted benign, then we have $W_{\tau,i}/W_{\tau,i-1}
\isom W(\delta_{\alpha_{\tau(i)}}\tilde{\delta}_{\tau(i)}\prod_{\sigma\in\mathcal{P}}\sigma(x)^{k_{i,\sigma}})$
for any $\tau\in\mathfrak{S}_n$ and $1\leqq i\leqq n$.
\end{rem}
\begin{rem}
When $n=2$ and $K=\mathbb{Q}_p$ and $W$ is crystalline, this definition is same as that of Kisin in [Ki08b].
When $n=2$ and $K=\mathbb{Q}_p$, the condition (3) is induced from the other weaker condition that $\mathrm{End}_{E[G_K]}(V)=E$.
For the tangent space theorem (so for the application to Zariski density), the author thinks that
this condition (3) is the most important and subtle condition to verify.
But, in general case, i.e. $n\geqq 3$ and $K=\mathbb{Q}_p$ case, or $n\geqq 2$ and $K\not=\mathbb{Q}_p$ case, it seems difficult to induce
the condition (3) from some other weaker conditions. In [Ch09], Chenevier calls the condition (3) to be
"generic".
\end{rem}
\begin{lemma}\label{21}
Let $W$ be a twisted benign $E$-$B$-pair.
If $W_1$ is a saturated sub $E$-$B$-pair of $W$, then
$W_1$ and $W/W_1$ are also twisted benign $E$-$B$-pairs.
\end{lemma}
\begin{proof}
This follows from the definition and the fact that
sub objects and quotients of crystabelline $E$-$B$-pairs
are crystabelline $E$-$B$-pairs.
\end{proof}
\subsection{Deformations of benign $B$-pairs}
\begin{lemma}\label{22}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$.
Then $W$ satisfies $\mathrm{End}_{G_K}(W)=E$ and
for any $\tau\in \mathfrak{S}_n$, $\mathcal{T}_{\tau}:0\subseteq W_{\tau,1}\subseteq
\cdots\subseteq W_{\tau,n-1}\subseteq W_{\tau,n}=W$ satisfies all the condition
in Proposition $\ref{19}$.
\end{lemma}
\begin{proof}
We prove $\mathrm{End}_{G_K}(W)=E$, other statements are trivial from
(1) of Definition $\ref{26}$. We prove this by induction on $n$, the rank of $W$.
If $n=1$, $\mathrm{End}_{G_K}(W)=\mathrm{H}^0(G_K, B_E)=E$.
We assume that the lemma is proved for rank $n-1$ case, let
$W$ be a rank $n$ twisted benign $E$-$B$-pair.
We take an element $\tau\in\mathfrak{S}_n$ and consider the filtration
$\mathcal{T}_{\tau}:0\subseteq W_{\tau,1}\subseteq\cdots \subseteq W_{\tau,n-1}\subseteq W_{\tau,n}=W$.
Then, by the above lemma, $W_{\tau,n-1}$ is twisted benign of rank $n-1$, so by induction hypothesis,
we have $\mathrm{End}_{G_K}(W_{\tau,n-1})=E$.
Let $f:W\rightarrow W$ be a non-zero morphism of $E$-$B$-pairs. Then, by definition of twisted benign $B$-pairs
and by Proposition $\ref{i}$, we have $\mathrm{Hom}_{G_K}(W_{\tau,n-1}, W/W_{\tau,n-1})=0$. So $f(W_{\tau,n-1})
\subseteq W_{\tau,n-1}$. Because $\mathrm{End}_{G_K}(W_{\tau,n-1})=E$, we have $f|_{W_{\tau,n-1}}=a\cdot\mathrm{id}_{W_{\tau,n-1}}$
for some $a\in E$. If $a=0$, then $f:W\rightarrow W$ factors through a non-zero morphism $f':W/W_{\tau,n-1}\rightarrow W$.
Because $\mathrm{Hom}_{G_K}(W/W_{\tau,n-1}, W_{\tau,n-1})=0$ by definition and by Proposition $\ref{i}$,
the natural map $\mathrm{Hom}_{G_K}(W/W_{\tau,n-1},W)\hookrightarrow \mathrm{Hom}_{G_K}(W/W_{\tau,n-1}, W/W_{\tau,n-1})=E$
is injective, so the composition of $f'$ with the natural projection $W\rightarrow W/W_{\tau,n-1}$ induces an isomorphism
$W/W_{\tau,n-1}\isom W/W_{\tau,n-1}$. This implies that the short exact sequence
$0\rightarrow W_{\tau,n-1}\rightarrow W\rightarrow W/W_{\tau,n-1}\rightarrow 0$ splits.
Then we can choose a $\tau'\in\mathfrak{S}_n$, such that $W_{\tau',1}=W/W_{\tau,n-1}$ , and
this $\tau'$ doesn't satisfy the condition (3) in the definition of benign $B$-pairs. It's contradiction.
So the above $a$ must not be zero. In this case, consider the map $f-a\cdot\mathrm{id}_W\in \mathrm{End}_{G_K}(W)$.
Then the same argument as above implies that $f=a\cdot\mathrm{id}_W$. So $\mathrm{End}_{G_K}(W)=E$.
\end{proof}
\begin{corollary}\label{23}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$.
Then the functor $D_{W}$ is pro-representable by $R_W$ which is
formally smooth of dimension $n^2[K:\mathbb{Q}_p]+1$.
For any $\tau\in\mathfrak{S}_n$, the functor $D_{W,\mathcal{T}_{\tau}}$ is pro-representable
by a quotient $R_{W,\mathcal{T}_{\tau}}$ of $R_W$ which is formally smooth
of dimension $\frac{n(n+1)}{2}[K:\mathbb{Q}_p]+1$.
\end{corollary}
\begin{proof}
This follows from Proposition $\ref{19}$.
\end{proof}
Next, we want to consider the relation between $R_{W}$ and $R_{W,\mathcal{T}_{\tau}}$ for
all $\tau\in \mathfrak{S}_n$.
In particular, we want to compare the tangent space of $R_{W}$ and the sum of tangent spaces
of $R_{W,\mathcal{T}_{\tau}}$ for all $\tau\in \mathfrak{S}_n$.
For this, first we need to recall the potentially crystalline deformation functor.
\begin{defn}
Let $W$ be a potentially crystalline $E$-$B$-pair.
Then we define potentially crystalline deformation functor $D_{W}^{\mathrm{cris}}$ which is a sub functor of $D_W$
defined by, for any $A\in \mathcal{C}_E$, $D_{W}^{\mathrm{cris}}(A):=\{[W_A]\in D_{W}(A)| W_A $ is potentially crystalline$\}$.
\end{defn}
\begin{lemma}\label{24}
Let $W$ be a potentially crystalline $E$-$B$-pair.
If $\mathrm{End}_{G_K}(W)=E$, then $D_{W}^{\mathrm{cris}}$ is pro-representable by
a quotient $R_{W}^{\mathrm{cris}}$ of $R_W$ which is formally smooth of dimension
equal to $\mathrm{dim}_E(D_{\mathrm{dR}}(\mathrm{ad}(W))/\mathrm{Fil}^0D_{\mathrm{dR}}(\mathrm{ad}(W))) +
\mathrm{dim}_E(\mathrm{H}^0(G_K, \mathrm{ad}(W)))$.
\end{lemma}
\begin{proof}
For pro-representability, by Proposition $\ref{12}$, it suffices to relatively representability of
$D_{W}^{\mathrm{cris}}\hookrightarrow D_W$ as in the proof of Proposition $\ref{15}$.
But in this case, the condition (1) and (2) is trivial and (3) follows from the fact
that the sub object of potentially crystalline $E$-$B$-pair is again potentially crystalline.
Formally smoothness follows from Proposition 3.1.2 and Lemma 3.2.1 of [Ki08a] by using deformations
of filtered ($\varphi, G_K$)-modules. For dimension, we take a finite Galois extension $L$ of $K$ such that
$W|_{G_L}$ is crystalline. Then, by the same argument as in the proof of Lemma $\ref{20.5}$,
any $W_A\in D^{\mathrm{cris}}_W(A)$ is crystalline when restricted to $G_L$. So it's easy to check
that the map $D_{W}(E[\varepsilon])\isom \mathrm{H}^1(G_K, \mathrm{ad}(W))$
induces isomorphism $D_{W}^{\mathrm{cris}}(E[\varepsilon])\isom \mathrm{Ker}(\mathrm{H}^1(G_K,
\mathrm{ad}(W))\rightarrow \mathrm{H}^1(G_L, B_{\mathrm{cris}}\otimes_{B_e} \mathrm{ad}(W)_e))$. Then again by the same way as in
Lemma $\ref{20.5}$, the natural map $\mathrm{H}^1(G_K, B_{\mathrm{cris}}\otimes_{B_e}\mathrm{ad}(W)_e)\rightarrow \mathrm{H}^1(G_L, B_{\mathrm{cris}}\otimes_{B_e}\mathrm{ad}(W)_e)$ is
injective. So we have isomorphism $D^{\mathrm{cris}}_W(E[\varepsilon])\isom \mathrm{Ker}(\mathrm{H}^1(G_K, \mathrm{ad}(W))
\rightarrow \mathrm{H}^1(G_K, B_{\mathrm{cris}}\otimes_{B_e}\mathrm{ad}(W)_e))$. We can calculate the dimension of this
group in the same way as in the proof of Proposition 2.7 [Na09].
\end{proof}
\begin{corollary}\label{25}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$.
Then $R_{W}^{\mathrm{cris}}$ is formally smooth of dimension
$\frac{(n-1)n}{2}[K:\mathbb{Q}_p]+1$.
\end{corollary}
\begin{proof}
This follows from Lemma $\ref{22}$ and $\mathrm{dim}_E(D_{\mathrm{dR}}(\mathrm{ad}(W))/\mathrm{Fil}^0D_{\mathrm{dR}}(\mathrm{ad}(W)))
=\frac{(n-1)n}{2}[K:\mathbb{Q}_p]$, which follows from the condition (2) in Definition $\ref{26}$.
\end{proof}
\begin{defn}\label{34}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$ such that $W|_{G_{K_m}}$ is crystalline.
For any $\tau\in\mathfrak{S}_n$, we define a rank one saturated
crystabelline $E$-$B$-pair $W'_{\tau}\subseteq W$ such that $D^{K_m}_{\mathrm{cris}}(W'_{\tau})=K_0\otimes_{\mathbb{Q}_p}E
e_{\tau(n)} \subseteq D^{K_m}_{\mathrm{cris}}(W)$. Then we define a sub functor
$D^{\mathrm{cris}}_{W,\tau}$ of $D_W$ by $D^{\mathrm{cris}}_{W,\tau}(A):=\{[W_A]\in D_W(A)| $ there exists rank one crystabelline
saturated sub $A$-$B$-pair $W'_A\subseteq W_A$ such that $W'_A\otimes_A E=W'_{\tau} \}$
\end{defn}
\begin{lemma}\label{27}
Under the above condition. The functor $D^{\mathrm{cris}}_{W,\tau}$ is pro-representable by
a quotient $R^{\mathrm{cris}}_{W,\tau}$ of $R_W$ which is formally smooth and it's dimension
is equal to $n(n-1)[K:\mathbb{Q}_p] +1$.
\end{lemma}
\begin{proof}
The proof of relatively representability of $D^{\mathrm{cris}}_{W,\tau}\hookrightarrow D_W$
and the proof of formally smoothness are easily followed from combination of proofs
of Proposition $\ref{15}$ and Proposition $\ref{17}$ and Lemma $\ref{24}$.
Here, we only prove dimension formula.
Let $\mathrm{ad}_{\tau}(W):=\{f\in\mathrm{ad}(W)| f(W'_{\tau})\subseteq
W'_{\tau}\}$. Then we have the following short exact sequence,
\begin{equation*}
0\rightarrow \mathrm{Hom}(W/W'_{\tau}, W)\rightarrow \mathrm{ad}_{\tau}(W)
\rightarrow \mathrm{ad}(W'_{\tau})\rightarrow 0.
\end{equation*}
From this, by taking long exact sequence and by using Proposition $\ref{i}$, we have
the following short exact sequence,
\begin{equation*}
0\rightarrow \mathrm{H}^1(G_K, \mathrm{Hom}(W/W'_{\tau},W))\rightarrow
\mathrm{H}^1(G_K, \mathrm{ad}_{\tau}(W))\rightarrow \mathrm{H}^1(G_K,
\mathrm{ad}(W'_{\tau}))\rightarrow 0.
\end{equation*}
We define $\mathrm{H}^1_{f,\tau}(G_K, \mathrm{ad}_{\tau}(W))$
by the inverse image of $\mathrm{H}^1_f(G_K,\mathrm{ad}(W'_{\tau}))$
in $\mathrm{H}^1(G_K,\allowbreak \mathrm{ad}_{\tau}(W))$. So we have
a short exact sequence,
\begin{equation*}
0\rightarrow \mathrm{H}^1(G_K, \mathrm{Hom}(W/W'_{\tau},W))
\rightarrow \mathrm{H}^1_{f,\tau}(G_K, \mathrm{ad}_{\tau}(W))
\rightarrow \mathrm{H}^1_f(G_K, \mathrm{ad}(W'_{\tau}))\rightarrow 0.
\end{equation*}
Then, in the same way as in Lemma $\ref{24}$, we can show that the natural map $D_{W}(E[\varepsilon])\rightarrow \mathrm{H}^1(G_K, \mathrm{ad}(W))$
induces an isomorphism $D^{\mathrm{cris}}_{W,\tau}(E[\varepsilon])\isom \mathrm{H}^1_{f,\tau}(G_K,
\allowbreak \mathrm{ad}_{\tau}(W))$.
By using Theorem $\ref{h}$ and Proposition $\ref{i}$, we can calculate that
$\mathrm{dim}_E\mathrm{H}^1(G_K, \mathrm{Hom}\allowbreak (W/W'_{\tau},W))=n(n-1)[K:\mathbb{Q}_p]$.
Because $\mathrm{ad}(W'_{\tau})=B_E$ is trivial, so we have $\mathrm{H}^1_{f}(G_K, \mathrm{ad}(W'_{\tau}))
=1$ by Proposition 2.7 of [Na09]. So we have $\mathrm{dim}_E\mathrm{H}^1_{f,\tau}(G_K, \mathrm{ad}_{\tau}(W))\allowbreak =n(n-1)[K:\mathbb{Q}_p] +1$.
This proves that $R_{W,\tau}^{\mathrm{cris}}$ is dimension $n(n-1)[K:\mathbb{Q}_p]+1$.
\end{proof}
\begin{lemma}\label{28}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$.
Let $W_A$ be a deformation of $W$ over $A$ which is potentially crystalline,
then $[W_A]\in D_{W,\mathcal{T}_{\tau}}(A)$ and $[W_A]\in D^{\mathrm{cris}}_{W,\tau}(A)$ for
any $\tau\in \mathfrak{S}_n$.
\end{lemma}
\begin{proof}
Let $W_A$ be as above. If $W|_{G_{K_m}}$ is crystalline,
then $W_A|_{G_{K_m}}$ is crystalline by the proof of Lemma $\ref{24}$.
So it suffices to show that $D^{K_m}_{\mathrm{cris}}(W_A)\isom
K_0\otimes_{\mathbb{Q}_p}Ae_1\oplus K_0\otimes_{\mathbb{Q}_p}Ae_2\oplus\cdots
\oplus K_0\otimes_{\mathbb{Q}_p}Ae_n$ such that $K_0\otimes_{\mathbb{Q}_p}Ae_i$ is
preserved by ($\varphi,G_{m}$) and $\varphi^f(e_i)=\tilde{\alpha_i}e_i$ for a lift $\tilde{\alpha_i}\in A^{\times}$
of $\alpha_i\in E^{\times}$ for any $1\leqq i\leqq n$. For proving this claim, first we note that
$D^{K_m}_{\mathrm{cris}}(W_A)$ is a free $K_0\otimes_{\mathbb{Q}_p}A$-module of rank $n$ and
$D^{K_m}_{\mathrm{cris}}(W_A)\otimes_A E\isom D^{K_m}_{\mathrm{cris}}(W)$ by
Proposition 1.3.4 and Proposition 1.3.5 [Ki09].
Then, for any $\sigma:K_0\hookrightarrow E$, we can write $D^{K_m}_{\mathrm{cris}}(W_A)_{\sigma}=
Ae_{1,\sigma}\oplus \cdots \oplus Ae_{n,\sigma}$ such that
$\varphi^f(e_{i,\sigma})\equiv \alpha_ie_{i,\sigma}$ (mod $m_A$) for any $1\leqq i\leqq n$.
Then, by easy linear algebra, we can take an $A$-basis $e'_{1,\sigma}, e'_{2,\sigma}, \cdots, e'_{n,\sigma}$ of
$D^{K_m}_{\mathrm{cris}}(W_A)_{\sigma}$ such that $\varphi^f(e'_{i,\sigma})=\tilde{\alpha}_ie'_{i,\sigma}$
for a lift $\tilde{\alpha}_i\in A^{\times}$ of $\alpha_i$ for any $i$. Because $\varphi$ and $G_{m}$ commute and $\alpha_i\not=\alpha_j$,
so $Ae'_{i,\sigma}$ is preserved by $G_{m}$. If we take $e_i:=e'_{i,\sigma}+\varphi(e'_{i,\sigma})+\cdots
\varphi^{f-1}(e'_{i,\sigma})\in D^{K_m}_{\mathrm{cris}}(W_A)$, then we have $D^{K_m}_{\mathrm{cris}}(W_A)=
K_0\otimes_{\mathbb{Q}_p}Ae_1\oplus \cdots \oplus K_0\otimes_{\mathbb{Q}_p}Ae_n$ satisfying
the property of the claim.
\end{proof}
\begin{lemma}\label{29}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$ and $\tau\in \mathfrak{S}_n$ such that
$W|_{G_{K_m}}$ is crystalline.
Let $[W_A]\in D_{W,\mathcal{T}_{\tau}}(A)$ be a trianguline deformation over $A$
with a lifting of triangulation $0\subseteq W_{1,A}\subseteq W_{2,A}\subseteq \cdots W_{n,A}=W_A$.
If $W_{i,A}/W_{i-1,A}$ is Hodge-Tate for any $1\leqq i\leqq n$, then $W_A|_{G_{K_m}}$ is crystalline.
\end{lemma}
\begin{proof}
First, we prove that $(W_{i,A}/W_{i-1,A})|_{G_{K_m}}$ is crystalline with Hodge-Tate weight $\{k_{i,\sigma}\}_{\sigma\in \mathcal{P}}$.
Because $W_{i,A}/W_{i-1,A}$ is written as successive extensions of $W_{\tau,i}/W_{\tau,i-1}$, $W_{i,A}/W_{i-1,A}$
has Hodge-Tate weight $\{k_{i,\sigma}\}_{\sigma\in\mathcal{P}}$. Then, twisting $W_A$ by a crystalline character $\delta_{\alpha_{\tau(i)}}^{-1}
\prod_{\sigma\in\mathcal{P}}\sigma(x)^{-k_{i,\sigma}}:K^{\times}\rightarrow A^{\times}$, we may assume that
$W_{i,A}/W_{i-1, A}$ is an $\mathrm{\acute{e}}$tale Hodge-Tate $A$-$B$-pair of rank one with Hodge-Tate weight zero and is
a deformation of $\mathrm{\acute{e}}$tale potentially unramified $E$-$B$-pair $W(\tilde{\delta}_{\tau(i)})$. Then, by Sen's theorem
([Se73] or Proposition 5.24 of [Be02]),
$W_{i,A}/W_{i-1,A}$ is potentially unramified, so there exists $\delta:K^{\times}\rightarrow A^{\times}$ a unitary
homomorphism such that $\delta|_{\mathcal{O}_K^{\times}}$ has a finite image and $W_{i,A}/W_{i-1,A}\isom W(\delta)$ and
$\delta$ is a lift of $\tilde{\delta}_{\tau(i)}$. Then, because $(1+m_A)\cap A^{\times}_{tor}=\{1\}$,
$\delta|_{\mathcal{O}_K^{\times}}=\tilde{\delta}_{\tau(i)}|_{\mathcal{O}_K^{\times}}:\mathcal{O}_K^{\times}\rightarrow A^{\times}$, so $W_{i,A}/W_{i-1,A}|_{G_{K_m}}$ is crystalline.
Next, we prove that $W_A|_{G_{K_m}}$ is crystalline by induction on rank $W$.
When $n=1$, we just have proved this.
Assume that the lemma is proved for rank $n-1$ case.
Then $W_{n-1, A}|_{G_{K_m}}$ is crystalline. If we put $W_A/W_{n-1,A}\isom W(\delta_{A,n})$, then we have
$[W_A]\in \mathrm{H}^1(G_K, W_{n-1,A}(\delta_{A,n}^{-1}))$. By considering Hodge-Tate weight of
$W_{A}$ and condition (3) of Definition $\ref{26}$, we get $\mathrm{Fil}^0D_{\mathrm{dR}}(W_{n-1,A}(\delta_{A,n}^{-1}))=0$.
By comparing dimensions by using Proposition 2.7 of [Na09], we have $\mathrm{H}^1_f(G_K,W_{A,n-1}(\delta_{A,n}^{-1}))=\mathrm{H}^1(G_K,
W_{A,n-1}(\delta_{A,n}^{-1}))$. So $W_A|_{G_{K_m}}$ is crystalline.
\end{proof}
\begin{lemma}\label{30}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$ and $W_1$ be a rank one crystabelline sub $E$-$B$-pair of $W$.
Then the saturation $W_{1}^{sat}:=(W^{sat}_{1,e}, W^{+, sat}_{1,\mathrm{dR}})$ of $W_1$ in $W$ is
crystabelline and the natural map $\mathrm{Hom}_{G_K}( W_{1}^{sat}, W)
\rightarrow \mathrm{Hom}_{G_K}(W_1,W)$ induced by natural
inclusion $W_1\hookrightarrow W_{1}^{sat}$ is isomorphism
between one dimensional $E$-vector spaces.
\end{lemma}
\begin{proof}
Because $W_{1,e}=W^{sat}_{1,e}$ by Lemma 1.14 of [Na09], so $W_{1}^{sat}$ is crystabelline.
By definition of twisted benign $E$-$B$-pairs, the Hodge-Tate weight of $W_{1}^{sat}$
is $\{k_{1,\sigma}\}_{\sigma\in\mathcal{P}}$. Then consider the following short exact sequence
of complexes of $G_K$-modules defined as in p.890 of [Na09],
\begin{equation*}
0\rightarrow \mathrm{C}^{\bullet}(W\otimes (W_{1}^{sat})^{\vee})\rightarrow \mathrm{C}^{\bullet}(W
\otimes W_{1}^{\vee})\rightarrow
(W\otimes W_{1}^{\vee})^+_{\mathrm{dR}})/
(W\otimes(W_{1}^{sat})^{\vee})^+_{\mathrm{dR}}) [0]\rightarrow 0.
\end{equation*}
From this, we have the exact sequence,
\[
\begin{array}{ll}
0\rightarrow \mathrm{H}^0(G_K, W\otimes (W_1^{sat})^{\vee})&\rightarrow
\mathrm{H}^0(G_K, W\otimes W_1^{\vee}) \\
&\rightarrow \mathrm{H}^0(G_K, (W\otimes W_1^{\vee})^+_{\mathrm{dR}}/
(W\otimes(W_1^{sat})^{\vee})^+_{\mathrm{dR}})\rightarrow \cdots
\end{array}
\]
By the condition (3) in Definition $\ref{26}$, we have $\mathrm{dim}_E\mathrm{H}^0(G_K, (W\otimes (W_1^{sat})^{\vee})^+_{\mathrm{dR}})
=\mathrm{dim}_E\mathrm{H}^0(G_K, (W\otimes W_1^{\vee})^+_{\mathrm{dR}})=n[K:\mathbb{Q}_p]$, So, by a usual argument
of de Rham representations, we have $\mathrm{H}^0(G_K, (W\otimes W_1^{\vee})^+_{\mathrm{dR}}/(W\otimes (W_1^{sat})^{\vee})^+_{\mathrm{dR}})=0$. So the map $\mathrm{H}^0(G_K, W\otimes (W_1^{sat})^{\vee}))\isom \mathrm{H}^0(G_K, W\otimes (W_1)^{\vee})
$ is isomorphism. Finally, for dimension, by the condition (1) in Definition $\ref{26}$ and Proposition $\ref{i}$ of [Na09],
we have $\mathrm{dim}_E\mathrm{H}^0(G_K, W\otimes (W_1^{sat})^{\vee})=1$. These prove the lemma.
\end{proof}
\begin{lemma}\label{31}
Let $W$ be a twisted benign $E$-$B$-pair and let $W_A$ be a deformation
of $W$ over $A$. If there exists a rank one crystabelline sub $A$-$B$-pair
$W_{1,A}\subseteq W_A$ such
that $W_1:=W_{1,A}\otimes_A E\hookrightarrow W_A\otimes_{A}E$ remains injective.
Then the saturation $W_{1,A}^{sat}$ of $W_{1,A}$ in $W_A$ as an $E$-$B$-pair is
a crystabelline $A$-$B$-pair and $W_A/W^{sat}_{1,A}$ is an $A$-$B$-pair and
$W_{1,A}^{sat}\otimes_A E\isom W_1^{sat} (\subseteq W)$.
\end{lemma}
\begin{proof}
First, by Proposition 2.14 of [Na09], there exists $\{l_{\sigma}\}_{\sigma\in\mathcal{P}}
\in \prod_{\sigma\in\mathcal{P}}\mathbb{Z}_{\leqq 0}$ such that $W_1^{sat}\isom W_1\otimes W(\prod_{\sigma\in\mathcal{P}}
\sigma(x)^{l_{\sigma}})$. Then we claim that the inclusion
$\mathrm{Hom}_{G_K}(W_{1,A}\otimes W_A(\prod_{\sigma\in\mathcal{P}}\sigma(x)^{l_{\sigma}}), W_A)
\rightarrow \mathrm{Hom}_{G_K}(W_{1,A}, W_A)$ induced by the natural inclusion $W_{1,A}\hookrightarrow W_{1,A}
\otimes W_A(\prod_{\sigma\in\mathcal{P}}\sigma(x)^{l_{\sigma}})$ is isomorphism and
that these groups are rank one free $A$-modules. By the same argument as in the last lemma,
the cokernel of $\mathrm{Hom}_{G_K}(W_{1,A}\otimes W_A(\prod_{\sigma\in\mathcal{P}}\sigma(x)^{l_{\sigma}}), W_A)
\rightarrow \mathrm{Hom}_{G_K}(W_{1,A}, W_A)$ is contained in $\mathrm{H}^0(G_K, (W_A\otimes W_{1,A}^{\vee})^+_{\mathrm{dR}}/
(W_A\otimes W_{1,A}^{\vee}\otimes W_A(\prod_{\sigma\in \mathcal{P}}\sigma(x)^{-l_{\sigma}}))^+_{\mathrm{dR}})$. This is zero
from the proof of the last lemma. So the natural inclusion $\mathrm{Hom}_{G_K}(W_{1,A}\otimes W_A(\prod_{\sigma\in\mathcal{P}}\sigma(x)^{l_{\sigma}}), W_A)
\rightarrow \mathrm{Hom}_{G_K}(W_{1,A}, W_A)$ is isomorphism. Next, we prove, by induction on length of A,
that $\mathrm{Hom}_{G_K}(W_{1,A}, W_A)$ is a free $A$-module of rank one.
When $A=E$, this claim is proved in the above Lemma $\ref{30}$.
Assume that $A$ is length $n$ and, for a small extension $A\rightarrow A'$, assume that the claim is
proved for $W_{A'}:=W\otimes_A A'$. Put $I$ the kernel of $A\rightarrow A'$, $W_{1,A'}:=W_{1,A}\otimes_A A'$. From the following exact sequence,
\begin{equation*}
0\rightarrow I\otimes_E \mathrm{Hom}_{G_K}(W_1, W)
\rightarrow \mathrm{Hom}_{G_K}(W_{1,A}, W_A)\rightarrow \mathrm{Hom}_{G_K}(W_{1,A'}, W_{A'})
\end{equation*}
and induction hypothesis, we have
$\mathrm{length}\mathrm{Hom}_{G_K}(W_{1,A}, W_A)\leqq
\mathrm{length}A$. On the other hand, the fact that given inclusion $\iota:W_{1,A}\hookrightarrow W_A$ remains
injective after tensoring $E$ and the fact that $\mathrm{dim}_E\mathrm{Hom}_{G_K}(W_1, W)=1$ and induction hypothesis
imply that the map $A\rightarrow \mathrm{Hom}_{G_K}(W_{1,A}, W_A): a\mapsto a\cdot\iota $ is injective. So we have
$\mathrm{length}(A)= \mathrm{length} \mathrm{Hom}_{G_K}(W_{1,A}, W_A)$. These prove the claim for $A$.
From these claims, the given inclusion $\iota:W_{1,A}\hookrightarrow W_A$ factors through a map $\tilde{\iota}:W_{1,A}':=W_{1, A}\otimes W_A(\prod_{\sigma}\sigma(x)^{l_{\sigma}})
\rightarrow W_A$ and this map is injective because the injection of morphisms of $B$-pair determined from the injection
of $W_e$-part of $B$-pairs and $W_{1,A,e}=(W_{1,A}\otimes W_A(\prod_{\sigma\in\mathcal{P}}\sigma(x)^{l_{\sigma}}))_e$. Then we claim that this $\tilde{\iota}$ gives $W_{1,A}^{sat}\isom W_{1,A}'$ and claim that this satisfies all the properties in this lemma. By induction on length $A$,
we assume that this claim is proved for $A'$. First, we prove $W_{A}/W_{1,A}'$ is an $E$-$B$-pair. For proving this, by Lemma 2.1.4 of [Ber08],
it suffices to show that $W_{A,\mathrm{dR}}^{+}/W_{1,A,\mathrm{dR}}^{'+}$ is a free $B_{\mathrm{dR}}^+$-module. By snake lemma, we have
\begin{equation*}
0\rightarrow I\otimes_E W^+_{\mathrm{dR}}/W^{sat +}_{1,\mathrm{dR}}\rightarrow
W^+_{A,\mathrm{dR}}/W^{' +}_{1, A,\mathrm{dR}}\rightarrow W^+_{A',\mathrm{dR}}/(W_{1,A,\mathrm{dR}}^{' +}\otimes_A A')\rightarrow 0.
\end{equation*}
From this and induction hypothesis, $W^+_{A,\mathrm{dR}}/W^{' +}_{1,A,\mathrm{dR}}$ is a free $B^+_{\mathrm{dR}}$-module.
Finally, we prove $A$-flatness of $W_{A}/W_{1,A}^{'}$. This follows from the fact that the map
$\tilde{\iota}\otimes \mathrm{id}_E:W_{1,A}^{'}\otimes_A E\hookrightarrow W_{A}\otimes_A E$ is saturated (we can see this
from the proof of the above first claim). So $W_A/W_{1,A}^{'}$ is an $A$-$B$-pair. We finish the proof of this lemma.
\end{proof}
\begin{lemma}\label{32}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$.
For any $\tau\in\mathfrak{S}_n$, we have $D_{W,\mathcal{T}_{\tau}}(E[\epsilon])
+D^{\mathrm{cris}}_{W,\tau}(E[\varepsilon])=D_{W}(E[\varepsilon])$.
\end{lemma}
\begin{proof}
By Corollary $\ref{23}$ and Lemma $\ref{24}$ and Lemma $\ref{27}$, we have $\mathrm{dim}_ED_{W}(E[\varepsilon])\allowbreak +\mathrm{dim}_ED_W^{\mathrm{cris}}(E[\varepsilon])
=\mathrm{dim}_E D_{W,\mathcal{T}_{\tau}}(E[\varepsilon])+\mathrm{dim}_ED^{\mathrm{cris}}_{W,\tau}(E[\varepsilon])$.
So it suffices to show that $D_{W,\mathcal{T}_{\tau}}(A)\cap D^{\mathrm{cris}}_{W,\tau}(A)=D_{W}^{\mathrm{cris}}(A)$ for
any $A\in \mathcal{C}_E$. First, the fact that $D^{\mathrm{cris}}_W(A)\subseteq D_{W,\mathcal{T}_{\tau}}(A)\cap D_{W,\tau}(A)$
follows from Lemma $\ref{28}$.
We prove that $D_{W,\mathcal{T}_{\tau}}(A)\cap D^{\mathrm{cris}}_{W,\tau}(A)\subseteq D_{W}^{\mathrm{cris}}(A)$ by induction on
rank $n$ of $W$. When $n=1$, this is trivial.
Let $W$ be rank $n$ and assume that the lemma is proved for rank $n-1$.
Let $[W_A]\in D_{W,\mathcal{T}_{\tau}}(A)\cap D^{\mathrm{cris}}_{W,\tau}(A)$.
Then, by definition of $D_{W,\mathcal{T}_{\tau}}$ and $D^{\mathrm{cris}}_{W,\tau}$, there exists an $A$-triangulation
$0\subseteq W_{1,A}\subseteq W_{2,A}\subseteq \cdots \subseteq W_{n-1,A}\subseteq W_{n,A}=W_A$
such that $W_{i,A}\otimes_A E\isom W_{\tau,i}$ for any $i$ and there exists saturated crystabelline rank one $A$-$B$-pair
$W'_{1,A}\subseteq W_{A}$ such that $W'_{1,A}\otimes_A E\isom W'_{\tau}$.
First, we claim that the composition of $W'_{1,A}\hookrightarrow W_A$ and $W_A\rightarrow W_A/W_{1,A}$
is injective. By snake lemma, $\mathrm{Ker}(W'_{1,A}\rightarrow W_A/W_{1,A})$ is a sub $E$-$B$-pair of
$W_{1, A}$. By the condition (1) of Definition $\ref{26}$ and by Proposition 2.14 of [Na09], we have
$\mathrm{Hom}_{G_K}(\mathrm{Ker}(W'_{1, A}\rightarrow W_A/W_{1,A}), W_{1,A})=0$, so $W'_{1,A}\rightarrow
W_A/W_{1,A}$ is injective. Then, by Lemma $\ref{31}$, the saturation $(W'_{1,A})^{sat}$ of $W'_{1,A}$ in $W_A/W_{1,A}$
is a rank one crystabelline $A$-$B$-pair satisfying the same conditions as those of $W'_{1,A}\hookrightarrow W_A$.
So, by induction hypothesis, $W_A/W_{1,A}$ is crystabelline. Then, by the condition (3) of Definition $\ref{26}$, Hodge-Tate weight of
$W_A/W_{1,A}$ is $\{k_{2,\sigma},k_{3,\sigma},\cdots,k_{n,\sigma}\}_{\sigma}$ (with multiplicity $[A:E]$).
Moreover, by (3) of Definition $\ref{26}$, crystabelline $W'_{1,A}$ has Hodge-Tate weight $\{k_{1,\sigma}\}_{\sigma\in\mathcal{P}}$
(with multiplicity). Because $k_{1,\sigma}\not= k_{i,\sigma}$ for any $i\not=1$ and there is no extension between different
Hodge-Tate weight objects by a theorem of Tate, these imply that $W_A$ is a Hodge-Tate $E$-$B$-pair. So,
by Lemma $\ref{29}$, $W_A$ is crystabelline. So we have that $[W_A]\in D_{W}^{\mathrm{cris}}(A)$.
\end{proof}
The following is the main theorem of this article, which is a crucial theorem for applications to Zariski density.
This theorem was discovered by Chenevier ([Che08], [Ch09]) for any rank case when $K=\mathbb{Q}_p$.
\begin{defn}
For $R=R_W$ or $R_{W,\mathcal{T}_{\tau}}$, we put $t(R):=\mathrm{Hom}_E(m_R/m^2_R,E)$ the tangent
space of $R$. Then, for any $\tau\in \mathfrak{S}_n$, we have a canonical inclusion $t(R_{W,\mathcal{T}_{\tau}})\hookrightarrow
t(R_W)$.
\end{defn}
\begin{thm}\label{33}
Let $W$ be a twisted benign $E$-$B$-pair of rank $n$.
Then we have $$\sum_{\tau\in\mathfrak{S}_n}t(R_{W,\mathcal{T}_{\tau}})=t(R_W).$$
\end{thm}
\begin{proof}
We prove this by induction on $n$.
When $n=1$, then the theorem is trivial.
Assume that the theorem is true for rank $n-1$ case.
Let $W$ be a rank $n$ twisted benign $E$-$B$-pair.
We take an element $\tau\in\mathfrak{S}_n$.
Then we consider the sub functor $D_{W,\tau}$ of $D_W$ by
$D_{W,\tau}(A):=\{[W_A]\in D_{W}(A)| $ there exists a rank one sub $A$-$B$-pair $W_{1,A}\subseteq W_A$ such
that the quotient $W_A/W_{1,A}$ is an $A$-$B$-pair and $W_{1,A}\otimes_A E\isom W'_{\tau}\}$ ($W'_{\tau}$ is defined in
Definition $\ref{34}$).
Then $D^{\mathrm{cris}}_{W,\tau}$ is a sub functor of $D_{W,\tau}$
and we can show that $D_{W,\tau}(E[\varepsilon])\isom \mathrm{H}^1(G_K, \mathrm{ad}_{\tau}(W))$, here $\mathrm{ad}_{\tau}(W)
:=\{f \in \mathrm{ad}(W)| f(W'_{\tau})\subseteq W'_{\tau} \}$. So, by Lemma $\ref{32}$, we have $\mathrm{H}^1(G_K, \mathrm{ad}_{\tau}(W))+
\mathrm{H}^1(G_K, \mathrm{ad}_{\mathcal{T}_{\tau}}(W))=\mathrm{H}^1(G_K, \mathrm{ad}(W))$.
Because, for any $\tau'\in\mathcal{S}_{\tau,n}:=\{\tau'\in\mathfrak{S}_n| \tau'(1)=\tau(n)\}$, we have
$D_{W,\mathcal{T}_{\tau'}}\subseteq D_{W,\tau}$. So we have natural maps $\mathrm{H}^1(G_K, \mathrm{ad}_{\mathcal{T}_{\tau'}}(W))
\rightarrow \mathrm{H}^1(G_K, \mathrm{ad}_{\tau}(W))$. Then it suffices to prove that
the map $\oplus_{\tau'\in\mathcal{S}_{\tau,n}}\mathrm{H}^1(G_K, \mathrm{ad}_{\mathcal{T}_{\tau'}}(W))\rightarrow
\mathrm{H}^1(G_K, \mathrm{ad}_{\tau}(W))$ is surjective. We prove this surjection as follows.
First, by definition, we have the following short exact sequences of $E$-$B$-pairs for any $\tau'\in\mathcal{S}_{\tau,n}$,
\begin{equation}
0\rightarrow \mathrm{Hom}(W/W'_{\tau}, W)\rightarrow \mathrm{ad}_{\tau}(W)\rightarrow \mathrm{ad}(W'_{\tau})
\rightarrow 0,
\end{equation}
\begin{equation}
0\rightarrow \{f\in\mathrm{ad}_{\mathcal{T}_{\tau'}}(W)| f(W'_{\tau})=0\}\rightarrow \mathrm{ad}_{\mathcal{T}_{\tau'}}(W)\rightarrow
\mathrm{ad}(W'_{\tau})\rightarrow 0.
\end{equation}
Moreover, we have
\begin{equation}
0\rightarrow \mathrm{Hom}(W/W'_{\tau}, W'_{\tau})\rightarrow \mathrm{Hom}(W/W'_{\tau}, W)\rightarrow \mathrm{ad}(W/W'_{\tau})\rightarrow 0,
\end{equation}
\begin{equation}
0\rightarrow \mathrm{Hom}(W/W'_{\tau}, W'_{\tau})\rightarrow \{f\in\mathrm{ad}_{\mathcal{T}_{\tau'}}(W)|f(W'_{\tau})=0\}
\rightarrow \mathrm{ad}_{\mathcal{T}_{\bar{\tau}'}}(W/W'_{\tau})\rightarrow 0.
\end{equation}
Here, for any $\tau'\in\mathcal{S}_{\tau,n}$, we put $\mathcal{T}_{\bar{\tau}'}:0\subseteq W_{\tau',2}/W'_{\tau}\subseteq W_{\tau',3}/W'_{\tau}
\subseteq \cdots \subseteq W_{\tau',n-1}/W'_{\tau}\subseteq W/W'_{\tau}$ the induced triangulation from $\mathcal{T}_{\tau'}$
on $W/W'_{\tau}$. We have $\mathrm{H}^2(G_K, \mathrm{Hom}(W/W'_{\tau}, W'_{\tau}))=0$ by the condition $(1)$ of Definition $\ref{26}$ and
Proposition $\ref{i}$. We have $\mathrm{H}^2(G_K, \mathrm{ad}_{\mathcal{T}_{\tau}}(W))=0$ from the proof of Proposition
$\ref{19}$. So from the short exact sequence (4) above, we have $\mathrm{H}^2(G_K, \{f\in\mathrm{ad}_{\tau'}(W)|f(W'_{\tau})=0\})=0$.
From this and from (1) and (2) above, for proving the surjection of $\oplus_{\tau'\in \mathcal{S}_{\tau,n}}\mathrm{H}^1(G_K,\mathrm{ad}_{\mathcal{T}_{\tau'}}(W))\allowbreak
\rightarrow \mathrm{H}^1(G_K, \mathrm{ad}_{\tau}(W))$, it suffices to prove
the surjection of $\oplus_{\tau'\in \mathcal{S}_{\tau,n}}\mathrm{H}^1(G_K, \{f\in\mathrm{ad}_{\tau'}(W)|f(W'_{\tau})=0\})
\rightarrow \mathrm{H}^1(G_K, \mathrm{Hom}(W/W'_{\tau}, W))$. By (3) and (4) above and by $\mathrm{H}^2(G_K, \mathrm{Hom}(W/W'_{\tau}, W'_{\tau}))=0$,
this surjection follows from the surjection of $\oplus_{\tau'\in\mathcal{S}_{\tau,n}}\mathrm{H}^1(G_K, \mathrm{ad}_{\mathcal{T}_{\bar{\tau}'}}(W/W'_{\tau}))\rightarrow \mathrm{H}^1(G_K, \mathrm{ad}(W/W'_{\tau}))$, which is the induction hypothesis. So we have finished the proof of this theorem
\end{proof}
|
\section{Introduction}
In a previous paper \cite{CZ}\ we have discussed how functions of position,
defined on a discrete lattice of time points, may be analytically interpolated
in time by functions of both $x$ and $t$, defined on a continuum of time
points, framed by the lattice, through the use of solutions to Schr\"{o}der's
nonlinear functional equation \cite{S}. \ For a more pedagogical discussion to
augment this Introduction, we encourage readers to look at that companion paper.
If the effect of the first discrete time step is given as the map $x\mapsto
f_{1}\left( x,s\right) $, for some parameter $s$, Schr\"{o}der's functional
equation i
\begin{equation}
s\Psi\left( x,s\right) =\Psi\left( f_{1}\left( x,s\right) ,s\right) \ ,
\label{SE
\end{equation}
with $\Psi$ to be determined. \ So,
\begin{equation}
f_{1}\left( x,s\right) =\Psi^{-1}\left( s\Psi\left( x,s\right) ,s\right)
\ ,
\end{equation}
where the inverse function $\Psi^{-1}$ obeys Poincar\'{e}'s equation
\begin{equation}
\Psi^{-1}\left( sx,s\right) =f_{1}\left( \Psi^{-1}\left( x,s\right)
,s\right) \ . \label{PE
\end{equation}
The $n$th iterate of (\ref{SE}) gives
\begin{equation}
s^{n}\Psi\left( x,s\right) =\Psi\left( f_{1}\left( \cdots f_{1}\left(
f_{1}\left( x,s\right) ,s\right) \cdots,s\right) ,s\right) \ ,
\end{equation}
with $f_{1}$ acting $n$ times, and thus the $n$th order functional composition
--- the so-called \textquotedblleft splinter\textquotedblright\ of the
functional equation --
\begin{equation}
f_{n}\left( x,s\right) \equiv f_{1}\left( \cdots f_{1}\left( f_{1}\left(
x,s\right) ,s\right) \cdots,s\right) =\Psi^{-1}\left( s^{n}\Psi\left(
x,s\right) ,s\right) \ .
\end{equation}
A continuous interpolation between the integer lattice of time points is then,
for \emph{any} $t$
\begin{equation}
f_{t}\left( x,s\right) =\Psi^{-1}\left( s^{t}\Psi\left( x,s\right)
,s\right) \ . \label{Interpolation
\end{equation}
This can be a well-behaved and single-valued function of $x$ and $t$ provided
that $\Psi^{-1}\left( x,s\right) $ is a well-behaved, single-valued function
of $x$,\ even though $\Psi\left( x,s\right) $ might be, and typically is,
multi-valued. \
As discussed in \cite{CZ}, the interpolation can be envisioned as the
trajectory of a particle,
\begin{equation}
x\left( t\right) =f_{t}\left( x,s\right) \ , \label{Trajectory
\end{equation}
where the particle is moving under the influence of a potential according to
Hamiltonian dynamics.\ \ The velocity of the particle is then found by
differentiating (\ref{Interpolation}) with respect to $t$,\
\begin{equation}
\frac{dx\left( t\right) }{dt}=\left( \ln s\right) s^{t}\Psi\left(
x\right) \left( \Psi^{-1}\left( s^{t}\Psi\left( x\right) \right)
\right) ^{\prime}\ ,
\end{equation}
where any dependence of $\Psi$ on $s$ is implicitly understood. \ Therefore,
the velocity will inherit and exhibit any multi-valuedness possessed by
$\Psi\left( x\right) $. \
Indeed, suppose that Hamiltonian dynamics have been specified, trajectories
have been computed, and $f_{1}=x\left( t=1\right) $ has emerged\ in terms of
initial velocity and initial position $x\left( t=0\right) \equiv x$. \ A
solution of the functional equation (\ref{SE}) can then be constructed, and
expressed in very physical terms, as just an exponential of \emph{the time
elapsed along any such particle trajectory} that passes through position $x$.
\ That solution i
\begin{equation}
\Psi\left( x\right) =s^{T\left( x\right) }\Psi_{0}\ ,\ \ \ T\left(
x\right) =\int^{x}\frac{dy}{v\left( y\right) }\ . \label{GeneralPsi
\end{equation}
Here $v$ is the velocity as a function of the position along the trajectory.
\ (Hence $v\left( x\left( t\right) \right) $, if you wish. \ Of course,
for a Hamiltonian system with time translational invariance, we may supplant
$x\left( t\right) \rightarrow x\left( t+\Delta t\right) $ in all these
expressions, as well as those to follow.) \ For a particle moving in a
potential $V\left( x\right) $ at fixed energy, $E$, the velocity can be
expressed in the usual way as $v\left( x\right) =\pm\sqrt{E-V\left(
x\right) }$, with suitably chosen mass units. \
When written in a form more closely related to Schr\"{o}der's functional
equation,\ the solution (\ref{GeneralPsi}) is simpl
\begin{equation}
\Psi\left( f_{t}\left( x\right) \right) =s^{\int_{x}^{f_{t}\left(
x\right) }\frac{dy}{v\left( y\right) }}~\Psi\left( x\right) \ .
\end{equation}
At $t=1$, $\int_{x}^{f_{1}\left( x\right) }\frac{dy}{v\left( y\right)
=1$, and Schr\"{o}der's equation (\ref{SE}) re-emerges. \ But, as this
construction clearly shows, one must be careful at turning points where $v=0$,
especially if these are encountered at finite times along the trajectory.
\ Typically these turning points produce branch points in $\Psi$ so that it is
multi-valued. \
An interesting physical effect of such branch points, when they are
encountered in finite times, is the possibility to switch from one branch of
the underlying analytic potential function to another, thereby changing the
functional form of the Hamiltonian governing the motion along seemingly
identical intervals of a real-valued trajectory. \ We will therefore call
these \emph{quasi}-Hamiltonian systems, and we stress that when such switches
occur this is \emph{not} standard textbook Hamiltonian dynamics. \ Further
explanation will be given below in the context of the logistic map.
The interpolation (\ref{Interpolation}) can also be viewed as a holographic
specification on the $x,t$ plane \cite{L}, determining $f_{t}\left( x\right)
$ in the surrounded \textquotedblleft bulk\textquotedblright\ from the total
data given at the bounding times, $\left\{ x\right\} \cup\left\{
f_{1}\left( x\right) \right\} $. \ In this point of view, \emph{fixed
points} in $x$ complete the boundary data and facilitate the solution of
Schr\"{o}der's equation through power series in $x$,\ hence leading to
$V\left( x\right) $ \emph{which was not known a priori} \cite{CZ}. \
In this holographic approach, the potential first appears as a quadratic in
$\Psi/\Psi^{\prime}$
\begin{equation}
V\left( x\right) =-\left( \ln s\right) ^{2}\left( \frac{\Psi\left(
x\right) }{\Psi^{\prime}\left( x\right) }\right) ^{2}\ , \label{GeneralV
\end{equation}
up to an additive constant, where any dependence of $\Psi$ on $s$ is again
implicit. \ The $x$ dependence of the potential,
\begin{equation}
V\left( x\right) \equiv-v^{2}\left( x\right) \ ,
\end{equation}
follows from that of the velocity profile of the interpolation, defined and
given by \cite{LogicalConnections
\begin{equation}
v\left( x\right) \equiv\left. \frac{df_{t}\left( x\right)
{dt}\right\vert _{t=0}=\left( \ln s\right) \Psi\left( x\right) \left(
\Psi^{-1}\left( \Psi\left( x\right) \right) \right) ^{\prime}=\frac{\ln
s}{\frac{d}{dx}\ln\Psi\left( x\right) }\ . \label{Generalv
\end{equation}
Monotonic motions between two fixed points, such as occur for the Ricker model
\cite{R}, provide the most elementary examples \cite{CZ}. \
However, situations where one or both of the fixed points are absent were not
fully addressed in our previous work. \ This is precisely the situation that
occurs when turning points are encountered at finite times in the particle
dynamics interpretation, and leads to an intriguing modification in the
physical picture involving the potential and its effect on the particle
trajectories. \ We consider here a specific example of such a situation
involving the well-studied logistic map of chaotic dynamics, namely
\textit{\cite{M,CE,F,K,C}}
\begin{equation}
x\mapsto sx\left( 1-x\right) \ . \label{LogisticMap
\end{equation}
Applying our functional methods to this example, the resulting interpolations
from a discrete lattice of time points to a time continuum then allow us to
appreciate analytic features of the logistic map, and to derive the governing
differential evolution laws --- indeed, subtly time-translation-invariant
Hamiltonian dynamical laws --- of the underlying physical system, hence to
obtain potentials that were \emph{not} previously known.
For the all-familiar logistic map illustrating transition to chaos, the
functional interpolation reveals that there are well-defined expressions for
the continuous time evolution of this map for \emph{all} parametric values of
the map, whether chaotic or not. \ We obtain agreement with the explicit
closed-form solutions of (\ref{SE}) for the special values of the parameters:
$s=-2,\ 2,$\ and\ $4$. \ (These explicit solutions have been known for almost
a century and a half \cite{S}.) \ \ Moreover, as indicated, from
(\ref{GeneralV})\ we can now find the potentials needed to produce these
explicit, continuously evolving trajectories in the language of Hamiltonian dynamics.
A new feature for the potentials so obtained for the $s=4$ case in particular,
and also for other values of $s>2$, is that they must change at discrete
intervals of the envisioned Hamiltonian particle's motion, to be consistent
with the evolution trajectories: \ \emph{Every time the particle hits a
turning point, the potential changes }\cite{Footnote1}. \ We therefore call
the corresponding $V$s \textquotedblleft switchback
potentials\textquotedblright\ and to echo our previous general remarks, we
will call the dynamical system as a whole \textquotedblleft
quasi-Hamiltonian.\textquotedblright\ \ For $s=4$ the switchback potentials
deepen successively and thus lead to more rapid, higher frequency motion,
resulting in the familiar chaotic behavior of the discrete $s=4$ logistic map
that they interpolate. \ The familiar frequency-cascade-and-folding behavior
of the chaotic discrete map is thus understood from --- indeed,
\emph{explained} by --- the subtle properties of the switchback potentials.
\section{The Logistic Map}
Consider in detail the logistic map (\ref{LogisticMap}) on the unit interval,
$x\in\left[ 0,1\right] $. \ Schr\"{o}der's equation for this map is \
\begin{equation}
s\Psi\left( x,s\right) =\Psi\left( sx\left( 1-x\right) ,s\right) \ .
\label{SFE
\end{equation}
The inverse function satisfies the corresponding Poincar\'{e} equation
\begin{equation}
\Psi^{-1}\left( sx,s\right) =s\Psi^{-1}\left( x,s\right) \left(
1-\Psi^{-1}\left( x,s\right) \right) \ . \label{PFE
\end{equation}
As originally obtained by Schr\"{o}der, there are three closed-form solutions
known, for $s=-2,$ $2,$ and $4$
\begin{align}
\Psi\left( x,-2\right) & =\frac{\sqrt{3}}{6}\left( 2\pi-3\arccos\left(
x-\frac{1}{2}\right) \right) \ ,\ \ \ \Psi^{-1}\left( x,-2\right)
=\frac{1}{2}-\cos\left( \frac{2x}{\sqrt{3}}+\frac{\pi}{3}\right)
\ ,\nonumber\\
\Psi\left( x,2\right) & =-\frac{1}{2}\ln\left( 1-2x\right)
\ ,\ \ \ \Psi^{-1}\left( x,2\right) =\frac{1}{2}\left( 1-e^{-2x}\right)
\ ,\nonumber\\
\Psi\left( x,4\right) & =\left( \arcsin\sqrt{x}\right) ^{2
\ ,\ \ \ \Psi^{-1}\left( x,4\right) =\left( \sin\sqrt{x}\right) ^{2}\ .
\label{ExactCases
\end{align}
Note that, while the $\Psi$ are multi-valued, the inverse functions are all
single-valued. \
More generally, consider a power series for any $s$
\begin{equation}
\Psi^{-1}\left( x,s\right) =x+x\sum_{n=1}^{\infty}x^{n}c_{n}\left(
s\right) \ . \label{PsiInverseSeries
\end{equation}
The Poincar\'{e} equation then leads to a recursion relation for the
$s$-dependent coefficients
\begin{equation}
c_{n+1}=\frac{1}{1-s^{n+1}}\sum_{j=0}^{n}c_{j}c_{n-j}\ ,
\label{PsiInverseSeriesCoefs
\end{equation}
with $c_{0}=1$, $c_{1}=\frac{1}{1-s}$, $c_{2}=\frac{2}{\left( 1-s\right)
\left( 1-s^{2}\right) }$, etc. \ The explicit coefficients are easily
recognized for $s=-2,$ $2,$ and $4$, and immediately yield the three
closed-form cases. \ Similarly let
\begin{equation}
\Psi\left( x,s\right) =x+x\sum_{n=1}^{\infty}\left( -x\right) ^{n
d_{n}\left( s\right) \ . \label{PsiSeries
\end{equation}
Then, as a consequence of Schr\"{o}der's equation, $d_{1}=1/\left(
1-s\right) $, and for $n\geq2$
\begin{equation}
d_{n}=\frac{1}{1-s^{n}}\sum_{k=1}^{\left\lfloor \frac{n+1}{2}\right\rfloor
}\binom{n+1-k}{k}s^{n-k}d_{n-k}\ . \label{PsiSeriesCoefs
\end{equation}
where $\left\lfloor \cdots\right\rfloor $ is the (integer-valued) floor
function. \ In principle, these series solve (\ref{PFE}) and (\ref{SFE})\ for
any $s$, within their radii of convergence. \
From extensive numerical studies we believe the radius of convergence for
(\ref{PsiSeries})\ depends on $s$ as follows
\begin{equation}
R_{\Psi}=\left\{
\begin{array}
[c]{ccc
\left\vert 1-\dfrac{1}{s}\right\vert & \text{if} & 0\leq s\leq2\ ,\\
& & \\
\dfrac{s}{4} & \text{if} & 2\leq s\leq4\ .
\end{array}
\right. \label{PsiSeriesRadius
\end{equation}
Preliminary work further suggests that these conjectural results can actually
be established analytically by comparison with known convergent series, at
least in some cases.
The radius of convergence for the inverse function series
(\ref{PsiInverseSeries}) is intriguing for $0\leq s\leq1$. \ A fit by
Mathematica to radii obtained numerically from explicit series coefficients
$c_{n}\left( s\right) $ up to $n=200$, for 20 different values of $s$ in
this range, using the functional form $\frac{1}{4}\left( 1-s\right) ^{a}\exp
bs$, gives $a=0.78$ and $b=0.62$. \ Thus $R_{\Psi^{-1}}\left( s\right) $
falls from $1/4$ to zero as $s$ goes from $0$ to $1$. \ However, for the
parameters of most interest, namely $2\leq s\leq4$, numerical studies suggest
the series all have \emph{infinite} radius of convergence, indicating the
$\Psi^{-1}$ are entire functions for these parameters. \ This is true for
$s=2$ and $s=4$, of course, as is evident from (\ref{ExactCases}). \ The
numerics provide compelling evidence that this is also true for all
intermediate $s$.
Given convergent series, the functional equations can then be used to continue
the series solutions outside their radii of convergence and thereby obtain
accurate determinations of the various branches of $\Psi$, and of the
large-but-finite argument behavior of $\Psi^{-1}$, in a manner familiar from,
say, the $\Gamma$ and $\zeta$ functions. \ (See \cite{CV}\ for details, and
the Appendix for a numerical example.) \ Comparison of the results from these
series-plus-functional-equation methods with the known exact solutions
(\ref{ExactCases}) shows perfect agreement.
The first few terms for $\Psi$ and $\Psi^{-1}$ for generic $s$ are given
explicitly by \cite{s=0
\begin{gather}
\Psi\left( x,s\right) =x+\frac{x^{2}}{s-1}+\frac{2s}{\left( s+1\right)
}\frac{x^{3}}{\left( s-1\right) ^{2}}+\frac{s\left( 1+5s^{2}\right)
}{\left( s^{2}+s+1\right) \left( s+1\right) }\frac{x^{4}}{\left(
s-1\right) ^{3}}\nonumber\\
+\frac{2s^{3}\left( 3+2s+7s^{3}\right) }{\left( s^{2}+1\right) \left(
s^{2}+s+1\right) \left( s+1\right) ^{2}}\frac{x^{5}}{\left( s-1\right)
^{4}}\nonumber\\
+\frac{2s^{3}\left( 1+3s+14s^{3}+14s^{4}+7s^{5}+21s^{7}\right) }{\left(
s^{4}+s^{3}+s^{2}+s+1\right) \left( s^{2}+1\right) \left( s^{2
+s+1\right) \left( s+1\right) ^{2}}\frac{x^{6}}{\left( s-1\right) ^{5
}+O\left( x^{7}\right) \ ,\\
\Psi^{-1}\left( x,s\right) =x+\frac{x^{2}}{1-s}+\frac{2}{\left( s+1\right)
}\frac{x^{3}}{\left( s-1\right) ^{2}}+\frac{5+s}{\left( s+1\right) \left(
s^{2}+s+1\right) }\frac{x^{4}}{\left( 1-s\right) ^{3}}\nonumber\\
+\frac{2\left( 7+3s+2s^{2}\right) }{\left( s^{2}+1\right) \left(
s^{2}+s+1\right) \left( s+1\right) ^{2}}\frac{x^{5}}{\left( s-1\right)
^{4}}\nonumber\\
+\frac{2\left( 21+14s+14s^{2}+8s^{3}+3s^{4}\right) }{\left( s^{4
+s^{3}+s^{2}+s+1\right) \left( s^{2}+1\right) \left( s^{2}+s+1\right)
\left( s+1\right) ^{2}}\frac{x^{6}}{\left( 1-s\right) ^{5}}+O\left(
x^{7}\right) \ ,
\end{gather}
from which we infer that $\Psi\left( x,s\right) /x$ and $\Psi^{-1}\left(
x,s\right) /x$ are actually series in $x/\left( 1-s\right) $ with
$s$-dependent coefficients that are analytic near $s=1$.
The trajectories interpolating the splinter (integer $t$) of the logistic map
are then
\begin{gather}
x\left( t\right) =\Psi^{-1}\left( s^{t}\Psi\left( x,s\right) ,s\right)
=s^{t}~x+\frac{s^{t}\left( 1-s^{t}\right) }{s-1}~x^{2}+\frac{2s^{t}\left(
1-s^{t}\right) \left( s-s^{t}\right) }{\left( s+1\right) \left(
s-1\right) ^{2}}~x^{3}\nonumber\\
+\frac{s^{t}\left( 1-s^{t}\right) \left( s-s^{t}\right) \left(
1+5s^{2}-\left( s+5\right) s^{t}\right) }{\left( s+1\right) \left(
s^{2}+s+1\right) \left( s-1\right) ^{3}}~x^{4}\nonumber\\
+\frac{2s^{t}\left( 1-s^{t}\right) \left( s-s^{t}\right) \left(
s^{2}-s^{t}\right) \left( 7s^{3}+2s+3-s^{t}\left( 2s^{2}+3s+7\right)
\right) }{\left( s+1\right) ^{2}\left( s^{2}+1\right) \left(
s^{2}+s+1\right) \left( s-1\right) ^{4}}~x^{5}+O\left( x^{6}\right) \ .
\end{gather}
The trajectories are single-valued functions of the time so long as $\Psi
^{-1}$ is single-valued, and in fact, they exist even for $s\rightarrow1$ as
formal series solutions. $\ $Explicitly,
\begin{gather}
\lim\limits_{s\rightarrow1}\Psi^{-1}\left( s^{t}\Psi\left( x,s\right)
,s\right) =x-t~x^{2}+t\left( t-1\right) ~x^{3}-\frac{1}{2}t\left(
t-1\right) \left( 2t-3\right) ~x^{4}\nonumber\\
+\frac{1}{3}t\left( t-1\right) \left( t-2\right) \left( 3t-4\right)
~x^{5}-\frac{1}{12}t\left( t-1\right) \left( t-2\right) \left(
12t^{2}-41t+31\right) ~x^{6}\nonumber\\
+\frac{1}{30}t\left( t-1\right) \left( t-2\right) \left( 30t^{3
-171t^{2}+302t-157\right) ~x^{7}+O\left( x^{8}\right) \ .
\end{gather}
For the three special cases,\ $s=-2,\ 2,$ and $4$, there are closed-form
results for various quantities of interest. \ For example, for $s=4$, the
trajectory and velocity are given b
\begin{equation}
\left. x\left( t\right) \right\vert _{s=4}=\Psi^{-1}\left( 4^{t
~\Psi\left( x,4\right) ,4\right) =\left( \sin\left( 2^{t}\arcsin\sqrt
{x}\right) \right) ^{2}\ , \label{ExactTrajectory4
\end{equation}
(as originally presented in \cite{S}, p 306) and b
\begin{subequations}
\begin{align}
\frac{d}{dt}x\left( t\right) & =\left( 2^{1+t}\ln2\right) \sin\left(
2^{t}\arcsin\sqrt{x}\right) \cos\left( 2^{t}\arcsin\sqrt{x}\right)
\arcsin\sqrt{x}\\
& =\left( \ln4\right) \sqrt{x\left( t\right) \left( 1-x\left( t\right)
\right) }\arcsin\sqrt{x\left( t\right) }\ .
\end{align}
The last expression evinces a continuous time-translational invariance.
\ However, this velocity function has branch points (i.e. turning points) at
$x\left( t\right) =0$ and $x\left( t\right) =1$, so some care is needed to
determine which branch of the function is involved, particularly when the
turning points are encountered at finite times. \ The system is therefore a
quasi-Hamiltonian one, as we have previously indicated, and as we shall
discuss in more detail.
With this caveat in mind, we may thus deduce the velocity profile $v\left(
x\right) =\left. \frac{dx\left( t\right) }{dt}\right\vert _{t=0}$, the
effective potential, and the force for the underlying quasi-Hamiltonian
system
\end{subequations}
\begin{align}
v\left( x\right) & =\left. \frac{dx\left( t\right) }{dt}\right\vert
_{s=4,t=0}=\left( \ln4\right) \sqrt{x\left( 1-x\right) }\arcsin\sqrt
{x}\ ,\\
V\left( x\right) & =-v^{2}\left( x\right) =\left( \ln4\right)
^{2}x\left( x-1\right) \arcsin^{2}\sqrt{x}\ ,\\
F\left( x\right) & =-\frac{d}{dx}V\left( x\right) =\left( \ln4\right)
^{2}\left( \arcsin\sqrt{x}\right) \left( \sqrt{x\left( 1-x\right)
}-\left( 2x-1\right) \arcsin\sqrt{x}\right) \ .
\end{align}
Note $x=0$ is an unstable fixed point for the system. \ Also, the system is
time-translationally invariant, so these same expressions hold at all times
everywhere along a trajectory, \emph{provided} due care is taken to determine
which branch of the various functions is in effect. Thus $V\left( x\left(
t\right) \right) =\left( \ln4\right) ^{2}x\left( t\right) \left(
x\left( t\right) -1\right) \arcsin^{2}\sqrt{x\left( t\right) }$, etc.
Similar closed-form results hold for the other two solvable cases, $s=-2$ and
$s=2$. \ In principle, there are also potentials underlying the trajectories
for other $s$ by dint of the series constructions. \ (We illustrate one other
case, $s=3$, in the Appendix.)
From the general expression for $x\left( t\right) $ in terms of $\Psi$ and
$\Psi^{-1}$, and with use of the chain rule, the potential may be expressed
entirely in terms of $\left( \ln\Psi\right) ^{\prime}$, as given\ in the
Introduction by (\ref{GeneralV}). \ The Schr\"{o}der auxiliary function is
then recognized as just an exponential of the time function
\begin{equation}
T\left( x\right) =\int^{x}\frac{dy}{\sqrt{-V\left( y\right) }}\ ,
\end{equation}
computed along a zero-energy trajectory, as given by (\ref{GeneralPsi}).
\section{Novel potentials and switchback effects}
The effective potentials for all three closed-form cases are somewhat unusual:
\ \
\begin{align}
V\left( x,s=4\right) & =\left( \ln4\right) ^{2}x\left( x-1\right)
\arcsin^{2}\sqrt{x}\ ,\\
V\left( x,s=2\right) & =-\left( \ln\sqrt{2}\right) ^{2}\left(
1-2x\right) ^{2}\ln^{2}\left( 1-2x\right) \ ,\\
V\left( x,s=-2\right) & =\frac{1}{36}\left( \ln\left( -2\right)
\right) ^{2}\left( 2x+1\right) \left( 2x-3\right) \left( 2\pi
-3\arccos\left( x-\frac{1}{2}\right) \right) ^{2}\ .
\end{align}
Another way to express the potential for $s=4$ is similar in form to that for
$s=-2$, namely
\begin{equation}
V\left( x,s=4\right) =\left( \ln2\right) ^{2}x\left( x-1\right) \left(
\pi-\arccos\left( 2x-1\right) \right) ^{2}\ .
\end{equation}
Indeed, it is well-known that the logistic maps for $s=4$ and $s=-2$ are
intimately related through the functional conjugacy of the underlying
Schr\"{o}der equations. \ But note the potential for $s=-2$ is in fact
complex, since $\ln\left( -2\right) =\ln2+i\pi$, as are the trajectories
under real time evolution for this case. \ Of course, since the complexity of
$V\left( x,s=-2\right) $\ is solely a multiplicative factor, if we switch to
complex time, $\tau=\left( \ln2+i\pi\right) \times t$, the
\begin{equation}
\frac{d^{2}x}{d\tau^{2}}=-\frac{\partial}{\partial x}\frac{V\left(
x,s=-2\right) }{2\left( \ln2+i\pi\right) ^{2}
\end{equation}
again yields real trajectories $x\left( \tau\right) $. \ As trivial as this
observation is, nevertheless it does raise several questions about complex
$x$, and about the behavior of $V$ in the complex plane. \ We discuss only one
aspect of this behavior here, and consider non-principal values for the
multi-valued functions $v\left( x,s=4\right) $ and $V\left( x,s=4\right)
$. \ In particular, \emph{all} branches of the $\arcsin$ function are
important to understand the behavior of the explicit trajectories
(\ref{ExactTrajectory4}). \
\textbf{Switchbacks on the road to chaos. \ }An interesting \emph{new}
feature, which we shall call the \emph{switchback effect}, appears for a
particle moving in the $s=4$ effective potential. \ This is a distinguishing
feature that we encounter for the $s=4$ chaotic map, but not, say, for the
non-chaotic $s=2$ map. \ The effect is traceable to the fact that, while the
trajectory is single-valued as a function of the time, the velocity and hence
the effective potential are \emph{not} single-valued as functions of
\emph{position}. \ Moreover, the branch points of the multi-valued functions
are encountered by zero-energy trajectories at \emph{finite} times for the
$s=4$ case, unlike, say, for the $s=2$ case. \ For the latter case, it takes
an infinite time for a particle with zero energy to reach a turning point.
Switchbacks are essentially transitions from one branch of the
position-dependent velocity function to another, and occur at the turning
points encountered at finite times by the interpolating particle trajectories,
(\ref{ExactTrajectory4}). \ The effect is easily seen upon viewing animations
of these trajectories \cite{Animations}.
Consider a particle with zero total energy that starts in the $V\left(
x,s=4\right) $ potential given above, in the region $0<x<1$. \ If the
particle is initially moving to the left, it will continue towards $x=0$,
taking an infinite amount of time to reach that turning point. \ But if the
particle is initially moving to the right, it will reach the $x=1$ turning
point in a \emph{finite }amount of time that depends on its initial $x$, as
given by
\begin{equation}
\Delta t_{0}\left( x\right) =\frac{1}{\ln4}\int_{x}^{1}\frac{dy
{\sqrt{y\left( 1-y\right) }\left( \arcsin\sqrt{y}\right) }=\frac{1}{\ln
2}~\ln\left( \frac{\pi/2}{\arcsin\sqrt{x}}\right) \ .
\end{equation}
So, for example, starting from the midpoint $x=1/2$, it takes unit time to
reach the turning point at $x=1$.
Upon reaching $x=1$, the explicit form of the time-dependent solution
(\ref{ExactTrajectory4}) exhibits the classical counterpart of\emph{ a sudden
transition} that keeps $E=0$, but \emph{changes the potential} for the return
trip towards $x=0$, exactly as follows from the particle moving on a different
branch of the $\arcsin$ function. \ Explicitly, the potential deepens t
\begin{equation}
V\left( x,s=4\right) \Longrightarrow V_{1}\left( x\right) =\left(
\ln4\right) ^{2}x\left( x-1\right) \left( -\pi+\arcsin\sqrt{x}\right)
^{2}\ ,
\end{equation}
with the particle's speed changing accordingly as a function of $x$. $\ $The
return velocity profile is now negative, and given b
\begin{equation}
v_{1}\left( x\right) =\left( \ln4\right) \sqrt{x\left( 1-x\right)
}\left( -\pi+\arcsin\sqrt{x}\right) \ .
\end{equation}
The $\arcsin$ in this last expression, as well as in $V_{1}$, is understood to
be the principal value.
Moving in this modified negative potential, it now takes the zero-energy
particle a \emph{finite} amount of time to travel from $x=1$ down to $x=0$, as
given by $\Delta t_{1}=1$. \ Upon reaching $x=0$, the exact solution
(\ref{ExactTrajectory4}) exhibits another sudden transition that keeps $E=0$,
but again alters the potential for the return trip towards $x=1$. \ The
potential become
\begin{equation}
V_{1}\left( x\right) \Longrightarrow V_{2}\left( x\right) =\left(
\ln4\right) ^{2}x\left( x-1\right) \left( \pi+\arcsin\sqrt{x}\right)
^{2}\ ,
\end{equation}
with corresponding changes in the velocity profile. \ Again, a finite amount
of time is needed for the particle to go from $x=0$ to $x=1$ in the potential
$V_{2}$, as given by $\Delta t_{2}=\ln\left( \frac{3}{2}\right) /\ln2$. \
Upon reaching $x=1$ again, the switchback process continues. \ The total
energy remains at zero, but the potential for the second return trip further
deepens
\begin{equation}
V_{2}\left( x\right) \Longrightarrow V_{3}\left( x\right) =\left(
\ln4\right) ^{2}x\left( x-1\right) \left( -2\pi+\arcsin\sqrt{x}\right)
^{2}\ ,
\end{equation}
etc. \ In order for the particle to follow the interpolating trajectory
(\ref{ExactTrajectory4}) specified by the Schr\"{o}der equation for the $s=4$
case, it has to move under the influence of successively deepening potentials. \
At later times, the effective potential seen by the particle on its zigzag
path between the turning points will depend on the total number of previous
encounters with those turning points. \ In this sense, the particle remembers
its past. \ Let $P$ be the total number of turning points previously
encountered on the particle's trajectory. \ The motion of the particle\ before
encountering the next, $\left( P+1\right) $st, turning point is then
completely determined, for $s=4$, by the effective potential (again, $\arcsin$
here is understood to be the principal value
\begin{equation}
V_{P}\left( x\right) =\left( \ln4\right) ^{2}x\left( x-1\right) \left(
\left( -\right) ^{P}\left\lfloor \frac{1+P}{2}\right\rfloor \pi+\arcsin
\sqrt{x}\right) ^{2}\ , \label{Indexed V
\end{equation}
where $\left\lfloor \cdots\right\rfloor $ is again the floor function. \ That
is to say, the potential deepens as $P$ increases. \ The corresponding
velocity profile speeds up:
\begin{equation}
v_{P}\left( x\right) =\left( \ln4\right) \sqrt{x\left( 1-x\right)
}\left( \left( -\right) ^{P}\left\lfloor \frac{1+P}{2}\right\rfloor
\pi+\arcsin\sqrt{x}\right) \ . \label{Indexed v
\end{equation}
The particle will either be traveling to the left, with $v_{P}\left(
x\right) <0$ for odd $P$, or traveling to the right, with $v_{P}\left(
x\right) >0$ for even $P$, with its speed increasing with $P$. \ As mentioned
previously, this effect is clearly seen upon viewing numerical animations of
the $s=4$ trajectories \cite{Animations}. \ It is also instructive to plot
$E\left( t\right) =v^{2}\left( x\left( t\right) \right) +V_{P}\left(
x\left( t\right) \right) $ versus $t$ for various $P$, with $t\geq
t_{P\text{ on}}$, to check $E=0$, as well as to see how the energy would
\emph{not} be conserved if the potentials were \emph{not} switched. \ See
Figure 1
\begin{figure}[ptb
\centering
\includegraphics[
height=2.0689in,
width=5.1664in
{1.eps
\caption{$E\left( t\right) $ for initial $x=1/2$, using potentials $V_{P}$,
with $P=0$, $1$, $2$, $3$, \& $4$.
\end{figure}
The time for the zero-energy particle to traverse the complete unit interval
in $x$ while moving through the $V_{P}$ potential (the $P$th \textquotedblleft
pseudo-half-cycle\textquotedblright) is always finite, for $P\neq0$
\begin{equation}
\Delta t_{P}=\left\vert \int_{0}^{1}\frac{dx}{v_{P}\left( x\right)
}\right\vert =\frac{1}{\ln2}~\left\vert \ln\left( 1+\frac{1}{2\left(
-\right) ^{P}\left\lfloor \frac{1+P}{2}\right\rfloor }\right) \right\vert
=\frac{1}{\ln2}\ln\left( \frac{P+1}{P}\right) \ .
\end{equation}
This \emph{transit time}\ decreases monotonically as $P$ increases, with
$\Delta t_{P}\underset{P\rightarrow\infty}{\sim}\frac{1}{P\ln2}$. \ Starting
from an initial $x$, with initial $v>0$, the times at which changes in the
potential occur, i.e. the times at which the particle encounters turning
points, are obtained by summing these transit times and then adding $\Delta
t_{0}\left( x\right) $. \ The transit time sums are simple enough
\begin{equation}
\sum_{N=1}^{P-1}\Delta t_{N}=\frac{\ln P}{\ln2}\ .
\end{equation}
Thus, a particle beginning at $x$, with initial $v>0$, will encounter its
$P$th turning point, and the potential $V_{P}$ will switch on, at time
$t_{P\text{ on}}\left( x\right) =\Delta t_{0}\left( x\right) +\frac{\ln
P}{\ln2}$. \ The potential $V_{P}$ will remain in effect for a time span equal
to the particle's transit time, $\Delta t_{P}=t_{P+1\text{ on}}\left(
x\right) -t_{P\text{ on}}\left( x\right) $, i.e. until the next potential
$V_{P+1}$ switches on. \
The potential index $P$ is in fact related to the action for the
pseudo-half-cycle \cite{Morse}. \ For even $P$, we have $I\propto\int_{0
^{1}\cdots$, while for odd $P$, we have $I\propto\int_{1}^{0}\cdots$. \ Thus
(in the units chosen, the mass is $m=2$
\begin{equation}
I_{P}=2\left( -1\right) ^{P}\int_{0}^{1}v_{P}\left( x\right) dx=\left(
4\left\lfloor \frac{1+P}{2}\right\rfloor +\left( -1\right) ^{P}\right)
I_{0}\ , \label{PthCrossingAction
\end{equation}
where for the first pass through the complete unit interval the action is
\begin{equation}
I_{0}\equiv2\int_{0}^{1}v\left( x\right) dx=\left( 4\ln2\right) \int
_{0}^{1}\sqrt{x}\sqrt{1-x}\arcsin\sqrt{x}dx=\frac{1}{8}\left( \ln2\right)
\pi^{2}\ .
\end{equation}
Moreover, $2\int_{0}^{1}\left( -v_{2k+1}\left( x\right) +v_{2k+2}\left(
x\right) \right) dx=8\left( k+1\right) I_{0}$ for integer $k$, so after
the initial crossing, the $\left( k+1\right) $st zigzag $1\rightarrow
0\rightarrow1$ (\textquotedblleft pseudo-full-cycle\textquotedblright)
contributes to the action: $\
\begin{equation}
I_{k+1\ \leftrightarrows}=8\left( k+1\right) I_{0}\ .
\end{equation}
The progressive deepening of the potential is also evident in the expanding
phase-space trajectories for the evolving particle, where the switchbacks
enhance the velocity after each encounter with a turning point ($v=0$). \ For
example, see Figure 2
\begin{figure}[ptb
\centering
\includegraphics[
height=3.0095in,
width=4.5363in
{2.eps
\caption{Phase-space trajectory $\left( x\left( t\right) ,v\left(
t\right) \right) $ for $s=4$, with $x\left( 0\right) =1/2$ and $v\left(
0\right) =\frac{1}{4}\pi\ln2$. \ The solid portion of the curve is for $0\leq
t\leq3$, while the dashed extension is for $3\leq t\leq3.5$. \ The dotted
portion of the curve is for negative times, $-\infty\leq t\leq0$, with
$x\left( -\infty\right) =v\left( -\infty\right) =0$.
\end{figure}
This Figure provides a good opportunity to clarify the differences between
conventional Hamiltonian and quasi-Hamiltonian systems. \ The apparent failure
to be a conventional Hamiltonian system is due entirely to the fact that a
given phase-space trajectory seems to repeatedly cross itself at two points
($x\left( t\right) =0$, and $x\left( t\right) =1$, as shown in the Figure)
and at various times. Thus a vector field on the phase space needed to
describe the motion in the conventional Hamiltonian formalism would appear to
be singular at those two points. \
These phase-space singularities are tied to the branch point and Riemann
surface sheet structure of the relevant analytic functions. \ In fact, an
alternate way to visualize the motion is as a trajectory on the sheets of a
Riemann surface. \ Consider the particle moving on the complex $x$ plane, and
not just the real line segment $\left[ 0,1\right] $, the endpoints of which
are now branchpoints. \ There are cuts from $+1$ to $+\infty$ and from $0$ to
$-\infty$. \ The particle first moves along the real axis, approaches the cut
at $+1$, and then goes around the branchpoint such that\ $\sqrt{x
\rightarrow\sqrt{x}$, $\sqrt{1-x}\rightarrow-\sqrt{1-x}$, and $\arcsin
\rightarrow\pi-\arcsin$. \ The particle returns along the real axis to the
origin and encircles the branchpoint at $0$, such that $\sqrt{x
\rightarrow-\sqrt{x}$, $\sqrt{1-x}\rightarrow\sqrt{1-x}$, and $\arcsin\sqrt
{x}\rightarrow-\arcsin\sqrt{x}$. \ The particle then goes back to $+1$ and
goes around it once more such that again $\sqrt{x}\rightarrow\sqrt{x}$,
$\sqrt{1-x}\rightarrow-\sqrt{1-x}$, and $\arcsin\rightarrow\pi-\arcsin$. \ The
trajectory continues in this way, flipping signs and adding $\pi$s according
to the formulae (\ref{Indexed V}) and (\ref{Indexed v}) \cite{Footnote2}.
For particles with initial $v>0$, a third way to picture the dynamics is in
terms of the total distance traveled by the particle. \ In this point of view,
the successive $V_{P}$ patch together to form a continuous potential $V\left(
X\right) $ on the real half-line, $X\geq0$, as shown in Figure 3. \ Indeed,
the half-line may be thought of as a \emph{covering manifold} of the unit
interval in $x$, with the previous multi-valued functions of $x$ now
single-valued functions of $X$. \ To avoid stagnation at any of the cusps of
$V\left( X\right) $, either physical \emph{or} conceptual, in this picture
the $E=0$ right-moving trajectories $X\left( t\right) $ may be considered as
limits of positive energy right-movers, $X\left( t\right) =\lim
_{E\downarrow0}X_{E}\left( t\right) $. \ This limiting process corresponds
to the encirclement of the branch points in the Riemann surface picture
\begin{figure}[ptb
\centering
\includegraphics[
height=3.0369in,
width=4.5645in
{3.eps
\caption{$V(X)$ where the total distance traveled by the particle is
$X-x\left( t=0\right) .$
\end{figure}
Also, an \textquotedblleft extended\textquotedblright\ phase space trajectory
involving $X\left( t\right) $ and $dX\left( t\right) /dt$ gives some
improvement over the situation depicted in Figure 2, with the trajectory no
longer self-intersecting. \ However, the corresponding phase-space vector
field would still be singular with cusps appearing in $dX\left( t\right)
/dt$ periodically in $X\left( t\right) $, similar to the cusps in the
potential $V\left( X\right) $. \ It is again possible to avoid the cusps by
considering the extended phase space trajectories as limits of positive energy
trajectories, but we leave that for the interested reader to pursue.
Finally, in terms of the coordinate transformation introduced in \cite{S},
namely, the angle
\begin{equation}
\theta=\arcsin\sqrt{x}\ ,
\end{equation}
the motion is completely unraveled, at least for $s=4$. \ In that case $E=0$
implies $\left( \frac{d\theta\left( t\right) }{dt}\right) ^{2}-\left(
\ln2\right) ^{2}\theta^{2}\left( t\right) =0$, and the trajectory for all
$t$ is just that of a particle moving in a repulsive quadratic potential
(inverted SHO). \ This is the easiest example solvable in closed-form by the
Schr\"{o}der method \cite{CZ}. \ The exponentially growing solution,
\ $\theta\left( t\right) =2^{t}\theta\left( 0\right) $, whose values in
the real-line cover of the circle are all physically distinct, immediately
leads to $x\left( t\right) $ for the particle moving through the various
$V_{P}\left( x\right) $, as is evident from (\ref{ExactTrajectory4}).
To amplify our discussion of the physics encoded in the switchback potentials,
we note that the chaotic behavior of the $s=4$ motion (i.e. the sensitive
dependence on initial conditions) follows entirely from the behavior of the
inverse Schr\"{o}der function. \ The sine function in (\ref{ExactTrajectory4})
discards any integer multiple of $2\pi$. \ So, for large and increasing times
the trajectory $\left. x\left( t\right) \right\vert _{s=4}$ only depends on
digits farther to the right of the decimal in the numerical value of the
initial $\arcsin\sqrt{x}$, hence it is sensitive to those initial conditions.
\ Of course, this latter fact is well-known in the context of the discrete
$s=4$ map, and has been discussed time and again in the literature for the
case where $t$ is an integer (for example, see \cite{K}). \ As is evident from
(\ref{ExactTrajectory4})\ the discussion carries over \emph{mutatis mutandis}
to the continuous time case. \ The only point we add here is to emphasize that
it is $\Psi^{-1}$ which is responsible for the effect. \ Insofar as $\Psi
^{-1}$ follows from $\Psi$ and the latter follows directly from the potential
(as well as the other way round, as discussed in our Introduction) the chaotic
behavior follows directly from the specific multi-valued form of the potential.
For other values of $s$, from (\ref{Trajectory}) and (\ref{Interpolation}) we
again have $x\left( t\right) =\Psi^{-1}\left( s^{t}\Psi\left( x,s\right)
,s\right) $, so the distinction between regular and chaotic behavior is again
determined entirely by the asymptotic behavior of $\Psi^{-1}$, hence the
asymptotic branch structure of $\Psi$, and hence the form of the switchback
potentials. \ The link between the analytic potential function and chaotic
behavior is arguably somewhat subtle, inasmuch as it requires some computation
to determine $\Psi$ and $\Psi^{-1}$, but the link is a direct one. \ Having
said that, there remains a case-by-case need for inspection of $\Psi^{-1}$ to
determine chaos or its absence. \ At this stage and in this paper, we do
\emph{not} purport to produce a universal analysis of all models --- we only
present methods that may facilitate such an analysis.
\section{Conclusion}
In summary, in this paper we\bigskip
--- Use a functional equation to convert nonlinear map equations into
dynamical systems in the continuous time domain, thereby facilitating the
study of such systems.
--- Illustrate this \emph{holographic} continuum interpolation technique with
the well-known model of the logistic map, explicitly for the chaotic case
$s=4$, an example which foreshadows an analysis for generic $s$.
--- Construct an intricate sequence of switchback Hamiltonians which drive the
system to just those phase-space trajectories resulting from such an analytic
interpolation of the logistic map, and produce the familiar chaotic behavior
for $s=4$.\bigskip
For nontrivial systems such as the chaotic $s=4$ logistic map illustrated
here, in order to produce the global features of the trajectories the
potential of the underlying Hamiltonian system \emph{must} change at each
switchback of the motion, as occurs when turning points are reached in finite
time. \ Therefore we have called these systems \emph{quasi-Hamiltonian}. \ The
trajectories dictate and completely determine the succession of switchback
potentials, in a richly extended analogy of inverse scattering techniques.
For values of $s$ between $2$ and $4$, a\emph{ }numerical analysis pursuant to
the techniques in this paper leads to quasi-Hamiltonian systems involving
switchbacks among sequences of potentials, eventually producing trajectories
that move between elaborate sets of turning points, in general. \ This
analysis is reported in full in a companion paper \cite{CV}. \ However, to
make the present paper more self-contained, and perhaps more convincing as to
the generality of the switchback effect, we have included some details for one
other case ($s=3$) in an Appendix.
To conclude, we encourage experimenters to look in various settings for the
particular type of continuous interpolates that we have described, including
fluid dynamical experiments, ecological systems, and, more generally, any
situation where an inherently continuous parameter (time, scale, etc.) has
been routinely sampled only at predetermined discrete values.
\begin{acknowledgments}
\textit{We thank D. Callaway for incisive questions and encouragement, D.
Sinclair for his emphasis on an appropriate title, and an anonymous reviewer
for suggestions to improve our presentation of the material. \ This work was
supported in part by NSF Award 0855386, and in part by the U.S. Department of
Energy, Division of High Energy Physics, under contract DE-AC02-06CH11357.}
\end{acknowledgments}
\section{Appendix: \ Some numerics for the $s=3$ case.}
The coefficients in the series expansion of $\Psi$ may be computed recursively
for any $s$, from (\ref{PsiSeriesCoefs}). \ While not immediately
recognizeable as the series expansion of any well-known function, nonetheless,
the coefficients may be used to estimate the radius of convergence of the
series (\ref{PsiSeries}) by the simple ratio test: \ $R_{\Psi}=\lim
_{n\rightarrow\infty}R_{\Psi}\left( n\right) $, with $R_{\Psi}\left(
n\right) \equiv\left\vert d\left( n-1\right) /d\left( n\right)
\right\vert $. \ To illustrate this consider the case $s=3$. \ For this case
numerical application of the ratio test leads to the data in Figure 4
\begin{figure}[ptb
\centering
\includegraphics[
trim=0.000000in 0.000000in 0.000000in 1.150459in,
height=2.1129in,
width=5.056in
{4.eps
\caption{$R_{\Psi}\left( n\right) -s/4$ for $s=3$, plotted for $2\leq
n\leq400$, to show convergence towards $R_{\Psi}=3/4$.
\end{figure}
From the numerical data we infer that $R_{\Psi}=3/4=\left. s/4\right\vert
_{s=3}$, as indicated earlier in (\ref{PsiSeriesRadius}). \ While this
numerical demonstration is not a proof of convergence, of course, we believe
it provides compelling evidence for convergence. \
Other values of $s$ produce numerical data qualitatively the same as the $s=3$
case, with only one noteworthy difference: \ While for $s>1$, the convergence
of the numerical data to the expressions in (\ref{PsiSeriesRadius}) is from
above, for $s<1$ the convergence is from below. \
A similar numerical calculation for the coefficients in the expansion of the
inverse function (\ref{PsiInverseSeries}) suggests $\Psi^{-1}\left( x\right)
$ is an entire function of $x$, for $s=3$, much like the closed-form results
in (\ref{ExactCases}). \ That is to say, $R_{\Psi^{-1}}=\infty$ for $s=3.$
Combining the series data for $\Psi$, $\Psi^{-1}$, and $V$ with the functional
equations obeyed by them leads to an accurate numerical determination of the
functions and the switchback potentials, for any $s$. \ The methods are
described in detail in \cite{CV}. \ Numerical results for\ $\left. \Psi
^{-1}\left( \frac{1}{2}~s^{t}\right) \right\vert _{s=3}$, with $-4\leq
t\leq9$, are shown in Figure 5 along with the expected asymptote, $\Psi_{\ast
}^{-1}=2/3$. \ This interval in $t$ corresponds to plotting $\Psi^{-1}\left(
x\right) $ for $0\lesssim x\lesssim10^{4}$
\begin{figure}[ptb
\centering
\includegraphics[
height=3.4147in,
width=5.056in
{5.eps
\caption{$\left. \Psi^{-1}\left( \frac{1}{2}~s^{t}\right) \right\vert
_{s=3}$, in blue, for $-4\leq t\leq9$, from series plus functional methods.
\end{figure}
Numerical results for the first few potentials in the sequence, for $s=3$, are
shown in Figure 6
\begin{figure}[ptb
\centering
\includegraphics[
height=3.7011in,
width=5.9502in
{6.eps
\caption{First four potentials in the sequence for $s=3$: \ $V_{0}$ red,
$V_{1}$ green, $V_{2}$ blue, and $V_{3}$\ orange.
\end{figure}
For this particular case, the sequence of potentials does not deepen
progressively, but rather the sequence converges, albeit very slowly, onto the
stable fixed point of the map, namely, $x_{\ast}=\left. 1-\frac{1
{s}\right\vert _{s=3}=2/3$, shown in the Figure as a small circle on the
$x$-axis. \ Hence the zero-energy particle trajectory will also converge onto
this fixed point, rigorously conserving $E=0$ in the process. \
Although the particle trajectory is not chaotic for $s=3$, nevertheless the
potentials are somewhat more exotic than the $s=4$ case in the sense that
there are an \emph{infinite} number of branch points for the $s=3$ case.
\ These may be obtained by iterating the action of the map on $x=1/2$. \ That
is to say, $\frac{1}{2}\rightarrow\frac{3}{4}=\allowbreak\frac{3}{2^{2
}\rightarrow\frac{9}{16}=\allowbreak\frac{3^{2}}{2^{4}}\rightarrow\frac
{189}{256}=\allowbreak3^{3}\frac{7}{2^{8}}\rightarrow\frac{37\,989
{65\,536}=\allowbreak3^{4}7\frac{67}{2^{16}}\rightarrow\cdots$. \ These branch
points are shown as the turning points of the potential sequence,
\begin{array}
[c]{ccccc
0.75, & 0.562\,5, & 0.738\,281\,, & 0.579\,666\,, & \cdots
\end{array}
$, represented by colored dots on the $x$-axis in Figure 6. \ The branch
points may also be seen by flipping the graph of $\Psi^{-1}$ in Figure 5 about
the SW-NE diagonal, in the usual way, to exhibit the various branches of
$\Psi$. \
This complicated branch structure for $s=3$ extends to other cases with
$2<s<4$ and most likely accounts not only for the failure to render them in
terms of simple, known functions, but also for some of the associated
wide-spread mystique that surrounds $\Psi$ and $\Psi^{-1}$ in these
cases.\ \ Of course, $s=2$ and $s=4$ only have branch points at $x=1/2$, and
at $x=0$ and $1$, respectively, as evident in the closed-form results
(\ref{ExactCases}) and as discussed in the text, so these cases are relatively
bland by comparison.
|
\section{Introduction}
We consider systems of two max-plus linear inequalities
\begin{equation}
\label{e:sysmain}
\begin{split}
a_{11}\otimes x_1\oplus\ldots\oplus
a_{1n}\otimes x_n &\leq
b_{11}\otimes x_1\oplus\ldots\oplus b_{1n}\otimes x_n,\\
a_{21}\otimes x_1\oplus\ldots\oplus a_{2n}\otimes x_n &\leq b_{21}\otimes x_1\oplus\ldots\oplus b_{2n}\otimes x_n.
\end{split}
\end{equation}
Here $\otimes:=+,$ $\oplus:=\max,$ and
$a_{ij},b_{ij},x_j\in\mathbb{R}\cup\{-\infty\}$ for $i=1,2$ and $j=1,\ldots,n$.
General systems of max-linear inequalities (equivalently, equalities)
were tackled by Butkovi\v{c} and Heged\"{u}s \cite{BH-84} who established
an elimination method for finding basic solutions of such systems,
starting with basic solutions of
just one equation or inequality and adding all other constraints one by one.
This algorithm served as a proof that solution sets to
max-linear systems have finite bases, and it did not
seem to be efficient enough for practical implementation.
But at present, Allamegeon, Gaubert and Goubault \cite{AGG-09} have come
up with a nouvelle approach to the scheme of \cite{BH-84}, in which
every step of adding new constraint is dramatically
improved by using a max-plus analogue of double description
method, and also a certain criterion of minimality estblished
in \cite{BSS-07,GK-07,NS-07II}, see also \cite{Ser-08}, which allows to efficiently test the extremality of a generator.
The idea of the present paper is that
when the number of inequalities is small,
the basic solutions can be written out explicitly.
However as shown by Wagneur, Truffet, Faye and Thiam
\cite{WTFT}, even in the case of two inequalities \eqref{e:sysmain}
the number of generators is large and the problem
to establish a systematic classification and
to resolve the extremality
by writing out explicit conditions is nontrivial. This goal
is achieved in the present paper by 1) representing the set of
all solutions as the union of cones generated
by certain Kleene stars (Section
\ref{s:gather}), 2) selecting basic solutions by means
of the above mentioned criterion of minimality \cite{BSS-07,GK-07,NS-07II} (Section \ref{s:extr}) which
we call the multiorder principle \cite{Ser-08}.
This leads to explicit description of basic solutions
and to a procedure which finds all of them in no more than $O(n^3)$
operations.
\section{Gathering the generators}
\label{s:gather}
We work with the analogue of linear algebra
developed over the max-plus semiring $\mathbb{R}_{\max,+}$ which is
the set of real numbers with adjoined minus infinity $\overline{\mathbb{R}}=\mathbb{R}\cup\{-\infty\}$ equipped with the operations
of ``addition'' $a\oplus b:=\max(a,b)$ and
``multiplication'' $a\otimes b:=a+b$.
Zero $\mathbf{0}$ and unity $\mathbf{1}$ of this semiring are equal,
respectively, to $-\infty$ and $0$.
The operations of the
semiring are extended to the nonnegative matrices and vectors in the
same way as in conventional linear algebra. That is if $A=(a_{ij})$,
$B=(b_{ij})$ and $C=(c_{ij})$ are matrices of compatible sizes with
entries from $\overline{\mathbb{R}}$,
we write $C=A\oplus B$ if $c_{ij}=a_{ij}\oplus b_{ij}$ for all $i,j$ and $%
C=A\otimes B$ if $c_{ij}=\bigoplus_k a_{ik}\otimes
b_{kj}=\max_{k}(a_{ik}\otimes b_{kj})$ for all $i,j$.
The notation $\otimes$ will be often omitted.
The main geometrical object of this max-plus linear algebra is a subset $K\in\RR^n$ closed under the operations of componentwise maximization $\oplus$ and ``multiplication''
$\otimes$ by scalars (which means addition in the conventional sense). Such subsets are called
{\em max-plus cones} or just {\em cones}
if there is no mix up with the ordinary convexity.
A vector $x\in\Rp^n$ is a {\em (max-linear) combination} of $y_1,\ldots, y^m\in S$. A set $S\subseteq\Rp^n$ is {\em generated} by
$y^1,\ldots,y^m\in S$ if each $x\in S$
is a max-linear combination of $y^1,\ldots,y^m$. When
vectors arise as columns (resp. rows) of matrices, it will
be convenient to represent them as max-linear combinations
of the {\em column unit vectors}
\begin{equation}
\label{col-units}
e_i=(\overbrace{\mathbf{0}\ldots \mathbf{0}}^{i-1}\ \mathbf{1}\ \mathbf{0}\ldots \mathbf{0})',
\end{equation}
respectively the {\em row unit vectors} $e'_i$, which are their
transpose.
The following series is called the {\em Kleene star} of $A$:
\begin{equation}
\label{def-kls}
A^*=I\oplus A\oplus A^2\ldots,
\end{equation}
where $I$ is the max-plus unity matrix, which has all diagonal entries $\mathbf{1}$ and all off-diagonal entries $\mathbf{0}$.
When $A^*$ has finite entries (in other words, converges)
it is easily shown that $A\otimes x\leq x$ is equivalent to
$A^*\otimes x=x$. We also have the following.
\begin{proposition}
\label{klstar-gen}
Let $A\in\Rp^n$ be such that $A^*$ has finite entries.
Then $\{x\mid A\otimes x\leq x\}$ is generated by the
columns of $A^*$.
\end{proposition}
This section will be based on the following two
observations. In the formulation we use the {\em row unit vectors} $e'_i$. We denote by $A_{i\cdot}$, resp. $A_{\cdot i}$,
the $i$th row, resp. the $i$th column, of $A$.
\begin{proposition}
\label{verysimple}
Let $A\in\Rp^{n\times n}$ have rows
\begin{equation}
\label{aicdot}
A_{i\cdot}=
\begin{cases}
e'_k\oplus \bigoplus_{l\neq k} a_{kl} e'_l, & \text{if $i=k$},\\
e'_i, \text{otherwise},
\end{cases}
\end{equation}
for $i=1,\ldots,n$. Then the set
$\{x\mid A\otimes x\leq x\}$ is generated by
the columns of $A$.
\end{proposition}
\begin{proof}
In this case $A^*=A$, after which Proposition \ref{klstar-gen} is applied.
\end{proof}
\begin{proposition}
\label{alsosimple}
Let $A\in\Rp^{n\times n}$ have rows
\begin{equation}
\label{aicdot2}
A_{i\cdot}=
\begin{cases}
e'_k\oplus \bigoplus_{l\in L_1} a_{kl} e'_l, & \text{if $i=k$},\\
e'_m\oplus \bigoplus_{l\in L_2} a_{ml} e'_l, & \text{if $i=m$},\\
e'_i, \text{otherwise},
\end{cases}
\end{equation}
where $L_1=\{l\neq k\mid a_{kl}\neq \mathbf{0}\}$,
$L_2=\{l\neq m\mid a_{kl}\neq \mathbf{0}\}$ and $k\neq m$.
\begin{itemize}
\item If $a_{km}a_{mk}\leq \mathbf{1}$ then $\{x\mid A\otimes x\leq x\}$
is generated by the columns of $A^*$.
\item If $a_{km}a_{mk}>\mathbf{1}$ then $\{x\mid A\otimes x\leq x\}$
is generated by $e_i$ for $i\notin L_1\cup L_2\cup\{k\}\cup\{m\}$.
\end{itemize}
\end{proposition}
\begin{proof}
In the first case $A^*$ is finite and we apply Proposition
\ref{klstar-gen}. For the second case observe that on one hand,
if $x_i\neq \mathbf{0}$ for some $i\in L_1\cup L_2\cup\{k\}\cup\{m\}$
then $x_k\neq \mathbf{0}$ and $x_m\neq \mathbf{0}$ which makes $A\otimes x\leq x$
impossible. On the other hand, any $x$ such that $x_i=\mathbf{0}$
for all $i\in L_1\cup L_2\cup\{k\}\cup\{m\}$ satisfies
$A\otimes x\leq x$.
\end{proof}
Now we proceed with writing out a generating set for solutions of
\eqref{e:sysmain}.
We denote $J_1:=\{i\mid a_{1i}\leq b_{1i},\ b_{1i}\neq \mathbf{0}\}$,
$J_2:=\{i\mid a_{2i}\leq b_{2i},\ b_{2i}\neq \mathbf{0}\}$, $I_1:=\{i\mid
a_{1i}>b_{1i}\}$ and $I_2:=\{i\mid a_{2i}>b_{2i}\}$. With this,
system \eqref{e:sysmain} is equivalent to
\begin{equation}
\label{e:sysred}
\begin{split}
\bigoplus_{i\in I_1} a_{1i}x_i& \leq \bigoplus_{i\in J_1} b_{1i} x_i,\\
\bigoplus_{i\in I_2} a_{2i}x_i& \leq \bigoplus_{i\in J_2} b_{2i} x_i.
\end{split}
\end{equation}
The solution set to \eqref{e:sysred} is the union of $S^{kl}$ defined by
\begin{equation}
\label{e:skldef}
S^{kl}=\{x\mid \bigoplus_{i\in I_1} a_{1i}x_i\leq b_{1k}x_k,\ \bigoplus_{i\in I_2} a_{2i}x_i\leq b_{2l}x_l\},
\end{equation}
for $k\in J_1$ and $l\in J_2$.
Further we represent $S^{kl}$ defined by \eqref{e:skldef}
in the form
\begin{equation}
\label{skl-eig}
S^{kl}=\{x\mid A^{kl}\otimes x\leq x\},
\end{equation}
where we have to describe $A^{kl}$. There are two cases: $k=l$ and $k\neq l$. We denote $\gamma_{ki}^1:=b_{1k}^{-1}a_{1i}$
and $\gamma_{ki}^2:=b_{2k}^{-1}a_{2i}$. We also denote
$\overline{I}_1:=\{1,\ldots,n\}\backslash I_1$ and
$\overline{I}_2:=\{1,\ldots,n\}\backslash I_2$. Observe
that $J_1\subseteq \overline{I}_1$ and
$J_2\subseteq \overline{I}_2$ (the containment may not be
strict in general).
If $k=l$, then the $k$th row of $A^{kl}$ is
\begin{equation}
\label{rows-akl1}
e'_k\oplus\bigoplus_{i\in I_1\cap\overline{I}_2} \gamma_{ki}^1 e'_i\oplus
\bigoplus_{i\in\overline{I}_1\cap I_2} \gamma_{ki}^2 e'_i\oplus\bigoplus_{i\in I_1\cap I_2} (\gamma_{ki}^1\oplus\gamma_{ki}^2)e'_i.
\end{equation}
and all other rows are row unit vectors.
If $k\neq l$ then the $k$th and the $l$th rows of $A^{kl}$ are
given by
\begin{equation}
\label{rows-akl2}
e'_k\oplus\bigoplus_{i\in I_1} \gamma_{ki}^1 e'_i, \quad e'_l\oplus\bigoplus_{i\in I_2}\gamma_{li}^2 e'_i,
\end{equation}
all other rows being row unit vectors.
Now we collect the generators of $\{x\mid A^{kl}\otimes x\leq x\}$
considering several special cases.
{\em Case 1.} $k=l\in J_1\cap J_2$.
The $k$th row of $A^{kl}$ is given by \eqref{rows-akl1} and
all other rows of $A^{kl}$ are unit vectors. By Proposition
\ref{verysimple}, $S^{kl}$ is generated by the
columns of $A^{kl}$. These are:
\begin{equation}
\label{cols-akl1}
\begin{split}
& e_i,\ i\in\overline{I}_1\cap\overline{I}_2,\\
&\gamma_{ki}^1e_k\oplus e_i,\ k\in J_1\cap J_2,\ i\in I_1\cap\overline{I}_2,\\
&\gamma_{li}^2e_l\oplus e_i,\ l\in J_1\cap J_2,\ i\in\overline{I}_1\cap I_2,\\
&(\gamma_{ki}^1\oplus\gamma_{ki}^2)e_k\oplus e_i,\ k\in J_1\cap J_2,
\ i\in I_1\cap I_2.
\end{split}
\end{equation}
{\em Case 2.} $k\neq l$, $k\in J_1\cap\overline{I}_2$, $l\in J_2\cap\overline{I}_1$.\\
Rows $k$ and $l$ of $A^{kl}$ are given by \eqref{rows-akl2},
all other rows being the unit vectors.
As $l\in\overline{I}_1$ and $k\in\overline{I}_2$, we obtain
$A^{kl}_{kl}=\gamma_{kl}^1=\mathbf{0}$ and $A^{kl}_{lk}=\gamma_{lk}^2=\mathbf{0}$
and hence $(A^{kl})^*=A^{kl}$.
Making transpose of \eqref{rows-akl2},
we obtain the columns of $(A^{kl})^*=A^{kl}$.
By Proposition \ref{alsosimple} part 1. they generate $S^{kl}$:
\begin{equation}
\label{cols-akl2}
\begin{split}
& e_i,\ i\in\overline{I}_1\cap\overline{I}_2,\\
& \gamma_{ki}^1e_k\oplus e_i,\ k\in J_1\cap\overline{I}_2,\ i\in I_1\cap\overline{I}_2,\\
& \gamma_{li}^2e_l\oplus e_i,\ l\in J_2\cap\overline{I}_1,\ i\in\overline{I}_1\cap I_2,\\
& \gamma_{ki}^1 e_k\oplus\gamma_{li}^2 e_l\oplus e_i,\
k\in J_1\cap\overline{I}_2,\ l\in J_2\cap\overline{I}_1,\ i\in I_1\cap I_2.
\end{split}
\end{equation}
{\em Case 3.} $k\in J_1\cap I_2$, $l\in J_2\cap\overline{I}_1$.\\
Rows $k$ and $l$ of $A^{kl}$ are given by
\eqref{rows-akl2}. However, $(A^{kl})^*\neq A^{kl}$, since
$k\in I_2$ implying that
$A_{lk}^{kl}=\gamma_{lk}^2\neq \mathbf{0}$.
Note that $(A^{kl})^*$ is always finite,
since $A_{kl}^{kl}=\mathbf{0}$ implying that the associated digraph of $A^{kl}$ does not contain any cycles with nonzero weight except
for the loops $(i,i)$.
For $i\in I_1$,
we obtain
$(A^{kl})^*_{li}=\gamma_{li}^2\oplus\gamma_{lk}^2\gamma_{ki}^1$.
More precisely,
$(A^{kl})^*_{li}=\gamma_{lk}^2\gamma_{ki}^1$ for $i\in I_1\cap\overline{I}_2$, and
$(A^{kl})^*_{li}=\gamma^2_{li}\oplus\gamma^2_{lk}\gamma_{ki}^1$
for $i\in I_1\cap I_2$.
The $l$th row of $(A^{kl})^*$ is given by
\begin{equation}
\label{rows-akl3}
e'_l\oplus\bigoplus_{i\in\overline{I}_2\cap I_1}
\gamma_{lk}^2\gamma_{ki}^1e'_i\oplus
\bigoplus_{i\in I_2\cap I_1}
(\gamma_{li}^2\oplus\gamma_{lk}^2\gamma_{ki}^1)e'_i
\oplus\bigoplus_{i\in\overline{I}_1\cap I_2}\gamma_{li}^2 e'_i.
\end{equation}
The $k$th row of $(A^{kl})^*$ is the same as in \eqref{rows-akl2} and all other rows are unit vectors.
We obtain the columns of $(A^{kl})^*$:
\begin{equation}
\label{cols-akl3}
\begin{split}
& e_i,\ \text{for $i\in\overline{I}_1\cap\overline{I}_2$}\\
& e_i\oplus\gamma_{ki}^1e_k\oplus \gamma_{lk}^2\gamma_{ki}^1 e_l,\
k\in J_1\cap I_2,\ l\in J_2\cap\overline{I}_1,\ i\in\overline{I}_2\cap I_1\\
& e_i\oplus\gamma_{ki}^1e_k\oplus(\gamma_{li}^2\oplus\gamma_{lk}^2\gamma_{ki}^1)e_l,\
k\in J_1\cap I_2,\ l\in J_2\cap\overline{I}_1,\ i\in I_1\cap I_2,\\
& e_i\oplus\gamma_{li}^2 e_l,\ l\in J_2\cap\overline{I}_1,\ i\in\overline{I}_1\cap I_2.
\end{split}
\end{equation}
{\em Case 4.} $k\in J_1\cap\overline{I}_2$, $l\in J_2\cap I_1$.\\
Rows $k$ and $l$ are given by \eqref{rows-akl2}, and by analogy with Case 3 we obtain
that the $l$th row of $(A^{kl})^*$ is the same as in \eqref{rows-akl2}, but the $k$th row
is given by
\begin{equation}
\label{rows-akl4}
\begin{split}
e'_k \oplus\bigoplus_{i\in\overline{I}_1\cap I_2} \gamma_{kl}^1\gamma_{li}^2e'_i\oplus
\bigoplus_{i\in I_2\cap I_1} (\gamma_{ki}^1\oplus\gamma_{kl}^1\gamma_{li}^2)e'_i
\oplus\bigoplus_{i\in\overline{I}_2\cap I_1}\gamma_{ki}^1 e'_i.
\end{split}
\end{equation}
We obtain the columns of $(A^{kl})^*$:
\begin{equation}
\label{cols-akl4}
\begin{split}
& e_i,\ i\in\overline{I}_1\cap\overline{I}_2\\
& e_i\oplus\gamma_{li}^2e_l\oplus \gamma_{kl}^1\gamma_{li}^2 e_k,\
k\in J_1\cap\overline{I}_2,\ l\in J_2\cap I_1,\ i\in\overline{I}_1\cap I_2\\
& e_i\oplus\gamma_{li}^2e_l\oplus(\gamma_{ki}^1\oplus\gamma_{kl}^1\gamma_{li}^2)e_k,\
k\in J_1\cap\overline{I}_2,\ l\in J_2\cap I_1,\ i\in I_1\cap I_2,\\
& e_i\oplus\gamma_{ki}^1 e_k,\
k\in J_1\cap\overline{I}_2,\ i\in\overline{I}_2\cap I_1.
\end{split}
\end{equation}
{\em Case 5.} $k\in J_1\cap I_2$, $l\in J_2\cap I_1$.\\
If $\gamma_{lk}^2\gamma_{kl}^{1}\leq \mathbf{1}$, then
the $l$th row of $(A^{kl} )^*$ is given by \eqref{rows-akl3}
and the $k$th row of $(A^{kl})^*$ is given by
\eqref{rows-akl4}. By Proposition \ref{alsosimple} part 1
the columns of $(A^{kl})^*$ generate $S^{kl}$.
If $\gamma_{lk}^2\gamma_{kl}^1>\mathbf{1}$, then by Proposition
\ref{alsosimple} part 2, $S^{kl}$
is generated by $e_i$ for $i\in\overline{I}_1\cap\overline{I}_2$.
If $\gamma_{lk}^2\gamma_{kl}^1\leq \mathbf{1}$, then
$(A^{kl})^*$ is finite and its columns are:
\begin{equation}
\label{cols-akl5}
\begin{split}
& e_i,\ i\in\overline{I}_1\cap\overline{I}_2,\\
& e_l\oplus\gamma_{kl}^1e_k,\ e_k\oplus\gamma_{lk}^2e_l,\
k\in J_1\cap I_2,\ l\in J_2\cap I_1,\\
& e_i\oplus\gamma_{li}^2e_l\oplus \gamma_{kl}^1\gamma_{li}^2 e_k,\
k\in J_1\cap I_2,\ l\in J_2\cap I_1,\ i\in\overline{I}_1\cap I_2,\\
& e_i\oplus(\gamma_{li}^2\oplus\gamma_{lk}^2\gamma_{ki}^1)e_l\oplus(\gamma_{ki}^1\oplus\gamma_{kl}^1\gamma_{li}^2)e_k,\
k\in J_1\cap I_2,\ l\in J_2\cap I_1,\ i\in I_1\cap I_2,\\
& e_i\oplus\gamma_{ki}^1e_k\oplus\gamma_{lk}^2\gamma_{ki}^1 e_l,\
k\in J_1\cap I_2,\ l\in J_2\cap I_1,\ i\in\overline{I}_2\cap I_1.
\end{split}
\end{equation}
\section{Identifying the basic solutions}
\label{s:extr}
A set $S\subseteq\Rp^n$ is said to be {\em independent}
if no vector in this set is generated by other vectors in this
set. If such independent set generates a cone $K$ then it is called
a {\em basis} of $K$. It can be shown \cite{BSS-07,Wag-91}
that if a basis of $K$ exists, then it consists of all
{\em extremals} (normalized in some sense): a vector
$x\in K$ is an extremal if $x=y\oplus z$ and $y,z\in K$ imply
$y=x$ or $z=x$. This also means that the basis of any cone is
essentially unique: any two bases are obtained
from each other by multiplying their
elements by scalars. Importantly,
any finitely generated cone has a basis \cite{BSS-07,Wag-91}.
The notion of extremal defined above is a max-plus analogue
of the notion of extremal ray (or extremal) of a convex cone. It is also a special case of the join irreducible element of a lattice.
The extremality is most conveniently expressed by the following
multiorder principle \cite{BSS-07,GK-07,NS-07II,Ser-08}
which we formulate
here only for finitely generated case. For any $i=1,\ldots,n$ we introduce the
relation
\begin{equation}
\label{leqi}
x\leq_i y\Rightarrow xx_i^{-1}\leq yy_i^{-1},\ \text{$x_i\neq \mathbf{0}$
and $y_i\neq \mathbf{0}$.}
\end{equation}
A vector $y\in K$ minimal with respect to $\leq_i$ will
be called {\em $i$-minimal}.
\begin{proposition}[Multiorder Principle]
Let $K\subseteq\Rp^n$ be generated by a finite
set $S\subseteq\Rp^n.$ Then $y\in S$ belongs to the basis
of $K$ (equivalently, is an extremal of $K$)
if and only if it is $i$-minimal for some $i\in\{1,\ldots, n\}$.
\end{proposition}
\begin{proof}
If $y$ is not $i$-minimal for any $i$, then for each
$i\in\operatorname{supp}(y)$ there exists $z^i$ such that $z^i\leq_i y$.
Then it can be verified that
\begin{equation}
\label{e:maxcomb11}
y=\bigoplus_{i\in\operatorname{supp}(y)} z^i (z_i^i)^{-1} y_i.
\end{equation}
Conversely if $y=\bigoplus_k \alpha_k z^k$ for some $z^k\in S$,
then for each $i\in\operatorname{supp}(y)$ there is $k(i)$ such that
$y_i=\alpha_{k(i)}z_i^{k(i)}$ and as
$y_j\geq \alpha_{k(i)} z_j^{k(i)}$ for all $j$ it follows
that $z^{k(i)}\leq_i y$ and $y$ is not $i$-minimal for any $i$.
\end{proof}
Next we classify all generators obtained in \eqref{cols-akl1},
\eqref{cols-akl2}, \eqref{cols-akl3},
\eqref{cols-akl4} and \eqref{cols-akl5} and give
procedures for checking their extremality.
We start with unit vectors and combinations of two unit
vectors.
\vskip 1cm
$S_1$. $e_i,$ $\forall i\in\overline{I}_1\cap\overline{I}_2$.\\
$S_{2A1}$. $\phi_{ik}=\gamma_{ki}^1 e_k\oplus e_i,$ $\forall k\in J_1\cap\overline{I}_2,$ $i\in I_1\cap\overline{I}_2$.\\
$S_{2A2}$. $\phi_{ik}=\gamma_{ki}^2 e_k\oplus e_i,$ $\forall k\in J_2\cap\overline{I}_1,$ $i\in I_2\cap\overline{I}_1$.\\
$S_{2B}$. $\phi_{ik}=(\gamma_{ki}^1\oplus\gamma_{ki}^2)e_k\oplus e_i,$ $\forall k\in J_1\cap J_2,$ $i\in I_1\cap I_2$.\\
$S_{2C}$. $\phi_{lk}=\gamma_{kl}^1e_k\oplus e_l$ and $\phi_{kl}:=\gamma_{lk}^2e_l\oplus e_k,$ $\forall k\in J_1\cap I_2$, $l\in J_2\cap I_1$
such that $\gamma_{kl}^1\gamma_{lk}^2\leq \mathbf{1}$.\\
All vectors in $S_1$, $S_{2A}$ and $S_{2B}$ belong to the basis.
Vectors in $S_{2C}$ belong to the basis
whenever they exist. For this, we determine the sets
\begin{equation}
\label{w-def}
\begin{split}
W&:=\{(k,l)\mid k\in J_1\cap I_2,\; l\in J_2\cap I_1,\;
\gamma_{kl}^1\gamma_{lk}^2\leq \mathbf{1}\}\\ \overline{W}&:=\{(k,l)\mid k\in J_1\cap
I_2,\; l\in J_2\cap I_1,\; \gamma_{kl}^1\gamma_{lk}^2>\mathbf{1}\}
\end{split}
\end{equation}
Then, $\phi_{kl},\phi_{lk}\in S_{3C}$ exist whenever $(k,l)\in W$.
Note that if $\gamma_{kl}^1\gamma_{lk}^2=\mathbf{1}$ then $\phi_{kl}$ and
$\phi_{lk}$ are multiples of each other so that one of them can be
removed.
We proceed with combinations of three unit vectors.
Denote $K_1=\{i\mid a_{1i}=b_{1i}=\mathbf{0}\}$ and
$K_2=\{i\mid a_{2i}=b_{2i}=\mathbf{0}\}$. Note that
$\{1,\ldots,n\}=I_1\cup J_1\cup K_1=I_2\cup J_2\cup K_2$.
\vskip 1cm
$S_{3A}$. $\psi_{ikl}=\gamma_{ki}^1e_k\oplus\gamma_{li}^2e_l\oplus e_i$ for $k\in J_1\cap\overline{I}_2$,
$l\in J_2\cap\overline{I}_1$, $i\in I_1\cap I_2$.\\
For all $i\in I_1\cap I_2$ determine the sets
\begin{equation}
\label{l-def}
\begin{split}
L_1(i)&:=\{k\in J_1\cap J_2\mid \gamma_{ki}^1<\gamma_{ki}^2\},\\
L_2(i)&:=\{l\in J_1\cap J_2\mid \gamma_{li}^2<\gamma_{li}^1\}.
\end{split}
\end{equation}
Then, $\psi_{ikl}\in S_{3A}$ belongs to the basis whenever
\begin{equation}
\label{3acond}
k\in(J_1\cap K_2)\cup L_1(i),\ l\in(J_2\cap K_1)\cup L_2(i).
\end{equation}
\vskip 1cm
\noindent $S_{3B1}$. $\psi_{ikl}=\gamma_{kl}^1\gamma_{li}^2e_k\oplus\gamma_{li}^2e_l\oplus e_i,$ $\forall k\in J_1\cap \overline{I}_2$, $l\in J_2\cap I_1$,
$i\in I_2\cap \overline{I}_1$.\\
$S_{3B2}$. $\psi_{ikl}=\gamma_{lk}^2\gamma_{ki}^1e_l\oplus\gamma_{ki}^1e_k\oplus e_i,$ $\forall k\in J_1\cap I_2$, $l\in J_2\cap\overline{I}_1$,
$i\in\overline{I}_2\cap I_1$.\\
For all $i\in I_2\cap\overline{I}_1$, $l\in J_2\cap I_1$, determine the sets
\begin{equation}
\label{mil-def}
M_1(i,l):=\{t\in J_1\cap J_2\mid \gamma_{tl}^1\gamma_{li}^2<\gamma_{ti}^2\}.
\end{equation}
For all $i\in I_1\cap\overline{I}_2$, $k\in J_1\cap I_2$, determine the sets
\begin{equation}
\label{mik-def}
M_2(i,k):=\{t\in J_1\cap J_2\mid \gamma_{tk}^2\gamma_{ki}^1<\gamma_{ti}^1\}.
\end{equation}
A vector in $\psi_{ikl}\in S_{3B1}$
(resp. $\psi_{ikl}\in S_{3B2}$)
belongs to the basis if and only if
the following two conditions are satisfied:\\
1. $i\in I_2\cap K_1$ or $(i,l)\in\overline{W}$
(resp. $i\in I_1\cap K_2$ or $(k,i)\in\overline{W}$),\\
2. $k\in M_1(i,l)$ or $k\in J_1\cap K_2$
(resp. $l\in M_2(i,k)$ or $l\in J_2\cap K_1$).
\vskip 1cm
\noindent
$S_{3C1}$. $\psi_{ikl}=(\gamma_{ki}^1\oplus\gamma_{kl}^1\gamma_{li}^2)e_k\oplus\gamma_{li}^2e_l\oplus e_i,$ $\forall k\in J_1\cap \overline{I}_2$, $l\in J_2\cap I_1$,
$i\in I_2\cap I_1$.\\
$S_{3C2}$. $\psi_{ikl}=(\gamma_{li}^2\oplus\gamma_{lk}^2\gamma_{ki}^1)e_l\oplus\gamma_{ki}^1e_k\oplus e_i,$ $\forall k\in J_1\cap I_2$,
$l\in J_2\cap\overline{I}_1$,
$i\in I_2\cap I_1$.\\
For all $i\in I_1\cap I_2$, $l\in J_2\cap I_1$, $k\in J_1\cap I_2$, determine the sets
\begin{equation}
\label{n-def}
\begin{split}
N_1(i,l):&=\{t\in J_1\cap J_2\mid \gamma_{ti}^1\oplus\gamma_{tl}^1\gamma_{li}^2<\gamma_{ti}^1\oplus\gamma_{ti}^2\}=\\
&=\{t\in L_1(i)\mid \gamma_{tl}^1\gamma_{li}^2<\gamma_{ti}^2\},\\
N_2(i,k):&=\{t\in J_1\cap J_2\mid \gamma_{ti}^2\oplus\gamma_{tk}^2\gamma_{ki}^1<\gamma_{ti}^1\oplus\gamma_{ti}^2\}=\\
&=\{t\in L_2(i)\mid \gamma_{tk}^2\gamma_{ki}^1<\gamma_{ti}^1\}.
\end{split}
\end{equation}
Then, $\psi_{ikl}\in S_{3C1}$ (resp. $\psi_{ikl}\in S_{3C2}$)
belongs to the basis if and only if
$k\in (J_1\cap K_2)\cup N_1(i,l)$
(resp. $l\in (J_2\cap K_1\cup N_2(i,k)$).
\vskip 1cm
$S_{3D1}$. $\psi_{ikl}=\gamma_{kl}^1\gamma_{li}^2e_k\oplus\gamma_{li}^2e_l\oplus e_i,$ $\forall k\in J_1\cap I_2$, $l\in J_2\cap I_1$,
$i\in I_2\cap \overline{I}_1$ such that $\gamma_{kl}^1\gamma_{lk}^2\leq \mathbf{1}$.\\
$S_{3D2}$. $\psi_{ikl}=\gamma_{lk}^2\gamma_{ki}^1e_l\oplus\gamma_{ki}^1e_k\oplus e_i,$ $\forall k\in J_1\cap I_2$, $l\in J_2\cap I_1$,
$i\in\overline{I}_2\cap I_1$ such that $\gamma_{kl}^1\gamma_{lk}^2\leq \mathbf{1}$.\\
$S_{3E}$. $\psi_{ikl}=(\gamma_{li}^2\oplus\gamma_{lk}^2\gamma_{ki}^1)e_l\oplus(\gamma_{ki}^1\oplus\gamma_{kl}^1\gamma_{li}^2)e_k\oplus e_i$,
$\forall k\in J_1\cap I_2$, $l\in J_2\cap I_1$, $i\in I_1\cap I_2$ such that $\gamma_{kl}^1\gamma_{lk}^2\leq \mathbf{1}$.\\
Provided that $(k,l)\in W$, vector $\psi_{ikl}\in S_{3D1}$
(resp. $\psi_{ikl}\in S_{3D2}$) belongs to the
basis if and only if $i\in K_1\cap I_2$ or
$(i,l)\in\overline{W}$
(resp. $i\in I_1\cap K_2$ or $(k,i)\in\overline{W}$), and
$\psi_{ikl}\in S_{3E}$ always belong to the basis.
Below we explain why the above procedure yields the basis.
We denote by $S_1$ the set of all generators $e_i$ for $i\in\overline{I}_1\cup\overline{I}_2$,
by $S_2$ the set of all $2$-generators $\phi_{ik}$ and $\phi_{kl}$, and by $S_3$ the set
of all $3$-generators $\psi_{ikl}$.
$S_1,S_2$:\\
The supports of all generators in $S_1\cup S_2$ are different, except for the
pairs of generators in $S_{2C}$, which exist if and only if $\gamma_{kl}^1\gamma_{lk}^2\leq \mathbf{1}$,
and are multiples of each other if and only if
$\gamma_{kl}^1\gamma_{lk}^2=\mathbf{1}$. Removing one vector from every such proportional
pair in $S_{2C}$ yields an independent set. Evidently, vectors in $S_1\cup S_2$ cannot be generated
with help of vectors in $S_3$, and this completes the explanation.
For the rest of the cases, first note that
the supports of all generators in $S_3$ are different and hence the set $S_3$ is independent.
It can only happen that the vectors in $S_3$ are linear combinations of the vectors in $S_1$ and $S_2$.
$S_{3A}$:\\
A vector $\psi_{ikl}\in S_{3A}$ may be a combination of vectors in $S_1$ and $S_{2B}$, as the
supports of some generators in these sets are contained in the support of a vector in $S_{3A}$.
By the minimality principle, a vector $\psi_{ikl}$ is extremal if and only if it is $i$-, $k$- or $l$-minimal.
Then, $\psi_{ikl}$ can be neither $k$- nor $l$-minimal since for all $k,l\in\overline{I}_1\cap\overline{I}_2$ the only
minimal generators are $e_k$ and $e_l$. The $i$-minimality of $\psi_{ikl}\in S_{3A}$ can be prevented only by
$\phi_{ki}\in S_{2B}$ or $\phi_{li}\in S_{2B}$. Condition \eqref{3acond} describes the situation when this
does not happen.
$S_{3B}$:\\
A vector $\psi_{ikl}\in S_{3B}$ can be a max combination of vectors in
$S_1$, $S_{2A}$ and $S_{2C}$ due to the inclusion of supports. Again, $\psi_{ikl}$ can
be neither $k$- nor $l$-minimal, since it can be represented as a combination
of $e_i$ and a vector from $S_{2A1}$ (resp. $S_{2A2}$) in the
case of $S_{3B1}$ (resp. $S_{3B2}$):
\begin{equation}
\begin{split}
\gamma_{kl}^1\gamma_{li}^2 e_k\oplus\gamma_{li}^2 e_l\oplus e_i&=\gamma_{li}^2(\gamma_{kl}^1e_k\oplus e_l)\oplus e_i.\\
\gamma_{lk}^2\gamma_{ki}^1 e_l\oplus\gamma_{ki}^1 e_k\oplus e_i&=\gamma_{ki}^1(\gamma_{lk}^2e_l\oplus e_k)\oplus e_i.
\end{split}
\end{equation}
Next we describe the 2-generators which can prevent the $i$-minimality of $\psi_{ikl}\in S_{3B1}$ (resp.
$\psi_{ikl}\in S_{3B2}$).\\
1. $\phi_{il},\phi_{li}\in S_{2C}$ (resp. $\phi_{ki},\phi_{ik}\in
S_{2C}$). These 2-generators do not arise
only if $i\in K_1$ for $S_{3B1}$ (resp. $i\in K_2$ for
$S_{3B2}$), for then there is no vector in $S_{2C}$ whose support is
a subset of the support of $\psi_{ikl}$, or if the corresponding
pair $\phi_{il},\phi_{li}\in S_{2C}$ (resp. $\phi_{ki},\phi_{ik}\in
S_{2C}$) does not exist meaning $(i,l)\in\overline{W}$ (resp. $(k,i)\in\overline{W}$).\\
2. $\phi_{ik}\in S_{2A2}$ (resp. $\phi_{il}\in S_{2A1}$).
These vectors do not arise only if $k\in K_2$ (resp. $l\in K_1$),
because then $k\notin J_2$ (resp. $l\notin J_1$) unlike in
the case of $S_{2A2}$ (resp. $S_{2A1}$). Otherwise, $\phi_{ik}$
(resp. $\phi_{il}$) are not dangerous, i.e., they do not precede
$\psi_{ikl}$ with respect to $\leq_i$ only if $k\in M_1(i,l)$
(resp. $l\in M_2(i,k)$), see \eqref{mil-def} and \eqref{mil-def}.
$S_{3C}$:\\
A vector $\psi_{ikl}\in S_{3C}$ can be a max combination of vectors in
$S_1$, $S_{2A}$ and $S_{2B}$. Again, $\psi_{ikl}$ can be neither $k$- nor $l$- minimal.
Indeed,
\begin{equation}
\label{case4-lindep}
\begin{split}
\psi_{ikl}&=\gamma_{ki}^1e_k\oplus e_i\oplus\gamma_{li}^2(\gamma_{kl}^1e_k\oplus e_l),\ S_{3C1}\\
\psi_{ikl}&=\gamma_{li}^2e_l\oplus e_i\oplus\gamma_{ki}^1(\gamma_{lk}^2e_l\oplus e_k),\ S_{3C2}
\end{split}
\end{equation}
where the vectors in brackets belong to $S_{2A1}$ and $S_{2A2}$ respectively. The first vector
cannot be $k$-minimal since $k\in\overline{I}_1\cap\overline{I}_2$, and it cannot be $l$-minimal as it to loses
$\gamma_{kl}^2e_k\oplus e_l\in S_{2A1}$.
The second vector cannot be $l$-minimal since $l\in\overline{I}_1\cap\overline{I}_2$, and it cannot be
$k$-minimal as it loses to $\gamma_{lk}^2e_l\oplus e_k\in S_{2A2}$. The remaining possibility of being
$i$-minimal can be destroyed by vectors from $S_{2B}$, and this does not happen if and only if
the given conditions are satisfied.
$S_{3D}$, $S_{3E}$:\\
A vector $\psi_{ikl}\in S_{3D}$ cannot be a
max combination of other vectors of $S_2$ than those
in $S_{2C}$. It is not a max combination of vectors in
$S_{2C}$ only if $i$ is not suitable for existence
of vectors in $S_{2C}$. This happens if
$i\in K_1\cap I_2$ or $(i,l)\in\overline{W}$ for the case $\psi_{ikl}\in S_{3D1}$,
and $i\in I_1\cap K_2$ or $(k,i)\in\overline{W}$ for the case $\psi_{ikl}\in S_{3D2}$.
Finally, the vectors in $S_{3E}$ cannot be combinations of vectors in $S_2$, since only vectors in $S_{2C}$ have relevant supports (and yet not enough). So the vectors in $S_{3E}$ are in the basis whenever they exist.
We note that the complexity of the above procedure id $O(n^3)$,
which is due to the computation of the sets $M_1(i,l)$
\eqref{mil-def}, $M_2(i,k)$ \eqref{mik-def}, $N_1(i,l)$ and
$N_2(i,k)$ \eqref{n-def}, and checking conditions for all
combinations of three unit vectors.
We conclude the paper with two examples. The second example is taken from \cite{WTFT}, Example 4.2.
{\em Example 1.} To illustrate the sets of generators
constructed in the paper on a simple example, we consider
the following system of two inequalities with four variables:
\begin{equation}
\label{e:sysex}
\begin{split}
& 4\otimes x_3\oplus 2\otimes x_4 \leq x_1\oplus 2\otimes x_2,\\
& 3\otimes x_1\oplus x_3 \leq x_2.
\end{split}
\end{equation}
We have $I_1=\{3,4\}$, $J_1=\overline{I}_1=\{1,2\}$,
$I_2=\{1,3\}$, $J_2=\{2\}$, $\overline{I}_2=\{2,4\}$.
We compute\\
$S_1$: just $e_2$, since $\overline{I}_1\cap\overline{I}_2=\{2\}$;\\
$S_{2A1}$: just $\gamma_{24}^1e_2\oplus e_4=e_2\oplus e_4$,
since $J_1\cap\overline{I}_2=\{2\}$ and $I_1\cap\overline{I}_2=\{4\}$;\\
$S_{2A2}$: just $\gamma_{21}^2e_2\oplus e_1=3e_2\oplus e_1$,
since $J_2\cap\overline{I}_1=\{2\}$ and $I_2\cap\overline{I}_1=\{4\}$;\\
$S_{2B}$: $(\gamma_{23}^1\oplus\gamma_{23}^2)e_2\oplus e_3=2e_2\oplus e_3$,
since $J_1\cap J_2=\{2\}$ and $I_1\cap I_2=\{3\}$;\\
$S_{2C}$: empty, since $J_2\cap I_1$ is empty;\\
$S_{3A}$: trivializes to $S_{2B}$;\\
$S_{3B1}$: empty, since $J_2\cap I_1$ is empty;\\
$S_{3B2}$: just $\gamma_{21}^2\gamma_{14}^1 e_2\oplus \gamma_{14}^1 e_1\oplus e_4=5e_2\oplus 2e_1\oplus e_4$,
since $J_1\cap I_2=\{1\}$, $J_2\cap \overline{I}_1=\{2\}$,
$\overline{I}_2\cap I_1=\{4\}$;\\
$S_{3C1}$: empty, since $J_2\cap I_1$ is empty;\\
$S_{3C2}$: just $(\gamma_{23}^2\oplus \gamma_{21}^2\gamma_{13}^1) e_2\oplus \gamma_{13}^1 e_1\oplus e_3$, which is $7e_2\oplus 4e_1\oplus e_3$,
since $J_1\cap I_2=\{1\}$, $J_2\cap \overline{I}_1=\{2\}$,
$I_2\cap I_1=\{3\}$;\\
$S_{3D1}$, $S_{3D2}$ and $S_{3E}$: empty, since $J_2\cap I_1$ is empty.
In this example, the basis consists of four generators in
$S_1$, $S_{2A1}$, $S_{2A2}$ and $S_{2B}$: $e_2$,
$e_2\oplus e_4$, $3e_2\oplus e_1$ and $2e_2\oplus e_3$.
Indeed, the remaining two generators in $S_3$ are redundant:
1) $5e_2\oplus 2e_1\oplus e_4$ ($S_{3B2}$) is a combination
of $e_2\oplus e_4$ ($S_{2A1}$) and $3e_2\oplus e_1$ ($S_{2A2}$),
2) $7e_2\oplus 4e_1\oplus e_3$ ($S_{3C2}$) is a combination
of $3e_2\oplus e_1$ ($S_{2A2}$) and $2e_2\oplus e_3$ ($S_{2B}$).
{\em Example 2.}
To compare our results with the approach of \cite{WTFT},
we consider \cite{WTFT}, Example 4.2, which is a system
of two inequalities with seven variables:
\begin{equation}
\label{e:sysex2}
\begin{split}
& x_4\oplus 4\otimes x_5\oplus 2\otimes x_6\oplus 6\otimes x_7\leq x_1\oplus 1\otimes x_2\oplus 5\otimes x_3,\\
& 5\otimes x_2\oplus 6\otimes x_3\oplus 2\otimes x_7\leq
3\otimes x_1\oplus x_4\oplus 2\otimes x_5\oplus 4\otimes x_6.
\end{split}
\end{equation}
In this case $I_1=\{4,5,6,7\}$, $J_1=\{1,2,3\}=\overline{I}_1$,
$I_2=\{2,3,7\}$, $J_2=\{1,4,5,6\}=\overline{I}_2$. We compute the
generators comparing them with those in the table of
\cite{WTFT} page 365:\\
$S_1:$ just $e_1$, since $\overline{I}_1\cap\overline{I}_2=\{1\}$. This is $x_1$ in the table of \cite{WTFT}.\\
$S_{2A1}:$ Combining $J_1\cap\overline{I}_2=\{1\}$ and
$I_1\cap\overline{I}_2=\{4,5,6\}$ we obtain $e_1\oplus e_4$, $4e_1\oplus e_5$ and $2e_1\oplus e_6$. Vector $e_1\oplus e_4$ corresponds to $x_3$, and the remaining two vectors are $x_5$ and $x_{10}$ in the table of \cite{WTFT}.\\
$S_{2A2}:$ Combining $J_2\cap \overline{I}_1=\{1\}$ and
$I_2\cap \overline{I}_1=\{2,3\}$ we obtain $2e_1\oplus e_2$ and $3e_1\oplus e_3$. These correspond to $x_4$ and $x_7$ in the table of \cite{WTFT}.\\
$S_{2B}:$ just $6e_1\oplus e_7$, combining
$J_1\cap J_2=\{1\}$ with $I_1\cap I_2=\{7\}$. This is
$x_2$ in the table of \cite{WTFT}.\\
$S_{2C}.$ To compute these we need to combine
$J_1\cap I_2=\{2,3\}$ with $J_2\cap I_1=\{4,5,6\}$. For each
$k=2,3$ and $l=4,5,6$ we need to check whether $\gamma_{kl}^1\gamma_{lk}^2\leq \mathbf{1}$, and each time
this condition is satisfied we have two vectors (or just
one vector if $\gamma_{kl}^1\gamma_{lk}^2=\mathbf{1}$). In our case
the condition is satisfied only with $k=3$ and $l=6$. This yields
two vectors $2e_6\oplus e_3$ and $e_3\oplus 3e_6$, which are
$x_6$ and $x_{11}$ in the table of \cite{WTFT}.\\
$S_{3A}:$ trivializes to $S_{2B}$.\\
$S_{3B1}:$ We need to combine $J_1\cap \overline{I}_2=\{1\}$,
$J_2\cap I_1=\{4,5,6\}$ and $I_2\cap\overline{I}_1=\{2,3\}$.
For $i=2,3$, $l=4,5,6$ and $k=1$, each time when
$\gamma_{il}^1\gamma_{li}^2>\mathbf{1}$, we have to verify
whether $\gamma_{kl}^1\gamma_{li}^2<\gamma_{ki}^2$ holds.
Each time when both conditions are satisfied, we have an independent vector of the basis. Here it never happens.\\
$S_{3B2}:$ We need to combine $J_1\cap I_2=\{2,3\}$,
$J_2\cap \overline{I}_1=\{1\}$ and $\overline{I}_2\cap I_1=\{4,5,6\}$.
For $k=2,3$, $i=4,5,6$ and $l=1$, each time when
$\gamma_{ki}^1\gamma_{ik}^2>\mathbf{1}$, we have to verify
whether $\gamma_{lk}^2\gamma_{ki}^1<\gamma_{li}^1$ holds.
Each time when both conditions are satisfied, we have an independent vector of the basis. Here it happens with
1) $l=1$, $k=3$ and $i=4$ leading to $3e_1\oplus e_3\oplus 5e_4$
which corresponds to $x_8$ of \cite{WTFT}, 2)
$l=1$, $k=3$ and $i=5$ leading to $3e_1\oplus e_3\oplus 1e_5$,
which corresponds to $x_9$ of \cite{WTFT}. \\
$S_{3C1}:$ Here we combine $J_1\cap\overline{I}_2=\{1\}$ with
$J_2\cap I_1=\{4,5,6\}$ and $I_1\cap I_2=\{7\}$. Since $\gamma_{ki}^1>\gamma_{ki}^2$ with $k=1$ and $i=7$, no vector belongs to the basis
in this case.\\
$S_{3C2}:$ We combine $J_1\cap I_2=\{2,3\}$, $J_2\cap\overline{I}_1=\{1\}$
and $I_1\cap I_2=\{7\}$. For each $k=2,3$, $l=1$ and $i=7$
we have to verify $\gamma_{li}^2\oplus \gamma_{lk}^2\gamma_{ki}^1<\gamma_{li}^2\oplus \gamma_{li}^1$.
This happens for $k=3$, $l=1$ and $i=7$ and yields the vector
$4e_1\oplus 1e_3\oplus e_7$, which corresponds to $x_{12}$
of \cite{WTFT}.\\
$S_{3D1}:$ We combine $J_1\cap I_2=\{2,3\},$
$J_2\cap I_1=\{4,5,6\}$, $I_2\cap\overline{I}_1=\{2,3\}$.
For $k=2,3$ and $l=4,5,6$, the condition
$\gamma_{lk}^1\gamma_{kl}^2\leq\mathbf{1}$ holds only for
$k=3$ and $l=6$, so it remains to verify $\gamma_{il}^1\gamma_{li}^2>\mathbf{1}$ for $i=2$ and $l=6$.
This condition holds and we obtain $\gamma_{36}^1\gamma_{62}^2e_3
\oplus\gamma_{62}^2 e_6\oplus e_2$ which is proportional with
$2e_2\oplus e_3\oplus 3e_6$.
Note that the max-linear combination of $e_2,e_3,e_6$ given for $x_{13}$ in the table of \cite{WTFT} is an error, since $Ax_{13} \not\le Bx_{13}$. \\
$S_{3D2}:$ We combine $J_1\cap I_2=\{2,3\},$
$J_2\cap I_1=\{4,5,6\}$, $\overline{I}_2\cap I_1=\{4,5,6\}$.
For $k=2,3$ and $l=4,5,6$, the condition
$\gamma_{lk}^1\gamma_{kl}^2\leq\mathbf{1}$ holds only for
$k=3$ and $l=6$, so it remains to verify $\gamma_{ki}^1\gamma_{ik}^2>\mathbf{1}$ for $i=4,5$ and $k=3$.
This condition holds in both cases and yields $\gamma_{63}^2\gamma_{34}^1e_6
\oplus\gamma_{34}^1 e_3\oplus e_4$ which is proportional with
$2e_6\oplus e_3\oplus 5e_4$, and $\gamma_{63}^2\gamma_{35}^1e_6
\oplus\gamma_{35}^1e_3\oplus e_5$ proportional with
$2e_6\oplus e_3\oplus 1e_5$.\\
$S_{3E}:$ We combine $J_1\cap I_2=\{2,3\},$ $J_2\cap I_1=\{4,5,6\}$ and $I_1\cap I_2=\{7\}$. As
$\gamma_{kl}^1\gamma_{lk}^2<\mathbf{1}$ only for $k=3$ and $l=6$,
we have only one generator, namely $3e_6\oplus 1e_3\oplus e_7$.
\vskip 1cm
Thus the basis consists of $e_1,$ $8$ combinations of $2$ unit vectors and $7$ combinations of $3$ unit vectors.
The $2$-combinations are: $e_1\oplus e_4$, $4e_1\oplus e_5$ and $2e_1\oplus e_6$ ($S_{2A1}$), $2e_1\oplus e_2$ and $3e_1\oplus e_3$ ($S_{2A2}$),
$6e_1\oplus e_7$ ($S_{2B}$), $2e_6\oplus e_3$ and
$e_3\oplus 3e_6$ ($S_{2C}$).
The $3$-combinations are: $3e_1\oplus e_3\oplus 5e_4$,
$3e_1\oplus e_3\oplus 1e_5$ ($S_{3B2}$), $4e_1\oplus 1e_3\oplus e_7$ ($S_{3C2}$), $2e_2\oplus e_3\oplus 3e_6$ ($S_{3D1}$),
$2e_6\oplus e_3\oplus 5e_4$ and $2e_6\oplus e_3\oplus 1e_5$
($S_{3D2}$), $3e_6\oplus 1e_3\oplus e_7$ ($S_{3E}$).
We note that all vectors that we have found, are solutions of the
system, and moreover, all $3$-generators turn both inequalities into equalities, which in analogy with the convex analysis also suggests that they must be extremals (the $2$-generators correspond to the intersections with coordinate planes). Actually vectors in $S_{3B2}$ and $S_{3C2}$ are different from
$x_8,x_9$ and $x_{12}$ from the table of \cite{WTFT} page 365,
to which they correspond in terms of supports. For these, $x_8=4e_1\oplus e_3\oplus 4e_4$
is a combination of $3e_1\oplus e_3\oplus 5e_4$ (from $S_{3B2}$),
$3e_1\oplus e_3$ (from $S_{2A2})$ and $e_1$,
$x_9=4e_1\oplus 1e_3\oplus e_5$ is a combination of $3e_1\oplus e_3\oplus 1e_5$ (from $S_{3B2}$)
and $3e_1\oplus e_3$, and $x_{12}=5e_1\oplus 1e_3\oplus e_7$ is a
combination of $4e_1\oplus 1e_3\oplus e_7$ (from $S_{3C2}$)
and $e_1$. The remaining generator in the table of
\cite{WTFT} is $x_{13}=e_2\oplus 2e_3\oplus 1e_6$. It is in error, since it violates the second inequality of \eqref{e:sysex2}, but in terms of support, it corresponds to
$2e_2\oplus e_3\oplus 3e_6$ from $S_{3D1}$. Also, there are
three combinations which are not in the table of \cite{WTFT},
from $S_{3D2}$ and $S_{3E}$.
|
\section{Introduction}
The Sagittarius (Sgr) dwarf galaxy is the closest Milky Way companion
($\sim 24$ kpc from the sun) and is currently being disrupted in the
Galaxy's tidal field. Both leading and trailing tidal streams have
been detected using many different tracer populations, including A
stars~\citep{Ya00}, RR Lyraes~\citep{Vi01,Wa09}, red clump
stars~\citep{Ma98,Ma99} and main sequence turn-off
stars~\citep{MD01,Ne02}. Perhaps the most spectacular views of the Sgr
stream are given by the M giants (Majewski et al. 2003). A portion of
the stream around the North Galactic Cap has been traced to much
fainter magnitudes using data from the Sloan Digital Sky Survey (SDSS,
see Belokurov et al. 2006a ``The Field of Streams'') revealing a
prominent bifurcation into the so-called streams A and B. There has
been considerable effort in modelling the disruption of Sgr since the
stream stars are useful tracers for the shape of the Milky Way's dark
halo potential. Nonetheless, this work has yielded ambiguous results.
Depending on which datasets are fit, oblate~\citep{Jo05},
prolate~\citep{He04}, spherical~\citep{Fe06}, or even
triaxial~\citep{La09} halos are preferred.
One reason for this ambiguity is that the structure, size and nature
of the Sgr progenitor are very uncertain. The original luminosity and
mass of Sgr is larger than its luminosity and mass today by
amounts depending on the mass density (including any dark matter) and
orbital history of the system. A neat illustration of this is provided
by the work of Jiang \& Binney (2000), who used N-body simulations to
show that the current configuration of the core of Sgr can be produced
by a very wide variety of initial conditions. At one extreme, the Sgr
dwarf might initially possess a mass in excess of $\sim 10^{11} {\Msun}$
and fall into the Galaxy from a distance in excess of 200 kpc. At the
other extreme, it might be only $\sim 10^9 {\Msun}$ and
start off at distances similar to its present apocenter of 60 kpc.
In this paper, we analyse the number of stars and the luminosity of
the present-day Sgr core and debris. By re-assembling Sgr, we provide
new constraints on its parameters prior to disruption. We describe
the data sets used for this study in \S 2 and analyse the
colour-magnitude distributions of the stream and core as well as the
luminosity density along the leading and trailing arms in \S 3. We
discuss the implications of our work for estimating the mass of
Sagittarius and conclude in \S 4.
\section{Data}
We trace the Sgr stream using main sequence, red giant and horizontal
branch stars from the SDSS (York et al. 2000), together with M giants
in the Two Micron All-Sky Survey (2MASS). SDSS is an imaging and
spectroscopic survey that covers one-quarter of the celestial sphere
including an area of $\sim8400$ deg$^2$ around the North Galactic Cap,
which contains a portion of the leading arm of the Sgr stream. There
are also some imaging stripes that sparsely sample low Galactic
latitude fields ($\sim3200$ deg$^2$) which contain parts of the
trailing arm. The imaging data are collected in five passbands
($u,g,r,i,z$), for which there are model isochrones~\citep{Gi04,Do08}
and photometric
transformations~\footnote{http://www.sdss.org/dr4/algorithms/sdssUBVRITransform.html}
readily available. Here, we use positions and magnitudes of stars
extracted from Data Release 7~\citep{Ab09} and corrected for
extinction using the maps of \cite{Sc98}. The bright limit in the
$u$-band is about 14 mag, whilst the faint limit is about 22. The
final SDSS sample contains $\sim 3,740,000$ stars along the
Sagittarius stream and $\sim 3,200,000$ stars in a comparison field in
the North Galactic Cap region as well as $\sim 294,000$ stars along the stream and
$\sim 255,000$ stars in the comparison fields in the southern
stripes. We additionally use the positions and $J,H,K$ magnitudes of
$\sim 1,450,000$ stars in the 2MASS All-Sky Point Source Catalog. This
is complete down to roughly $K=15$ mag.
Finally, we use a dataset from \cite{Bel06} to study the properties of
the Sgr core. It comprises of two fields ($\sim 299,000$ stars on and
$\sim 57,000$ stars off the Sgr core) in two passbands $B$ and $V$. The
sample goes approximately $3$ magnitudes below the main sequence
turn-off and includes stars up to the red clump, blue horizontal
branch and M giants.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figure1.eps}
\caption{Left: M giants selected from 2MASS
tracing the Sgr stream in Galactic coordinates. The box marks
the central parts of the Sgr dwarf and the arrow indicates
the direction of motion of the dwarf. The leading and
trailing arms are visible at positive and negative Galactic
latitude respectively. The gray-scale shows the SDSS
footprint. Right: Zoom-in of the area outlined by the box,
with the fields of \citet{Bel06} marked with squares. The
photometric center of Sgr is marked with a star.}
\label{fig:2mass_boxes}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figure2.eps}
\caption{``Field of Streams'' \citep{Be06} with the Sgr debris
bifurcated into a lower (A) and an higher (B) declination
stream. Selection boxes along the stream (used in \S 3.3)
are shown, together with comparison boxes at the same
Galactic latitude but reversing the sign in longitude,
(i.e., effectively reflecting the boxes in
$\ell=180^\circ$). The stream boxes are $5^\circ$ wide in
right ascension and are chosen to cover the entire width of
the stream in declination.}
\label{fig:fos_boxes}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figure3a.ps}
\includegraphics[width=\textwidth]{figure3b.ps}
\caption{Upper: Density of SDSS BHBs in the plane of distance
versus right ascension for the Sgr stream (left), comparison
field (middle) and their difference (right). Note the
prominent linear feature associated with streams A and B,
which is shadowed by the blue stragglers some 2 magnitudes
fainter.The solid lines bracket the distance gradient in
Sgr's BHBs. Lower: Comparison of distances to the Sgr stream
in 2MASS (open symbols) and SDSS (filled symbols, from
Belokurov et al. 2006) datasets with the uncertainty in
distance shown in gray. The leading arm is marked with
circles, the trailing with triangles. Note the good match
between distances derived from different indicators. Arrows
mark locations of data-set overlaps.}
\label{fig:distances}
\end{figure*}
\section{All-Sky Views of the Sgr Stream}
We begin by giving an overview of different observations of Sgr and
the terminology that we will employ throughout the paper. We refer to
the still bound portion of the dwarf as the core or remnant, which we
define as stars within 10 kpc of the center. These stars follow a King
profile (see e.g., Majewski et al. 2003). The luminosity of Sgr is
usually quoted as the light integrated inside this radius, but it is
not obvious that all of these stars are still gravitationally
bound. The tidal debris from the disruption wraps around the Galaxy
(possibly multiple times) in what is usually referred to as the
leading and trailing tidal streams or arms. The portion of the debris
that is seen around the North Galactic Cap is bifurcated into stream A and stream
B. The working hypothesis for this paper is that both stream A and
stream B are part of the leading arm. Part of the leading arm is also
detected in the 2MASS survey. The trailing arm is in part covered by
the southern stripes of the SDSS and more fully covered in 2MASS.
The best panorama of the Sgr stream is given by M giants selected from
2MASS, as first realized by \citet{Ma03}. Their choice of cuts on
$J,H,K$ magnitudes picks out bright red stars at the typical distances
of Sgr. For Figure~\ref{fig:2mass_boxes}, and in the following
analysis, we adopt the selection criteria introduced by Majewski et
al. (2003), namely
\begin{eqnarray}
\label{eq:thecuts}
J-K>0.85 \nonumber \\
J-H<0.561(J-K)+0.36 \nonumber \\
J-H>0.561(J-K)+0.22\\
|b| > 10^\circ \nonumber\\
E(B-V)<0.555\nonumber
\end{eqnarray}
Additionally, for the left panel of Figure~\ref{fig:2mass_boxes} only,
we remove some of the reddest stars with $J-K>1.25$ and limit the
magnitude (heliocentric distance) range to $10 <K<12$ ($7 < D < 70$
kpc) for the trailing arm and to $11 < K <13$ ($5 < D < 110$ kpc) for
the leading arm. This is in order to isolate the tidal debris of the
Sgr. Notice too that the Magellanic Clouds are very
prominent. Superficially, Sgr already looks to be intermediate
between the Large and Small Cloud in size and luminosity.
The right panel of Figure~\ref{fig:2mass_boxes} shows a zoom-in of the
Sgr core, made using the cuts of eqn~(\ref{eq:thecuts}), together with
an additional color-magnitude (CMD) box. This is derived by shifting
the ridgeline of Sgr M giants ($K=-8.650(J-K)+20.374$, Equation 5 of
Majewski et al. 2003) by $\pm3.5 \sigma$ or $\pm1.26$ mag in the color
range $0.95<J-K<1.1$. Using this selection gives a clear view of the
Sgr core, the photometric center of which is marked. Also shown are
the on-core and off-core fields of \citet{Bel06}.
The SDSS view of the Sgr stream is confined primarily to the North
Galactic Cap and hence is shown in equatorial coordinates in
Figure~\ref{fig:fos_boxes} to minimize distortion. The ``Field of
Streams''~\citep{Be06} shows the Sgr stream over a 100$^\circ$ arc,
with an unprecedented level of detail. The datasets provided by 2MASS
and SDSS are complementary -- the 2MASS survey picks up the trailing
arms with good clarity, but it loses the leading arm around the North
Galactic Cap.
\subsection{Distances}
There is information about the distances to debris at a number of
points along the stream~~\citep{Be06,Ne07,Wa09}. Heliocentric
distances to the trailing arm were originally computed by
\citet{Ma03}, who showed that there was little distance gradient along
this part of the stream. There has been more controversy about
distances to the leading arm, mainly because there are much steeper
distance gradients.
Given the crucial role that distances play, we begin by seeking
independent confirmation of earlier results. For the leading arm in
the Field of Streams, we select blue horizontal branch (BHB)
candidates from the SDSS in the on and off-stream fields using the
criteria given in Figure 10 of \citet{Ya00}. This gives us $\sim
18700$ candidates in the Sgr stream and $\sim 8500$ candidates in the
comparison field. We note that these numbers do not include all
genuine BHBs, and suffer from some contamination from blue stragglers
(BS). The upper panels of Figure~\ref{fig:distances} show the density
of BHBs in the two fields, together with their difference. The
distances to BHBs are calculated assuming an absolute magnitude of
$M_g = 0.7$, which is appropriate for our color range ($u-g > 0.9$,
see Table 2 of \cite{Si04}). Running diagonally across the plot is the
signal from the BHBs in the Sgr stream. It is accompanied by a fuzzy
shadow of BSs two magnitudes fainter. The Galactic halo foreground
appears to be slightly asymmetric, most likely due to the presence of
the Virgo Overdensity~\citep{Ju08} in parts of the stream
field. However, the subtraction works well and leaves the Sgr stream
clearly visible. The stream signal is roughly linear in the plane of
distance versus right ascension and can be described by a straight
line with gradient 0.386. The solid lines in the two leftmost panels
show the straight line fit offset by $\pm 5$ kpc. The lower panel
compares the BHB distances (straight line) to the leading arm with
those originally inferred by \citet{Be06} using the $i$ magnitude of
the subgiant branch in the ``Field of Streams''. It is reassuring that
there is excellent agreement between the two independently calculated
distances, especially as the BHB's luminosity is almost independent of
age and metallicity.
Also shown are the distances to 2MASS M giants in both leading and
trailing arms. Using the cuts described above, we create a density
distribution of M giant stars in Sgr's orbital plane as shown in
Figure~\ref{fig:missing}. Distances to pixels with highest density
along the stream are then computed with a typical uncertainty of 6
kpc. This is a combination of the uncertainty in the color-absolute
magnitude relation for M giants and the intrinsic width of the stream.
Finally, we also show two detections of the trailing debris in the
SDSS southern stripes. The distances are computed using both BHBs and
main-sequence turn-off stars. The overlaps of the SDSS and 2MASS
datasets will play an important role in what follows.
As a useful summary of the data, we list distances and right
ascensions in Table~\ref{tab:distances}. Recently, \citet{Ya09}
reviewed the detections of the Sgr stream in SDSS data, and their
results are in good agreement with ours.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure4.ps}
\caption{Density distribution of 2MASS M giants in the orbital
plane of Sgr as defined in Majewski et al. (2003, see
Figure 11). The center of Sgr is marked with a star. The
solid red line marks the ridge line, whilst the dashed lines
indicate the $3 \sigma$ boundary of the tidal arms. The
circles are the SDSS observations
around the North Galactic Cap. The triangles mark the stream detections in
the SDSS southern stripes. The inset shows the number of M
giants as a function of heliocentric distance in the
dot-dashed slice of the stream. A Gaussian fit to the
distribution yields a width of $6$ kpc. The Galactic Center
and the Sun positions are marked.} \label{fig:missing}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figure5.ps}
\caption{Comparison of the Hess diagrams of the Sgr core
(left) and stream (right). Note that the populations in the
core and the stream are very similar. Boxes are overlaid to
identify characteristic populations used later in the paper,
including main sequence (MS), red giant branch (RGB), red
clump (RC), red horizontal branch (RHB), blue horizontal
branch (BHB), blue stragglers (BS), and luminous red giant
branch (LRGB). The black points in the left hand panel are M
giant stars selected in the core using the cuts described
above. The dashed box shown in the left panel is used to
scale up the density in the control field. Overplotted in
the left panel are Dartmouth isochrones ~\citep{Do08} of
varying age and metallicity (solid $15$ Gyr, [Fe/H]$=0.0$;
dashed $11$ Gyr, [Fe/H]=$-0.5$; dash-dotted $8$ Gyr,
[Fe/H]$=-1.0$). We are confident that there are no turn-offs
of older stellar populations in the Sagittarius core that
were not detected in the Bellazzini et al. data.}
\label{fig:hess_dif_sample}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure6.ps}
\caption{LF of the Sgr remnant (black), together with two
possible extrapolations at the faint end (dashed and
dotted). Also shown are the LFs of Ursa Minor (black
dashed), the globular cluster M15 (red solid), together with
a mean globular cluster LF (red dashed).}
\label{fig:core_lf}
\end{figure}
\begin{deluxetable*}{lcccccccccccccccccccccc}
\tablecolumns{22}
\tablecaption{Heliocentric Distances to parts of the Sgr stream. SDSS and 2MASS observations are summarized in the top and bottom halves respectively.}
\tablehead{}
\startdata
RA - Stream A &215&210&205&200&195&190&185&180&175&170&165&160&155&150&145&140&135&130&125\\
Distance (kpc)&47&47&46&43&40&38&35&33&30&28&27&25&24&21&20&19&18&17&16\\ \\
RA - Stream B &215&210&205&200&195&190&185&180&175&170&165&160&155&150&145\\
Distance (kpc)&45&44&45&41&38&35&32&30&28&26&26&23&22&19&16\\ \\
RA - Trailing &5&35\\
Distance (kpc)&19&25\\ \\
\hline \\
RA - Leading &252 &246 &240 &234 &229 &223 &217 &212\\
Distance (kpc)&36&38&40&41&43&44&44&44\\ \\
RA - Trailing&286&292&298&...&316&322&328&333&...&18&22&27&31&35&39&44&48&52&57&61\\
Distance (kpc)&24&23&22&...&21&21&20&19&...&20&20&21&22&23&24&25&26&27&28&29\\
\enddata
\label{tab:distances}
\end{deluxetable*}
\begin{deluxetable*}{cccc}
\tablecolumns{4}
\tablecaption{Single exponential fits to the A (top half) and B (bottom half)
streams (see Figure~\ref{fig:boxcounts}).
The left column gives the slope of the exponential together with its uncertainty. The middle column gives the luminosity per kpc of the fits at RA$=220^\circ$. The right column indicates the number of data points used in the exponential fit. Where applicable the values in brackets indicate the parameters derived using all data points.}
\tablehead{
\colhead{Sample}&
\colhead{Exponent ($\times 10^{-3}$)}&
\colhead{Luminosity ($L_{\odot} \times10^4$)}&
\colhead{Data Points}
}
\startdata
Total & 7.0 $\pm$ 0.2 & 27.00 & 12\\
MS & 6.6 $\pm$ 0.3 & 15.92 & 12\\
BS & 4.5 $\pm$ 0.8 & 0.24 & 12\\
RGB & 6.4 $\pm$ 0.5 &2.01 & 12\\
RC & 7.7 (3.5)$\pm$ 1.0 (0.4) & 2.14 (1.66)& 6\\
RHB & 8.1 $\pm$ 0.6 &3.18 & 12\\
BHB & 6.2 $\pm$ 1.0 &0.33 & 12\\
LRGB & 8.7 (-2.7)$\pm$ 0.7 (0.5) &2.12 (1.1)& 4\\
\\
\hline
\\
Total & 4.5 $\pm$ 0.6 &15.01 & 12\\
MS & 4.9 $\pm$ 0.6 &9.20 & 12\\
BS & 2.7 $\pm$ 0.7 &0.18 & 12\\
RGB & 4.1 $\pm$ 0.7 &1.47 & 12\\
RC & 4.4 (0.5) $\pm$ 1.2 (0.4)& 1.2 (0.92) & 0\\
RHB & 5.1 $\pm$ 1.0 &1.47 & 12\\
BHB & 4.5 $\pm$ 1.2 &0.24 & 12\\
LRGB & 4.4 (-1.0) $\pm$ 1.2 (0.5) &1.2 (1.2) & 0
\enddata
\label{tab:powerlaws}
\end{deluxetable*}
\subsection{Sagittarius Core}
In order to derive the luminosity of Sgr, we need to disentangle its
stars from the Galactic fore and background. We start by re-visiting
the calculation of the luminosity of the remnant. The original value
of \citet{IGI} is based on data from Schmidt photographic plates. With
an estimate of the extent of the remnant, together with a CMD of a
$2.5^\circ \times 2.5^\circ$ central field, they calculated an
absolute magnitude of $M_V \sim -13$. This relies on the similarity of
the CMDs of Sgr on the one hand and Fornax and the SMC on the
other. \citet{Ma98} acquired data on 24 fields of size $0.25^\circ
\times 0.25^\circ$ across the face of Sgr and built a number
density profile of main sequence stars. Then, using a surface
brightness normalization from ~\citet{Ma95}, this can be integrated to
give a total luminosity of the remnant as $M_V = -13.3$. However, due
to the size of the field of view ($0.25^\circ \times 0.25^\circ$) in
~\citet{Ma95}, the luminosity function (LF) does not extend brighter
than the red clump and does not include BHBs and BSs. So, the shape of
the surface brightness profile is well-constrained, as confirmed by
\citet{Ma03}, but the overall normalization is more uncertain.
The data of \citet{Bel06} cover a larger area than those of
\citet{Ma95} at the same location, and hence give a clearer view of
all the stellar populations. The core data are obtained in a
$1^{\circ} \times 1^{\circ}$ field approximately $2^{\circ}$ east of
the Sgr center located at $(l,b)\approx (6.5^{\circ},-16.5^{\circ})$
and along the major axis. The control field is $0.5^{\circ} \times
0.5^{\circ}$ at $(l,b)\approx (-6.0^{\circ} \times -14.5^{\circ})$.
For both fields, we fix the reddening at $E(B-V) = 0.11$, which is the
mean value over the core field, as judged from the maps of
\citet{Sc98}. The left panel of Figure~\ref{fig:hess_dif_sample}
shows the CMD of the Sgr core. This is a Hess difference that has been
constructed by subtracting from the core Hess diagram the control
field Hess diagram. To scale the control field density, we minimize
the residual counts in the dashed box shown in
Figure~\ref{fig:hess_dif_sample}. There are a number of readily
identifiable populations in the CMD, including the main sequence (MS),
red giant branch (RGB), red clump (RC), red horizontal branch (RHB),
blue horizontal branch (BHB), blue stragglers (BS) and the luminous
red giant branch (LRGB), together with M Giants selected in the same
manner as in the right-hand panel of Figure \ref{fig:2mass_boxes}. We
use the CMD boxes shown to build the LF of the core and calculate
luminosity density profiles along the stream. It is reassuring that
our core CMD looks very similar to Figure 1 of \citet{Bel06}, which
has been constructed from the same data, but with a different
algorithm.
Figure~\ref{fig:core_lf} shows the observed LF of the Sgr remnant as
derived from integrating the light in the CMD boxes. This drops
dramatically beyond $M_V \sim 5$, and a number of possible
extrapolations at the faint end are shown. However, most of the
luminosity is in stars brighter than the turn-off. \cite{Ma95}
computed the luminosity of the Sgr remnant by integrating the LF
between the red clump and the upper-main sequence to get a surface
brightness of 25.8 mag arcsec$^{-2}$. They account for the missing
flux in fainter stars by extrapolating their LF to $M_V =13.2$ to get
$0.5$ mag correction. They do not explicitly account for stars
brighter than the red clump or bluer than the main-sequence turn-off.
Repeating the calculation for our LF derived using the Bellazzini et
al. (2006) data, we get a similar surface brightness in stars between
the red clump and the upper main sequence, namely $25.7$ mag
arcsec$^{-2}$. However, we also find the BHBs, BSs, LRGBs and M giants
contribute an additional $0.3$ mag. Two possible extrapolations at the
faint end are shown, which contribute $0.5$ mag (dot-dashed) and
$0.15$ (dotted) respectively. This can be compared to LFs of a typical
globular cluster (red dashed, Cox 1999), the dSph Ursa Minor (black
dashed, Wyse et al. 2002) and one of the most luminous globular
clusters, M15 (red solid, Piotto et al. 1997). Taking our lead from
the dSph LF suggests adding $0.5$ mag for faint stars to arrive at a
surface brightness of $24.9$ mag arcsec$^{-2}$ (compared to a value of
$25.3$ given in Mateo et al. 1995).
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figure7.ps}
\caption{Left and right panels show the luminosity per unit
length in a variety of stellar populations in streams A and
B respectively. The luminosities for the RGB, RC, RHB, BHB,
BS and LRGB stars are derived directly from the Hess
difference diagrams. The numbers are normalized to $1$ kpc
distance along the stream. The luminosities in the MS stars
are corrected for missing flux using the derived LF. Error
bars are determined assuming Poisson noise. A single
exponential is fit to each population. However, the RC and
LRGB populations show significant noise and hence for stream
A only the first few data points follow an exponential
fit. In stream B the noise dominates these two populations
and the exponential shown is estimated without using any
data points as described in the main text. The fits to these
two populations are therefore shown as dashed lines. The
data points used in determining the fits are shown as filled
symbols and those excluded as open symbols. The behavior of
streams A and B is very similar, suggesting that both A and
B may be parts of the same leading arm. Finally, we list at
the side of the figure the mean of the fractional
contribution to the total luminosity across the stream for
each population.}
\label{fig:boxcounts}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.475\textwidth]{figure8a.eps}
\includegraphics[width=0.475\textwidth]{figure8b.eps}
\caption{Top: Location of 2MASS M giants in the orbital plane
defined by the Sgr orbit; see equation 9 of Majewski et
al. (2003) for the definition of latitude and longitude
($B_\odot$,$\Lambda_\odot$). The red box encloses the Sgr
leading (left) and trailing (right) streams, whilst the
foreground box is chosen to have the same area. Bottom:
Number counts ($N_{\rm MG}$) of M giants in the stream in
bins of $5^\circ \times 20^\circ$ as a function of longitude
in the orbital plane. Crosses (pluses) are counts in the
central (foreground) boxes. The curves are estimates of the
underlying smooth foreground. The red circles and triangles
are the resulting profiles. The profile of the trailing arm
derived by Majewski et al. (2003) is plotted in grey in the
bottom right panel.}
\label{fig:2MASSprofiles}
\end{figure*}
\subsection{The Density Profile of the Sagittarius Stream}
The CMD of streams A and B of the Sgr leading arm is already shown in
Figure 4 of \citet{Be06} and is known to look similar to that of the
core. Here, we refine the calculation of the leading arm CMD in three
ways. First, we use the latest data release DR7 of the
SDSS~\citep{Ab09}. Second, we use a different comparison field,
defined by reflecting the boundaries of the stream across the
$\ell=180^{\circ}$ line (see Figure \ref{fig:fos_boxes}). This
assumes that the Galactic field star population is symmetric across
$\ell = 180^\circ$ and that the control fields therefore probe a
similar field star population as is present around the Sgr stream.
Third, in order to combine portions of the stream at different
distances, we offset each star in the stream and control field using a
smooth mapping between its right ascension and distance (see Figure
\ref{fig:distances}). Rather than showing the difference of two Hess
diagrams, here we choose to show the ratio. The advantage of this is
that it accounts naturally for the fact that the depth of the CMD
varies along the stream. This procedure helps to enhance fainter
features like the horizontal branch and luminous part of the red giant
branch, as can be seen in the right panel of
Figure~\ref{fig:hess_dif_sample}. This also explains the differences
in relative density between the two panels. There are some small but
significant differences in the location of CMD features. For example,
the stream's CMD is slightly bluer than the core's. This could be due
to metallicity gradients, although the bulk of it is probably due to
slight underestimation of extinction in the core (which is of course
very difficult to measure in low latitude fields). Nonetheless, the
size of our CMD boxes means that regardless of the slight color
mismatch, Sgr member stars will still be selected.
We count the numbers of stream stars in the population boxes marked in
Figure~\ref{fig:hess_dif_sample}, and subtract the counts in the
comparison field. Given the distance to each field, this produces a
Hess difference in absolute magnitude versus color, $M_V$ versus
$B-V$. Due to the substantial distance gradient along the stream, the
faintest population box -- main sequence stars -- is probed to
different limiting magnitudes. To correct this, we make use of the
fact that the stream's population is similar to that of the Sgr core,
at least below the subgiant branch. We extend each field's LF down to
$M_V=13.2$ by requiring that the slope of the stream's LF at $M_V =
3.4$ matches that of the core's extrapolated LF. This value is chosen
as it corresponds to the limiting SDSS magnitude $V \sim 22$ in the
most distant leading arm field.
The Hess difference is integrated over each CMD box to give the total
luminosity. Since each field has the same angular extent but is at a
different heliocentric distance, the luminosity is normalized by the
length of the segment in kpc. The two panels of
Figure~\ref{fig:boxcounts} show the luminosity profiles as a function
of right ascension for multiple populations in streams A and B,
together with exponential fits. The choice of fitting law is motivated
by the behavior of the total luminosity. We expect that all stellar
populations behave in a similar way, and this is largely confirmed by
our results. There are populations not so well described by an
exponential law, but these deviations may be caused by noise in the
data. Each of our fields probes at least some way down the
main-sequence, so the only population that is affected by our LF
correction is the main-sequence sample itself. However, as shown in
Figure \ref{fig:boxcounts}, the RGB, RHB, BHB and BSs all have slopes
that match the slope in the corrected main sequence sample, which
suggests that the observed trend is not due to the correction we
apply. Notice that the slopes of all the stellar populations in both
streams A and B are positive at a statistically significant
level. This implies that the corresponding densities of the
populations are all fading with decreasing right ascension, that is
increasing distance from the progenitor (as also noticed by Yanny et
al. 2009). This is consistent with both A and B being part of the same
leading stream.
Stream A is roughly a factor two brighter in luminosity than B. About
$60\%$ ($50\%$) of the signal is contributed by the main sequence
stars when the dot-dashed (dotted) extrapolation of Figure
\ref{fig:core_lf} is used. The total luminosities in the fields are on
average a factor of $1.3$ brighter when using the steeper
extrapolation. BHBs and BSs contribute little to the total luminosity,
but can be very cleanly selected and so give a good indication of the
underlying trend. By contrast, the counts corresponding to the RC and
LRGB stars are much noisier. This is because the foreground
contamination by the thick disc and halo is most severe, and varies
considerably in the bright part of the CMD. In fact, the exponential
fall-off is only present in the first few bins of stream A,
corresponding to right ascensions between $180^\circ$ and
$205^\circ$. The dashed lines for stream A therefore show exponential
fits using only the data at higher right ascension (shown as filled
symbols). For the fainter and therefore even noisier stream B, even
this option is unavailable and so we assume a constant ratio of BHBs
to RCs and LRGBs in the two streams to derive the dashed lines in the
right-hand panel. This assumes that streams A and B have similar
metallicities, which is supported by the work of Yanny et al. (2009).
Table \ref{tab:powerlaws} gives the slope of each exponential fit
together with its uncertainty, as well as the luminosity for each
component at a right ascension of $220^\circ$. We also indicate the
number of data points used in determining the exponential fit
To obtain as complete a profile of the stream as possible, we will
need to exploit the complementarity of the 2MASS and SDSS datasets.
2MASS provides a powerful probe of the trailing arm, for which
\cite{Ma03} have already deduced a profile. There are also parts of
the leading arm in 2MASS, for which we build the profile as
illustrated in Figure~\ref{fig:2MASSprofiles}. The upper panel shows
the locations of M giants, selected using the same method as the
zoom-in of Figure \ref{fig:2mass_boxes} after offsetting stars to the
distance of the core using Table~\ref{tab:distances}. The results are
shown in the heliocentric coordinate system aligned with the Sgr
orbital plane, as defined by equation 9 of \citet{Ma03}. There are
two selection boxes in red to pick out the stream signal and the
foreground. The resulting profiles of the raw number counts in
$5^\circ$ bins, the foreground and the corrected number counts are
shown in the lower panel. Applying the same procedure to M giants in
the trailing arm, we recover a profile similar to that of
\citet{Ma03}, although rising in the outer parts.
Before we can begin reassembly, we need to do three more
things. First, we measure distance along the stream from the
progenitor to the observed fields using
Table~\ref{tab:distances}. Secondly, we combine streams A and B in the
SDSS dataset into a single profile of the leading arm. Thirdly, in
order to combine SDSS and 2MASS observations, we convert M giant
number counts into total luminosity densities along the stream. We
measure the relationship between the number of M giants ($N_{\rm MG}$)
and the total luminosity at the locations where the \citet{Bel06} and
SDSS observations overlap with the 2MASS coverage (see the lower panel
of Figure~\ref{fig:distances}). Here we can measure the total
luminosity using the Hess difference method and count 2MASS M giants.
The uncertainty in the ratio of the two is determined assuming Poisson
statistics. There are clearly substantial gradients between the
remnant and both the leading and trailing arms. This is illustrated
in Figure~\ref{fig:fields}. We use a straight line fit, which has a
gradient of $280 \pm 30 {L_\odot}$ kpc$^{-1}$ per M giant. Alternatively,
assuming a step function constant in the remnant and constant in the
stream does not affect the following analysis substantially.
Figure~\ref{fig:luminosity_dist_profile} shows the resulting
luminosity profile as a function of distance along the stream for both
the leading (circles) and trailing (triangles) arms. Points derived
from 2MASS data are represented by open symbols, whereas those from
SDSS by filled symbols. The profiles of the leading and trailing arms
are different by a factor of $\sim 3$ at small distances from the
remnant, primarily as a consequence of the fact that the trailing arm
is much more extended. Notice too the bump in the leading arm at
$\sim 40$ kpc distance along the stream. Such features are common in
simulations -- such as in Figure 1 of \citet{La05} -- as a consequence
of the pile-up of material stripped at disk-crossings. There is also
an evident break in the profile of the trailing arm at $\sim 10$ kpc
distance along the stream, similar to that seen by \citet{Ma03}. It is
natural to associate this break with the transition between the core
and the streams. Even though the profiles have a complex structure,
it is useful to give a simple fit. For the trailing arm, we fit a
broken exponential and find indices $-0.207$ and $0.008$ either side of
a break radius at $9.6$ kpc. For the leading arm, the break radius
lies behind the disk and we only measure the profile after the break
as $-0.007$.
\subsection{The Luminosity of Sgr}
Now we begin the reassembly by calculating the total luminosity in the
Sgr remnant and tidal debris. We estimate the remnant's luminosity by
summing the observations of the trailing arm within the profile break
to get $2.0\pm0.3\times10^7 L_{\odot}$. This is doubled on the
assumption of the symmetric structure of the remnant, giving us a
total magnitude of $M_V = -14.2\pm0.1$, which is within 2 $\sigma$ of
the value of \citet{Ma98} of $M_V = -13.4\pm0.5$.
For the debris, we give two estimates. The first is a lower limit set
by the data alone, namely $2.6\pm0.1 \times 10^7 L_\odot$ for the
leading arm and $8.2\pm0.7 \times 10^6 L_\odot$ for the trailing. The
combined value already contains more than half the core's luminosity,
and is well constrained by the deep SDSS coverage of the leading arm.
The second is an estimate of the luminosity contained within the
entire debris stream, not simply the portions we observe. This is
based on integrating the leading arm profile from $25$ kpc
(corresponding to the first 2MASS observation) to $300$ kpc, which is
triple the length to the last SDSS observation. This corresponds to a
full wrap, as judged by the simulations~\citep[see
e.g.,][]{La05,Fe06}. This gives a luminosity of $3.5\pm0.1 \times 10^7
L_\odot$. This nominal error is so small because the fit is
constrained by the SDSS observations which have small uncertainties
due to the large number of stars in the SDSS sample. This also ensures
that potentially larger photometric errors on individual stars are
averaged out. To account for the region between the core and $25$
kpc, we average two possible extreme cases. If we assume that the
core is symmetric, then the leading arm profile needs to dip down
sharply to meet the trailing arm profile at $10$ kpc. This would
contribute $0.6\times10^7 L_\odot$ to the luminosity of the leading
arm. Alternatively, since we do not actually observe any indication
of such a dip, the profile may continue rising until joining up with
the core's profile at approximately $5$ kpc. This would contribute
$1.6\times10^7 L_\odot$. Therefore, for the missing inner portion of
the leading arm, we take as a compromise the average value
$1.1\pm0.5\times 10^7 L_\odot$ where the large uncertainty reflects
the range of possible values. The total leading arm luminosity (given
by adding the extrapolated exponential and the estimate for the region
between the core and the first observations)is then doubled to account
for the trailing arm. We do not integrate the profile of the trailing
arm since it is rising, which most likely reflects its approach to
appocenter. Integrating a rising profile is obviously very sensitive
to the limits of integration. At present it is difficult to make a
reasonable assumption about the behavior of the trailing arm far from
the core. This finally gives for the total debris luminosity
$9.2\pm0.7 \times 10^7 L_\odot$. This will be reduced to
$6.6\pm0.7\times10^7 L_\odot$ if the fainter LF in
Figure~\ref{fig:core_lf} is used. It is evident that the formal error
from Poisson noise, typically of order of $10\%$, is much smaller than
uncertainties induced by extrapolations. To account for the missing
data, we were forced to extrapolate both the luminosity function and
parts of the density profiles.
The various luminosities used in the final answer are summarized in
Table \ref{tab:streamlum}. The total luminosity of Sgr progenitor is
$13.2\times10^7 L_\odot$ or $M_V \sim -15.5$ with the brighter
LF. If we repeat the above calculation with the fainter LF, we arrive
at $9.6\times10^7 L_\odot$ or $M_V \sim -15.1$. This implies
that the progenitor is fainter than the SMC, but comparable to the
brighter M31 dSphs like NGC 147 ($L=9.4\pm2.0\times10^7 L_\odot$,$M_V
= -15.1$) and NGC 185 ($L=13.5\pm2.0\times10^7 L_\odot$, $M_V =
-15.5$).
The systematic uncertainties clearly outweigh the Poisson noise but we
are confident that we have atleast found a firm lower limit
($9.6\times10^7 L_\odot$) to the luminosity. The sources of systematic
uncertainty include the interpolation of the inner slope of the
leading arm. This is over quite a modest range and unlikely to be
seriously in error. In addition considerable uncertainty derives from
the extrapolation of the LF as shown above. Perhaps the greatest
unknown is the behavior of the profiles at large distances. We have
assumed that the exponential behavior continues to 300 kpc, but there
may be processes that cause additional clumps and bumps.
\begin{deluxetable}{lccc}
\tablecolumns{3} \tablecaption{Luminosity of various parts of the
Sgr stream together with their associated Poisson
uncertainties.} \tablewidth{0.5\textwidth} \tablehead{{}&
\multispan{2}{Luminosity (in $10^7 L_{\odot}$)} \\
{Stream Portion} & {Faint LF} & {Bright LF}} \startdata
Core (Bellazzini et al.) &$2\times1.5\pm0.3$ & $2\times2.0\pm0.3$ \\
Leading Arm (SDSS) & $2.0\pm0.1$& $2.6\pm0.1$ \\
Trailing Arm (2MASS) & $0.6\pm0.1$& $0.8\pm0.1$ \\
Leading Arm (25 to 300 kpc) & $2.5\pm0.1$& $3.5\pm0.1$ \\
Leading Arm (Core to 25 kpc) & $0.8\pm0.5$& $1.1\pm0.5$ \\
\enddata
\label{tab:streamlum}
\end{deluxetable}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figure9.ps}
\caption{Conversion factor between total luminosity and
numbers of M giants at different locations along the
stream. The dash-dotted line shows the change in the
conversion factor along the stream assuming that it is
constant within the core, and constant along the stream. The
dashed line shows the conversion factor assuming it varies
linearly throughout the core and stream.}
\label{fig:fields}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figure10.ps}
\caption{Comparison between the leading and trailing arms as
seen in 2MASS and SDSS. We plot here the total luminosity
per unit kpc along the stream versus the distance along the
stream. For the 2MASS data, this is obtained from the number
counts of M giants by the scaling described in the main
text. The symbols have the same meaning as in the lower
panel of Figure~\ref{fig:distances}. For the trailing
stream, the dashed line shows the best broken exponential fit
to the data. For the leading stream, the slope of the inner
part is fixed which determines the break point.}
\label{fig:luminosity_dist_profile}
\end{figure}
\section{Discussion and Conclusions}
Using SDSS and 2MASS data, we have provided for the first time the
luminosity profiles of both the leading and trailing arms of the Sgr
stream. Prior to this work, the only available profile was that of
the trailing arm derived from M giants by \citet{Ma03}. This was
essentially flat, and so cannot be extrapolated to derive useful
constraints on the mass lost during disruption. Here, our profiles for
both the leading and trailing arms have gradients. It is the gradient
in the leading arm that enables us to derive useful constraints.
The luminosity in the leading stream clearly declines with increasing
distance from the progenitor. Although the beginning of the tail is
absent because it lies behind the disk, the profile still shows mild
evidence for a bump. This is most likely a pile-up at apogalacticon.
The trailing stream can be traced all the way from the remnant to the
anti-Center, and shows a clear break around the tidal radius as we
move from the core to the tail. At first glance, this looks very
different to the leading arm profile. But in reality it may just be a
stretched-out version of the leading profile, as the trailing arm
extends to much greater Galactocentric distances. In this picture, the
rising part of the profile is interpreted as the approach to an
apogalacticon pile-up. We speculate that a similar but steeper rise
would be observed in the leading arm, making the core profile
symmetric, were it not obscured by the disk.
It is worth emphasizing that a firm lower limit $M_V \sim -15$ for the
Sgr progenitor luminosity is given by the data with minimal
assumptions. The most important is that the mass is equally
apportioned between leading and trailing in both the remnant and the
tails. Thanks to deep and wide SDSS coverage, the leading profile is
measured with excellent accuracy. It is the tightness of the
datapoints in the leading arm that is providing this firm lower limit.
For example, integrating the exponential profile out to $200$ or $300$
kpc makes little difference to the final answer.
Our results may have implications for the scenario proposed by
Fellhauer et al. (2006), who argued that stream A is material in the
leading arm that was stripped recently, whereas stream B is debris in
the trailing arm that was stripped long ago. Here, we have explicitly
assumed that streams A and B are both parts of the same leading arm.
This is supported by recent findings of \citet{Ya09}, who noted that
the stars in streams A and B have very similar metallicities and
velocities. Also, if stream B is old and trailing, then its density
(about one half of A) is larger than the young trailing arm (about a
third of A), which seems counter-intuitive. So, although Fellhauer et
al's (2006) interpretation has not been ruled out, there are possible
causes for concern.
The luminosity profiles offer a new, and largely unexplored,
constraint on the disruption process. The fact that the leading arm
profile is not flat, but decreases, directly limits the stellar mass
lost in the disruption. The relative numbers of counts in the leading
and trailing arms can be used to constrain the geometry of the
orbit. The stars move along the tail with a drift velocity determined
by the properties of the progenitor and the
orbit~\citep{De04}. Therefore, the time of onset of destruction of the
stellar component is in principle recoverable.
Our estimate of the progenitor luminosity is $9.6 - 13.2 \times 10^7
L_\odot$ or $M_V \sim -15.1 - -15.5$ with roughly $70\%$ of the light
in the tidal tails. The nearest look-alikes in the Local Group are
probably the M31 satellites NGC 147 and NGC 185. These two dSphs
bracket properties of Sgr in terms of number of globular clusters,
metallicity, gas content and velocity dispersion~\citep[see
e.g. Tables 12.4 and 12.6 of ][]{Be00}. Their mass-to-light ratios are
comparatively modest at $\sim 5$~\citep{Ge09}. If we accept the
comparison, then this suggests that the mass within the luminous
radius of the Sgr progenitor was $\sim 10^9 {M_\odot}$. This is right at
the lower mass end of the range of models considered by \citet{Ji00}
who investigated the range of progenitor properties which could
reproduce the structure of the core remnant observed today.
We can use the estimate that $70\%$ of Sgr's luminosity now resides in
the tidal debris to draw conclusions about the properties of the
progenitor's dark matter halo. Pe\~narrubia et al. ($2008$,ab) have
modelled the disruption of dwarf galaxies with baryons following a
King profile embedded in a dark matter Navarro-Frenk-White (NFW)
halo. The simulations suggest that the evolution of the parameters
of the King profile is governed primarily by the amount of mass that
has been stripped, rather than the details of the dwarf's orbit or
the segregation of baryons and dark matter. Using as the present
parameters of the bound core those determined by Majewski et
al. ($2003$), we find that prior to disruption the King profile had
a central velocity dispersion $\sigma_0= 23$ kms$^{-1}$, projected
central stellar density $\Sigma_0= 5.7 \times 10^6 M_\odot$
kpc$^{-2}$ and a core radius $R_{\rm C} = 1.5$ kpc. This implies
that prior to mass stripping the central velocity dispersion,
projected central stellar density and core radius were,
respectively, a factor $2.0$, $5.3$, and $1.2$ higher than at
present. Subsequently, we apply the method of Pe\~narrubia et
al. ($2008$a), to estimate the properties of the NFW dark matter
halo in which Sgr was embedded prior to mass stripping implied by
this King profile. These authors solve the Jeans equations in order
to find a set of $r_{\rm c,max}-v_{\rm c,max}$ parameters (the
maximum circular velocity of the halo and the radius at which this
is achieved which fully specify the NFW halo) that would allow a
given velocity dispersion of the stellar component. This analyses
yields a family of NFW models depending on the assumed spatial
segregation of the stellar component within the dark matter halo. To
break this degeneracy the authors appeal to the results of
cosmological N-body simulations which show a strong correlation
between an NFW halo's $v_{\rm c,max}$ and $r_{\rm c,max}$. Using the
values for the King profile described above we derive a peak
circular velocity $v_{\rm c,max}=35$ kms$^{-1}$ and scale radius
$r_{\rm S}=3.5$ kpc ($r_{\rm c,max} \approx 2.17 r_{\rm S}$). The
maximum circular velocity and scale radius are consistent with the
values for the NFW haloes of other classical dSphs (see Figure $5$
of Pe\~narrubia et al. $2008$a), but already towards the high mass
end, with $M_{\rm vir} \approx 1.0 \times 10^{10} M_\odot$. The
Draco dSph, found by Pe\~narrubia et al. to have the most massive
dark matter halo, has $M_{\rm vir}\approx 6.3\times10^9
M_{\odot}$. As shown by Pe\~narrubia et al. ($2006$), at this mass,
dynamical friction can reduce the orbits apo- and pericenters by a
factor of $>2$ over a Hubble time. We note that $9.6 - 13.2 \times
10^7 L_\odot$ is likely a lower bound for the progenitor's
luminosity since we have only considered one wrap of debris and have
neglected the possibility of additional pile-ups at previous
apocenters in the orbit. The amount of light contained in these will
depend on the details of the mass loss process. With additional
light in the debris Sagittarius quickly becomes more massive than
any of the other known Milky Way dwarfs (see Figure
\ref{fig:l_vs_mvir}).
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figure11.ps}
\caption{Virial mass of the NFW halo at $z=0$ as a
function of luminosity remaining in the core. The shaded
region shows the range of luminosities determined in this
paper. If we are missing light from additional wraps or
apocenter pile-ups the mass quickly exceeds that of the
other classical dwarfs.The dashed line indicates the
virial mass of Draco (the most massive classical dwarf
according to Pe\~narrubia et al. $2008$a).}
\label{fig:l_vs_mvir}
\end{figure}
\acknowledgements
MNO is funded by the Gates Cambridge Trust, the Isaac Newton
Studentship fund and the Science and Technology Facilities Council
(STFC), whilst VB acknowledges financial support from the Royal
Society. We thank Gerry Gilmore and Mike Irwin, as well as an
anonymous referee, for valuable comments.
Funding for the SDSS and SDSS-II has been provided by the Alfred P.
Sloan Foundation, the Participating Institutions, the National Science
Foundation, the U.S. Department of Energy, the National Aeronautics
and Space Administration, the Japanese Monbukagakusho, the Max Planck
Society, and the Higher Education Funding Council for England. The
SDSS Web Site is http://www.sdss.org/.
This publication makes use of data products from the Two Micron All
Sky Survey, which is a joint project of the University of
Massachusetts and the Infrared Processing and Analysis
Center/California Institute of Technology, funded by the National
Aeronautics and Space Administration and the National Science
Foundation.
|
\section{INTRODUCTION}
Classical random walks (CRW) play an important role in classical algorithms
and have many applications in many realms of science \cite{Barber70}. While
in the quantum worlds, quantum random walks (QRW) are markedly different due
to the phenomenon of interference \cite{Knight03}, which might lead to
extensive applications in quantum information science. The recent
presentation of several quantum algorithms based on QRW has shown the
exponential speed up over classical algorithms \cite{Childs01} as the most
prominent algorithms suggested by Shor \cite{Shor} and Grover \cite{Grover}.
Most recently, Childs has shown that, in principle, any quantum algorithm
can be recast as a QRW algorithm \cite{Childs09}. Meanwhile, numerous
schemes of QRW have been proposed by using various systems, including ion
traps \cite{Travag02}, microwave or optical cavities \cite{Sanders03},
nuclear magnetic resonance (NMR) system\cite{Du03}, optical lattices \cite%
{Dur02}, linear optical elements \cite{Jeong04,Zou06,Do05,Zhang07},
quantum accelerator modes \cite{Ma06} and superconducting circuit quantum
electrodynamics \cite{Xue08}. All these results are advancing forward
towards the ultimate goals of quantum computation.
CRW can take many different forms. For simplicity, we consider the discrete
time random walk on a line. It can be understood as a "walker" standing at
the starting point ($x=0$), holding a coin, decides which side (right or left) he
should take his one step toward by flipping the coin (suppose with head to
right and tail to left). He flips the coin, walks one step to right or left
based on the flipping results. After repeating this process $n$ times, he will
randomly walk to a position $x$. This is a simple model of CRW on a line. We
can get the probability $P_{n}(x)$ of the "walker" being found at $x$ after $%
n$ steps:
\begin{equation}
P_{n}(x) =\frac{1}{2^{n}}\binom{n}{\frac{x+n}{2}}.
\end{equation}
From Eq. 1, we can easily calculate that the mean value is zero and the
standard deviation of this distribution is $\sigma _{c}=\sqrt{n}$, which we
will find being distinctly different to QRW later. QRW occurs on quantum
particles, where their motion described by wave-function and the "coin" changed into "quantum coin" \cite{Kempe03}. Under this version, "quantum walker" does not walk one
step determinately to one position. So there must be interference between
wave-function at different positions. Suppose we define the two level states
as $\left\vert \downarrow \right\rangle $ and $\left\vert \uparrow
\right\rangle $, which correspond to the two sides of "quantum coin", and
one dimensional discrete position $\left\vert x\right\rangle $ ($x$ is
integer) as the walk space. Then we can perform a Hadamard translation ($%
\hat{H}$) to the "quantum coin" as a "quantum coin" tossing. The direction
of "quantum walker" walks to every one step is decided by the results of
tossing "quantum coin", for $\left\vert \downarrow \right\rangle $ steps to $%
\left\vert x-1 \right\rangle $ and $\left\vert \uparrow \right\rangle $
steps to $\left\vert x+1 \right\rangle $. The operator of QRW can be written
as:
\begin{equation}
\hat{U}=e^{i\hat{p}\hat{\sigma} _{z}}\hat{H}.
\end{equation}
Where $\hat{p}$ is the pseudo-momentum operator of the particle confined to
one dimension, and $\hat{\sigma} _{z}$ is the pauli-$z$ operator acting on
the qubit \cite{Travag02}. Then the state of the system after $n$ steps is
\begin{equation}
\left\vert \Psi _{n}\right\rangle =\left( e^{i\hat{p}\hat{\sigma} _{z}}\hat{H%
}\right) ^{n}\left\vert \Psi _{in}\right\rangle
\end{equation}%
For different initial states of "quantum coin", the results of QRW are very
different. We can also calculate the mean value which is nonzero and the
standard deviation of the probability distributions $\sigma _{q}\approx\frac{%
\sqrt{2}}{2}t$ \cite{Cart03}. These are two main different features between
CRW and QRW.
Photons are well known to be extremely powerful for carrying and
manipulating quantum information. QRW have been experimentally demonstrated
in linear optical systems \cite{Do05, Zhang07}. Ref. \cite{Do05} realized a
QRW by employing photon's polarizations and space modes. But the number of
linear optical elements grows quickly when steps increasing and the results
are not very perfect. Ref. \cite{Zhang07} carried out an experiment of QRW
by using photon's up-down routes and orbital angular momentum (OAM). This
setup is simpler than \cite{Do05} because there are only two space modes and
the photons "walk" on OAM space. However, these two experiments both have a
critical shortcoming, that is, they both need Mach-Zehnder interferometers.
The instability of interferometer directly leads the precision decreasing
when steps of random walk increasing. In this paper, we design an
experimental scheme of QRW by using linear optical system. We employ
photon's spin angular momentum (SAM) as "quantum coin" and OAM space as
one-dimensional walk space. By using q-plates \cite{Mar}, our scheme is very
simple and efficient for there is no Mach-Zehnder interferometer in the setup. This
paper is organized as follow: first we briefly introduce the OAM of photons
and the q-plates; then we will give a detailed description and discussion of
our experimental scheme; last is the conclusion.
\section{ORBITAL ANGULAR MOMENTUM and Q-PLATE}
As we know, light beams have polarizations and helical modes. For photons,
it is corresponding to spin angular momentum (SAM) and orbital angular
momentum (OAM), respectively. SAM is associated with the circular
polarization of light, each photon carries $\pm\hbar$ angular momentum
depending on the handedness of the polarization. OAM is associated with
transverse amplitude and phase profile in the cross section orthogonal to
the propagation axis, and one photon carries an intrinsic orbital angular
momentum given by $m\hbar $ if the wave function contains phase term $%
exp(im\phi )$ \cite{Allen92}. Since the quantum nature of OAM was
demonstrated by Mair et al. \cite{Mair01}, more and more studies have
focused on encoding the high-dimensional quantum information by using OAM of
a single photon \cite{OAM}. These works open up a new possibility for
generation and manipulation of higher-dimensional quantum states and
implementation of higher-dimensional quantum information processing
protocols, by using OAM as a quantum information carrier.
When light propagates through a medium, the SAM and OAM are not always
simultaneous independent coupling with the matter \cite{He}. In vacuum or in a
homogeneous and isotropic medium, the SAM and OAM are separately conserved.
But in a medium which is inhomogeneous and anisotropic, they are not. In
appropriate condition, the exchange with matter of the SAM affects the
direction of the exchange of the OAM, and it is possible that the angular
momentum only converting between SAM and OAM. A device called q-plate is used to
conduct this function. Q-plate is a slab of liquid crystal (or uniaxial
birefringent medium) with special structure, and it is essentially a
retardation wave plate whose optical axis is aligned nonhomogeneously in the
transverse plane in order to create a topological charge $q$ in its
orientation. If we assume $q=1$, for example, each photon being converted
from right circular to left circular change its spin z-component angular
momentum from $-\hbar $ to $+\hbar $, and the orbital z-component angular
momentum changes $-2\hbar $. Therefore, the total variation of angular
momentum is nil, and there is no net transfer of angular momentum
to the q plate. The plate acts only as a medium for the conversion between
spin and orbital angular momentum. If $q\neq 1$, it will exchange an angular
momentum of $\pm 2\hbar (q-1)$ with each photon. Using Dirac marks, suppose
the initial state is $|\uparrow _{S},m_{L}\rangle $ or $|\downarrow
_{S},m_{L}\rangle $, which refers to the photon carrying SAM with $\hbar $
(left circular polarization) or $-\hbar $ (right circular polarization) and
OAM with $m\hbar $. The action of a tuned q-plate of charge $q=1$ on this
state can be summarized as follows:
\begin{equation}
(|\uparrow _{S},m_{L}\rangle ,|\downarrow _{S},m_{L}\rangle )\overset{q-plate%
}{\longrightarrow }(|\downarrow _{S}(m+2)_{L}\rangle ,|\uparrow
_{S},(m-2)_{L}\rangle )
\end{equation}%
In this way, a photon in an OAM state $m=0$ is transformed into $m=\pm 2$,
or a photon in $m=\pm 2$ to $m=\pm 4$ or $m=0$, depending on the
polarization. This spin-controlled-orbital transform is very suitable for
implementing quantum walks. Here we use the character of infinity space of
OAM and the properties of q-plates to plan out an implementation of
one-dimensional discrete time QRW, where SAM refer to the quantum "coin" and
$m$ of OAM refer to the photon "positions" when random walk.
We will give a detailed explanation later.
\begin{figure}[tbh]
\includegraphics[width=10cm]{setup.eps}
\caption{Experimental scheme of one-dimensional quantum walks on
spin-orbital angular momentum space of photons. One half wave plate (HWP)
sandwiched by two quarter wave plates (HWP) can realize any single bit
unitary operation, which is used as a initial state preparation. Gray box,
which is made up by two HWPs, one QWP and one q-plate (QP), performs the QRW
operation $\hat{U}$. An n-steps QRW can be achieved by applying $n$ times of
gray box. At end of the setup is the OAM analysis system, which contains a
computer-generated hologram (Holo), a single mode fiber (SMF) and a single
photon counting module (SPCM). When we change different holograms (with
different $\Delta m$), we can get the probabilities of the photons being in
each OAM state. }
\end{figure}
\section{EXPERIMENTAL SCHEME}
QRW on the OAM space of photons have been studied both theoretically \cite%
{Zou06} and experimentally \cite{Zhang07}. However, the steps of QRW can
hardly be increased in the experiment, because of the stability of
Mach-Zehnder interferometer decreasing quickly when the arms of
interferometer increasing. We plan out a QRW scheme without interferometers
which can achieve a multi-step QRW stably and can be easily implemented in
the lab. In our scheme, we can use the q-plates with $q=0.5$ \cite{Exp1}
where the OAM index $m$ changes to $m\pm1$ when passing it.
Our experimental scheme is shown in Fig. 1. The single photon source (can be
realized by deep attenuating the coherent laser or by nonlinear process)
emits one photon with linear polarization (suppose horizontal polarization)
and OAM $m=0$. The state can be writen as $|H,0\rangle =|(R+L)/\sqrt{2}%
,0\rangle $, where H is short for horizontal polarization, R and L are short
for right circle polarization and left circle polarization. We can also use
a set of wave plates made by one half wave plate (HWP) sandwiched by two
quarter wave plates (QWP) \cite{Englert}, to prepare different initial polarization
states as quantum coin states if needed. Then we will toss the quantum coin
and decide which direction should step to. This process is achieved by the
gray box shown in Fig. 1. Hadamard gate ($\hat{H}$) is a kind of quantum
coin tossing, which can realize translation as following:
\begin{eqnarray}
&&\text{ \ \ \ }\hat{H}|R\rangle =\frac{1}{\sqrt{2}}(|R\rangle +|L\rangle )
\\
&&\text{\ \ \ \ }\hat{H}|L\rangle =\frac{1}{\sqrt{2}}(|R\rangle -|L\rangle ).
\notag
\end{eqnarray}%
In the gray box, a QWP at $0^\circ$ and a HWP at $22.5^\circ$ make up a $%
\hat{H}$ gate for $|R\rangle $ and $|L\rangle $ states. After flipping the
coin, then the photon will step to $m+1$ when coin state is $|R\rangle $ or
to $m-1$ when coin state is $|L\rangle $. Here the photon's QRW positions
are denoted by the OAM numbers of photon. From Eq. (4), we know a q-plate
can easily realize this process combining with a HWP at $0^\circ$ for spin
reverse. So, the QRW operator $\hat{U}$ is achieved by the gray box. Then we
can apply this operation $n$ times as a $n$ steps QRW. If the initial state
is $\left\vert \Psi _{in}\right\rangle =|R,0\rangle $, we can deduce three
steps for example:
\begin{eqnarray}
&&\left\vert \Psi _{in}\right\rangle \overset{\hat{U}}{\longrightarrow }%
\frac{1}{\sqrt{2}}\left( \left\vert R,-1\right\rangle +\left\vert
L,1\right\rangle \right) \\
&&\text{ \ \ \ }\overset{\hat{U}}{\longrightarrow }\frac{1}{2}\left(
\left\vert R,-2\right\rangle +\left\vert R,0\right\rangle +\left\vert
L,0\right\rangle -\left\vert L,2\right\rangle \right) \notag \\
&&\text{\ \ \ \ }\overset{\hat{U}}{\longrightarrow }\frac{1}{2\sqrt{2}}%
\left( \left\vert R,-3\right\rangle +\left\vert L,-1\right\rangle +2\left\vert
R,-1\right\rangle -\left\vert R,1\right\rangle +\left\vert L,3\right\rangle
\right). \notag
\end{eqnarray}%
From the third step, we can see the distribution of probabilities of photons
being found at different positions (photon's OAM) is asymmetric, which is
very different to CRW. At the end of this setup, we should detect the
probabilities of the photons being in a OAM state $|m\rangle $. This is
achieved by using different computer-generated holograms (Holo) \cite{Bazh90} with $\Delta
m=\pm 1,\pm 2,\cdots,\pm n$ and a single mode fiber (SMF). We know that the
single mode fiber can only pass $TEM_{00}$ mode light and has a strong
absorption to other higher order modes. So, to detect photon in OAM state $%
|m\rangle $, we can use a hologram of $\Delta m=-m$ to change the photon
into OAM state $|0\rangle $, which then can pass through the SMF and
detected by a single photon count modulator (SPCM).
The efficiency of our setup is
very high due to high transmitted efficiency of q-plate \cite{Mar} and linear
optical elements. We know that the
different OAM photons will spatially separate in their radial pattern by
free propagation. This may induce imperfection when take a long free
propagation. However, this can be resolved by placing the q-plates in the
common focal plane of two lens (the lens are not appeared in our setup) \cite{Slu09}. It is might be inconvenient in changing the different
holograms for position detecting. For $n$ steps QRW, the detection needs to
change $2n$ different holograms. But this is not very hard to our scheme,
because it can keep a long time stable for all the different OAM modes are
in one spatial mode and no use of Mach-Zehnder interferometer.
There is another work of experimental realization of QRW by taking advantage of OAM of photons as walk space (see Ref. \cite{Zhang07}). Comparing with the scheme of Ref. \cite{Zhang07}, our scheme has two main improvements: the
first one is the efficiency, because the transmitted efficiency of q-plate is higher than computer-generated holograms;
the second is the stability, our scheme is inherently stable for no need of Mach-Zehnder interferometer, because we know that Mach-Zehnder interferometer is very sensitive with environment especially
for long arms. So, by employing SAM as quantum coin and using
q-plate, as we analyzed above, the scheme of this paper is more efficient and
simple. QRW on a line can be understood as an interference phenomenon
\cite{Knight03}, although there is no Mach-Zehnder interferometer in our
scheme, there still exists interference. From Eq. (6), we can see that in
the third step of QRW, the spin interference between state $\left\vert
R,0\right\rangle $ and $\left\vert L,0\right\rangle $ occurs on the third
Hadamard gate.
\section{CONCLUSION}
By employing photons SAM and OAM, we put forward a simple and efficient
scheme of discrete-time QRW on a line. We have introduced an optical device,
the gray box, by which the QRW operation $\hat{U}$ can be realized. By applying
$n$ times of gray box, it is very simple to realize a $n$ steps QRW. The efficiency of
our scheme is high and the stability is very good, which are suitable for
implementing multi-step QRW. We think our scheme is very simple to do some
QRW algorithm experimentally.
\section{ACKNOWLEDGEMENT}
The authors thank Professor L. Marrucci for helpful
discussion on q-plates, and acknowledge the National Fundamental Research Program (2010CB923102),
National Natural Science Foundation of China (Grant
No. 60778021, 10674106 and 10774117) for supporting this work.
|
\section{Introduction}
\label{sec:Intro}
Since its early measurements, the large value of
the top quark mass ($M_{\mathrm{top}}$) has represented a really striking property of this particle, giving
to the top quark a special position within the standard model (SM)
and suggesting also possible links to new physics\,\cite{sm_ext}.
In fact, apart from being itself a fundamental parameter of the SM, $M_{\mathrm{top}}$ is
by far the largest mass among the ones of the observed fermions, and this
makes the top quark contribution dominant in higher order
corrections to many observables.
Therefore $M_{\mathrm{top}}$ plays a central role in checking the
consistency of theoretical predictions of the SM.
The higher order corrections apply also to the $W$ boson propagator, and therefore
affect the calculated value of the $W$ mass, $M_{W}$.
As the latter depends logarithmically on the mass of the
Higgs boson, precise measurements of $M_{W}$ and $M_{\mathrm{top}}$ allow setting indirect
constraints on the value of the mass of this fundamental, but still unobserved
particle\,\cite{higgs}.
Moreover, possible contributions due to some unknown physics might also be constrained.
Finally, the present value of $M_{\mathrm{top}}$ makes the Yukawa coupling to the Higgs field of
${\cal O}(1)$ and this could indicate a special role of the top quark in
the mechanism of electroweak symmetry breaking.
All these reasons make the accurate knowledge of
$M_{\mathrm{top}}$ a really important issue, but the same is true for
the measurement of the $t\bar{t}$ production cross section ($\sigma_{t\bar{t}}$),
both as a test for physics contributions beyond the SM and as a test
of current next-to-leading-order (NLO) QCD calculations~\cite{theo-xsec}.
Usually, measurements of $\sigma_{t\bar{t}}$ rely upon event counting and
are performed assuming an {\it a priori} value for $M_{\mathrm{top}}$. The technique used here
allows the simultaneous measurement of
both these important and related properties of the top quark.
At the Tevatron Collider at Fermi National Accelerator Laboratory, top quarks are produced mostly in pairs.
In the SM the top quark decays into a $W$ boson and a $b$ quark almost 100\% of the time,
and the topology of the final state resulting from
a $t\bar{t}$ event depends on the hadronic or leptonic decay of the two
final-state $W$ bosons.
In this paper, we consider events characterized by a multijet topology ({\em all-hadronic} mode) with
no energetic leptons. This $t\bar t$ final state has the advantage of a large branching ratio
($\approx 4/9$) and of
having no undetectable final-state particles.
The major challenge of this channel is the
large background from QCD multijet production, which dominates the signal by 3 orders
of magnitude after the application of a specific online event selection (trigger).
To increase the purity of the candidate sample, requirements based on the kinematical and topological
characteristics of SM $\mbox{$t\bar{t}$}$ events are expressed in terms of an
artificial neural network and applied to the data.
Further improvement is then obtained from the requirement of at least one jet identified as originating from
a $b$ quark using a secondary vertex $b$-tagging algorithm.
Simulations predict that a clear $t\bar{t}$ signal
will thus become visible over background in the selected
data sample, and the measurement of the top quark mass and the $t\bar{t}$
cross section is made possible in spite of the overwhelming
QCD multijet production.
A reconstructed top quark mass is determined by fitting the
kinematics of the six leading jets in the event to a $t\bar{t}$ final state.
This variable, denoted as $m_{t}^{\mathrm{rec}}$, does not strictly represent a
measurement of $M_{\mathrm{top}}$, but its distribution obtained
by a sample of $t\bar{t}$ events is sensitive to $M_{\mathrm{top}}$ itself.
The jet energy scale ($\mathrm{JES}$) is a factor representing
the set of corrections needed to obtain a better estimate of the energy of a parton
starting from a jet reconstructed by clusters in the calorimeter.
The default $\mathrm{JES}$ used in simulated events is obtained
by a tuning to the data, but possible discrepancies between data and simulation lead
to an uncertainty on this value. The strong correlation existing between
the $m_{t}^{\mathrm{rec}}$ distribution and the $\mathrm{JES}$ implies therefore a corresponding
uncertainty on $M_{\mathrm{top}}$.
However, the $\mathrm{JES}$ can be calibrated
using the selected samples of $t\bar{t}$ candidate events, where
a second variable, $m_{W}^{\mathrm{rec}}$, is reconstructed by the four-momenta of
the jets assigned to the $W$ bosons. This variable is related to the well-known value
of the $W$-boson mass, and the $\mathrm{JES}$
can be adjusted in such a way that both the $m_{t}^{\mathrm{rec}}$ and the $m_{W}^{\mathrm{rec}}$
distributions for simulated events match the observed data.
The inclusion of this procedure, usually referred to as {\em in situ}
calibration, enables a significant reduction of the systematic uncertainty
associated with the inaccurate knowledge of the
$\mathrm{JES}$, and represents an important improvement of the work described
in this paper with respect to the previous CDF analysis by a
similar method\,\cite{ahprd}.
The $m_{t}^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$ distributions
are reconstructed in two separate samples of selected data events, defined
by the presence of exactly one and two or more $b$-tagged jets respectively.
The data are then compared to corresponding distributions expected from background
and $t\bar t$ events simulated with various values of the top quark mass
and of the $\mathrm{JES}$ to fit for these parameters.
In addition, the fitted signal yields are used to derive
a measurement of the $t\bar{t}$ production cross section.
The results reported here are based on data taken between March 2002 and April 2008, corresponding
to an integrated luminosity of $2.9$~fb$^{-1}$.
This measurement complements other recent determinations of the top quark mass and $t\bar{t}$
cross section by
CDF and D0\,\cite{latestCDFD0mass,latestCDFD0xsec} in other final states, and improves
the latest CDF measurements in the same channel\,\cite{ahprd, latestCDFah}.
The organization of the paper is as follows: Section\,\ref{sec:detector} contains a brief description
of the CDF\,II detector. The trigger and the neural-network-based sample selection are discussed in
Sec.\,\ref{sec:dataset}, along with the identification of jets initiated by $b$ quarks ($b$ jets).
Sections\,\ref{sec:MonteCarlo} and\,\ref{sec:bkg} present the simulated signal samples
and
the data-driven method we use for estimating the background from multijet data.
Section\,\ref{sec:temp} describes how the fundamental variables
$m_{t}^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$ are reconstructed,
while in Sec.\,\ref{sec:cutoptimization}
we present the final requirements to define the samples of events
used in the measurement.
The parametrization of the dependence of the distributions of reconstructed variables
on the values of the top quark mass and the jet energy scale are described in Sec.\,\ref{sec:temppar}.
The fit to the experimental distributions and its calibration are described
in Secs.\,\ref{sec:like} and\,\ref{sec:sanity}, respectively.
Section\,\ref{sec:syst} details the study of
the systematic uncertainties on the mass measurement, that is then reported
in Sec.\,\ref{sec:meas}.
We describe in Sec.\,\ref{sec:xsec} the measurement of the $t\bar{t}$ cross section.
\section{The CDF\,II Detector}
\label{sec:detector}
The CDF\,II detector\,\cite{CDFdetector} is an azimuthally and forward-backward symmetric apparatus designed
to study $p\bar p$ collisions at the Tevatron. A cylindrical coordinate system is used
where $\theta$ is the polar angle to the
proton beam direction at the event vertex, $\phi$ is the azimuthal angle about the beam axis, and
pseudorapidity is defined as \mbox{$\eta = - \ln \left[ \tan(\theta/2)\right] $}.
We define transverse energy as $\mbox{$E_T$} = E \sin\theta$ and transverse momentum as $p_T = p\sin\theta$
where $E$ is the energy measured by calorimeters, and $p$ is the magnitude of the momentum measured
by a tracking system.
The detector consists of a magnetic spectrometer surrounded by calorimeters and muon chambers.
The charged particle tracking system is immersed in a 1.4~T solenoidal magnetic field with axis
parallel to the beamline. A set of silicon microstrip detectors provides charged particle tracking in
the radial range from 1.5 to 28~cm, while
a 3.1~m long open-cell drift chamber, the central outer tracker (COT), covers the radial range
from 40 to 137 cm.
In combination the silicon and COT detectors provide excellent tracking up to about
pseudorapidities $|\eta|\le 1.1$, and with decreasing precision up to $|\eta| \le 2.0$.
Segmented electromagnetic and hadronic calorimeters
surround the tracking system, and measure the energy deposit of particles interacting in the calorimeters.
The electromagnetic and hadronic calorimeters
are lead-scintillator and iron-scintillator
sampling devices, respectively, covering the range $|\eta|\le 3.6$.
They are segmented in the central region ($|\eta|<1.1$)
in towers of 15$^\circ$ in azimuth and 0.1
in $\eta$, and the forward region ($1.1<|\eta|<3.6$) in towers of 7.5$^\circ$ for
$|\eta|<2.11$ and 15$^\circ$ for $|\eta|>2.11$,
while the coverage in $|\eta|$ increases
gradually from 0.1 to 0.6\,.
The electromagnetic calorimeters\,\cite{ecal,pem} are instrumented
with proportional chambers (at large angles) or
scintillating strip detectors (in the forward regions), which measure
the transverse profile of electromagnetic showers
at a depth corresponding to the expected shower maxima.
Drift chambers located outside the central hadronic calorimeters and behind a 60~cm iron shield
detect muons with $|\eta| \le 0.6$\,\cite{CMU}. Additional drift chambers and scintillation
counters detect muons in the region $0.6<|\eta|<1.5$. Multicell gas Cherenkov counters\,\cite{CLC} with a
coverage of $3.7<|\eta|<4.7$ measure the average number of inelastic $p\bar p$ collisions and thereby
are used to determine the luminosity.
\section{\boldmath Multijet Event Selection and $b$ tagging}
\label{sec:dataset}
The final state of all-hadronic $\mbox{$t\bar{t}$}$ events is characterized by the presence of at least
six jets from the decay of the two top quarks, where additional jets might come from initial- or
final-state radiation (ISR or FSR). Events having such a topology are collected using a
multijet trigger
which relies on calorimeter information.
Subsequently, jets are identified during event reconstruction
by grouping clusters of energy in the calorimeter using a
fixed-cone algorithm with a radius of 0.4 in $\eta$-$\phi$ space\,\cite{jets}.
After a preliminary selection of multijet events, a neural-network selection
based on relevant kinematical variables is used to further improve the purity of the sample.
\subsection{Multijet Trigger}
The CDF trigger system has three levels.
The first two levels consist of
special-purpose electronic circuits and the third one of conventional programmable digital processors.
At level 1, the trigger requires the presence of
at least one calorimeter tower with transverse energy
$E_{T}^{\mathrm{tow}}\ge 10$~GeV. At level 2 the total
transverse energy, obtained as the sum over all calorimeter
towers, $\sum E_{T}^{\mathrm{tow}}$, must be $\ge 175$~GeV.
Moreover, the presence of least four clusters of towers,
each with transverse energy $E_{T}^{\mathrm{clus}}\ge~15$~GeV, is
required.
Finally, the third trigger level confirms the level 2 selection using a more accurate
determination of the jet energy, requiring four or more reconstructed jets with
$\mbox{$E_T$}\ge 10$~GeV.
Approximately $14\,\times 10^{6}$ events satisfy the trigger requirements,
corresponding to an events signal-over-background ratio (S/B) of about 1/1200, assuming
a theoretical cross section of $6.7$~pb for a top quark mass of 175 GeV/$c^2$\,\cite{theo-xsec}.
\subsection{Preselection and Topology Requirements}
\label{sec:presel}
Events satisfying the trigger requirements are reconstructed in terms of their final-state
observables (tracks, vertices, charged leptons, and jets). We retain only those events that are
well contained in detector acceptance, requiring the primary event vertex\,\cite{vertex}
to lie inside the luminous region ($|z|<60$~cm).
We remove
events having well identified energetic electrons or muons as defined in\,\cite{ljetsKIN},
namely electrons with $E_{T} > 20$\,GeV and muons with $p_{T} > 20$\,GeV/$c$.
In order to have jets matching as accurately as possible to the hard scattering partons,
we correct jet energies for detector response and multiple interactions\,\cite{JESNIM}.
First, we consider the $\eta$-dependence of detector response
and energy loss in the uninstrumented regions.
Then, after accounting for the small extra energy deposited by multiple
collisions in the same
beam-beam bunch crossing, a correction for calorimeter nonlinearity is applied so that
the jet energies are equal, on average, to the energy of the particles incident
on the jet cone.
The total uncertainty on the estimate of the original parton energy,
where all uncertainties
for the individual corrections are added in quadrature,
varies from 8\% to 3\% with jet transverse energy increasing from 15\,GeV to 50 GeV,
and remains approximately
constant at 3\% above 50 GeV. Jets with $|\eta|\le 2$ and $\mbox{$E_T$}\ge 15$\,GeV,
after all corrections are applied, are selected for further analysis.
As the uncertainty on the missing transverse energy, $\mbox{$\protect \raisebox{0.3ex}{$\not$}\Et$}$\,\cite{met},
increases proportionally to $\sqrt{\sum\mbox{$E_T$}}$\,\cite{metjetPRL},
its significance is defined as $\frac{\mbox{$\protect \raisebox{0.3ex}{$\not$}\Et$}}{\sqrt{\sum\mbox{$E_T$}}}$,
where the $\mbox{$\protect \raisebox{0.3ex}{$\not$}\Et$}$ is corrected for
any identified muons,
while $\sum\mbox{$E_T$}$ is obtained by summing the $E_T$'s of all the selected jets.
We then require that $\frac{\mbox{$\protect \raisebox{0.3ex}{$\not$}\Et$}}{\sqrt{\sum\mbox{$E_T$}}}$ be $<3$\,${\rm GeV}^{\frac{1}{2}}$
to select events with small $\mbox{$\protect \raisebox{0.3ex}{$\not$}\Et$}$.
At this stage, called preselection,
we are left with about $8.2\, \times \,10^{6}$ events.
As the topology of the candidate events is determined by the jet multiplicity,
we define the signal region by selecting events with a number
of jets $6\le N_{\rm jets}\le 8$ and
we also require jet pairs to be
separated by at least 0.5 units in the $\eta$-$\phi$ space.
The number of events passing these
additional requirements is $1.671\, \times \,10^{6}$
with an expected S/B of approximately 1/430.
\subsection{Neural-Network-based Kinematical Selection}
\label{sec:sel}
To further improve the purity of the signal sample,
we use a multivariate
approach and take advantage of the distinctive features of signal and
background events through a neural network, which takes into account the
correlations between the kinematical
variables which enter as input nodes in the network.
The network uses the {\sc mlpfit} package\,\cite{mlp} as implemented by {\sc root}\,\cite{root} through
the {\em TMultiLayerPerceptron} class.
\par
A first set of 11 global variables, summarized in Table\,\ref{tab:nnvar}, have already been proven to be
effective\,\cite{ahprd} in reducing the QCD background.
Studies performed for this analysis on the jet development in the calorimeter
have indicated that a good
discrimination between quark-initiated and gluon-initiated jets can be accomplished
with $\eta$ moments ($M_{\eta}$) and $\phi$ moments ($M_{\phi}$) of a jet,
which are defined as
\begin{equation}
M_\eta = \sqrt{ \left[\sum_{\mathrm{tow}} \frac{E_{T}^{\mathrm{tow}}}{E_T}\eta^2_{\mathrm{tow}} \right] - \eta^2 }
\end{equation}
and
\begin{equation}
M_\phi = \sqrt{\left[\sum_{\mathrm{tow}} \frac{E_{T}^{\mathrm{tow}}}{E_T}\phi^2_{\mathrm{tow}} \right] - \phi^2 }\,,
\end{equation}
where $E_T$, $\eta$, and $\phi$ are, respectively, the transverse energy, the pseudorapidity and
the azimuthal angle of the jet, while $E_{T}^{\mathrm{tow}}$ is the transverse energy deposited
in the calorimeter towers belonging to the jet.
We remove possible biases coming from $E_T$ distributions, which might differ in signal
and background events, by deconvoluting the $E_T$
dependence through a rescaling of all moments to a common reference value of $E_T=50$\,GeV.
We obtain what we call scaled moments\,:
\begin{equation}
M^s_\eta = M_\eta \times \frac{f^\eta_q(50~ {\rm GeV})}{f^\eta_q(E_T)}
\end{equation}
and
\begin{equation}
M^s_\phi = M_\phi \times \frac{f^\phi_q(50~ {\rm GeV})}{f^\phi_q(E_T)}\,,
\end{equation}
where $f^\eta_q(E_T)$ and $f^\phi_q(E_T)$ are the functions that fit the profiles
of $M_\eta$ $vs$ $E_T$ and of
$M_\phi$ $vs$ $E_T$ in quark-initiated jets from simulated $t\bar t$ events.
These scaled moments are quite different for jets coming from a quark or a gluon
in simulated $t\bar t$ events. Such a behavior has been verified in data events
where the jet origin is well known.
To take advantage of the large number of jets in a $t\bar t$ event, we consider the geometric average
of the $\eta$ moments and of the $\phi$ moments, see Fig.\,\ref{fig:shapes}, evaluated using all jets
which are not identified as coming from a heavy quark by the criteria
explained in Sec.\,\ref{sec:btag}.
The 13 variables are used as inputs to a neural network with two hidden layers with 20 and 10 hidden nodes,
respectively, and one output node.
The network is trained on same-size samples of signal and background events
with $6\le N_{\rm jets}\le 8$ (about half a million events).
In order to model the signal we use the {\sc pythia}\,v6.2\,\cite{Pythia}
leading-order (LO) Monte Carlo generator with parton showering followed by a simulation of
the CDF\, II detector. The reference top quark mass chosen for the training is
$M_{\mathrm{top}}=175$\,GeV/$c^2$. The background is obtained from the multijet data
events themselves, since the signal fraction is expected to be very small before applying
the neural-network selection.
The value of the output node, $N_{\mathrm{out}}$, is the quantity we use as a discriminator between
signal and background, and is shown in Fig.\,\ref{fig:nnout} for the $6\le N_{\mathrm{jets}}\le 8$ sample.
\begin{table}[hbtp]
\begin{center}
\caption{Input variables to the neural network.}\label{tab:nnvar}
\begin{tabular}{cl}
\hline\hline
\parbox[c][5mm][c]{15mm}{Variable} & ~~~~~~~~~~~~~~~~~~~~Description \\
\hline
\parbox[c][5mm][c]{15mm}{ $ \sum \mbox{$E_T$} $ } & Scalar sum of selected jets $E_T$\\
\parbox[c][5mm][c]{15mm}{ $ \sum _3\mbox{$E_T$} $ } & As above, except the two highest-$E_T$ jets\\
\parbox[c][5mm][c]{15mm}{ $ C $ } & Centrality\\
\parbox[c][5mm][c]{15mm}{ $ A $ } & Aplanarity\\
\parbox[c][5mm][c]{15mm}{ $ M_{2j}^{\mathrm{min}}$ } & Minimum dijet invariant mass\\
\parbox[c][5mm][c]{15mm}{ $ M_{2j}^{\mathrm{max}}$ } & Maximum dijet invariant mass\\
\parbox[c][5mm][c]{15mm}{ $ M_{3j}^{\mathrm{min}}$ } & Minimum trijet invariant mass\\
\parbox[c][5mm][c]{15mm}{ $ M_{3j}^{\mathrm{max}}$ } & Maximum trijet invariant mass\\
\parbox[c][5mm][c]{15mm}{ $ E_T^{\star, 1}$ } & $E_T\sin^2\theta^\star$ for the highest-$E_T$ jet\\
\parbox[c][5mm][c]{15mm}{ $ E_T^{\star, 2}$ } & $E_T\sin^2\theta^\star$ for the next-to-highest-$E_T$ jet\\
\parbox[c][5mm][c]{15mm}{ $ \langle E_T^\star\rangle$ } & Geometric mean over the remaining jets\\
\parbox[c][5mm][c]{15mm}{ $ \langle M_{\eta}^s\rangle$} & Geometric mean over the untagged jets\\
\parbox[c][5mm][c]{15mm}{ $ \langle M_{\phi}^s\rangle$} & Geometric mean over the untagged jets\\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[width=8.5cm]{./Fig01.eps}
\caption{
Geometric average of the $\eta$ scaled moments ($\langle M^{s}_{\eta} \rangle$, upper plot) and of the
$\phi$ scaled moments ($\langle M^{s}_{\phi} \rangle$, lower plot)
for QCD multijet (solid histogram) and
simulated $\mbox{$t\bar{t}$}$ (dashed histogram) events with $6\le N_{\mathrm{jets}}\le 8$.
}
\label{fig:shapes}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=8.0cm]{./Fig02.eps}
\caption{
Neural-network output $N_{\mathrm{out}}$ for QCD multijet (solid histogram) and
simulated $\mbox{$t\bar{t}$}$ (dashed histogram) events with $6\le N_{\mathrm{jets}}\le 8$.
Histograms are normalized to unity. The neural-network implementation
that we use in the
{\em TMultiLayer-Perceptron} produces
an output which is not strictly bound between 0 and 1.
}
\label{fig:nnout}
\end{figure}
\subsection{\boldmath Tagging $b$ quarks}
\label{sec:btag}
In order to enrich the $t\bar t$ content in the event sample, we use a $b$-tagging algorithm based on
secondary vertex reconstruction
as described in detail in\,\cite{vertex,massTMT}. The algorithm identifies
a jet likely to contain a hadron with a $b$ quark
by reconstructing
its decay vertex with at least two
high-quality tracks with hits in the silicon vertex detector.
A $b$-tagged jet ({\em tag}, in brief) must have
an associated secondary vertex with a displacement from the primary vertex
in the transverse
plane larger than 7.5 times the transverse-displacement resolution.
This is evaluated for each secondary vertex, but its typical
value is about 190\,$\mu$m.
The tagging efficiencies for jets coming from the fragmentation of $b$ or $c$ quarks
are corrected in simulated events according to the efficiency seen in the data,
by a factor $0.95\pm 0.04$, both for $b$ jets
and $c$ jets. These factors are described in detail in\,\cite{vertex}.
\section{Event Simulation}
\label{sec:MonteCarlo}
The standard model $t\bar{t}$ events used to study the event selection
and to check the performance of the method (Sec.\,\ref{sec:sanity}) are simulated using
{\sc pythia}\,v6.2\,\cite{Pythia}. Samples generated with input values of the top quark mass,
$M_{\mathrm{top}}^{\mathrm{in}}$, ranging from 160 to 190 GeV$/c^{2}$ are considered and, for each sample, the event selection
is repeated by varying the $\mathrm{JES}$ from its default value\,\cite{JESNIM}.
The displacement, denoted as $\Delta\mathrm{JES}$, is measured relative to
the uncertainty, $\sigma_{\mathrm{JES}}$, on the default value itself, so that
the value of $\mathrm{JES}$ applied to simulated events is increased by
$\Delta\mathrm{JES} \cdot \sigma_{\mathrm{JES}} $ with respect to the default.
To test the method, input values $\Delta\mathrm{JES}^{\mathrm{in}}$ ranging from $-3$ to $+3$
are considered.
Different generators and different values for the model parameters are
used to estimate the systematic uncertainties, as described in Sec.\,\ref{sec:syst}.
\section{Background Estimate}
\label{sec:bkg}
The background for the $\mbox{$t\bar{t}$}$ multijet final state comes mainly from QCD
production of
heavy-flavor quark pairs ($b\bar b$ and $c\bar c$) and
from false tags of light-flavor quark jets.
Other standard model processes such as $W/Z$+jets have a smaller
production cross section and small acceptance due to the selection requirements.
Given the large theoretical uncertainties on the QCD multijet production cross section,
a more accurate background estimate is
obtained from the data, rather than from Monte Carlo simulations.
A tag rate per jet, defined as the probability of tagging a jet whose tracks
are reconstructed in the vertex detector ({\it fiducial} jet), is then evaluated in
a sample of events with exactly four jets passing the preselection
and therefore still dominated by the background (S/B $\approx 1/5000$).
The rate is parametrized in terms
of variables sensitive to both the tagging efficiency for heavy-flavored objects
and the probability of false tags\,: the jet $\mbox{$E_T$}$, the number of tracks reconstructed in the
silicon vertex detector and associated with the jet, $N_{\rm trk}^{\mathrm{jet}}$,
and the number of primary vertices reconstructed in
the event, $N_{\rm vert}$~\cite{ahprd}.
By definition, the tag rate estimates the probability that a fiducial jet
having, on average, the characteristics of jets from background events is tagged.
Its average value is about $3.7\%$, with negligible uncertainty.
However, direct exploitation of the tag rate to predict the number of background events
with exactly a given number of tags would give incorrect numbers.
This happens because, by construction, this rate is the ratio between the
number of tagged jets and the number of fiducial jets in a whole sample of events.
Possible correlations among jets in the same event are not considered.
As heavy-flavor quarks come in pairs in QCD background, the probability
to tag a pair of jets
in the same event is therefore larger than the simple product of the tag probabilities
of individual jets given by the tag rate.
To account for this we introduce correction factors to obtain
a better estimate for the number of 1-tag and $\ge 2$-tag background events.
These factors are derived in a control sample dominated by the background
(events with six, seven or eight jets and $N_{\mathrm{out}} \le 0.25$, with S/B\,$\approx 1/1300$ for
one tag and S/B$\approx 1/400$ for $\ge 2$ tags)
as the ratio between the observed number of events with $n$ tags (with $n = 1,\,2,\,3$)
and the average expectation obtained
by using the tag rate to evaluate the probability for each event to have
the same number, $n$, of tagged jets.
These factors represent, therefore, average corrections
to the probability for a possible {\em tag configuration},
that is for the assumption that among the fiducial jets in an event of the sample
selected before the $b$-tagging requirements ({\em pretag} sample)
only a given subset is actually tagged when the algorithm is applied.
Their average values are $0.94$, $1.48$, and $2.46$ for events with one, two, and three tagged
jets, with relative statistical uncertainties of $0.4\%$, $1.1\%$, and $5.1\%$ respectively.
Similarly to the tag rate, these corrections should be valid for events with the
characteristics of background events.
The accuracy of our modeling of the background processes is verified in control samples,
i.e. on events with
higher values of $N_{\mathrm{out}}$ and therefore with a larger
fraction of signal events and with possible different kinematics and background composition.
As the background prediction is performed using the data in the pretag sample,
the presence of $t\bar{t}$ events must also be taken into account.
Therefore a correction is applied to derive a better evaluation, $n_{(b,\,\mathrm{exp})}$, of the background
normalization from the raw estimate $n_{(b,\,\mathrm{raw})}$ directly
obtained by the corrected tag rate matrix.
This correction must subtract the contribution, $n_{t\bar{t}}^{tr} $,
coming from applying the matrix to signal events and included in $n_{(b,\,\mathrm{raw})}$.
Denoting by $N_{\mathrm{obs}}$ the number of events observed in the data sample,
by $n_{t\bar{t}}$ the number of signal events in this sample and
assuming that the excess of events with respect to the expected background
is totally due to the signal, the correction can be written as\,:
\begin{eqnarray}
n_{(b,\,\mathrm{exp})} & = & n_{(b,\,\mathrm{raw})} - n_{t\bar{t}}^{tr} \nonumber \\
& = & n_{(b,\,\mathrm{raw})} - \frac{ n_{t\bar{t}}^{tr} }{ n_{t\bar{t}} } \cdot n_{t\bar{t}} \nonumber \\
& = & n_{(b,\,\mathrm{raw})} - \frac{ n_{t\bar{t}}^{tr} }{ n_{t\bar{t}} } \cdot (N_{\mathrm{obs}} - n_{(b,\,\mathrm{exp})} ) \,,~~~~~
\end{eqnarray}
which, with $R_{t\bar{t}} \equiv n_{t\bar{t}}^{tr} / n_{t\bar{t}} $, gives
\begin{eqnarray}
n_{(b,\,\mathrm{exp})} & = & \frac{ n_{(b,\,\mathrm{raw})} - R_{t\bar{t}} \cdot N_{\mathrm{obs}} }{ 1 - R_{t\bar{t}} } \,.
\end{eqnarray}
$R_{t\bar{t}}$ can be inferred from simulated events and amounts to
$0.314 \pm 0.003$\,(stat) [$0.067 \pm 0.0014$\,(stat)] for 1-tag ($\geq 2$-tag) events.
Further possible discrepancies between the observed and expected number of
events are considered as due to the modeling of the background
and accounted for as a systematic uncertainty.
\section{Mass reconstruction}
\label{sec:temp}
The simultaneous measurement of the top quark mass and the $\mathrm{JES}$
is based on the reconstruction, event by event,
of both the top quark and the $W$ masses through a constrained fitting technique.
The shapes of the distributions obtained by this procedure are sensitive to
the values of both $M_{\mathrm{top}}$ and $\mathrm{JES}$. Therefore, for simulated events,
they are built using samples corresponding to the different input values of $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$
listed in Sec.\,\ref{sec:MonteCarlo}.
Moreover, given the different resolution in the reconstructed top quark mass and the $W$ boson mass,
and also the different S/B which can be achieved by requiring events with exactly one or $\ge 2$ tags,
two sets of distributions are separately derived in these samples.
\subsection{Reconstructed top quark mass}
\label{subsec:Mtoptemplates}
For each event we determine a reconstructed top quark mass, $m_t^{\mathrm{rec}}$,
from the four-momenta of selected jets.
Sixteen equations can be considered to connect the four-momenta of the two top quarks and their
decay products according to the
\mbox{$t \bar t \rightarrow b \bar b \, W^+ W^- \rightarrow b \bar b \, q_1 \bar q_2 \, q_3 \bar q_4$} hypothesis\,:
\begin{eqnarray}
p_t^\mu & = & p_{W^+}^\mu + p_b^\mu \,, \\% \nonumber \\
p_{\bar t}^\mu & = & p_{W^-}^\mu + p_{\bar b}^\mu\,, \\% \nonumber \\
p_{W^+}^\mu & = & p_{q_1}^\mu + p_{\bar q_2}^\mu \,, \\% \nonumber \\
p_{W^-}^\mu & = & p_{q_3}^\mu + p_{\bar q_4}^\mu \,,
\end{eqnarray}
with $\mu = 0,\,1,\,2,\,3$.
There are 13 unknown quantities, i.e., the unknown top quark mass and the three-momenta of
the top quarks and of the $W$ bosons,
so the kinematics of the events are overconstrained.
The fit is performed using only the six highest-$E_T$ jets (leading jets) of the
event and considering their possible
assignments to quarks of a $t \bar t$ final state. The total number of different permutations
giving two doublets of jets corresponding to the $W$ bosons and two triplets of jets corresponding
to the top quarks is 90. Since we require the presence of $b$ tags, assigning
the tagged jets only to $b$ quarks reduces this number to 30 for 1-tag events and 6
in case of two or more $b$ tags\,\cite{twotagfit}.
For each permutation
the kinematics of the event is reconstructed minimizing the following $\chi^2$ function\,:
\begin{eqnarray}
\chi^{2} & = & ~\frac{ \big(m_{jj}^{(1)}-M_W \big)^2 }{ \Gamma^2_W } +
\frac{ \big(m_{jj}^{(2)}-M_W \big)^2 }{ \Gamma^2_W } \nonumber \\
& + & \frac{ \big(m_{jjb}^{(1)}-m_{t}^{\mathrm{rec}} \big)^2 }{ \Gamma^2_t } +
\frac{ \big(m_{jjb}^{(2)}-m_{t}^{\mathrm{rec}} \big)^2}{\Gamma^2_t}~~~~~~~ \nonumber \\
& + & \sum^6_{i=1} \frac{ \big(p^{\mathrm{fit}}_{T,i}-p^{\mathrm{meas}}_{T,i} \big)^2}{\sigma^2_i} \,.
\end{eqnarray}
The minimization procedure is performed with respect to seven parameters, i.e., the reconstructed top quark mass
$m_{t}^{\mathrm{rec}}$ and the transverse momenta $p^{\mathrm{fit}}_{T,i}$ of
the six jets, which are constrained
to the measured value $p^{\mathrm{meas}}_{T,i}$ within their known resolution $\sigma_{i}$.
The invariant masses of the jet doublets assigned to light-flavor quarks coming from a $W$, $m_{jj}^{(1,2)}$,
and of the trijet systems including one doublet and one of the jets assigned to $b$ quarks, $m_{jjb}^{(1,2)}$,
are evaluated by the trial momenta of jets at each step of the minimization.
On the contrary, the measured mass $M_W$ and the natural width $\Gamma_W$ of the $W$ boson
as well as the assumed natural width of the
top quark, $\Gamma_t$, are kept constant to $80.4\,$GeV/$c^2$, $2.1$ GeV/$c^2$ and $1.5$\,GeV/$c^2$
respectively\,\cite{pdg,widt}.
The permutation of jets which gives the lowest $\chi^2$ value is selected, and the
corresponding fitted value of $m_{t}^{\mathrm{rec}}$ enters
an invariant mass distribution ({\em template})
which will be used for the $M_{\mathrm{top}}$ measurement.
\subsection{\boldmath Reconstructed $W$ mass}
\label{subsec:Mwtemplates}
Reconstructing the mass of $W$ bosons by using dijet systems represents a possibility to
obtain a variable,
in principle, insensitive to $M_{\mathrm{top}}$ which allows, therefore,
an independent determination of $\mathrm{JES}$.
To build the $m_{W}^{\mathrm{rec}}$ distributions we use the same procedure and $\chi^2$ expression
considered for $m_{t}^{\mathrm{rec}}$,
but now the $W$-boson mass is also left as a free parameter in the fit
(i.e. $M_{W}$ becomes $m_{W}^{\mathrm{rec}}$).
Again, for each event, the value of $m_{W}^{\mathrm{rec}}$ corresponding to the permutation of the
jet-to-parton assignments with the lowest $\chi^2$ enters the distribution.
Using different fits in the reconstruction of $m_t^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$ can lead to selecting
different assignments of jets to partons for the two variables in the same event.
This is not
a problem as the same procedure is followed both on data and simulated events.
Reconstructing the top quark
mass using a constant value of $M_{W}$, as decribed in Sec.\,\ref{subsec:Mtoptemplates}, improves the
resolution of the distributions and therefore the determination of the {\em true} value of $M_{\mathrm{top}}$.
The correlations between the values of $m_t^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$ in the same event are taken into
account in the calibration of the likelihood fit used for the measurement (Sec.\,\ref{sec:like}).
\subsection{Background templates}
\label{sec:bkgtemplates}
In order to reconstruct data-driven background templates we apply the kinematical fitter
to the sample of events passing the neural-network selection, but before the requirement
of tagged $b$ jets.
The same procedures described in Secs.~\ref{subsec:Mtoptemplates}
and~\ref{subsec:Mwtemplates} are repeated
on these events assigning fiducial jets to $b$ quarks and then looping over all possible
assignments of other jets to the remaining quarks, performing the fit
for each permutation and selecting the reconstructed $m_t^{\mathrm{rec}}$ and $m_W^{\mathrm{rec}}$ values
corresponding to the best $\chi^{2}$. These values then enter the templates weighted
by the {\em corrected} probability
of the assumed tag configuration; see Sec.~\ref{sec:bkg}.
As for the normalization, the
background distributions also need to be corrected
for the presence of signal in the pretag sample by subtracting
the contribution from $t\bar{t}$ events.
The shape of this contribution is obtained from simulated samples
and depends on the assumed $M_{\mathrm{top}}$ and $\mathrm{JES}$, while the normalization
is given by the difference $n_{(b,\,\mathrm{raw})} - n_{(b,\,\mathrm{exp})}$,
as described in Sec.\,\ref{sec:bkg}.
In order to check how well our modeling describes the background, we consider events in
control regions defined by the $N_{\mathrm{out}}$ value, in ranges where the signal presence after
tagging is still very low.
In these regions the templates, i.e. the main elements of our measurement, are reconstructed
by the procedure described in the previous sections, both for the signal
and the background, as well as other important distributions like $N_{\mathrm{out}}$ and
the $\chi^{2}$ of the fit used to build the $m_{t}^{\mathrm{rec}}$ templates.
These distributions are then compared to observed data, taking into account the
contribution from signal events.
The agreement is generally good in all the control regions,
and this confirms the reliability of the background model.
Figures~\ref{fig:MtopValid} and~\ref{fig:MwValid} show, as examples,
distributions of $m_{t}^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$
in one of the control regions
for 1-tag and $\ge 2$-tag events, where the sum of signal and background
is compared to the same distributions reconstructed in the data. In these plots the
integral of the signal distributions
corresponding to $M_{\mathrm{top}}=175$\,GeV/$c^{2}$ and the default
value $\Delta\mathrm{JES}$=0
have been normalized to the difference between the observed data and
the corrected expected background.
\begin{figure}[htbp!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig03.eps,width=8cm} \\
\epsfig{file=./Fig04.eps,width=8cm}
\end{tabular}
\caption{
Histograms of the reconstructed top quark mass $m_{t}^{\mathrm{rec}}$ for 1-tag events, upper plot,
and $\ge 2$-tag events, lower plot, are shown in a control region defined by
$0.75 \le N_{\mathrm{out}} < 0.85$.
Along with the data are plotted the
expected background and the signal contribution
for $M_{\mathrm{top}}=175$\,GeV/$c^{2}$ and the default value $\Delta\mathrm{JES} = 0$,
normalized to the difference between the data and the background.
The value of the purely statistical
$\chi^{2}$ probability is reported on each plot.
}
\label{fig:MtopValid}
\end{center}
\end{figure}
\begin{figure}[htbp!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig05.eps,width=8cm} \\
\epsfig{file=./Fig06.eps,width=8cm}
\end{tabular}
\caption{
Histograms of the reconstructed $W$ mass $m_{W}^{\mathrm{rec}}$ for 1-tag events, upper plot,
and $\ge 2$-tag events, lower plot, are shown in a control region defined by
$0.75 \le N_{\mathrm{out}} < 0.85$.
Along with the data are plotted the
expected background and the signal contribution
for $M_{\mathrm{top}}=175$\,GeV/$c^{2}$ and the default value $\Delta\mathrm{JES} = 0$,
normalized to the difference between the data and the background.
The value of the purely statistical
$\chi^{2}$ probability is reported on each plot.
}
\label{fig:MwValid}
\end{center}
\end{figure}
\section{Event samples}
\label{sec:cutoptimization}
In order to obtain the best performance from our method, we performed sets of {\em pseudoexperiments} (PEs)
to find out which requirements on the
values of $N_{\mathrm{out}}$ and of the $\chi^{2}$ used to obtain
the $m_{t}^{\mathrm{rec}}$ values minimize the statistical uncertainty on the top quark mass measurement.
The procedure is
similar to the one outlined in Sec.~\ref{sec:sanity},
with a binned version of the same likelihood.
It is applied separately to 1-tag and $\geq 2$-tag samples and considers many different combinations
of possible requirements.
The smallest values for the uncertainty are obtained using $\left(N_{\mathrm{out}} \ge 0.90,\,\chi^{2} \le 6\right)$
in the 1-tag sample and $\left(N_{\mathrm{out}} \ge 0.88,\,\chi^{2} \le 5\right)$ in the $\ge 2$-tag sample so that
we add these requirements to the prerequisites described in Sec.~\ref{sec:presel}.
The final definition of the samples used in our analysis is summarized in Table~\ref{tab:samples}.
\begin{table}[htbp]
\begin{center}
\caption{Final definition and requirements for selected event samples.}
\label{tab:samples}
\begin{tabular}{lccc}
\hline \hline
Event Sample~~~ & $b$-tags & \parbox[c][5mm][c]{20mm}{$N_{\mathrm{out}}$} & $m_{t}^{\mathrm{rec}}$ fit $\chi^{2}$ \\
\hline
One tag & $\equiv 1$ & \parbox[c][5mm][c]{20mm}{ $ \ge 0.90 $ } & $ \le 6 $ \\
$\ge 2$-tags & $ 2$ or $3$ & \parbox[c][5mm][c]{20mm}{ $ \ge 0.88 $ } & $ \le 5 $ \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\noindent After these selections, 3452 and 441 events are observed for the 1-tag and $\ge 2$-tag samples
respectively. We can evaluate the average number of background events expected in
the selected samples and their uncertainties, as described in Sec.~\ref{sec:bkg}.
The systematic uncertainties on the background normalizations are estimated
by assuming that the discrepancy between the observed number of events in the data and
the sum of the expected contributions from signal and background
(where, in this case, the theoretical cross section value of
$6.7$\,pb is considered for $t\bar{t}$ events production) is due to a bad evaluation
of the background.
This is done separately for 1-tag and $\ge 2$-tag samples, and the resulting
relative uncertainties on the expected number of events are
\mbox{$\sigma( n_{(b,\,\mathrm{exp})}^{1\,tag} ) = 2.9\%$}
and \mbox{$\sigma( n_{(b,\,\mathrm{exp})}^{\ge 2\,tags}) = 14.6\%$}, respectively.
The efficiencies of the full selection on $t\bar{t}$ events corresponding
to $M_{\mathrm{top}} = 175$\,GeV/$c^{2}$ and $\Delta\mathrm{JES} = 0$ are $3.6\%$ and $1.0\%$ for 1-tag
and $\ge 2$-tag events respectively.
These values are used to evaluate the expected signal contributions of Table\,\ref{tab:finalsamples},
where $\sigma_{t\bar{t}} = 6.7$\,pb is assumed.
In the same table, the observed number of events and the expected background in each sample are also summarized.
\begin{table}[htbp]
\begin{center}
\caption{Number of events observed in the selected data samples and corresponding
expected numbers of background and $t\bar{t}$ events. The signal contribution
is evaluated for $M_{\mathrm{top}} = 175$\,GeV/$c^{2}$, $\Delta\mathrm{JES} = 0$ and $\sigma_{t\bar{t}} = 6.7$\,pb.
}
\label{tab:finalsamples}
\begin{tabular}{lccc}
\hline \hline
Event sample & \parbox[c][5mm][c]{20mm}{Observed} & Background & \parbox[c][5mm][c]{17mm}{$t\bar{t}$}\\
\hline
One tag & \parbox[c][5mm][c]{20mm}{ $3452$ } & $2785 \pm 83$ & $693$ \\
$\ge 2$ tags & \parbox[c][5mm][c]{20mm}{ $ 441$ } & $ 201 \pm 29$ & $193$ \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\section{Likelihood fit}
\label{sec:likefit}
The technique described in Sec.\,\ref{sec:temp}
allows one to obtain sets of {\em observed} $m_t^{\mathrm{rec}}$ and $m_W^{\mathrm{rec}}$
values reconstructed in the data samples with 1 or $\ge 2$ tags
as well as to build signal and background distributions for the same
variables.
In order to measure the top quark
mass simultaneously with the $\mathrm{JES}$, a fit is performed where an
unbinned likelihood function is maximized
to find the values of $M_{\mathrm{top}}$, $\Delta\mathrm{JES}$,\, and the number of
signal ($n_{s}$) and background ($n_{b}$)
events for each tagging category which give
the probability density functions (p.d.f.'s)
best describing the data.
\subsection{Probability density functions}
\label{sec:temppar}
The signal templates are fitted by normalized combinations
of Gamma and Gaussian p.d.f.'s, and the dependence of the shape
on input $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$ is included writing
the parameters of the p.d.f.'s as linear functions of these
variables.
Figures~\ref{fig:pdftop} and~\ref{fig:pdfw}
show examples of the fitted p.d.f.'\,s superimposed on the
$m_{t}^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$
signal templates respectively for different $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$ values.
\begin{figure}[tbp!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig07.eps,width=8cm}
\\
\epsfig{file=./Fig08.eps,width=8cm}
\end{tabular}
\caption{
Histograms and corresponding fitted probability density functions for the
signal $m_{t}^{\mathrm{rec}}$ in
$\ge 2$-tag events
for a constant $\Delta\mathrm{JES}$ value ($\Delta\mathrm{JES}=0$),
varying the input top quark mass (upper plot)
and for a constant $M_{\mathrm{top}}$ value (175\,GeV/$c^2$), varying the input jet energy scale
(lower plot).
}
\label{fig:pdftop}
\end{center}
\end{figure}
\begin{figure}[htbp!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig09.eps,width=8cm}
\\
\epsfig{file=./Fig10.eps,width=8cm}
\end{tabular}
\caption{
Histograms and corresponding fitted probability density functions
for the signal $m_{W}^{\mathrm{rec}}$ in
$\ge 2$-tag events
for a constant $\Delta\mathrm{JES}$ value ($\Delta\mathrm{JES}=0$),
varying the input top quark mass (upper plot),
where the independence of $m_{W}^{\mathrm{rec}}$ on $M_{\mathrm{top}}$ is apparent,
and for a constant $M_{\mathrm{top}}$ value (175\,GeV/$c^2$), varying the input jet energy scale
(lower plot).
}
\label{fig:pdfw}
\end{center}
\end{figure}
The shape of distributions built for the background cannot depend on the characteristics
of signal events, and in particular on the value of top quark mass.
Moreover, as they are obtained from data, the shapes correspond to the
reference value of the jet energy scale. For these reasons
no dependence on $M_{\mathrm{top}}$ and $\mathrm{JES}$ is considered in the p.d.f.'s used
to fit the background templates.
Actually, a very weak dependence is introduced through the corrections to the shape
of the background distributions, performed to take into account the presence of signal events
in the pretag sample, as described in Sec.\,\ref{sec:bkgtemplates}.
These effects are taken into account as a systematic uncertainty.
Examples of background $m_{t}^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$ distributions and the
corresponding fitted p.d.f.'s are shown
in Fig.~\ref{fig:pdfbkg} for $\ge 2$-tag events. Discrepancies between the
fitted p.d.f.'s and the corresponding distributions are considered in the
calibration procedure, presented in Sec.\,\ref{sec:sanity}.
\begin{figure}[th!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig11.eps,width=8cm} \\
\epsfig{file=./Fig12.eps,width=8cm}
\end{tabular}
\caption{
Data-driven background histograms of $m_{t}^{\mathrm{rec}}$ (upper plot)
and $m_{W}^{\mathrm{rec}}$ (lower plot) for $\ge 2$-tag events.
The bands denote the 1$\sigma$ uncertainty on the bin contents
of the histograms, including both statistical and
systematic contributions.
The solid lines show the p.d.f.'s fitted to the histograms.
}
\label{fig:pdfbkg}
\end{center}
\end{figure}
\subsection{The Likelihood Function}
\label{sec:like}
The likelihood function ${\cal L}$ is divided into three main parts and can be written as
\begin{eqnarray}
\label{eq:totallike}
{\cal L} = {\cal L}_{1\,\mathrm{tag}} \times {\cal L}_{\ge 2 \,\mathrm{tags}} \times {\cal L}_{\Delta\mathrm{JES}_{\mathrm{constr}}}~~~ \,.
\end{eqnarray}
The ${\cal L}_{1}$ and ${\cal L}_{\ge 2 \, \mathrm{tags}}$ terms further consist of other factors\,:
\begin{eqnarray}
{\cal L}_{1,\ge 2\, \mathrm{tags}} = {\cal L}_{M_{\mathrm{top}}} \times {\cal L}_{\mathrm{JES}} \times {\cal L}_{\mathrm{Poiss}}
\times {\cal L}_{N^{\mathrm{bkg}}_{\mathrm{constr}}}\,,~~~
\end{eqnarray}
where the four terms on the right side
assume respectively the following form [the superscripts referring to the
tag sample are omitted and $f_{s} \equiv n_{s} / (n_{s} + n_{b})$, $f_{b} \equiv 1 - f_{s}$]\,:
{\footnotesize
\begin{eqnarray}
\label{eq:mtreclike}
\prod_{i=1}^{N_{\mathrm{obs}}} f_{s} \cdot P^{m_{t}^{\mathrm{rec}}}_{\mathrm{sig}}(m_{t,\,i}\,|\,M_{\mathrm{top}},\,\Delta\mathrm{JES})
+ f_{b} \cdot P^{m_{t}^{\mathrm{rec}}}_{\mathrm{bkg}}(m_{t,\,i})\,, ~~~~~
\end{eqnarray}
}
{\footnotesize
\begin{eqnarray}
\label{eq:mwreclike}
\prod_{i=1}^{N_{\mathrm{obs}}} f_{s} \cdot P^{m_{W}^{\mathrm{rec}}}_{\mathrm{sig}}(m_{W,\,i}\,| \, M_{\mathrm{top}},\, \Delta\mathrm{JES})
+ f_{b} \cdot P^{m_{W}^{\mathrm{rec}}}_{\mathrm{bkg}}(m_{W,\,i})\,, ~~~
\end{eqnarray}
}
{\small
\begin{eqnarray}
\label{eq:poisslike}
\frac{e^{-(n_s + n_b)} \cdot (n_s + n_b)^{N_{\mathrm{obs}}}}{N_{\mathrm{obs}}!} \,
\end{eqnarray}
}
{\small
\begin{eqnarray}
\label{eq:bkgconstlike}
\exp{\left[- ~ \frac{\left(n_{b}-n_{(b,\,\mathrm{exp})}\right)^2}{2\sigma_{n_{(b,\,\mathrm{exp})}}^2}\right]}\,.
\end{eqnarray}
}
In expression~(\ref{eq:mtreclike}) the probability to observe the set
$m_{t,\,i},~(i=1,...,N_{\mathrm{obs}})$ of $m_{t}^{\mathrm{rec}}$ values
reconstructed in the data is calculated by using the total probability density
function resulting from the combination
of the parametrized signal and background p.d.f.'s (Sec.\,\ref{sec:temppar}),
$P^{m_{t}^{\mathrm{rec}}}_{\mathrm{sig}}$ and $P^{m_{t}^{\mathrm{rec}}}_{\mathrm{bkg}}$ respectively,
as a function of the free parameters of the fit.
In term~(\ref{eq:mwreclike}) the same is done for the set of the observed
$W$ masses, $m_{W,\,i},~(i=1,...,N_{\mathrm{obs}})$,
and the $m_{W}^{\mathrm{rec}}$ p.d.f.
The term~(\ref{eq:poisslike}), ${\cal L}_{\mathrm{Poiss}}$, gives the probability to observe the
number of events selected in the data, given the average number of
signal ($n_{s}$) and background ($n_{b}$) events expected in the sample, as assumed at each step of the
likelihood fit. In the last term,~(\ref{eq:bkgconstlike}),
the parameter $n_{b}$ is constrained by a Gaussian to the {\em a priori} background estimate given in
Sec.~\ref{sec:cutoptimization}, i.e. $n_{(b,\,\mathrm{exp})} = 2785\, \pm \, 83$ for 1-tag events and
$n_{(b,\,\mathrm{exp})} = 201\, \pm \, 29$ for $\ge 2$-tag events.
Finally, the last term in expression~(\ref{eq:totallike}), ${\cal L}_{\Delta\mathrm{JES}_{\mathrm{constr}}}$,
is a Gaussian term constraining
$\Delta\mathrm{JES}$ to its {\em a priori} value\,:
\begin{eqnarray}
\exp{\left[- ~ \frac{\left( \Delta\mathrm{JES} - \Delta\mathrm{JES}_{\mathrm{constr}}\right)^{2}}{2}\right]}\,.
\end{eqnarray}
When the measurement is performed on data, the $\mathrm{JES}$ can be constrained
to the value independently measured in~\cite{JESNIM}.
Given the meaning of $\Delta\mathrm{JES}$, described in Sec.\,\ref{sec:MonteCarlo},
this means that, in this case, $\Delta\mathrm{JES}_{\mathrm{constr}} = 0$.
\section{Verification and Calibration of the Method}
\label{sec:sanity}
We want to investigate the possible presence of biases in the top quark mass and
jet energy scale measurements introduced
by our method, as well as to have an estimate of its statistical power
before performing the measurement on the actual data sample.
To do so, we run realistic PEs where {\em pseudodata} are extracted
from simulated signal and data-driven background distributions.
A set of 3000 PEs is performed for each simulated value of the top quark mass and
of the displacement in the jet energy scale (Sec.\,\ref{sec:MonteCarlo}).
Using the notation introduced in Sec.\,\ref{sec:MonteCarlo}, we refer to these input
values as $M_{\mathrm{top}}^{\mathrm{in}}$ and $\Delta\mathrm{JES}^{\mathrm{in}}$ and they represent the
{\em true} values we want to measure.
In each PE the actual numbers of signal ($N_{(s,\,obs)}$) and
background ($N_{(b,\,obs)}$) events in each
tagging category are generated with Poisson distributions
with mean $n_{(s,\,\mathrm{exp})} = N_{\mathrm{obs}} - n_{(b,\,\mathrm{exp})}$ and $n_{(b,\,\mathrm{exp})}$ respectively,
where $N_{\mathrm{obs}}$ are the observed number of events in the data samples
($N_{\mathrm{obs}} = 3452 $ for 1-tag and
$N_{\mathrm{obs}} = 441 $ for $\ge 2$-tag events).
A set of $N_{(s,\,obs)}$ and $N_{(b,\,obs)}$ mass values is then drawn
from $m_{t}^{\mathrm{rec}}$ and $m_{W}^{\mathrm{rec}}$ distributions of signal and background
and used as input to the likelihood fit (Sec.\,\ref{sec:likefit}) that returns
simultaneous measurements of $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$, denoted as
$M_{\mathrm{top}}^{\mathrm{out}}$ and $\Delta\mathrm{JES}^{\mathrm{out}}$.
The average of these measurements over the whole set of 3000 PEs represents the best
estimate of the input values obtained by the fitting procedure
and therefore can be used to study its behavior.
We fit the dependence of these averages
with respect to the input values over the whole range of
simulated $M_{\mathrm{top}}^{\mathrm{in}}$ and $\Delta\mathrm{JES}^{\mathrm{in}}$ as
{\small
\begin{eqnarray}
\langle M_{\mathrm{top}}^{\mathrm{out}} \rangle~~ & = & (A_{00} + A_{01} \cdot \Delta\mathrm{JES}^{\mathrm{in}}) \nonumber\\
& + & (A_{10} + A_{11} \cdot \Delta\mathrm{JES}^{\mathrm{in}}) \cdot (M_{\mathrm{top}}^{\mathrm{in}} - 175)
\,,~~~~~~~ \label{eq:MtopCalib}
\end{eqnarray}
}
{\small
\begin{eqnarray}
\langle \Delta\mathrm{JES}^{\mathrm{out}} \rangle & = & \left[ B_{00} + B_{01} \cdot (M_{\mathrm{top}}^{\mathrm{in}} - 175)\right] \nonumber \\
& + &
\left[ B_{10} + B_{11} \cdot (M_{\mathrm{top}}^{\mathrm{in}} - 175)\right] \cdot \Delta\mathrm{JES}^{\mathrm{in}}\,.~~~~~~~ \label{eq:JESCalib}
\end{eqnarray}
}
These relations can be inverted to obtain calibration functions to be applied to further measurements
and therefore, on average, a more reliable estimate of the true values (2D calibration).
The calibrated values resulting from a measurement giving $M_{\mathrm{top}}^{\mathrm{out}}$
and $\Delta\mathrm{JES}^{\mathrm{out}}$
are denoted as $M_{\mathrm{top}}^{\mathrm{corr}}$ and $\Delta\mathrm{JES}^{\mathrm{corr}}$,
while the respective uncertainties,
obtained by propagating through the calibration the uncertainties from the likelihood fit,
are $\delta M_{\mathrm{top}}^{\mathrm{corr}}$ and $\delta \Delta\mathrm{JES}^{\mathrm{corr}}$.
A second set of PEs is then performed to test the goodness of the procedure.
Table~\ref{tab:calib} shows the coefficients $A_{ij}$ and $B_{ij}$ obtained both
from calibrated and uncalibrated PEs compared to their ideal values in the absence
of any bias.
\begin{table}[htbp]
\begin{center}
\caption{Coefficients of expressions~(\ref{eq:MtopCalib}) and~(\ref{eq:JESCalib})
as obtained from calibrated and uncalibrated pseudoexperiments.
The ideal values in the absence of any bias are also shown.
}
\label{tab:calib}
\begin{tabular}{lccc}
\hline \hline
\parbox[c][4.5mm][c]{10mm}{} & \parbox[c][4.5mm][c]{32mm}{Uncalibrated} & \parbox[c][4.5mm][c]{22mm}{Calibrated} &
\parbox[c][4.5mm][c]{12mm}{Ideal} \\
& PEs & PEs & value \\
\hline
$A_{00}$ & \parbox[c][4.5mm][c]{30mm}{$175.47 \pm 0.01$} & $174.99 \pm 0.01 $ & $175$ \\
$A_{01}$ & \parbox[c][4.5mm][c]{30mm}{$-0.24 \pm 0.01$} & $ 0.00 \pm 0.01 $ & $0$ \\
\hline
$A_{10}$ & \parbox[c][4.5mm][c]{30mm}{$0.985 \pm 0.002$} & $1.000 \pm 0.002$ & $1$ \\
$A_{11}$ & \parbox[c][4.5mm][c]{30mm}{$0.009 \pm 0.001$} & $0.000 \pm 0.001$ & $0$ \\
\hline
$B_{00}$ & \parbox[c][4.5mm][c]{30mm}{$-0.026 \pm 0.003$} & $0.002 \pm 0.003$ & $0$ \\
$B_{01}$ & \parbox[c][4.5mm][c]{30mm}{$0.0009 \pm 0.0004$} & $-0.0001 \pm 0.0004$ & $0$ \\
\hline
$B_{10}$ & \parbox[c][4.5mm][c]{30mm}{$1.052 \pm 0.002$} & $0.999 \pm 0.002$ & $1$ \\
$B_{11}$ & \parbox[c][4.5mm][c]{30mm}{$-0.0016 \pm 0.0002$} & $0.0001 \pm 0.0002$ & $0$ \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
In Fig.~\ref{fig:linmtopjes2} examples of linearity plots
are shown for calibrated PEs. These plots, together with the numbers in Table~\ref{tab:calib},
show how the calibration removes any average bias.
\begin{figure}[th!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig13.eps,width=8cm} \\
\epsfig{file=./Fig14.eps,width=8cm}
\end{tabular}
\caption{
Examples of response linearity plots ($\langle M_{\mathrm{top}}^{\mathrm{corr}} \rangle$ vs $M_{\mathrm{top}}^{\mathrm{in}}$,
upper plot, and
$\langle \Delta\mathrm{JES}^{\mathrm{corr}} \rangle $ vs $\Delta\mathrm{JES}^{\mathrm{in}}$, lower plot) after the
2D calibration. The solid lines represent the linear functions
which best fit the response as a function
of the input values, while the dashed straight lines
represent the ideal behavior.
}
\label{fig:linmtopjes2}
\end{center}
\end{figure}
To check that the uncertainties $\delta M_{\mathrm{top}}^{\mathrm{corr}}$
and $\delta \Delta\mathrm{JES}^{\mathrm{corr}}$
are also unbiased,
we consider the width of $M_{\mathrm{top}}^{\mathrm{corr}}$
and $\Delta\mathrm{JES}^{\mathrm{corr}}$ pull distributions,
that is, the distributions of deviations of the calibrated values from
the true inputs in the PEs, divided by the uncertainties themselves.
We find that the uncertainties are both underestimated, and multiplicative correction
factors equal to $1.084$ for $\delta M_{\mathrm{top}}^{\mathrm{corr}}$ and to $1.115$
for $\delta \Delta \mathrm{JES}^{\mathrm{corr}}$ are needed.
After these corrections the average expected uncertainty on the measured
top quark mass and jet energy scale displacement
for true $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$ around $175$\,GeV/$c^2$ and $0$
are
\begin{eqnarray}
\delta M_{\mathrm{top}}^{\mathrm{corr}}\,(\mathrm{stat} + \mathrm{JES})~~~ & = & 2.0\,\mathrm{GeV}/c^{2}\,, \\
\nonumber \\
\delta \Delta\mathrm{JES}^{\mathrm{corr}}\,(\mathrm{stat} + M_{\mathrm{top}}) & = & 0.45 \,
\end{eqnarray}
\section{Systematic uncertainties}
\label{sec:syst}
The possible systematic uncertainties
on the top quark mass and the jet energy scale measurements
have been studied
and are summarized in this section.
These arise mostly from the measurement technique itself,
from uncertainties in the simulation of the $t \bar t$ events,
from mismodeling of the detector response, and from uncertainty on
the shapes of signal and background templates used to derive the p.d.f.'s and
to calibrate the measurement.
The uncertainties are usually evaluated by performing PEs, extracting pseudodata from
templates built using signal samples where the possible systematic effects
have been considered and included.
Corrections to the shape of the raw background templates
are performed as described in Sec.\,\ref{sec:bkgtemplates}
to obtain the corrected background templates corresponding
to the effect one wants to study.
On the contrary, nothing is changed in the
elements of the likelihood fit, because
it is the default procedure that we want to apply to real
data and that, therefore, we have to test in case of possible mismodeling of the data themselves.
The results from these PEs are then compared to the ones obtained by using default templates,
and the shifts in the average $M_{\mathrm{top}}^{\mathrm{corr}}$ and $\Delta\mathrm{JES}$$^{\mathrm{corr}}$
values are taken as
the estimate of the systematic uncertainties.
In some cases the statistical uncertainties on the shifts may be larger than the shifts themselves, and
therefore we use conservatively the former as the systematic uncertainty.
In the following, after the description of each effect, we also quote in parentheses
the values of the
corresponding uncertainties for the top quark mass and the jet energy scale respectively.
These values are then summarized in Table\,\ref{tab:obssyst}.
The 2D calibration removes the average biases, especially related to the
parametrization of the templates using smooth probability density functions.
Residual biases usually exist at single
($M_{\mathrm{top}}^{\mathrm{in}},\,\Delta\mathrm{JES}^{\mathrm{in}}$) points, and have to be taken into account.
We therefore consider the shift of the mean of the pull distributions with respect to 0
at each ($M_{\mathrm{top}}^{\mathrm{in}}$,\,$\Delta\mathrm{JES}^{\mathrm{in}}$)
point to evaluate this {\em residual bias} systematic uncertainty, which, given the definition of
pull in Sec.~\ref{sec:sanity}, is defined as a
function of the uncertainty on the calibrated measurements. To obtain the proper coverage of both
positive and negative biases we evaluate them separately, so that asymmetric uncertainties are
finally considered. They are generally given by
$(^{+0.37}_{-0.20}) \cdot \delta M_{\mathrm{top}}^{\mathrm{corr}}$
for $M_{\mathrm{top}}^{\mathrm{corr}}$ and
$(^{+0.43}_{-0.56}) \cdot \delta \Delta\mathrm{JES}^{\mathrm{corr}}$
for $\Delta\mathrm{JES}^{\mathrm{corr}}$. Specifying the values obtained
in the measurement on the data, described in Sec.~\ref{sec:meas}, we obtain
$^{+0.8}_{-0.4}$\,GeV/$c^{2}$ on $M_{\mathrm{top}}^{\mathrm{corr}}$, $ ^{+0.18}_{-0.24} $ on $\Delta\mathrm{JES}^{\mathrm{corr}}$.
\par
The uncertainties on the parameters of the 2D calibration give a small uncertainty on the corrected values
$M_{\mathrm{top}}^{\mathrm{corr}}$ and $\Delta\mathrm{JES}^{\mathrm{corr}}$ which can be evaluated by the calibration functions
and the values of $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$ fitted in the data
($ <0.1$\,GeV/$c^{2}$, $ <0.01 $).
\par
Many sources of systematic effects arise from uncertainties in modeling of
the hard interaction in simulated events. {\sc pythia} and {\sc herwig}\,\cite{Herwig}
Monte Carlo generators differ in their hadronization schemes and in their description of the
underlying event and multiple interactions. The default signal samples
have been generated with {\sc pythia} and therefore an uncertainty is obtained
by using a sample generated using {\sc herwig}
($0.3$\,GeV/$c^{2}$, $0.25 $).
\par
Jets coming from possible emission of hard gluons
might fall among the six leading jets and populate the tails in the top quark invariant mass distribution.
The amount of radiation from partons in the initial or final state
is set by parameters of the {\sc pythia} generator used to simulate signal events.
To study these effects, templates are built using samples where
the values of the parameters have been changed with respect to the default, to increase or
to decrease the amount of radiation~\cite{massTMT}
($0.1$\,GeV/$c^{2}$, $0.06 $).
\par
Since the default jet energy corrections are derived from data samples
deficient in heavy flavors~\cite{JESNIM},
an additional uncertainty comes from considering the different properties
of $b$ quarks. We account for the uncertainties on the $b$-quark semileptonic branching ratios,
fragmentation modeling, and calorimeter response to heavy-flavor hadrons
($0.2$\,GeV/$c^{2}$, $0.04 $).
\par
The different efficiency of the $b$-tagging algorithm
on data and simulated events is usually considered by introducing
a constant scale factor ($b$-tag SF).
The overall uncertainty on this parameter affects the cross section
measurement described in Sec.\,\ref{sec:xsec}.
However, such scale factor does not need to be considered regarding
the top quark mass measurement, because
it could slightly change only the population of the signal templates, but not their shape.
On the other hand, variations of the latter could be caused by the possible dependence of
the $b$-tag SF on the transverse energy of jets, which is then considered as a systematic effect
($0.1$\,GeV/$c^{2}$, $0.01 $).
\par
The uncertainty on the top quark mass coming from the likelihood fit
includes the uncertainty due to the jet energy scale. However,
as described in Sec.\,\ref{sec:presel}, this
uncertainty is the result of many independent effects with different
behavior with respect to properties of jets like $E_{T}$ and $\eta$\,\cite{JESNIM},
and therefore represents
a leading-order correction. Second-order effects can arise from
uncertainties on the single corrections applied to the jet energies.
To evaluate these possible effects, we build signal templates by varying separately by $\pm 1\, \sigma$
the single corrections and, for each one of these variations,
PEs were performed by using these templates and
not applying the
constraint ${\cal L}_{\Delta\mathrm{JES}_{\mathrm{constr}}}$ in the likelihood fit,
as this term is related to effects of the full correction.
The resulting uncertainties have been added in quadrature to obtain a {\em residual} $\mathrm{JES}$
uncertainty on the top quark mass ($0.5$\,GeV/$c^{2}$).
\par
The choice of parton distribution functions (PDF) in the proton used
in Monte Carlo generators can affect
the kinematics of simulated $t \bar t$ events
and thus the top quark mass measurement. We considered four sources of uncertainties\,:
the difference arising from the use of the default CTEQ5L~\cite{CTEQ} PDF
and one calculated from the MRST group, MRST72~\cite{MRST};
the uncertainty depending on the value of $\alpha_{s}$, evaluated
by the difference between the use of MRST72 and MRST75 PDF's;
the uncertainty depending on the differences between the LO
and NLO calculations of PDF'\,s, evaluated by the difference
between using default CTEQ5L (LO) and CTEQ6M (NLO) PDF'\,s;
and the uncertainties on PDF'\,s derived from experimental data uncertainties
($^{+0.3}_{-0.2}$\,GeV/$c^{2}$, $^{+0.05}_{-0.04}$).
\par
The probability to have multiple $p\bar{p}$ interactions during the same bunch crossing
is a function of the instantaneous luminosity. This is reflected in the
increasing number of primary vertices reconstructed in the events at higher luminosities.
We account for the fact that the simulated samples for the signal process
do not model the actual luminosity profile of the data by considering the signal distributions
for events with 1, 2, 3, and $\ge 4$ reconstructed vertices separately.
These distributions are then used
to obtain the templates by weighted averages, where the weights are evaluated as the
fractions of events with 1, 2, 3 and $\ge 4$ vertices observed in the data.
Moreover, a possible mismodeling of the dependence of the jet energy response
as a function of the reconstructed number of primary vertices in simulated events is considered
($0.2$\,GeV/$c^{2}$, $0.01 $).
\par
Uncertainties from modeling of color reconnection
effects~\cite{CR} are estimated by comparing
the results of two sets of PEs performed drawing
pseudodata from templates built using two different
samples of events simulated by {\sc pythia}.
The samples are generated with two different
sets of parameters, corresponding to two different
models of color reconnection ($0.4$\,GeV/$c^{2}$, $0.08 $).
\par
The shapes of the signal and background distributions are affected
by uncertainties due to the limited statistics of the simulated events
and data
samples used to build them. These uncertainties affect the results of a measurement which is
performed maximizing an
unbinned likelihood where parametrized p.d.f.'\,s, fitted to default templates, are evaluated.
We address this effect obtaining 100 sets of templates by statistical fluctuations of default ones and
performing pseudoexperiments drawing data from each of these sets separately.
From each set we obtain an average value for $M_{\mathrm{top}}^{\mathrm{corr}}$
and $\Delta\mathrm{JES}^{\mathrm{corr}}$,
and the spread of these values is taken as the systematic uncertainty
($0.3$\,GeV/$c^{2}$, $0.07 $).
\par
Besides the purely statistical effects, quoted above, the shape of the
background templates also has uncertainties
due to the corrections for the presence of signal events in the pretag sample, Sec.~\ref{sec:bkgtemplates},
and to the
systematic uncertainty on the background normalization, Sec.~\ref{sec:cutoptimization}.
We address this source of systematic uncertainty by the same technique used for the
statistical contributions, that
is by obtaining a set of 100 background templates where the content of each bin is separately
fluctuated by Gaussian distributions centered on the default bin content and with a
width equal to its systematic uncertainty, and taking the spread of results from PEs
as the systematic uncertainty
($0.1$\,GeV/$c^{2}$, $0.02 $).
\par
Table~\ref{tab:obssyst} shows a summary of all the systematic uncertainties and their
sum in quadrature, which gives a total systematic uncertainty of
$^{+1.2}_{-1.0}$\,GeV/$c^2$ for the $M_{\mathrm{top}}$ measurement and
$^{+0.34}_{-0.37}$ for the $\Delta\mathrm{JES}$.
\begin{table}[hbt!]
\caption[c 2]{
Systematic uncertainties
and their sizes for the top-quark mass and the jet
energy scale measurements.
The total uncertainty is obtained by the sum in quadrature of single contributions.
}
\label{tab:obssyst}
\begin{tabular}{lcc}
\hline
\hline
Source & \parbox[l][5mm][c]{17mm}{$\delta M_{\mathrm{top}}^{syst}$} & $\delta \Delta \mathrm{JES}^{syst}$ \\
& (GeV/$c^2$) & \\
\hline
Residual bias & \parbox[c][7mm][c]{17mm}{ $ ^{+0.8}_{-0.4}$} & $ ^{+0.18}_{-0.24} $ \\
2D calibration & \parbox[c][7mm][c]{17mm}{ $ <\,0.1 $} & $ <\, 0.01 $ \\
Generator & \parbox[c][5mm][c]{17mm}{ $ 0.3 $} & $ 0.25 $ \\
ISR/FSR & \parbox[c][5mm][c]{17mm}{ $ 0.1 $} & $ 0.06 $ \\
$b$-jet energy scale & \parbox[c][5mm][c]{17mm}{ $ 0.2 $} & $ 0.04 $ \\
$b$-tag SF $E_T$ dependence & \parbox[c][5mm][c]{17mm}{ $ 0.1 $} & $ 0.01 $ \\
Residual $\mathrm{JES}$ & \parbox[c][5mm][c]{17mm}{ $ 0.5 $} & ... \\
PDF & \parbox[c][7mm][c]{17mm}{ $ ^{+0.3}_{-0.2}$} & $ ^{+0.05}_{-0.04} $ \\
Multiple $p\bar{p}$ interactions ~~~ & \parbox[c][5mm][c]{17mm}{ $ 0.2 $} & $ 0.01 $ \\
Color reconnection & \parbox[c][5mm][c]{17mm}{ $ 0.4 $} & $ 0.08 $ \\
Statistics of templates & \parbox[c][5mm][c]{17mm}{ $ 0.3 $} & $ 0.07 $ \\
Background shape & \parbox[c][5mm][c]{17mm}{ $ 0.1 $} & $ 0.02 $ \\
\hline
Total & \parbox[c][7mm][c]{17mm}{$ ^{+1.2}_{-1.0} $} & $ ^{+0.34}_{-0.37}$ \\
\hline
\hline
\end{tabular}
\centering
\end{table}
\section{\boldmath Top Mass and $\mathrm{JES}$ measurements}
{\label{sec:meas}
After the kinematical selection with $N_{\mathrm{out}}\ge 0.90$ ($\ge 0.88$) and
$\chi^2 \le 6$ ($\le 5$), we are left with 3452 (441) events with one ($\ge 2$) tag(s).
The background amounts to $2785 \pm 83$ ($201 \pm 29$) for events with one ($\ge 2$) tag(s).
For these events a top quark mass has been reconstructed using the likelihood fit
described in Sec.~\ref{sec:like} and applied to the data sample. Once the calibration procedure
and corrections are applied, as described in Sec.~\ref{sec:sanity},
the best estimate of the top quark mass is
\begin{eqnarray}
M_{\mathrm{top}} = 174.8 \pm 2.4\, (\mathrm{stat}+ \mathrm{JES})\,\mathrm{GeV}/c^{2} \,,
\end{eqnarray}
while the value obtained for the jet energy scale displacement is
\begin{eqnarray}
\Delta\mathrm{JES} = -0.30 \pm 0.47\, (\mathrm{stat}+M_{\mathrm{top}}) \,.
\end{eqnarray}
We can also evaluate separately the purely statistical contributions obtaining
\begin{eqnarray}
M_{\mathrm{top}} = 174.8 \pm 1.7\,(\mathrm{stat}) \pm 1.6\,(\mathrm{JES})\,\mathrm{GeV}/c^{2} \,,
\end{eqnarray}
and
\begin{eqnarray}
\Delta\mathrm{JES} = -0.30 \pm 0.35\,(\mathrm{stat}) \pm 0.32\,(M_{\mathrm{top}}) \,.
\end{eqnarray}
The plot in Fig.\,\ref{fig:contours} shows the measured values together with
the log-likelihood contours corresponding to 1\,$\sigma$, 2\,$\sigma$, and 3\,$\sigma$ uncertainty
on the value of the top quark mass\,\cite{pdg}. The slope of the
major axis of the contours denotes that the measurements of $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$
have a negative correlation, and the value of the correlation coefficient
obtained from the likelihood fit is $-0.68$.
\begin{figure}[htbp!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig15.eps,width=8cm}
\end{tabular}
\caption{
Negative log-likelihood contours for the likelihood fit performed
for the $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$ measurements.
The minimum is also shown and corresponds to the values measured in the data.
The contours are drawn at values of 0.5, 2.0, and 4.5 of the increase of
the log-likelihood from the minimum value. These curves correspond
to 1\,$\sigma$, 2\,$\sigma$, and 3\,$\sigma$ uncertainty on the measurement of the top
quark mass.
}
\label{fig:contours}
\end{center}
\end{figure}
The plots in Fig.~\ref{fig:topmeas} show the $m_{t}^{\mathrm{rec}}$
distributions
for the data compared to the expected background and the signal for a top quark mass
of $175.0$\,GeV/$c^2$ and a jet energy scale displacement of $-0.3$,
that is, the values of simulated $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$ as close as possible to the measurements
in the data. The signal and background contributions are normalized to the respective
number of events as fitted in the data.
\begin{figure}[htbp!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig16.eps,width=8cm} \\%&
\epsfig{file=./Fig17.eps,width=8cm}
\end{tabular}
\caption{Histograms of $m_{t}^{\mathrm{rec}}$
as obtained in the data (black points) for 1-tag (upper plot) and $\ge 2$-tag events (lower plot)
are compared to the distributions from signal and background corresponding to
$M_{\mathrm{top}}= 175$~GeV/$c^{2}$ and $\Delta\mathrm{JES} = -0.3$.
The expected histograms are normalized to the measured values for the average number of signal
and background events. The values of the purely statistical
$\chi^{2}$ and of its probability are reported on each plot, showing the overall
agreement between the data and the distributions corresponding to the fitted values
of $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$.
}
\label{fig:topmeas}
\end{center}
\end{figure}
The plots in Fig.~\ref{fig:sigma} compare the measured statistical uncertainty,
just after the 2D calibration,
with the expected distribution from default pseudoexperiments
using as inputs $M_{\mathrm{top}} = 175.0$\,GeV/$c^{2}$
and $\Delta\mathrm{JES} = -0.3$. We find that the probability of achieving a better
sensitivity is $91.6\%$ for $M_{\mathrm{top}}$ and $81.2\%$ for $\Delta\mathrm{JES}$.
\begin{figure}[htbp!]
\begin{center}
\begin{tabular}{c}
\epsfig{file=./Fig18.eps,width=8cm} \\
\epsfig{file=./Fig19.eps,width=8cm}
\end{tabular}
\caption{Distributions of the uncertainties on the top quark mass (upper plot) and the
jet energy scale displacement
(lower plot)
as expected from default PEs performed using as input values $M_{\mathrm{top}}^{\mathrm{in}} = 175.0$\,GeV/$c^{2}$
and $\Delta\mathrm{JES}^{\mathrm{in}} = -0.3$.
The vertical lines indicate the uncertainties obtained in our reported
result.
}
\label{fig:sigma}
\end{center}
\end{figure}
Summarizing, the measured values for the top quark mass and the jet energy scale are
\begin{eqnarray}
M_{\mathrm{top}} = 174.8 \, \pm 2.4\,(\mathrm{stat} + \mathrm{JES})\, ^{+1.2}_{-1.0}\,(\mathrm{syst})\,\mathrm{GeV}/c^{2}\,,~~
\end{eqnarray}
\begin{eqnarray}
\Delta \mathrm{JES} = -0.30\, \pm 0.47\,(\mathrm{stat} + M_{\mathrm{top}})\,^{+0.34}_{-0.37}\,(\mathrm{syst}) \,,~~
\end{eqnarray}
which, isolating the purely statistical contributions and
adding the uncertainties from $\mathrm{JES}$ and $M_{\mathrm{top}}$ to the respective
systematic uncertainties, can also be written as
\begin{eqnarray}
M_{\mathrm{top}} = 174.8~\pm 1.7\,(\mathrm{stat})~ ^{+2.0}_{-1.9}\,(\mathrm{syst})~\mathrm{GeV}/c^{2}\,,~~
\end{eqnarray}
\begin{eqnarray}
\Delta \mathrm{JES} = -0.30~ \pm 0.35\,(\mathrm{stat})~ ^{+0.47}_{-0.49}\,(\mathrm{syst}) \,.~~
\end{eqnarray}
This measurement of the top quark mass has been used in the current world average\,\cite{latestCDFD0mass}.
\section{Cross section measurement}
\label{sec:xsec}
The procedure used to measure the top quark mass also returns the average number
of signal events expected, given the selected data samples.
These results can be turned into a measurement of the $t\bar t$ cross section,
as follows.
\subsection{The likelihood function}
\label{sec:sigmalike}
From the number of signal events, $n_{s}^{1\,\mathrm{tag}}$ and $n_{s}^{\ge 2 \,\mathrm{tags}}$, as obtained
from the mass likelihood fit, we derive a measurement
of the $t\bar{t}$ production cross section considering the efficiency for selecting
a $t\bar{t}$ event in the two tagging categories.
The cross section measurement is performed by maximizing a likelihood function which can be
divided into two parts\,:
\begin{eqnarray}
{\cal L} = {\cal L}_{1\,\mathrm{tag}} \times {\cal L}_{\ge 2\,\mathrm{tags}}\,,
\end{eqnarray}
where
each term can be expressed as\,:
\begin{eqnarray}
{\cal L}_{1,\ge 2\, \mathrm{tags}} = {\cal L}_{\sigma_{t\bar{t}}} \times {\cal L}_{\epsilon}\,,
\end{eqnarray}
where
\begin{eqnarray}
{\cal L}_{\sigma_{t\bar{t}}} & = &
\exp{\left[ - ~ \frac{ \left(\sigma_{t\bar{t}} \cdot \epsilon \cdot {\mathrm L} - n_{s}\right)^{2} }{2\sigma_{n_{s}}^{2}}\right]}
\end{eqnarray}
contains all the parameters of the fit, i.e. the production cross section $\sigma_{t\bar{t}}$,
the integrated luminosity ${\mathrm L}$, the signal efficiency $\epsilon$, and
the signal yield $n_{s} \pm \sigma_{n_{s}}$, as given by the mass measurement,
while ${\cal L}_{\epsilon}$ is a Gaussian term constraining the efficiency within
its statistical uncertainty.
The efficiencies are evaluated using a sample of about $4 \times 10^{6}$ $t\bar{t}$ events
generated with \mbox{$M_{\mathrm{top}} = 175$\,GeV/$c^{2}$} and assuming \mbox{$\Delta\mathrm{JES}= -0.3$},
i.e. the value we measured by the mass likelihood fit,
and are summarized along with signal yields and other parameters in Table\,\ref{tab:xsecpar}.
While studying the performance of the procedure, using pseudoexperiments produced assuming a given
input cross section, we observe the need to introduce a small correction.
The outcome of the fit needs
to be multiplied by a factor $k_\sigma=0.982\pm 0.008$ in order to obtain an unbiased measurement
of the cross section.
\begin{table}
\begin{center}
\caption{Input variables to the cross section evaluation.
For the signal yields, the first uncertainty is the purely statistical one.}
\label{tab:xsecpar}
\begin{tabular}{lc}
\hline\hline
Variable~~~~~~~~~~~~~~~~ & \parbox[c][5mm][c]{30mm}{ Input value } \\
\hline
Signal yield, one tag & \parbox[c][5mm][c]{30mm}{$ 643 \pm 59 \pm 54$ } \\
Signal yield, $\ge 2$ tags~~~~~~~~~~~ & \parbox[c][5mm][c]{30mm}{$ 216 \pm 21 \pm 14$ } \\
Efficiency, one tag & \parbox[c][5mm][c]{30mm}{$ (3.55 \pm 0.01)\%$ } \\
Efficiency, $\ge 2$ tags & \parbox[c][5mm][c]{30mm}{$ (1.00 \pm 0.01)\%$ } \\
Integrated luminosity & \parbox[c][5mm][c]{30mm}{$ 2874 \pm 172$\,pb$^{-1}$ } \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
From the maximization of the likelihood, we obtain a central value for the $t\bar{t}$ production cross
section
\begin{eqnarray}
\sigma_{t\bar{t}} =7.2 \pm 0.5 (\textnormal{stat})\pm 0.4(\textnormal{lum})\,\mathrm{pb}\,,
\end{eqnarray}
evaluated assuming $M_{\mathrm{top}}=175$ GeV/c$^2$ and $\Delta\mathrm{JES}=-0.3$, close to the values measured in Sec.\,\ref{sec:meas}.
The first uncertainty is the statistical one, while the second one derives from the 6\% uncertainty
on the integrated luminosity.
As the signal efficiencies depend strongly on the assumed values for $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$,
the measured $t\bar t$ cross section also has the same dependence.
For reference we report in Table\,\ref{tab:xsec-mass-jes} the cross
sections corresponding to other $(M_{\mathrm{top}},\Delta\mathrm{JES})$ points with a top quark mass
near the current CDF average.
In this case we assume $\Delta\mathrm{JES}=0$ and the systematic uncertainty on JES
is increased from 6.1\% to 9.2\%, corresponding to
changing the $\Delta$JES by $\pm 1$ rather than by $\pm 0.6$ units, that is the sum in quadrature of
the uncertainties on the measured jet energy scale, Sec.\,\ref{sec:meas}.
\begin{table}
\begin{center}
\caption{Cross section as evaluated assuming different values for $M_{\mathrm{top}}$ and $\Delta\mathrm{JES}$.}
\label{tab:xsec-mass-jes}
\begin{tabular}{ccc}
\hline\hline
\parbox[c][5mm][c]{30mm}{$M_{\mathrm{top}}$ (GeV/c$^2$)} & ~~~~~~$\Delta\mathrm{JES}$~~~~~~ & ~~~~~~$\sigma_{t\bar t}$ (pb) \\
\hline
\parbox[c][5mm][c]{30mm}{175.0} & -0.3 & 7.24 \\
\parbox[c][5mm][c]{30mm}{175.0} & 0.0 & 7.00 \\
\parbox[c][5mm][c]{30mm}{172.5} & 0.0 & 7.21 \\
\parbox[c][5mm][c]{30mm}{170.0} & 0.0 & 7.29 \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Systematic uncertainties}
\label{sec:sigmasyst}
Most of the sources of systematic uncertainties affecting
the measurement of $\sigma_{t\bar{t}}$ are the same
discussed for the measurement of the top quark mass.
We just need to evaluate their effects both on the signal
yields and on the signal efficiencies in order to derive
the effects on the cross section.
There are few other sources of systematic uncertainty specific to a cross section measurement which have
not been discussed in Sec.\,\ref{sec:syst}, because they
affect only the signal efficiencies.
These include the uncertainty on the calibration constant, $k_\sigma$,
on the $W \to \mathrm{hadrons} $ branching ratio (BR)~\cite{pdg},
on the trigger simulation and on the distribution of the primary vertex $z$-coordinate.
As for the effect of the JES uncertainty on the efficiency, we have evaluated it by changing the $\Delta$JES by
$\pm 0.6$ units with respect to the measured value $\Delta\mathrm{JES}=-0.3$.
Residual effects due to individual
levels of corrections have been accounted for, too.
The relative uncertainties $\Delta \sigma_{t\bar{t}} / \sigma_{t\bar{t}} $ for the individual
sources are summarized in Table\,\ref{tab:xsecsyst}.
Considering their sum in quadrature, the $t\bar{t}$ production cross section amounts to
\begin{eqnarray}
\sigma_{t\bar t} = 7.2 \, \pm \, 0.5({\rm stat}) \, \pm \, 1.0({\rm syst}) \, \pm \, 0.4({\rm lum})\,\mathrm{pb}\,,~~
\end{eqnarray}
assuming $M_{\mathrm{top}} = 175$\,GeV/$c^{2}$ and $\Delta\mathrm{JES}=-0.3$.
\begin{table}
\begin{center}
\caption{
Systematic uncertainties
and their relative sizes
for the cross section measurement.
The total uncertainty is obtained by the sum in quadrature of single contributions.
}
\label{tab:xsecsyst}
\begin{tabular}{lc}
\hline\hline
Source & \parbox[c][5mm][c]{25mm}{$\Delta \sigma_{t\bar{t}} / \sigma_{t\bar{t}} $ (\%)}\\
\hline
Calibration factor & \parbox[c][5mm][c]{25mm}{ 0.8 } \\
Generator & \parbox[c][5mm][c]{25mm}{ 4.2 } \\
ISR/FSR & \parbox[c][5mm][c]{25mm}{ 0.6 } \\
$b$-jet energy scale & \parbox[c][5mm][c]{25mm}{ 2.8 } \\
$b$-tag SF $E_T$ dependence & \parbox[c][5mm][c]{25mm}{ 5.4 } \\
PDF & \parbox[c][5mm][c]{25mm}{ 3.4 } \\
Multiple $p\bar{p}$ interactions~~~~~~~~~ & \parbox[c][5mm][c]{25mm}{ 2.5 } \\
Color reconnection & \parbox[c][5mm][c]{25mm}{ 0.8 } \\
Templates statistics & \parbox[c][5mm][c]{25mm}{ 0.8 } \\
Background shape & \parbox[c][5mm][c]{25mm}{ 0.3 } \\
Background normalization & \parbox[c][5mm][c]{25mm}{ 8.2 } \\
$\mathrm{JES}$ & \parbox[c][5mm][c]{25mm}{ 6.1 } \\
Residual $\mathrm{JES}$ & \parbox[c][5mm][c]{25mm}{ 2.1 } \\
Primary vertex & \parbox[c][5mm][c]{25mm}{ 0.2 } \\
BR($W\to \mathrm{hadrons}$) & \parbox[c][5mm][c]{25mm}{ 0.8 } \\
Trigger & \parbox[c][5mm][c]{25mm}{ 1.8 } \\
\hline
Total & \parbox[c][5mm][c]{25mm}{ 13.7 } \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\section{Conclusions}
\label{sec:fine}
Using a very effective neural-network-based kinematical selection and a
$b$ jet identification technique,
we measure the top quark mass to be
\begin{eqnarray}
M_{\mathrm{top}}=174.8\, \pm \, 2.4({\rm stat +JES}) \,^{+1.2}_{-1.0}({\rm syst})\,\mathrm{GeV}/c^{2}\,,~~
\end{eqnarray} \\
and the $t\bar t$ production cross section to be
\begin{eqnarray}
\sigma_{t\bar t} = 7.2 \, \pm \, 0.5({\rm stat}) \, \pm \, 1.0({\rm syst}) \, \pm \, 0.4({\rm lum})\,\mathrm{pb}\,.~~
\end{eqnarray}
These values represent the most precise measurements to date of $M_{\mathrm{top}}$
and $\sigma_{t\bar t}$ in
the all-hadronic decay channel.
The results are consistent with the measurements obtained
in other decay channels by CDF\,and D{0}
Collaborations\,\cite{latestCDFD0mass, latestCDFD0xsec}
and, as it concerns $\sigma_{t\bar t}$,
with the theoretical predictions evaluated at the
value of the top quark mass obtained in our
measurement\,\cite{theo-xsec}.
\bigskip
\section{Acknowledgments}
We thank the Fermilab staff and the technical staffs of the participating institutions
for their vital contributions.
This work was supported by the U.S. Department of Energy and National Science Foundation;
the Italian Istituto Nazionale di Fisica Nucleare;
the Ministry of Education, Culture, Sports, Science and Technology of Japan;
the Natural Sciences and Engineering Research Council of Canada;
the National Science Council of the Republic of China;
the Swiss National Science Foundation; the A.P. Sloan Foundation;
the Bundesministerium f\"ur Bildung und Forschung, Germany;
the World Class University Program, the National Research Foundation of Korea;
the Science and Technology Facilities Council and the Royal Society, United Kingdom;
the Institut National de Physique Nucleaire et Physique des Particules/CNRS;
the Russian Foundation for Basic Research;
the Ministerio de Ciencia e Innovaci\'{o}n, and Programa Consolider-Ingenio 2010, Spain;
the Slovak R\&D Agency; and the Academy of Finland.
|
\section{Introduction}
Theoretical explanation of the existence of self-sustained
spherically symmetric plasmoids is still a challenge in plasma
physics~\cite{Ste99}. A static bunch of charged particles confined
by its own internal forces does not seem to be stable. The
hydrodynamical pressure $p$ and the magnetic pressure $B^2/8\pi$
will try to expand a plasmoid (see Ref.~\cite{LanLif82p322})
making its existence impossible in the absence of external forces
such as gravity, etc.
However a spherical plasmoid can be implemented in the form of
spherically symmetric oscillations of electrons in
plasma~\cite{DvoDvo07}. This kind of plasma pulsation does not
have a magnetic field (see Sec.~\ref{PHIA} below) and thus does
not lose energy for radiation. This fact can explain the relative
stability of spherical plasmoid generated in natural conditions.
It is, however, known that the frequency of free electrons
oscillations in plasma cannot be less than the Langmuir frequency.
For an electrons density in plasma of $\sim
10^{15}\thinspace\text{cm}^{-3}$ the Langmuir frequency is about
$100\thinspace\text{GHz}$. It is rather difficult to create a
strong external field of such a high frequency in natural
conditions to generate a plasmoid. Therefore one should point out
a physical mechanism that acts at the initial stages of the
plasmoid evolution, during which one can encounter a relatively
low-frequency external field and which makes the plasmoid
generation possible. We suggest that plasma superconductivity can
be one such mechanisms.
The idea that a dense plasma can reveal superconducting properties
was put forward in Ref.~\cite{Dij80Zel08}. The superconductivity
seems to play a key role at the stages of the plasmoid formation.
However it is difficult to find the physical mechanism which would
be responsible for the appearance of the superconducting phase in
rather hot plasma.
It is known that the formation of a bound state of two electrons,
a Cooper pair, underlies the superconductivity phenomenon in
metals. A Cooper pair is formed when the electrostatic repulsive
field of an electron is shielded by the effective attractive
interaction due to the exchange of virtual phonons. Such a bound
state is destroyed when the temperature of a metal exceeds a few
kelvin degrees~\cite{Mad80p315}.
Although the temperature of plasmas in natural conditions is far
greater than a typical temperature of a superconducting metal one
can find physical processes leading to the appearance of the
effective attraction between electrons. A charged test particle,
e.g. an electron, moving in plasma is known to emit ion acoustic
waves. Therefore a test electron can be surrounded by a cloud of
positively charged ions. Under certain conditions this effective
potential can screen the repulsive interaction between two
electrons and result in the creation of a bound state (see
Ref.~\cite{NamAka85}). This phenomenon, as well as the exchange of
dust acoustic waves, can lead to the effective attraction of dust
particles in a dusty plasma~\cite{Shu01}. Note that the formation
of bound states of electrons in plasma due to the exchange of ion
acoustic waves is analogous to the Cooper pairs
formation~\cite{NamVlaShu95}.
In this work we study spherically symmetric plasma structures
using the method of classical electrodynamics. Note that in
Ref.~\cite{DvoDvo07} we considered a spherical plasmoid as quantum
oscillations of electrons in plasma. Firstly, in Sec.~\ref{PHIA},
we examine the electromagnetic potentials for the spherically
symmetric motion of charged particles and find a gauge in which
the vector potential is zero. Secondly, in Sec.~\ref{CLOSC}, we
study free and forced spherically symmetric oscillations of
electrons on the basis of the system of equations of classical
plasma hydrodynamics. In Sec.~\ref{WAKE}, using the methods of
Ref.~\cite{NamAka85} we calculate the scalar potential created by
a test electron participating in forced oscillations in plasma. We
examine the conditions when the effective potential is attractive
and consider the possibility of forming a bound state. Finally, in
Sec.~\ref{DISC}, we examine possible applications of the obtained
results to the description of natural spherical plasmoids.
\section{A description of
spherically symmetric oscillations of electrons in plasma based on
classical electrodynamics\label{CLPL}}
In this section we will study oscillations of elections in plasma
which have the spherical symmetry. The method of classical
electrodynamics will be used to describe this process. Firstly, in
Sec.~\ref{PHIA}, we will be interested in the various choices of
electromagnetic potentials for such a system. Then, in
Sec.~\ref{CLOSC}, we will obtain the exact solution of the
hydrodynamical equations which describes oscillations of electrons
density as well as the dispersion relation for these oscillations.
\subsection{Electromagnetic potentials in the system of
radially oscillating particles\label{PHIA}}
It is clear that the scalar potential of an electric field
$\varphi$ in the system making spherically symmetric pulsations
can depend only on the radial coordinate, $\varphi(r,t)$. The
vector potential $\mathbf{A}$ has only radial component which is
also a function of only radial coordinate, $A_r(r,t)$. Due to the
existence of the spatial dispersion in plasma, oscillations of
charged particles should vanish on big distances from the center
of the system, i.e. $A_r(r,t) \to 0$ at $r \to \infty$.
Suppose that we have found the potentials $\varphi$ and
$\mathbf{A}$. Now we can make the gauge transformation,
$\mathbf{A}' = \mathbf{A} + \nabla f$ and $\varphi' = \varphi +
(1/c)\partial f/\partial t$, where $f(r,t)=\int_r^\infty
A_r(r',t)\mathrm{d}r'$. This gauge transformation does not change
the electric field. Note that the magnetic field is identically
equal to zero for a spherically symmetric motion of charged
particles, $\mathbf{B}= \nabla \times \mathbf{A} = 0$. For the
chosen function $f$ we obtain that the vector potential can be
eliminated in all the space. Now the electric field has the form,
$\mathbf{E}=-\nabla\varphi'$. In the following we will omit the
prime in the definition of the scalar potential.
\subsection{Classical plasma hydrodynamics description
of a spherical plasmoid\label{CLOSC}}
In the first approximation we suggest that only electrons
participate in oscillations of plasma since the mobility of ions
is low. In the absence of collisions and other forms of
dissipation the system of the hydrodynamic equations for the
description of plasma oscillations can be presented in the
following way (see Ref.~\cite{Jac65p369}):
\begin{align}
\notag
& \frac{\partial n_e}{\partial t} + \nabla \cdot (n_e\mathbf{v}) = 0, \\
\notag
& \frac{\partial\mathbf{v}}{\partial t} + (\mathbf{v} \cdot \nabla)\mathbf{v} =
-\frac{e}{m}\mathbf{E}-\frac{1}{m n_e}\nabla p, \\
\notag
& \nabla \cdot \mathbf{E} = -4\pi e(n_e-n_i)+4\pi\rho_\mathrm{ext}(\mathbf{r},t), \\
\label{hydrodyn}
& \frac{\partial\mathbf{E}}{\partial t} =
4\pi e n_e \mathbf{v},
\end{align}
where $n_e$ is the electrons density, $n_i$ is the ions density,
$p$ is the plasma pressure, $m$ is the mass of the electron and
$e>0$ is the proton charge. In Eq.~\eqref{hydrodyn} we include the
possible external source $\rho_\mathrm{ext}$ and take into account
that the magnetic field is equal to zero (see Sec.~\ref{PHIA}).
Supposing that $n_i = n_0$ and the deviation of the electrons
density from the equilibrium value is small, $n_e-n_0 = n \ll
n_0$, we can linearize the system~\eqref{hydrodyn} and obtain the
single differential equation for the perturbation of the electrons
density,
\begin{equation}\label{closcil}
\frac{\partial^2 n}{\partial t^2}+\omega_e^2 n -
\frac{1}{m}
\left(
\frac{\partial p}{\partial n}
\right)_0 \nabla^2 n = \frac{4\pi e n_0}{m}\rho_\mathrm{ext},
\end{equation}
where $\omega_e=\sqrt{4\pi e^2 n_0/m}$ is the plasma frequency for
electrons and $(\partial p/\partial n)_0$ is the derivative taken
at $n_e=n_0$. The latter quantity depends on the equation of state
of electrons in plasma.
In the absence of the external source the spherically symmetric
solution of Eq.~\eqref{closcil} has the form
\begin{equation}\label{clsol}
n(r,t) = A \cos(\omega t)\frac{\sin\gamma r}{r},
\end{equation}
where $A$ is the constant chosen to satisfy the condition $|n| \ll
n_0$. The frequency of oscillations $\omega$ and the length scale
parameter $\gamma$ are related by the following identity:
\begin{equation}\label{cldisprel}
\omega^2=\omega_e^2+\frac{1}{m}
\left(
\frac{\partial p}{\partial n}
\right)_0\gamma^2,
\end{equation}
which shows that free oscillations with $\omega \geq \omega_e$
exist in plasma. However, if $\rho_\mathrm{ext} \sim \cos\Omega
t$, it is clear that forced oscillations with $\Omega<\omega_e$
can be also excited.
In Ref.~\cite{DvoDvo07} we studied spherically symmetric
oscillations of electrons in plasma using the quantum mechanical
approach. In that paper, we solved the non-linear Schr\"{o}dinger
equation for the wave function normalized on the number density of
electrons, $|\psi(\mathbf{r},t)|^2=n_e(\mathbf{r},t)$. It was
obtained that in the spherically symmetric case the density of
electrons has the form, $n_e(r,t)=n_0+A\cos(\omega t)\sin(\gamma
r)/r+\dotsb$, i.e. is similar to Eq.~\eqref{clsol}. However the
dispersion relation in Ref.~\cite{DvoDvo07} was different from
Eq.~\eqref{cldisprel}. Moreover quantum oscillations of electrons
reveal the typical size of the system, where the most intensive
oscillations happen, $L = \pi/\gamma =
\pi\sqrt{\hbar/2m\omega_e}$, at the critical frequency
$\omega=2\omega_e$. On the contrary, if we use the classical
electrodynamics method, the parameter $\pi/\gamma$, in principle,
can be arbitrary. Of course, Eq.~\eqref{cldisprel} is valid only
for quite long waves, when $\gamma \ll k_e$ (see
Ref.~\cite{Jac65p374}), where $k_e$ is the Debye wave number (see
the definition in Sec.~\ref{WAKE}).
\section{Formation of bound states of electrons at the initial
stages of the spherical plasmoid evolution\label{WAKE}}
Any particle in plasma has a two-fold life. On the one hand it is
a test particle moving through plasma and interacting with the
whole plasma rather than with separate plasma particles. On the
other hand any test particle is a part of plasma and hence it
contributes to self-consistent electromagnetic fields in plasma.
In this section we will use the method of test particles (see
Refs.~\cite{NamAka85,NamVlaShu95}) to calculate the effective
potential of a radially oscillating electron.
Let us study the electric field created by electrons participating
in spherically symmetric motion considering each electron as a
test particle with the charge $q$. Each of the test particles is
taken to interact with the rest of hot electrons, having the
temperature $T$, and with cold ions. The permittivity of this
plasma has the form,
\begin{equation}\label{perm}
\varepsilon(\mathbf{k},\omega)=
1+
\left(
\frac{k_e}{k}
\right)^2-
\left(
\frac{\omega_i}{\omega+\mathrm{i}0}
\right)^2,
\end{equation}
where $k_e=\sqrt{4\pi n_0 e^2/T}$ is the Debye wave number and
$\omega_i=\sqrt{4\pi (Z e)^2 n_0/M}$ is the plasma frequency for
ions with the mass $M$ and the charge $Ze$, $Z$ is the degree of
the ionization of an ion.
We suppose that a test particle makes harmonic oscillations around
the point $\mathbf{r}_0$ with the frequency $\Omega$ and the
amplitude $\mathbf{a}$: $\mathbf{r}'(t) = \mathbf{r}_0 +
\mathbf{a}\sin \Omega t$. To find the electric potential created
by one of the charged particles we should account for the Maxwell
equation for the electric displacement field, $\nabla \cdot
\mathbf{D} = 4 \pi q \delta^3(\mathbf{r}-\mathbf{r}')$.
Expressing the electric field as $\mathbf{E}(\mathbf{k},\omega) =
- \mathrm{i} \mathbf{k} \varphi(\mathbf{k},\omega)$, we obtain the
scalar potential of the system in the form
\begin{align}\label{varphi1}
\varphi(\mathbf{k},\omega) = &
\frac{4 \pi q}{k^2 \varepsilon(\mathbf{k},\omega)}
\notag
\\
& \times
\int \mathrm{d}t
e^{\mathrm{i} \omega t -
\mathrm{i} \mathbf{k} \cdot (\mathbf{r}_0 + \mathbf{a}\sin \Omega t)}.
\end{align}
We recall that the vector potential can be taken to be equal to
zero for the spherically symmetric system (see Sec.~\ref{PHIA}).
Let us decompose the exponential factor in the integrand of
Eq.~\eqref{varphi1} using the series of Bessel functions of the
$n$-th order,
\begin{equation}\label{bessel}
e^{-\mathrm{i} \mathbf{k} \cdot \mathbf{a} \sin \Omega t}=
\sum_{n=-\infty}^{+\infty}
J_n(\mathbf{k} \cdot \mathbf{a})e^{-\mathrm{i} n \Omega t}.
\end{equation}
Finally on the basis of Eqs.~\eqref{varphi1} and~\eqref{bessel} we
obtain the scalar potential in the form
\begin{align}\label{varphi2}
\varphi(\mathbf{r},t) = &
\frac{q}{2\pi^2}
\sum_{n=-\infty}^{+\infty}
\int \mathrm{d}\omega\mathrm{d}^3\mathbf{k}
e^{-\mathrm{i} \omega t + \mathrm{i} \mathbf{k} \cdot (\mathbf{r}-\mathbf{r}_0)}
\notag
\\
& \times
\delta(\omega-n\Omega)
\frac{J_n(\mathbf{k} \cdot \mathbf{a})}
{k^2\varepsilon(\mathbf{k},\omega)}.
\end{align}
To analyze Eq.~\eqref{varphi1} we present the reciprocal of the
permittivity~\eqref{perm} in the following way:
\begin{equation}\label{iaw}
\frac{1}{\varepsilon(\mathbf{k},\omega)}=
\frac{k^2}{k^2+k_e^2}
\left(
1+\frac{\omega_a^2}{\omega^2-\omega_a^2}
\right),
\end{equation}
where $\omega_a=k\omega_i/\sqrt{k^2+k_e^2}$ is the dispersion
relation for ion acoustic waves.
Using Eq.~\eqref{iaw} we present $\varphi$ in Eq.~\eqref{varphi2}
as a sum of two terms, $\varphi = \varphi_D + \varphi_W$, where
\begin{align}\label{debye1}
\varphi_D(\mathbf{r},t) = &
\frac{q}{2\pi^2}
\sum_{n=-\infty}^{+\infty}
\int \mathrm{d}^3\mathbf{k}
e^{-\mathrm{i} n \Omega t + \mathrm{i} \mathbf{k} \cdot (\mathbf{r}-\mathbf{r}_0)}
\notag
\\
& \times
\frac{J_n(\mathbf{k} \cdot \mathbf{a})}{k_e^2+k^2},
\end{align}
is the analog of the Debye-H\"{u}ckel screening potential and
\begin{align}\label{wake1}
\varphi_W(\mathbf{r},t) = &
\frac{q}{2\pi^2}
\sum_{n=-\infty}^{+\infty}
\int \mathrm{d}^3\mathbf{k}
e^{-\mathrm{i} n \Omega t + \mathrm{i} \mathbf{k} \cdot (\mathbf{r}-\mathbf{r}_0)}
\notag
\\
& \times
\frac{\omega_a^2}{(n\Omega)^2-\omega_a^2}
\frac{J_n(\mathbf{k} \cdot \mathbf{a})}{k_e^2+k^2},
\end{align}
is the wake potential due to the emission of ion acoustic
waves~\cite{NamAka85}.
To study the behaviour of the potentials $\varphi_D$ and
$\varphi_W$ we choose the specific coordinate system with
$\mathbf{r}_0 = 0$ and $\mathbf{a} = a\mathbf{e}_z$, where
$\mathbf{e}_z$ is the unit vector along the $z$-axis. We also
decompose the vectors $\mathbf{k}$ and $\mathbf{r}$ using the
cylindrical coordinates: $\mathbf{k} = (k_\rho,k_z,\phi_k)$ and
$\mathbf{r} = (\rho,z,\phi)$.
Taking into account the value of the following integral:
\begin{equation}
\int_0^{2\pi}\mathrm{d}\phi_k e^{\mathrm{i} k_\rho \cos(\phi_k-\phi)}
= 2\pi J_0(k_\rho \rho),
\end{equation}
we rewrite the potential $\varphi_D$ in the form,
\begin{align}\label{debye2}
\varphi_D(\rho,z,t) = &
2q
\sum_{n=0}^{\infty} (-1)^n
\int_0^\infty k_\rho \mathrm{d}k_\rho
\notag
\\
& \times
\frac{e^{-|z|\sqrt{k_\rho^2+k_e^2}}J_0(k_\rho \rho)}{\sqrt{k_\rho^2+k_e^2}}
\notag
\\
& \times
\Big\{
I_{2n}
\left(
a \sqrt{k_\rho^2+k_e^2}
\right)
\cos[2n \Omega t]
\notag
\\
& \pm
I_{2n+1}
\left(
a \sqrt{k_\rho^2+k_e^2}
\right)
\notag
\\
& \times
\sin[(2n+1) \Omega t]
\Big\},
\end{align}
where $I_n(x) = (\mathrm{i})^{-n} J_n(\mathrm{i}x)$ is the Bessel
function of the imaginary argument.
In Eq.~\eqref{debye2} the `$+{}$' sign stands for $z>0$ and
`$-{}$' for $z<0$.
The electromagnetic field of a charged linear oscillator in
vacuum, $\varepsilon=1$, was studied in Ref.~\cite{SokTer74}. The
scalar potential found in that book is different from
Eq.~\eqref{debye2} since in Ref.~\cite{SokTer74} the problem of
radiation of a linear oscillator was considered and the Lorentz
gauge for potentials was used. We study the case of the
spherically symmetric motion of plasma. As we demonstrated in
Sec.~\ref{PHIA} such a system does not have any magnetic field and
thus cannot emit radiation. In our situation it is more convenient
to use the gauge in which $\mathbf{A}=0$. That is why we get a
different expression for the scalar potential.
It is interesting to analyze Eq.~\eqref{debye2} in the static
limit. For the test particle at rest, i.e. at $a \to 0$, only the
term with the function $I_0(x)$ survives. It is possible to show
that in this limit Eq.~\eqref{debye2} transforms to
$\varphi_D(\rho,z,t) = (q/r) e^{-k_e r}$,
where $r=\sqrt{z^2+\rho^2}$. We can see that
one recovers the usual form of the Debye-H\"{u}ckel potential.
This analysis justifies our definition of $\varphi_D$.
Now we study the wake potential~\eqref{wake1} using the same
technique as for $\varphi_D$. We present the value of $\varphi_W$
in the cylindrical coordinates as
\begin{align}\label{wake2}
\varphi_W(\rho,z,t) & =
-2q
\int_0^\infty k_\rho \mathrm{d}k_\rho
\frac{e^{-|z|\sqrt{k_\rho^2+k_e^2}}J_0(k_\rho \rho)}{\sqrt{k_\rho^2+k_e^2}}
\notag
\\
& \times
I_0
\left(
a \sqrt{k_\rho^2+k_e^2}
\right)
\notag
\\
& + 2q
\sum_{n=1}^{\infty}
\left\{
\begin{matrix}
1 \\
(-1)^n \
\end{matrix}
\right\}
\frac{\omega_i^2}{\omega_i^2-(n\Omega)^2}
\notag
\\
& \times
\frac{k_n^3}{k_e^2+k_n^2}
\int_0^1 \mathrm{d}x
J_0(k_n \rho \sqrt{1-x^2})
\notag
\\
& \times
J_n(a k_n x)\sin(k_n|z|x - n \Omega t),
\end{align}
where
\begin{equation}\label{kn}
k_n = k_e \frac{n\Omega}{\sqrt{\omega_i^2-(n\Omega)^2}}.
\end{equation}
The upper multiplier in Eq.~\eqref{wake2} corresponds to $z>0$ and
the lower one -- to $z<0$. Comparing Eq.~\eqref{debye2} and
Eq.~\eqref{wake2} we can see that the terms containing the
integrals of $I_0(x)$ cancel in the sum of the two potentials.
To derive Eq.~\eqref{wake2} we suggest that $n\Omega<\omega_i$,
i.e. the wake potential appears only for non-rapid oscillations.
For example, we can consider the forced oscillations of electrons
in plasma described in Sec.~\ref{CLOSC} and suppose that $\Omega$
is a bit less than $\omega_i$. It means that only the first
harmonic is excited.
Let us study Eq.~\eqref{wake2} at the line of the test particle
oscillations, $\rho = 0$, and at big distances from the test
particle, $|z| \gg a$. Putting $n=1$ we obtain from
Eq.~\eqref{wake2}
\begin{equation}\label{wake3}
\varphi_W(z,t) \approx \mp
q \frac{\Omega^2}{\omega_i^2-\Omega^2}
\frac{k_1 a}{|z|}
\cos(k_1 |z| - \Omega t),
\end{equation}
where the `$-{}$' sign stands for $z>0$ and `$+{}$' for $z<0$.
Returning to the general Eq.~\eqref{debye2} for $\varphi_D$ we
also rewrite it for $\rho = 0$ as
\begin{align}\label{debye4}
\varphi_D(z,t) = &
\frac{2q}{|z|}
\sum_{n=0}^{\infty} (-1)^n
\int_{k_e|z|}^\infty \mathrm{d}x
e^{-x}
\notag
\\
& \times
\Big\{
I_{2n}
\left(
\frac{a}{|z|}x
\right)
\cos[2n \Omega t]
\notag
\\
& \pm
I_{2n+1}
\left(
\frac{a}{|z|}x
\right)
\notag
\\
& \times
\sin[(2n+1) \Omega t]
\Big\}.
\end{align}
Comparing Eqs.~\eqref{wake3} and~~\eqref{debye4} we can see that
at large distances the non-Coulomb wake potential transcends the
Debye-H\"{u}ckel potential. For example, the term with the
function $I_1(x)$ in Eq.~\eqref{debye4} at $|z| \gg a$ and $k_e a
\ll 1$ has the form
\begin{equation}\label{debye5}
\mp \frac{2qa}{z^2}e^{-k_e |z|}\sin(\Omega t),
\end{equation}
which has much smaller value than the wake potential~\eqref{wake3}
at $|z|>1/k_e$. The terms which contain the functions $I_n(x)$
with $n>1$ will give contributions to $\varphi_D$ smaller than
that in Eq.~\eqref{debye5}.
The wake potential is attractive when $\cos(k_1 |z| - \Omega t)>0$
for $z>0$ and when $\cos(k_1 |z| - \Omega t)<0$ for $z<0$. We
recall that we study the radial pulsation of plasma. Thus the
coordinate $z$ coincides with the radial direction. Suppose that a
test electron attracts another electron which is situated at the
distance $d$ away the center of the system during the certain
period of time, $\Delta t_1 = \pi/\Omega$. During the next half a
period of the wake potential variation, $\Delta t_2 = \pi/\Omega$,
the same test electron will attract a different electron situated
at the same distance $d$ but closer to center of the system. It
means that a test electron can always attract some charged
particles in plasma.
To form a bound state with another electron in plasma the energy
of interaction of a test electron $e\varphi_W$ should be greater
than the total energy of its oscillations, $E_\mathrm{osc} = m a^2
\Omega^2/2$. Suppose that two electrons are at the distance
$d=1/k_e$ and the amplitude of oscillations $a = 0.1d \ll d$.
Studying the plasma consisting of electrons, with temperature $T =
10^3\thinspace\text{K}$ and number density $n_0 =
10^{15}\thinspace\text{cm}^{-3}$, as well as of singly ionized
nitrogen atoms we obtain that the ratio
$|e\varphi_W|/E_\mathrm{osc}>1$ if $\Omega > 10^{-4}\omega_i$.
Taking into account that $\omega_i \sim
10^{10}\thinspace\text{s}^{-1}$ we obtain that the frequency of
the forced oscillations should be in the region
$10^6\thinspace\text{s}^{-1}<\Omega<10^{10}\thinspace\text{s}^{-1}$.
Thus we get that the effective attraction can take place in the
atmospheric plasma for the reasonable frequencies of an external
field.
\section{Discussion\label{DISC}}
We studied electrons oscillations in plasma using classical
electrodynamics. It is found that spherically symmetric
oscillations are possible in the classical case, although the
dispersion relation~\eqref{cldisprel} is different from the
previously found for quantum oscillations~\cite{DvoDvo07}. We
suggest that the radial pulsation of plasma underlies the rare
atmospheric electricity phenomenon called a ball lightning
(BL)~\cite{Ste99}.
Radial oscillations of electrons in plasma described in the
present work are unbounded and occure in all the space. It is the
case when a plasmoid, BL, propagates far away from any external
surface. The situation changes when one considers spherically
symmetric oscillations of plasma in a cavity inside a dielectric
material. This kind of oscillations was studied in
Ref.~\cite{SteYuVla93} in presence of external electric and
magnetic fields. The exact system of non-linear equations
describing oscillating spatial patterns was obtained and analyzed
both analytically and numerically. The solutions obtained in
Ref.~\cite{SteYuVla93} can correspond to BL passing through
microscopic cracks in a dielectric material, e.g. glass. There are
several reports of such events collected in Ref.~\cite{Sta85}. The
problem of the interaction of BL with external materials should be
analyzed in out future works.
As we demonstrated, in order to generate spherically symmetric
oscillations, one should excite them with rather high frequencies,
$\Omega > \omega_e$ in the classical case (or $\Omega > 2\omega_e$
in the quantum case). The electron plasma frequency is
$(10^{12}-10^{13})\thinspace\text{s}^{-1}$ for $n_0 =
(10^{15}-10^{17})\thinspace\text{cm}^{-3}$. Such high frequencies
are very difficult to obtain in any natural conditions.
We showed in Sec.~\ref{CLOSC} that forced oscillations are
possible with frequencies less than $\omega_e$. Of course, forced
oscillations will decay as soon as the external force is switched
off. There should be a mechanism which provides the smooth
transition from the generation regime of a spherical plasmoid with
the external harmonic source having $\Omega<\omega_e$ to a regime
with self-sustained oscillations having $\Omega \sim \omega_e$.
We suggest that this mechanism could be a formation of bound
states of electrons in plasma. The motion of bound states of
electrons can result in the appearance of a superconducting state
of plasma inside a spherical plasmoid. Previously the idea that
the superconductivity can exist in plasma was put forward in
Ref.~\cite{Mei84}. Without the existence of superconducting phase
electrons participating in radial oscillations will lose their
energy very quickly, because of the various friction mechanisms,
and will recombine into the initial neutral gas.
In Sec.~\ref{WAKE} we obtained that a test electron oscillating in
plasma with the frequency $\Omega\lessapprox\omega_i$ would emit
ion acoustic waves. Thus the test electron appears to be
surrounded by a cloud of `phonons' that shield its repulsive
potential. Under some conditions the effective interaction between
the test electron and other electrons in plasma turns out to be
attractive. This process can lead to the formation of bound states
of electrons.
Note that for the first time the role of `phonons', or acoustic
waves, for the description of the stability of a spherical
plasmoid was discussed in Ref.~\cite{VlaYak78}. However, in that
work, the phonons exchange between ions and neutral atoms was
considered in frames of the quantum theory. The applications to
the obtained results to the theory of BL were also studied in
Ref.~\cite{VlaYak78}.
The existence of the superconducting phase inside BL was
previously proposed in Ref.~\cite{Dij80Zel08}. However in these
papers the dense plasma of BL was already supposed to be in the
superconducting phase and the phenomenological consequences of
this phenomenon were described. No physical mechanisms for the
formation of the superconducting state were proposed. In the
present work we suggest that the exchange of ion acoustic waves
created in spherically symmetric oscillations of electrons in
plasma results in the effective attractive potential between
electrons. This mechanism can lead to the formation of bound
states of electrons and possibly to a superconducting state inside
a spherical plasmoid.
After one switches off the low frequency external field, bound
states of electrons will be destroyed. We assume that during the
superconducting stage of the plasmoid evolution a source of
internal energy should appear. It was suggested in
Ref.~\cite{nuclfus} that nuclear reactions can serve as the energy
source of BL. The recombination of charged particles will be
compensated by the processes of ionization and creation of new
charged particles owing to this internal source of energy of BL.
\section*{Acknowledgments}
The work has been supported by the CONICYT (Chile) through
Programa Bicentenario PSD-91-2006. The author is very thankful to
Sergey Dvornikov and Timur Rashba for helpful discussions.
|
\section{Introduction}
{\bf Invariants}\\[1pt]
Let $\mathcal{O} (V_n)^{\SLtwo}$ denote the algebra of invariants of binary
forms (forms in two variables) of degree $n$ with complex coefficients.
This algebra was extensively studied in the nineteenth century,
and for $n \leq 6$ the structure was clear and a finite basis was known.
Gordan \cite{Go} proved in 1868 that $\mathcal{O} (V_n)^{\SLtwo}$ has
a finite basis for all $n$. For $n=7$ the invariants were determined
by von Gall \cite{vG7} and Dixmier \& Lazard \cite{DiLa}
(see also Bedratyuk \cite{Be}).
The invariants for $n=8$ were found by von Gall \cite{vG8} and
Shioda \cite{Sh}.
The case $n=9$ was done by Cr\"oni \cite{Cr} and the present authors
\cite{BP}.
Here we consider the case $n=10$, and show that $\mathcal{O} (V_{10})^{\SLtwo}$
is generated by 106 (explicitly known) basic invariants, and give
the degrees.
\begin{proposition}\label{prop1}
The algebra $I$ of invariants of the binary decimic (form of degree $10$)
is generated by $106$ invariants. The nonzero numbers $d_m$ of
basic invariants of degree $m$ are \\
\smallskip
\begin{tabular}{@{\,}c|cccccccccccccccc@{\,}}
$m$ & $2$ &$4$ & $6$ & $8$ & $9$ & $10$ & $11$ & $12$ & $13$ & $14$ &
$15$ & $16$ & $17$ & $18$ & $19$ & $21$ \\
\hline
$d_m$ & $1$ & $1$ & $4$ & $5$ & $5$ & $8$ & $8$ & $12$ & $15$ & $13$ &
$19$ & $5$ & $5$ & $1$ & $2$ & $2$ \\
\end{tabular}
\end{proposition}
This list agrees with Sylvester \& Franklin \cite{SyFr} for degrees
less than 17. Sylvester predicted 3 basic invariants of degree 17
and none of degree higher than 17 for a total of 99 basic invariants.
Tom Hagedorn (unpublished) found 104 invariants,
cf. Olver \cite{Olver} (p. 40).
The existence of basic invariants of degree 21 seems to be new.
That the list is complete follows as a corollary from
the construction of a homogeneous system of parameters (hsop), see below.
\medskip\noindent
{\bf Systems of parameters}\\
A (homogeneous) {\it system of parameters} for a graded algebra $A$
is an algebraically independent set $S$ of homogeneous elements of $A$
such that $A$ is module-finite over the subalgebra generated by the set $S$.
Hilbert \cite{Hi2} showed the existence of a system of parameters
for algebras of invariants, cf.~Proposition \ref{hilbert} below.
Here we find an explicit system of parameters for $\mathcal{O}(V_{10})^{\SLtwo}$.
\begin{proposition}\label{hsop}
The algebra $I$ of invariants of the binary decimic has a
system of parameters of degrees $2$, $4$, $6$, $6$, $8$, $9$, $10$, $14$.
\end{proposition}
This is useful, since it provides an upper bound for the degrees
of basic invariants that is sufficiently low, so that a simple
computer search can find a basis for the invariants up to that degree.
\section{Finding the basic invariants}\label{finding}
A set of basic invariants of the algebra $I$ of invariants
is a minimal set of generators. The individual generators are
not uniquely determined, but their degrees are.
\medskip
The ring $I$ is graded: $I = \oplus_m I_m$, where $I_m$ is the
subspace of invariants, homogeneous of degree $m$.
If $y_1,\ldots,y_{n-2}$ is a system of parameters, where $y_i$ has
degree $d_i$, then the Poincar\'e series $P(t)$, defined by
$P(t) = \sum_m \dim I_m t^m$, can be written as a rational function
in $t$ with denominator $\prod (1-t^{d_i})$.
(Throughout this note, $\dim$ is vector space dimension, that is,
is $\dim_\mathbb{C}$.)
Now $P(t)$ is known: it was given as a series by Cayley \& Sylvester
(cf. \cite{Sy}) and as a rational function by Springer \cite{Spr}.
For $n=10$ we have
\begin{align*}
P(t) = \phantom{} & 1 + t^2 + 2t^4 + 6t^6 + 12t^8 + 5t^9 + 24t^{10} +
13t^{11} + 52t^{12} + 33t^{13} + 97t^{14} + \\
& 80t^{15} + 177t^{16} + 160t^{17} + 319t^{18} + 301t^{19} + 540t^{20} +
547t^{21} + 887t^{22} + \\
& 926t^{23} + 1429t^{24} + 1512t^{25} + 2219t^{26} + 2402t^{27} +
3367t^{28} + 3681t^{29} + \\
& 5015t^{30} + 5502t^{31} + 7294t^{32} + 8064t^{33} + 10419t^{34} +
11550t^{35} + \\
& 14664t^{36} + 16253t^{37} + 20287t^{38} + 22531t^{39} + 27682t^{40} +
30738t^{41} + \\
& 37319t^{42} + 41378t^{43} + 49671t^{44} + 55060t^{45} + 65390t^{46} +
72391t^{47} + \\
& 85250t^{48} + \cdots
\end{align*}
If we generate invariants of degree $m$, and have found $\dim I_m$
independent ones, then we have found all in degree $m$.
If we know that there is a hsop with degrees
$2$, $4$, $6$, $6$, $8$, $9$, $10$, $14$, then
$$
P(t) = a(t) / (1-t^2)(1-t^4)(1-t^6)^2(1-t^8)(1-t^9)(1-t^{10})(1-t^{14})
$$
where
\begin{align*}
a(t) = \phantom{} & 1 + 2t^6 + 4t^8 + 4t^9 + 7t^{10} + 8t^{11} + 15t^{12} +
15t^{13} + 20t^{14} + 27t^{15} + \\
& 29t^{16} + 35t^{17} + 40t^{18} + 44t^{19} + 47t^{20} + 55t^{21} +
52t^{22} + 57t^{23} + 56t^{24} + \\
& 57t^{25} + 52t^{26} + 55t^{27} + 47t^{28} + 44t^{29} + 40t^{30} +
35t^{31} + 29t^{32} + 27t^{33} + \\
& 20t^{34} + 15t^{35} + 15t^{36} + 8t^{37} + 7t^{38} + 4t^{39} + 4t^{40} +
2t^{42} + t^{48}.
\end{align*}
This means that all basic invariants have degree at most 48,
and we never have to consider subspaces of dimension larger
than 85250, which is doable.
\medskip
So, the procedure is to find basic invariants in some way,
and multiply them together so as to construct for each $m$ the
invariants in $I_m$ that are known already. Compute a basis for
the subspace of $I_m$ spanned by these known invariants,
and if this subspace has the same dimension as $I_m$ itself,
it is all of $I_m$ and we can go to the next $m$.
Since any invariant can be written as a linear combination
of bracket monomials, it seems reasonable to expect that one can
find a spanning set for $I_m$ by just randomly generating some bracket
monomials. This worked fine for the nonic, and for the decimic in degrees
below 21, but in degree 21 where $\dim I_{21} = 547$ and we quickly
generated a subspace of dimension 546, a few dozen attempts to
randomly generate an invariant outside this hyperplane failed.
Therefore, we reverted to the procedure with guaranteed success:
Gordan proved that a basis for the invariants can be found effectively
by computing transvectants, and this indeed yielded the 106th invariant.
(Immediately afterwards the random process also succeeded.)
\medskip
Various reductions simplify the calculations.
First of all, we did the computations modulo a small prime $p$,
e.g. $p=109$ worked. If the images of the invariants under
reduction mod $p$ are independent, then the invariants are independent.
Secondly, if the form is $\sum _{i=0}^{10} \binom{10}{i}a_ix^{10-i}y^i$, we took $a_4=a_7=a_9=0$ and $a_{10}=1$.
Again: if the images of the invariants under this substitution
are independent, then the invariants are independent.
Similar things work for the nonic. But here we have the invariant
$j_2 = a_0a_{10}-10a_1a_9+45a_2a_8-120a_3a_7+210a_4a_6-126a_5^2$ of degree 2.
After the substitutions this becomes $a_0+45a_2a_8-126a_5^2$, and the
substitution $a_0 = -45a_2a_8+126a_5^2$ maps $I_m$ onto $I_m/j_2 I_{m-2}$,
and $\dim I_m/j_2 I_{m-2} = \dim I_m - \dim I_{m-2}$.
Now six variables ($a_1,a_2,a_3,a_5,a_6,a_8$) are left, and the largest dimension
occurring is $\dim I_{48}/j_2I_{46} = 19860$, comparatively small.
(Compared to $\dim I_{48} = 85250$, this saves almost a factor 80
in computation time when an $O(N^3)$ rank algorithm is used.)
\medskip
The computation was done, and the result is: for $m \le 48$ the
values of $d_m$ are as listed in Proposition \ref{prop1}.
Consequently, if there is a system of parameters with degrees
2, 4, 6, 6, 8, 9, 10, 14, so that no basic invariant has degree
larger than 48, then Proposition \ref{prop1} follows.
\section{A system of parameters for $\mathcal{O} (V_{10})^{\SLtwo}$}
Let $V_n$ be the space of forms of degree $n$ (in the variables $x,y$).
A {\it covariant of order m and degree d} of $V_n$ is an
$\SLtwo$-equivariant homogeneous polynomial map
$\phi: V_n \rightarrow V_m$ of degree $d$.
The invariants of $V_n$ are the covariants of order 0.
The identity map is a covariant of order $n$ and degree 1.
Customarily, one indicates such a covariant $\phi$ by giving its
image of a generic element $f \in V_n$.
(In particular, the identity map is noted $f$.)
Let $V_{m,d}$ be the space of covariants of order $m$ and degree $d$.
Consider $f \in V_{10}$,
\[f=a_0x^{10}+10a_1x^9y+\ldots +10 a_9xy^9+a_{10}y^{10}, \]
and the following covariants
\begin{alignat*}{4}
k&=(f,f)_8 &\in\,& V_{4,2},\quad\quad&
m&=(f,k)_4 &\in\,& V_{6,3},\quad\quad\quad\quad\\
q&=(f,f)_6 &\in\,& V_{8,2},&
r&=(f,q)_8 &\in\,& V_{2,3},\\
k_q&=(q,q)_6 &\in\,& V_{4,4},&
k_m&=(m,m)_4\, &\in\,& V_{4,6},\\
m_q&=(q,k_q)_4~ &\in\,& V_{4,6},
\end{alignat*}
and invariants (the suffix indicates the degree)
\begin{alignat*}{4}
j&_2 &\,=\,& (f,f)_{10},&
A&_6 &\,=\,& (m,m)_6,\\
j&_4 &\,=\,& (k,k)_4,&
C&_6 &\,=\,& (r,r)_2,\\
j&_8 &\,=\,& (k,k_m)_4,&
j&_{14} &\,=\,& ((k_q,k_q)_2,m_q)_4,\\
j&_9 &\,=\,& ((m,k)_1,k^2)_8,\quad\quad&
A&_{14} &\,=\,& ({(k,k)_2}^2,(m,m)_2)_8,\\
j&_{10} &\,=\,& ((m,m)_2,k^2)_8.
\end{alignat*}
\smallskip
\begin{theorem}\label{explhsop}
The eight invariants $j_2$, $j_4$, $A_6$, $C_6$, $j_8$, $j_9$,
$j_{10}$, $j_{14}+A_{14}$ form a homogeneous system of parameters for
the ring $\mathcal{O}(V_{10})^{\SLtwo}$ of invariants of the binary decimic.
\end{theorem}
This is proved by invoking Hilbert's characterization of
homogeneous systems of parameters as sets that define the nullcone.
\subsection{The nullcone}
The {\em nullcone} of $V_n$, denoted $\mathcal{N}(V_n)$, is the set of binary forms
of degree $n$ on which all invariants vanish. It turns out (\cite{Hi2})
that this is precisely the set of binary forms of degree $n$
with a root of multiplicity $>\frac{n}{2}$. The elements of $\mathcal{N}(V_n)$
are called {\em nullforms}.
The nullcone $\mathcal{N}(V_n\oplus V_m)$ is
the set of pairs $(g,h)\in V_n\oplus V_m$ such that $g$ and $h$ have a
common root of multiplicity $>\frac{n}{2}$ in $g$ and of multiplicity
$>\frac{m}{2}$ in $h$. (In this note, this result can be taken as the
definition of the symbol $\mathcal{N}(V_n\oplus V_m)$.)
We have the following result, due to Hilbert \cite{Hi2},
formulated for the particular case of binary forms:
\begin{proposition} \label{hilbert}
For $n \ge 3$, consider $i_1,\ldots ,i_{n-2}\in \mathcal{O}(V_n)^{\SLtwo}$
homogeneous invariants of $V_n$.
The following two conditions are equivalent:
\begin{itemize}
\item[(i)] $\mathcal{N}(V_n)=\mathcal{V}(i_1,\ldots ,i_{n-2})$,
\item[(ii)] $\{i_1,\ldots ,i_{n-2}\}$ is a homogeneous system of parameters
of $\mathcal{O}(V_n)^{\SLtwo}$.
\end{itemize}
\end{proposition}
(Here $\mathcal{V}(J)$ stands for the vanishing locus of $J$.)
\medskip\noindent
We prove the above theorem by first finding a defining set
for the nullcone that is still too large, and then showing
that some elements are superfluous.
\medskip
\medskip
We need information on the invariants of $V_n$ for $n = 2,\,4,\,6,\,8$:
\begin{lemma}\label{hsops}
The following are systems of parameters of
$\mathcal{O}(V_n)^{\SLtwo}$ for $n = 2,\,4,\,6,\,8$.
\begin{itemize}
\item[(i)] If $n=2$: $(f,f)_2$ of degree $2$.
\item[(ii)] If $n=4$: $(f,f)_4$ and $((f,f)_2,f)_4$ of degrees $2$ and $3$.
\item[(iii)] If $n=6$: $(f,f)_6$, $(k,k)_4$, $((k,k)_2,k)_4$,
and $(m^2,(k,k)_2)_4$ of degrees $2$, $4$, $6$ and $10$,
where $k = (f,f)_4$ and $m = (f,k)_4$.
\item[(iv)] If $n=8$: $(f,f)_8$, $((f,f)_4,f)_8$, $(k,k)_4$, $(m,k)_4$,
$((k,k)_2,k)_4$, $((k,k)_2,m)_4$ of degrees $2$, $3$, $4$, $5$, $6$ and $7$,
where $k = (f,f)_6$ and $m = (f,k)_4$.
\end{itemize}
\end{lemma}
\begin{proof}
This is classical for $n=2$, $4$, $6$, see, e.g., \cite{Clebsch,GrYo,Schur},
and due to von Gall \cite{vG8} and Shioda \cite{Sh} for $n=8$.
\end{proof}
\begin{lemma} \label{jerzy} {\rm (Weyman \cite{We1})}
Let $f\in V_d$. If $d > 4k-4$ and all $(f,f)_{2k}$, $(f,f)_{2k+2}$, ...
vanish, then $f$ has a root of multiplicity $d-k+1$. If $d = 4k-4$ and
$((f,f)_{2k-2},f)_d$, $(f,f)_{2k}$, $(f,f)_{2k+2}$, ... vanish,
then $f$ has a root of multiplicity $d-k+1$. \qed
\end{lemma}
\begin{lemma} \label{covariants} Let $f\in V_{10}$ and $j_2=(f,f)_{10}$,
$k=(f,f)_8\in V_4$, $m=(f,k)_4 \in V_6$, $q=(f,f)_6\in V_8$. We have:
\begin{enumerate}
\item[(i)] If $j_2=0$, $k\neq 0$ and $(k,m)\in \mathcal{N}_{V_4\oplus V_6}$,
then $f$ has a root of multiplicity 6.
\item[(ii)] If $j_2=0$, $k=0$ and $0\neq q \in \mathcal{N}_{V_8}$,
then $f$ has a root of multiplicity 7.
\item[(iii)]
If $j_2=0$, $k=0$ and $q=0$, then $f$ has a root of multiplicity 8.
\end{enumerate}
\end{lemma}
\begin{proof}
The covariants $k$, $q$ and the invariant $j_2$ are:
\begin{align*}
j_2&=-252 a_5\supr{2} + 420 a_4 a_6 - 240 a_3 a_7 + 90 a_2 a_8 - 20 a_1 a_9 + 2 a_0 a_{10},\\
k&=(70 a_6\supr{2} - 112 a_5 a_7 + 56 a_4 a_8 - 16 a_3 a_9 + 2 a_2 a_{10})y^4+\\
&(56 a_5 a_6 - 112 a_4 a_7 + 80 a_3 a_8 - 28 a_2 a_9 + 4 a_1 a_{10}) xy^3+\\
&(168 a_5\supr{2} - 252 a_4 a_6 + 96 a_3 a_7 - 6 a_2 a_8 -8 a_1 a_9 + 2 a_0 a_{10}) x^2y^2+\\
&(56 a_4 a_5 - 112 a_3 a_6 + 80 a_2 a_7 - 28 a_1 a_8 + 4 a_0 a_9) x^3 y+\\
&(70 a_4\supr{2} - 112 a_3 a_5 + 56 a_2 a_6 - 16 a_1 a_7 + 2 a_0 a_8)x^4,\\
q&=(-20 a_7\supr{2} + 30 a_6 a_8 - 12 a_5 a_9 + 2 a_4 a_{10})y^8+\\
&(-40 a_6 a_7 + 72 a_5 a_8 - 40 a_4 a_9 + 8 a_3 a_{10})y^7 x+\\
&( -140 a_6\supr{2} + 168 a_5 a_7 - 40 a_3 a_9 + 12 a_2 a_{10})y^6x\supr{2}+\\
&( -168 a_5 a_6 + 280 a_4 a_7 - 120 a_3 a_8 +
8 a_1 a_{10}) y^5x^3+\\
&(-252 a_5\supr{2} + 280 a_4 a_6 + 40 a_3 a_7 -
90 a_2 a_8 + 20 a_1 a_9 + 2 a_0 a_{10}) y^4x^4+\\
&(-168 a_4 a_5 +
280 a_3 a_6 - 120 a_2 a_7 + 8 a_0 a_9) y^3x^5+\\
&(-140 a_4\supr{2} +
168 a_3 a_5 - 40 a_1 a_7 + 12 a_0 a_8) y^2x^6+\\
&(-40 a_3 a_4 +
72 a_2 a_5 - 40 a_1 a_6 + 8 a_0 a_7) y x^7+\\
&(-20 a_3\supr{2} +
30 a_2 a_4 - 12 a_1 a_5 + 2 a_0 a_6)x^8.
\end{align*}
\noindent
(i). If $(k,m) \in \mathcal{N}_{V_4\oplus V_6}$ then $k$ and $m$ have a common root,
of multiplicity 3 in $k$ and of multiplicity 4 in $m$.
Without loss of generality we consider the cases $k=x^4$, $x^4 \mid m$
and $k=x^3y$, $x^4 \mid m$.
\medskip\noindent
Case 1: $k=x^4$. Then $m$ becomes:
\[m=(f,x^4)_4=a_4 x^6 + 6 a_5 x^5 y + 15 a_6 x^4 y^2 + 20 a_7 x^3 y^3 +
15 a_8 x^2 y^4 + 6 a_9 x y^5 + a_{10} y^6.\]
From $x^4 \mid m$ it follows $a_7=\ldots =a_{10}=0$. We replace this in $k$ and because we supposed $k=x^4$ we obtain also $a_6=a_5=0$. But then $x^6 \mid f$, hence $f$ will have a root of multiplicity 6.
\medskip\noindent
Case 2: $k=x^3 y$. Then $m$ becomes:
\[
m\!=\!(f,x^3 y)_4\!=\!
-a_3 x^6 - 6 a_4 x^5 y - 15 a_5 x^4 y^2 - 20 a_6 x^3 y^3 -
15 a_7 x^2 y^4 - 6 a_8 x y^5 - a_9 y^6.
\]
From $x^4 \mid m$ it follows $a_6=\ldots =a_9=0$. We replace this in $k$ and $j_2$ and as we supposed $k=x^3y$ we obtain
\begin{align*}
168 a_5\supr{2} + 2 a_0 a_{10}=0,\\
-252 a_5\supr{2} + 2 a_0 a_{10}=0,
\end{align*}
which implies $a_5=0$. But then the coefficient of $x^3$ in $k$ becomes 0. Contradiction with our assumption.
\medskip\noindent
(ii). Without loss of generality we suppose $x^5 \mid q$.
We denote by $J$ the ideal generated by $j_2$, the coefficients of $k$
and the coefficients of $x^4 y^4,x^3y^5,\ldots, y^8$ in $q$.
Denote also by $p_1$, $p_2$ and $p_3$ the coefficients of
$x^7y$, $x^6y^2$ and $x^5y^3$, respectively, in $q$. We have
\[p_1\supr{4},p_2\supr{3},p_3\supr{2} \in J,\]
which implies that $x^8 \mid q$.
Consider now the ideal $J$ generated by $j_2$, the coefficients of $k$
and the coefficients of $x^7y,x^6y^2,\ldots, y^8$ in $q$.
Denote by $p_0$ the coefficient of $x^8$ in $q$.
We have $a_i p_0\in J$ for $i=10,9,8,7,6,5,4$.
Because $q \neq 0$ we find $a_{10}=\ldots=a_4=0$.
This means that $x^7 \mid f$, so $f$ will have a root of multiplicity 7.
\medskip\noindent
(iii). This follows from Lemma \ref{jerzy}.
\end{proof}
\begin{lemma} \label{kmv10}
Let $k\in V_4$ and $m\in V_6$, $k\neq 0$, $m\neq 0$, both of them nullforms.
If the transvectants $((m,m)_4,k)_4$, $((m,m)_2,k^2)_8$, $(m^2,k^3)_{12}$,
$((m,k)_1,k^2)_8$, and $((k,k)_2\supr{2},(m,m)_2)_8$ vanish, then
$(k,m)\in \mathcal{N}_{V_4\oplus V_6}$.
\end{lemma}
\begin{proof}
Suppose $(k,m)\notin \mathcal{N}_{V_4\oplus V_6}$.
Without loss of generality we suppose
\begin{align*}
k=&x^3(a_1 x+a_2y),\\
m=&y^4(b_1x^2+b_2xy+b_3y^2).
\end{align*}
We have
\[0=((m,m)_4,k)_4 \sim a_1b_1\supr{2}\]
\medskip\noindent
Case 1: $a_1=0$. Then
\begin{alignat*}{2}
0&=((m,m)_2,k^2)_8 &\,\sim\,& a_2\supr{2} b_1\supr{2},\\
0&=((m,k)_1,k^2)_8 &\,\sim\,& a_2\supr{3}b_3,\\
0&=((k,k)_2\supr{2},(m,m)_2)_8 &\,\sim\,& a_2\supr{4}(5b_2\supr{2}-12 b_1b_3)
\end{alignat*}
Because $k\neq 0$ we have $a_2 \neq 0$, but then it follows that
$b_1=b_3=b_2=0$. Contradiction with $m\neq 0$.
\medskip\noindent
Case 2: $a_1 \ne 0$, $b_1=0$. Then
\begin{alignat*}{2}
0&=((m,m)_2,k^2)_8 &\,\sim\,& a_1\supr{2} b_2\supr{2},\\
0&=((m,k)_1,k^2)_8 &\,\sim\,& a_2\supr{3}b_3,\\
0&=((k,k)_2\supr{2},(m,m)_2)_8 &\,\sim\,& a_2\supr{4}b_2\supr{2},\\
0&=(m^2,k^3)_{12} &\,\sim\,& a_1(a_2\supr{2} b_2\supr{2} - 11 a_1 a_2 b_2 b_3
+ 22 a_1\supr{2} b_3\supr{2})
\end{alignat*}
If $a_2\neq 0$ then $b_2=b_3=0$.
And if $a_2=0$ then $a_1\supr{2}b_2\supr{2}=a_1\supr{3} b_3\supr{2}=0$,
and again $b_2=b_3=0$. Contradiction with $m\neq 0$.
\end{proof}
After this preparation we can write down a defining set for the nullcone.
Define $k$, $m$, $q$, $j_2$, $j_4$, $A_6$, $j_8$, $j_9$, $j_{10}$, $j_{14}$,
$A_{14}$ as above (before Theorem \ref{explhsop}), and moreover
\begin{alignat*}{4}
j&_6 &\,=\,& ((k,k)_2,k)_4,\quad\quad&
A&_{12} &\,=\,& (m^2,k^3)_{12},\\
B&_6 &\,=\,& ((q,q)_4,q)_8.
\end{alignat*}
\begin{proposition} \label{nullconev10}
With notations as above, the nullcone $\mathcal{N} _{V_{10}}$ is defined by
\[
\mathcal{N} _{V_{10}}=\mathcal{V} (j_2,j_4,j_6,A_6,B_6,j_8,j_9,j_{10},A_{12},j_{14},A_{14}).
\]
\end{proposition}
\begin{proof}
Since $k \in V_4$ we can apply Lemma \ref{hsops}(ii) and conclude that
if $j_4=j_6=0$ then $k$ is a nullform.
Without loss of generality we consider three cases:
$k=0$, $k=x^4$ and $k=x^3y$.
\medskip\noindent
Case 1: $k=0$. Denote by $I=(j_2,k)$ the ideal generated by $j_2$ and
the coefficients of $k$. Define
\begin{alignat*}{4}
A&_4 &\,=\,& (q,q)_8,&
A&_{10} &\,=\,& (m_q,k_q)_4,\\
A&_8 &\,=\,& (k_q,k_q)_4,\quad\quad&
B&_{12} &\,=\,& ((k_q,k_q)_2,k_q)_4,
\end{alignat*}
Since $q \in V_8$, in order to show that $q$ is a nullform
it suffices by Lemma \ref{hsops}(iv) to show that each of
$A_4$, $B_6$, $A_8$, $A_{10}$, $B_{12}$ and $j_{14}$ vanishes.
Easy Gr\"obner basis computations show that
$A_4,A_8,A_{10} \in I$ and $B_{12} \in (I,B_6)$.
It follows that if $k=0$ and $j_2=B_6=j_{14}=0$ then $q$ is a nullform.
Now Lemma \ref{covariants} implies that $f$ is a nullform.
\medskip\noindent
Case 2: $k=x^4$. Then we have:
\begin{align*}
A_{12} &\sim a_{10}\supr{2},\\
j_{10} &\sim - a_9\supr{2} + a_8 a_{10},\\
j_8 &\sim 3 a_8\supr{2} - 4 a_7 a_9 + a_6 a_{10},\\
A_6 &\sim -10 a_7\supr{2} + 15 a_6 a_8 - 6 a_5 a_9 + a_4 a_{10}.
\end{align*}
If $A_{12}=j_{10}=j_8=A_6=0$ then it follows that $a_{10}=\ldots =a_7=0$.
If we substitute this in $k$ we obtain
\begin{align*}
k=\,&70 a_6\supr{2} y^4+56 a_5 a_6 xy^3+
(168 a_5\supr{2} - 252 a_4 a_6) x^2y^2+\\
&(56 a_4 a_5 - 112 a_3 a_6) x^3 y+
(70 a_4\supr{2} - 112 a_3 a_5 + 56 a_2 a_6)x^4,
\end{align*}
and as we supposed $k=x^4$ we get also $a_6=a_5=0$,
which implies that $f$ is a nullform.
\medskip\noindent
Case 3: $k=x^3y$. Then we have:
\begin{align*}
j_9 &\sim a_9,\\
A_{14} &\sim a_7a_9-a_8\supr{2},\\
j_{10} &\sim -5 a_7\supr{2} + 2 a_6 a_8 + 3 a_5 a_9,\\
A_6 &\sim -10 a_6\supr{2} + 15 a_5 a_7 - 6 a_4 a_8 + a_3 a_9.
\end{align*}
If $j_9=A_{14}=j_{10}=A_6=0$ then $a_9=\ldots =a_6=0$.
We substitute this in $k$ and $j_2$:
\begin{align*}
k=\,& 2 a_2 a_{10}y^4+4 a_1 a_{10} xy^3+
(168 a_5\supr{2} + 2 a_0 a_{10}) x^2y^2+\\
&56 a_4 a_5 x^3 y+
(70 a_4\supr{2} - 112 a_3 a_5 )x^4,\\
j_2=\,&-252 a_5\supr{2} + 2 a_0 a_{10}
\end{align*}
From $168 a_5\supr{2} + 2 a_0 a_{10}=-252 a_5\supr{2} + 2 a_0 a_{10}=0$
we find $a_5=0$, which contradicts $k=x^3y$.
\end{proof}
So far, we defined the nullcone using 11 invariants, but we need
a definition using 8 invariants. As a first step, replace the two
invariants of degree 14 by a single one.
Now for $f=x^2 y (2a_1 x^7 + 9a_8 y^7)$
all invariants from Proposition \ref{nullconev10} vanish, except $A_{14}$.
And for $f=y^3 (120a_3 x^7 + a_{10} y^7)$
all invariants from Proposition \ref{nullconev10} vanish, except $j_{14}$.
That means that the single invariant of degree 14 cannot be either
$j_{14}$ or $A_{14}$. However, as it turns out we can use $j_{14}+A_{14}$.
\subsection{Finding the system of parameters}
Proposition \ref{nullconev10} gives an explicit set of invariants
(and in particular an explicit set of degrees of invariants)
that define the nullcone. Having that, only a finite amount of work is left.
The final part of the construction of the system of parameters
was done by computer. All computations were carried out in the
ring $R$ generated by the 106 invariants found in Section \ref{finding}.
Or, more precisely, in the quotient $Q = R/j_2R$, reduced mod $p$,
where this time $p = 197$ (the different $p$ has no significance),
and again $a_4$, $a_7$ and $a_9$ were taken to be zero.
It was checked that the graded parts of the resulting ring have
the expected dimension (for degree up to 54), so that no collapse
occurred as a consequence of the reduction mod $p$ or the substitution
of variables.
The ideal generated in this ring by all invariants of degrees
4, 6, 8, 9, 10, 14 has full dimension 542 for its graded part
of degree 24. We know that $\dim I_{24} = 1429$ and $\dim I_{22} = 887$
and multiplication by $j_2$ is an injection, so $\dim I_{24}/j_2I_{22} = 542$.
It follows that the ideal generated by these invariants, together with $j_2$,
contains all of $I_{24}$, so that no invariants of degree 12 are needed
to define the nullcone (since their squares are in $I_{24}$, and they
themselves are in the radical).
With only $j_{14}+A_{14}$ instead of all invariants of degree 14
in the set of generators of the ideal, one finds full dimension 1148
for the graded part of degree 28, so this single invariant of degree 14
suffices.
With only $j_{10}$ instead of all invariants of degree 10, one finds
full dimension 221 in degree 20, so this single invariant of degree 10
suffices.
With only $j_9$ instead of all invariants of degree 9, one finds
full dimension 890 in degree 27, so this single invariant of degree 9
suffices.
With only $j_8$ instead of all invariants of degree 8, one finds
full dimension 2279 in degree 32, so this single invariant of degree 8
suffices.
That only leaves the invariants of degree 6. After some work it turned out
that with only $A_6$ and $C_6$ one finds full dimension 37892 in degree 54,
so these suffice, and we have constructed the homogeneous system of
parameters promised in Theorem \ref{explhsop}.
\medskip
Note that one knows what to expect if all is well:
the coefficients of the polynomial $a(t)$ from Section \ref{finding}
give for each degree the codimension of the set of invariants
in the ideal generated by the hsop in the space of all invariants
of that degree. Since 54 is the smallest multiple of 6 where $a(t)$
has zero coefficient, that explains why the computation had to extend
to there.
|
\section{Introduction}
\setcounter{equation}{0}
Aloha is one of the most common examples of a multiple access
protocol \cite{Schwartz87,Bertsekas01}.
The classical approach to Aloha adopts simplified packet
collision models in which simultaneous
transmissions are never successful. This makes this classical approach not well
adapted to a wireless MANET scenario, where it is the SINRs at
different receiver locations
which determine the set of successful transmissions.
The present paper contributes to the study of Spatial Aloha, a variant of Aloha well
adapted to MANETs. More precisely, it bears on the mathematical
analysis of Spatial Aloha in the context of large Mobile
Ad hoc Network (MANETs) with randomly located nodes. It focuses
on the SINR coverage scheme, where each transmission requires
that the receiver be covered by the transmitter with a minimum SINR.
The present paper uses the stochastic geometry approach proposed in \cite{allerton,jsac}
where the {\em space-time density of successful transmissions} was evaluated and optimized.
The present paper identifies a potential weakness of this SINR coverage scheme: in most practical
cases, and in particular as soon as thermal noise is bounded from
below by a positive constant, the mean delay for a typical
node to transmit a packet is infinite e.g. in the Rayleigh fading case.
\subsection{Main Paper Contributions}
The main {\em modeling advances} of the present paper are two-fold:
\begin{itemize}
\item We add a {\em time dimension} to the existing spatial analysis.
Time is slotted, and we hence focus on slotted Aloha.
We assume that the geographical locations of the MANET
nodes remain unchanged over time and that only the variables modeling the
MAC status (allowed to transmit or delayed) and the channel characteristics
(such as fading and thermal noise) vary over time.
In other words, we consider a full separation of the time scale of node
mobility on one side and the time scale of MAC and physical layer on the
other side, which makes sense in many practical situations, where the
former is much larger than the later. This is in
contrast with what happens in delay tolerant networks (DTNs) ; see
e.g.~\cite{GrossglauserTse2002,Jacquetinfo09}.
\item We propose some more {\em realistic receiver models} than the
fixed-distance receiver model introduced in \cite{allerton,jsac};
the new models are inspired by the fact that the routing schemes typically choose, as next-hop,
the closest possible receiver among some common set of potential receivers.
\end{itemize}
The main theoretical advances of the paper bear on the analysis of the {\em local
delays} in such MANETs; the local delay of a node is the random numbers of
slots required by this node to successfully transmit a packet to its
next-hop node.
We first perform a time analysis of these local delays given the
location of the MANET nodes. This analysis shows that each node has
finite-mean geometric conditional random delay and thus a positive
throughput to its next-hop receiver.
However, the spatial irregularities of the network imply that
these conditional throughputs vary from node to node; in a Poisson
configuration, one can find nodes which have an arbitrarily
small throughput and consequently an arbitrarily large delay.
In order to capture the performance of the whole MANET, one usually
considers its ``typical node''. The typical node statistics
are spatial (or large population) averages of the individual node
characteristics.
Our analysis shows in several practical cases,
including the Rayleigh fading and positive thermal noise case,
that the local delay of the typical MANET node
is heavy tailed and that its mean is infinite,
which however does {\em not} imply that the mean throughput of the
typical node is null. Moreover, in certain cases,
the mean local delay of the typical node exhibit a phase transition
phenomenon that we propose
to call the {\em wireless contention phase transition}: it is
finite when certain model parameters (as receiver distance, thermal
noise, Aloha medium access probability)
are below a threshold and infinite above.
On the theory side, we also comment on the connections between the
heavy tailedness alluded to the local delay of the typical MANET node
and the observations of
\cite{JelenkovicTanInfocom07,JelenkovicTan2007},
where it was shown that a finite population ALOHA model with
variable and unbounded size packets has power law transmission
delays. Although the
physical phenomena at hand are quite different here (our spatial MANET
model has fixed-packet-sizes) and there, we use
the so-called ``RESTART algorithm phenomenon'', which has recently received
a lot of attention (see e.g.~\cite{Asmussen,JelenkovicTan2009}), to establish
some links between our findings and the results of \cite{JelenkovicTanInfocom07,JelenkovicTan2007}.
More precisely, we argue that, in our spatial MANET model with fixed-packet-size Aloha MAC,
the delay of the typical node is heavy tailed and can have infinite mean due to a ``RESTART'' phenomenon,
where the {\em spatial irregularities in the MANET} play the same role as
the packet size variability in \cite{JelenkovicTanInfocom07} and \cite{JelenkovicTan2007}.
The main practical contributions of the paper bear on ways to
guarantee finite mean local delay of the typical node
by {\em increasing diversity} in the MANET.
The proposed solutions have the potential of
breaking the RESTART rigidity, e.g. by increasing
the variability of fading, by increasing mobility, by
adding appropriate receivers or, finally, by using adaptive coding,
which completely brakes out the
outage/RESTART logic.
The paper is organized as follows. In the remaining part of this
section we briefly present the related work.
We present our MANET model in Section~\ref{s.MANET}.
The mean local delays are introduced and evaluated in Section~\ref{s.TheLD}.
We analyze the phase-transition phenomenon in Section~\ref{s.Phase-Transition}.
Section~\ref{s.FiniteMean} focuses on the ways to make the mean delays finite.
We conclude our work in Section~\ref{s.conclusion}.
\subsection{Related Work}
\label{ss.RelatedWork}
As already mentioned, the present paper assumes a full time-scale
separation for the mobility on one side and for the MAC and physical layer
on the other side. This assumption makes a major difference between what
is done in this paper and what is done in DTNs, where one leverages node mobility
to contribute to the transport of packets. There is a large number of publications
on the throughput in DTNs and we will not review the literature on the
topic which is huge.
Let us nevertheless stress that there are some interesting connections between the
line of though started in (\cite{GrossglauserTse2002}), where it was first shown that
mobility increases capacity and what is done in Section~\ref{secalo:HMN}.
We show in this section that mobility helps in a way which is quite different from
that considered in \cite{GrossglauserTse2002}:
mobility may in certain cases break dependence and hence mitigate the RESTART phenomenon,
it may hence decrease the mean local delay of the typical node (or equivalently increase
its throughput), even if one {\em does not use mobility to transport packets}.
Among recent papers, that we are aware of, on the time-space analysis of MANETs,
we would quote \cite{Jacquetinfo09} and \cite{Haenggi-spaswin09}. The former focuses
on node motion alone and assumes that nodes within transmission range can transmit
packets instantaneously. The authors then study the speed at which some multicast
information propagates on a Poisson MANET where nodes have independent motion
(of the random walk or random way-point type).
The latter focuses on a first passage percolation problem,
however, the model used in \cite{Haenggi-spaswin09} is the so called protocol model;
its analysis significantly differs from that of our physical (SINR-based) model.
In particular, there is no notion of local delay.
\section{MANET model with Slotted Aloha}
\setcounter{equation}{0}
\label{s.MANET}
\subsection{Space-Time Scenarios}
\label{ss.ST-scenarios}
In what follows we describe a few space-time models considered in this
paper. In the most simple case this will consist in adding
the time-dimension to the Poisson Bipolar model introduced and
studied in~\cite{jsac}. We will go however beyond the simple receiver
model proposed there.
The idea consists in assuming
that the geographical locations of the MANET nodes remain unchanged over time
and that the MAC status (allowed to transmit or delayed)
and other characteristics of each node
(as fading and thermal noise) vary over time.
Time is discrete with a sequence of {\em time slots}, $n=0,1,\ldots$,
w.r.t. which all the nodes are perfectly synchronized.
More precisely, we assume that a snapshot of the MANET
can be represented by a marked Poisson point process
(P.p.p.) on the Euclidean plane, where the point process is homogeneous
with intensity $\lambda$ and represents the random locations of the nodes
and where the multidimensional mark of a
point/node carries information about its MAC status and other characteristics of the
channels in the successive time slots.
This marked Poisson p.p. is denoted by
$\widetilde\Phi=\{(X_i,(e_i(n),y_i,\mathbf{F}_i(n), W_i(n):n)\}$, where:
\begin{itemize}
\item $\Phi=\{X_i\}$ denotes the {\em locations of the potential
transmitters} of the MANET; $\Phi$ is always assumed Poisson with positive and finite intensity $\lambda$;
\item\label{i.ei}
$e_{i}(n)$ is the {\em MAC decision} of point $X_i$ of $\Phi$ at time $n$;
we will always assume that, given $\Phi$, the random variables
$e_i(n)$ are i.i.d. in $i$ and $n$; i.e.
in space and time, with $\Pr\{\,e_i(n)=1\,\}=1-\Pr\{\,e_i(n)=0\,\}=p$.
\item\label{i.yi} $y_i\in{\mathbb{R}}^2$ is the location of the
{\em receiver} of the node $i$; a few scenario for the choice of the
receivers are presented in Section~\ref{ss.receiver} below.
\item\label{i.Fi}
$\mathbf{F}_i(n)=\{F_i^j(n):j\}$ is the {\em virtual power}
emitted
by node $i$ (provided $e_i=1$)
towards receiver $y_j$ at time $n$.
By virtual power $F_i^j(n)$, we understand the product of the effective
power of transmitter $i$ and of the random fading from this
node to receiver $y_j$. The random (vector valued) processes
$\{F_i^j(n):n\}$ are assumed to be i.i.d. in $i$ and $j$ given
$\{X_i,y_i\}$, and we denote by $F$ the
generic marginal random variable $F_i^j(n)$.
In the case of constant effective transmission power $1/\mu$ and
Rayleigh fading (which is our default scenario),
$F$ is exponential with mean $1/\mu$
(see e.g.~\cite[p.~50 and~501]{TseVis2005}). However, one also consider also non exponential cases
to analyze other types of fading such as e.g. Rician
or Nakagami scenarios (cf~\cite[Sec.~23.2.4]{FnT2})
or simply the case without fading (when
$F\equiv 1/\mu$ is deterministic).
Note that we do not have specified yet the dependence between the
virtual powers $F_i^j(n)$ for different time slots $n$. This is
done in what follows.
\begin{itemize}
\item By the {\em fast fading} case we understand
the scenario when the random variables $F_i^j(n)$
are i.i.d. in~$n$ (recall, that the default option is that they are
also i.i.d. in $i,j$);~%
~\footnote{%
Note that our fast fading means that it remains constant over a slot
duration and can be seen as i.i.d. over different time slots.
This might not correspond to the terminology used in many papers of literature,
where fast fading means that the channel conditions fluctuate much over a given time slot.}
\item The {\em slow fading} case is that where $F_i^j(n)\equiv F_i^j$,
for all $n$.
\end{itemize}
\item\label{i.W} $W_i(n)$ represents the thermal noise at the receiver $y_i$
at time $n$. The processes $\{W_i(n):n \}$ are independent
in~$i$ given $\{y_i\}$, with the generic marginal random variable
denoted by $W$. For the time dependence,
one can consider both $W_i(n)$ independent in $n$
({\em fast noise}) and constant $W_i(n)\equiv W_i$ ({\em slow noise}).
\end{itemize}
\subsection{Receiver Models}
\label{ss.receiver}
We first recall the simple receiver model used in~\cite{jsac} and then
propose a few possible extensions.
\subsubsection{Bipolar Receiver Model} In this model the receivers
are {\em external to the MANET p.p.} $\Phi=\{X_i\}$. We assume that, given $\Phi$,
the random variables $\{X_i-y_i\}$ are i.i.d random vectors with
$|X_i-y_i|=r$; i.e. each {\em receiver is at distance} $r$ from its transmitter.
Note that the receivers of different MANET nodes are different (almost surely).
\subsubsection{Nearest Receiver models}
In practice, some routing algorithm
specifies the receivers(s) (relay node(s)) of each given transmitter.
In what follows focus on two {\em nearest receiver models}
where each transmitter selects its receiver as close by as possible
is some set of {\em potential receivers} $\Phi_0$, common to all MANET nodes
$\Phi$; i.e. $y_i=Y_i^*=\mathop{\text{\rm arg~min}}_{Y_i\in\Phi_0, Y_i\not=X_i}\{|Y_i-X_i|\}$.
Two incarnations are considered
\paragraph{Independent Poisson Nearest Receiver (IPNR) Model}
In this model we assume that the potential receivers form some
stationary P.p.p. $\Phi_0$, of intensity $\lambda_0$,
which is {\em independent} of (and in
particular {\em external} to) the MANET $\Phi$.
\paragraph{MANET Nearest Neighbor (MNN) Model}
All the nodes of the MANET are considered as potential
receivers; i.e. $\Phi_0=\Phi$.~%
\footnote{Both IPNR and MNN models require some additional
specifications on what
happens if two or more transmitters pick the same receiver
and, the MNN model, what
happens if the picked receiver is also transmitting.
Our analysis applies to the situation when the SINR threshold
$T>1$ (cf.~\ref{ss.SINR}), which excludes multiple receptions by
a given receiver and
simultaneous emission and reception.}
\subsection{Mean Path-loss Model}
\label{ss.OPL}
Below, we assume that the receiver
of node $i$ receives a power from the transmitter located at
node $j$ at time $n$ which is equal to $F_i^j(n)/l(|X_j-y_i|)$,
where \hbox{$|\cdot|$} denotes
the Euclidean distance on the plane and $l(\cdot)$
is the path loss function.
An important special case consists in taking
\begin{equation}\label{simpl.att}
l(u)=(Au)^{\beta} \quad \text{for $A>0$ and $\beta>2$.}
\end{equation}
Other possible choices of path-loss function avoiding the
pole at $u=0$ consist in taking e.g. $\max(1,l(u))$,
$l(u+1)$, or $l(\max(u,u_0))$.
\subsection{SINR Coverage}
\label{ss.SINR}
In this paper we mainly focus on the SINR coverage/outage scenario:
we will say that transmitter $\{X_i\}$
{\em covers} (or {\em is successfully received by}) its receiver
$y_i$ at time slot $n$ if
\begin{equation}\label{eq:SINR}
\mbox{SINR}_i(n)=\frac{F_i^i(n)/l(|X_i-y_i|)}{W_i(n)+I_i^1(n)}\ge T\,,
\end{equation}
where $I_i^1(n)=\sum_{X_j\in\widetilde\Phi^1(n),\, j\not=i}
F_j^i(n)/l(|X_j-y_i|)$
is the {\em interference} at receiver $y_i$ at time $n$;
i.e., the sum of the signal powers
received by $y_i$ at time $n$ from all the nodes in
$\Phi^1(n)=\{X_j\in\Phi: e_j(n)=1\}$ except $X_i$.
In mathematical terms, $I_i^1(n)$ is
an instance of {\em Shot-Noise (SN)} field generated by
of $\widetilde\Phi^1(n)\setminus\{X_i\}$.
Denote by $\delta_i(n)$ the indicator that~(\ref{eq:SINR}) holds,
namely, that location $y_i$ is covered by transmitter $X_i$ with
the required quality at time $n$.
\subsection{Typical MANET node}
\label{ss.Palm}
Let ${\mathbf P}^0$ denote the Palm
distribution of the P.p.p. $\Phi$ (cf~\cite[Sec.~10.2.2]{FnT1}).
Under this distribution,
the MANET nodes are located at $\Phi\cup\{X_0=0\}$,
where $\Phi$ is a copy of the stationary P.p.p.
(cf Slivnyak's theorem; ~\cite[Th.~1.4.5]{SKM95}).
Under ${\mathbf P}^0$, the other random objects/marks of the model, $(e_i(n),\mathbf{F}_i(n))$ and $W_i(n)$ as well as
$y_i(n)$ in the Bipolar receiver model, are i.i.d. given
$\Phi\cup\{X_0=0\}$, and have the same law as their original distribution.
(For more details on Palm theory cf. e.g.~\cite[Sections~1.4, 2.1
and 10.2]{FnT1}.)
In the IPNR model, the potential receiver p.p. $\Phi_0$ remains independent of
$\Phi\cup\{X_0=0\}$ and Poisson-distributed; in the MNN model the receiver p.p.
is still determined by the MANET configuration $\Phi\cup\{X_0=0\}$ (with the receiver
of a node being its nearest neighbor).
The node $X_0=0$, considered under ${\mathbf P}^0$, is called {\em the typical MANET node}.
\section{Local Delay}
\setcounter{equation}{0}
\label{s.TheLD}
The {\em local delay of the typical node
is the number of time slots needed for node $X_0=0$
(considered under the Palm probability ${\mathbf P}^0$ with respect to $\Phi$)
to successfully transmit}:
$$\mathbf{L} =\mathbf{L} _0=\inf\{n\ge1: \delta_0(n)=1\}\,.$$
This random variable depends on the origin of time (here 1) but
we focus on its law below, which does not depend on the chosen
time origin.
The main objective of this paper is to study~${\mathbf E}^0[\mathbf{L} ]$ under the
full separation of time scales described in the introduction.
Let ${\mathcal{S}}$ denote all the {\em static elements of the network model}:
i.e. the elements which are random but which do not vary with time $n$.
In all models, we have all locations $\Phi,\{y_i\} \in{\mathcal{S}}$. Moreover, in the
slow fading model, we have $\{\mathbf{F}_i(n)=\mathbf{F}_i\}\in{\mathcal{S}}$ and similarly in
the slow noise model, $\{W_i(n)=W_i\}\in{\mathcal{S}}$.
Given a realization of all the elements of ${\mathcal{S}}$, denote by
\begin{equation}
\label{chalopic}
\pi_c({\mathcal{S}})={\mathbf E}^0[e_0(1)\delta_0(1)\,|\,{\mathcal{S}}]
\end{equation}
the conditional probability, given ${\mathcal{S}}$, that $X_0$
is authorized by the MAC to transmit and that this transmission is successful
at time $n=1$.
Note that due to our time-homogeneity, this conditional probability
does not depend on $n$. The following result allows us to express ${\mathbf E}^0[\mathbf{L} ]$.
\begin{Lemma}\label{l.delay-basic}
We have
\begin{equation}\label{e.delay-basic}
{\mathbf E}^0[\mathbf{L} _0]
={\mathbf E}^0\Bigl[\frac1{\pi_c({\mathcal{S}})}\Bigr]\,.
\end{equation}
\end{Lemma}
{\medskip\noindent \bf Remark:~\nolinebreak}
One can interpret $\pi_c({\mathcal{S}})$ as the {\em (temporal) rate of
successful packet transmissions} (or the {\em throughput})
of node $X_0$ given all the static elements of the network.
Its inverse $1/\pi_c({\mathcal{S}})$ is the local delay of this node in this environment.
In many cases, this throughput is a.s. positive (so that we will have
a.s. finite delays)
for all static environments. If this last condition holds true, by
Campbell's formula (cf~\cite[Eq.~(10.14)]{FnT1}),
almost surely, all the nodes have
finite mean delays and positive throughputs. However, the spatial
irregularities of the network imply that this throughput varies from
node to node, and in a Poisson configuration, one can find
nodes which have an arbitrarily small throughput (and consequently an arbitrarily large
delay). The mean local delay of the typical node
${\mathbf E}^0[\mathbf{L} ]$ is the spatial average
of these individual mean local delays. A finite mean indicates that the
fraction of nodes in bad shape (for throughput or delay)
is in some sense not significant. In contrast, ${\mathbf E}^0[\mathbf{L} ]=\infty$
indicates that an important fraction of the nodes are in
a bad shape. This is why the finiteness of the mean local delay of the
typical node is an
important indicator of a good performance of the network.
\begin{proof}(of Lemma~\ref{l.delay-basic})
Since the elements that are not in ${\mathcal{S}}$ change over time in the
i.i.d. manner given a realization of the elements of ${\mathcal{S}}$,
the successive attempts of node $X_0$ to access to
the channel and successfully transmit at time $n\ge 1$ are independent (Bernoulli)
trials with probability of success $\pi_c({\mathcal{S}})$.
The local delay $\mathbf{L} =\mathbf{L} _0$ is then a geometric random variable (the number of trials
until the first success in the sequence of Bernoulli trials) with parameter $\pi_c({\mathcal{S}})$.
Its (conditional) expectation (given ${\mathcal{S}}$) is known to be
${\mathbf E}^0[\mathbf{L} \,|\,{\mathcal{S}}]=1/\pi_c({\mathcal{S}})$.
The result follows by integration with respect to the
distribution of ${\mathcal{S}}$.
\end{proof}
\begin{Exe}\label{exe.basic}
In order to understand the reasons for which ${\mathbf E}^0[\mathbf{L} ]$ may or
may not be finite, consider the following two extremal situations.
Suppose first that the whole network is independently re-sampled
at each time slot (including node locations
$\Phi$, which is {\em not} our default option).
Then ${\mathcal{S}}$ is empty (the $\sigma$-algebra generated by it is
trivial) and the temporal rate of successful transmissions is equal to
the space-time average rate $\pi_c({\mathcal{S}})={\mathbf E}^0[e_0(1)\delta_0(1)]=p\,p_c$,
where $p_c$ is the space-time probability of coverage
for the typical node in the corresponding
receiver model. Consequently, in this case of extreme variability (w.r.t. time),
we have ${\mathbf E}^0[\mathbf{L} ]=1/p_c<\infty$ provided $p_c>0$,
which holds true under very mild assumptions, e.g. for all considered
receiver and path-loss models with fast or slow Rayleigh fading model
and any noise model provided, $0<p<1$ (cf.~\cite{jsac} for the
evaluation of $p_c$ for the Poisson
Bipolar model).
On the other hand, if nothing varies over time (including MAC status,
which again is ruled out in our general assumptions), we
have $\pi_c({\mathcal{S}})=e_0(1)\delta_0(1)$ (because the conditioning on
${\mathcal{S}}$ determines $e_0(1)\delta_0(1)$ in this case).
In this case under very mild
assumptions (e.g. if $p<1$), this temporal rate
$e_0(1)\delta_0(1)$ is zero with positive probability, making ${\mathbf E}^0[\mathbf{L} ]=\infty$.
Note that in this last case, some nodes in the MANET
succeed in transmitting packets every time slot, whereas
others never succeed.
Having seen the above two extremal cases, it is not difficult to
understand that the mean local delay of the typical node very much depends on
how much the time-variability ``averages out'' the spatial
irregularities of the distribution of nodes in the MANET.
\end{Exe}
Note that by Jensen's inequality,
$${\mathbf E}^0[\mathbf{L} ]\ge\frac1{{\mathbf E}^0[\pi_c({\mathcal{S}})]}=\frac1{p_c}\,.$$
The inequality is in general strict and we may have
\hbox{${\mathbf E}^0[\mathbf{L} ]=\infty$} while $p_c>0$.
In the remaining part of this section we will study several particular
instances of space-time scenarios.
\subsection{Local Delays in Poisson Bipolar Model}
In the Poisson bipolar model, we assume a static repartition for the MANET
nodes $\Phi$ and for their receivers $\{y_i\}$. The MAC variables
$e_i(n)$ are i.i.d. in $i$ and $n$. All other elements
(fading and noise) have different time-scenarios.
\subsubsection{Slow Fading and Noise Case}
Let us first consider the situation where $\{\mathbf{F}_i\}$ and $W$ are static.
\begin{Prop}
Assume the Poisson Bipolar network model with slow fading and slow noise.
If the distribution of $F,W$ is such that
$\Pr\{\,WTl(r)>F\,\}>0$, then ${\mathbf E}^0[\mathbf{L} ]=\infty$.
\end{Prop}
\begin{proof}
We have
\begin{eqnarray*}
\pi_c({\mathcal{S}})&=&p{\mathbf E}^0[e_0(1)\delta_0(1)\,|\,
{\mathcal{S}}]\\
&=&p\Pr^0\{F^0_0\ge Tl(r)(W_0+I_0^1(0))\,|\,{\mathcal{S}}\}\\
&\le&p1\hspace{-0.30em}{\mbox{I}}(F^0_0\ge Tl(r)W)\,.
\end{eqnarray*}
The last indicator is equal to 0 with non-null probability
for our assumptions. Using~(\ref{e.delay-basic}), we conclude that
${\mathbf E}^0[\mathbf{L} ]=\infty$.
\end{proof}
\subsubsection{Fast Fading Case}
The following auxiliary result is useful when studying fast
Rayleigh fading.
\begin{Lemma}\label{l.SN-LT-condtional-inverse}
Consider the Poisson shot-noise $I=\sum_{X_i\in\Phi}G_i/l(|X_i|)$,
where $\Phi$ is some homogeneous Poisson p.p. with
intensity $\alpha$ on ${\mathbb{R}}^2$,
$G_i$ are i.i.d. random variables with Laplace transform
${\mathcal{L}}_G(\xi)$ and $l(r)$ is any response function (in our case it
is always some path-loss function).
Denote by ${\mathcal{L}}_I(\xi\,|\,\Phi)={\mathbf E}[e^{-\xi I}\,|\,\Phi]$ the
conditional Laplace transform of $I$ given $\Phi$. Then
$${\mathbf E}\Bigl[\frac1{{\mathcal{L}}_{I}(\xi\,|\,\Phi)}\Bigr]=
\exp\biggl\{-2\pi\alpha\int_0^\infty
v\Bigl(1-\frac1{{\mathcal{L}}_G(\xi/l(v))}\Bigr)\,\text{\rm d} v\biggr\}\,.$$
\end{Lemma}
\begin{proof}
By the independence of $G_i$ given $\Phi$, we have
\begin{eqnarray*}
{\mathbf E}[e^{-\xi I}\,|\,\Phi]&=&
\prod_{X_i\in\Phi}{\mathbf E}[e^{-\xi{\mathcal{L}}_G(\xi/l(|X_i|))}]\\
&=&\prod_{X_i\in\Phi}{\mathcal{L}}_G(\xi/l(|X_i|))\\
&=&\exp\Bigl\{\sum_{X_i\in\Phi}\log({\mathcal{L}}_G(\xi/l(|X_i|)))\Bigr\}\,.
\end{eqnarray*}
Taking the inverse of the last expression and using
the known formula for the Laplace transform of the Poisson p.p.
(it can be derived
from the formula for the Laplace functional of the Poisson p.p.;
see e.g.~\cite{daley}, and was already used e.g.~\cite{jsac}).
we obtain
\begin{equation}\label{e.SN-LT-condtional-inverse-nonhomo}
{\mathbf E}\Bigl[\frac1{{\mathcal{L}}_{I}(\xi\,|\,\Phi)}\Bigr]=
\exp\biggl\{-\alpha\int_{{\mathbb{R}}^2}
\Bigl(1-e^{-\log({\mathcal{L}}_G(\xi/l(|x|)))}\Bigr)\,\text{\rm d} x\biggr\}\,.
\end{equation}
Passing to polar coordinates completes the proof.
\end{proof}
Coming back to local delays,
let us consider now the situation where the random variables $\{\mathbf{F}_i(n)\}$
are i.i.d. in $n$. We consider only the Rayleigh fading case.
\begin{Prop}\label{p.delay-bipolar-fast-fading}
Assume the Poisson Bipolar network model with fast Rayleigh fading.
In the case of fast thermal noise, we have
$${\mathbf E}^0[\mathbf{L}]=
\frac1{p}
{\mathcal{D}}_W(Tl(r))
\exp\Bigl\{2\pi p\lambda\int_0^\infty
\frac{vTl(r)}{l(v)+(1-p)Tl(r)}\,\text{\rm d} v\Bigr\}\,,
$$
where
\begin{itemize}
\item ${\mathcal{D}}_W(s)={\mathcal{D}}_W^{slow}(s)={\mathcal{L}}_W(-s)$ for the slow noise
case,
\item ${\mathcal{D}}_W(s)={\mathcal{D}}_W^{fast}(s)=1/{\mathcal{L}}_W(s)$ for the fast noise
case.
\end{itemize}
\end{Prop}
\begin{proof}
In the fast Rayleigh fading case, we have
$\pi_c({\mathcal{S}})=\Pr\{F\ge Tl(r)(W+I^1)\,|\,\Phi\}$
for the fast noise case and $\pi_c({\mathcal{S}})=\Pr\{F\ge
Tl(r)(W+I^1)\,|\,\Phi,W\}$
for the slow noise model.
Using the assumption on $F$, we obtain
$$\pi_c({\mathcal{S}})={\mathcal{L}}_W(\mu Tl(r)){\mathbf E}[e^{-\mu Tl(r) I^1}\,|\,\Phi]$$
in the fast noise case
and
$$\pi_c({\mathcal{S}})=e^{-\mu WTl(r)}{\mathbf E}[e^{-\mu Tl(r) I^1}\,|\,\Phi]$$
for the slow noise case.
The result then follows from~(\ref{e.delay-basic})
and Lemma~\ref{l.SN-LT-condtional-inverse}
with $G=eF$. Note that in this case
${\mathcal{L}}_G(\xi)=1-p+p{\mathcal{L}}_F(\xi)$, which gives
${\mathcal{L}}_{eF}(\xi)=1-p+p\mu/(\mu+\xi)$.
\end{proof}
\subsection{Local Delay in the Nearest Receiver Models}
\label{ss.LD-NRM}
In this section we study the IPNR and the MNN receiver models.
We work out formulas for the mean local delay of the typical node
under the following conditions: {\em fast Rayleigh fading}
and fast or slow noise of arbitrary distribution.
\begin{Prop}
\label{p:delay-NN-main}
Assume fast Rayleigh fading.
\begin{itemize}
\item In the IPNR model, we have
\begin{equation}\label{e.delay-INR-main}
\!\!\!\!{\mathbf E}^0[\mathbf{L} ]\!=\!\frac{2\pi\lambda_0}{p}\!\int_0^\infty \!\!\!re^{-\pi\lambda_0r^2}
{\mathcal{D}}_W(\mu Tl(r))\,{\mathcal{D}}^{INR}_I(\mu Tl(r))\text{\rm d} r
\end{equation}
where
\begin{equation}\label{eq:ch9basicld}
{\mathcal{D}}_{I}^{INR}(s)=
\exp\left\{2\pi\lambda \int_0^\infty \frac{ps}{l(v)+ (1-p)s} v\, \text{\rm d}
v\right\}
\end{equation}
and ${\mathcal{D}}_W(s)$ is as in
Proposition~\ref{p.delay-bipolar-fast-fading}.
\item In the MNN model, we have
\begin{eqnarray}\label{e.delay-MNN-main}
\lefteqn{{\mathbf E}^0[\mathbf{L} ]=\frac{2\pi\lambda}{p(1-p)}}\\[0.5ex]
&&\times\int_0^\infty re^{-\pi\lambda r^2}
{\mathcal{D}}_W(\mu Tl(r))\,{\mathcal{D}}^{MNN}_I(r,\mu Tl(r))\,\,\text{\rm d} r\,, \nonumber
\end{eqnarray}
where
\begin{eqnarray}\label{eq:ch9basiclddsm}
\lefteqn{{\mathcal{D}}_{I}^{MNN}(r,s)=\exp \biggl\{
\lambda\pi \int_0^\infty \frac{ps}{l(v)+ (1-p)s} v\,\text{\rm d} v
}\nonumber\\
&&\hspace{-3em}+\lambda
\int_{\theta=-\frac \pi 2}^{\frac \pi 2}
\int_{v>2r \cos \theta }
\frac{ps}{l(v)+ (1-p)s} v\, \text{\rm d} v \text{\rm d} \theta
\biggr\}
\end{eqnarray}
and ${\mathcal{D}}_I(s)$ is as above (as in
Proposition~\ref{p.delay-bipolar-fast-fading}).
\end{itemize}
\end{Prop}
\begin{proof}
We condition on the location of the nearest
neighbor $y_0=Y^*_0$ of $X_0=0$ under ${\mathbf P}^0$.
In the IPNR model the distance from the origin to $Y_0^*$ is known to
have the following distribution (under both the Palm and the stationary law):
${\mathbf P}\{|Y_0^*|>r\}=e^{-\pi\lambda_0 r^2}$. Due to the
independence assumption, in this model, given the location of the receiver $Y_0^*$,
the distribution of the MANET nodes $\Phi$ remains unchanged under ${\mathbf P}^0$.
The remaining part of the proof follows the same lines as that of
Proposition~\ref{p.delay-bipolar-fast-fading}.
For the MNN model, recall from Section~\ref{ss.Palm} that under
${\mathbf P}^0$, the nodes of $\Phi\setminus\{X_0\}$
are distributed as those of the homogeneous Poisson p.p.
Thus the distance $|Y^*_0-X_0|=|Y^*_0|$ has the same distribution as in
the IPNR model with $\lambda_0=\lambda$.
However, in the MNN model, given some particular location of $y_0=Y^*_0$,
one has to take the following fact into account: there are
{\em no MANET nodes} (thus, in particular, no interferers)
in $B_0(|y_0|)$. Consequently, under ${\mathbf P}^0$, given $Y_0^*=y_0$,
the SN $I^1_0$ in~(\ref{eq:SINR}) is no longer
driven by the stationary Poisson p.p. of intensity $\lambda_1$,
but as the SN of $\Phi^1$ given that there are no nodes of $\Phi$ in
$B_0(|y_0|)$. Note that the location $y_0$ at which we evaluate this
last SN is on the boundary (and not in the center)
of the empty ball. By the strong Markov property of Poisson p.p.
(cf.~\cite{Zuyev2006}), the
distribution of a Poisson p.p. given that $B_0(|y_0|)$ is empty
is equal to the distribution of the (non-homogeneous)
Poisson p.p. with intensity equal to 0 in $B_0(|y_0|)$ and
$\lambda_1$ outside this ball. Putting these arguments together,
and exploiting the rotation invariance of the picture conclude the proof.
\end{proof}
Notice that the integrals in (\ref{eq:ch9basicld})
and (\ref{eq:ch9basiclddsm}) are finite for any of the path-loss
models suggested in Section~\ref{ss.OPL}.
However, the outer integrals (in $r$) in~(\ref{e.delay-INR-main})
and~(\ref{e.delay-MNN-main}) may be infinite. In order to study this
problem note first that we have the following bounds in the MNN model:
\begin{Remark}
\label{remal:dsmbounds}
In the MANET receiver case, we have the bounds
\begin{equation}
\label{eq:ch9basiclddsm-bound}
\Bigl({\mathcal{D}}_{I}^{INR}(s)\Bigr)^{1/2}
\le {\mathcal{D}}_I^{MNN}(r,s)\le
{\mathcal{D}}_{I}^{INR}(s)\,.
\end{equation}
\end{Remark}
\section{Wireless contention phase transition for the local
delay}
\label{s.Phase-Transition}
In this section we will show that under quite natural assumptions
them mean local delay of the typical node can be infinite.
For some models, it can exhibit the following phase transition:
${\mathbf E}^0[\mathbf{L} ]<\infty$ or ${\mathbf E}^0[\mathbf{L} ]=\infty$ depending on the
model parameters (as $p$, distance $r$ to the receiver, or the mean fading $1/\mu$).
\subsection{Bipolar Model}
We begin with the simple Bipolar receiver model.
Proposition~\ref{p.delay-bipolar-fast-fading} shows
that in the fast fading and noise case, ${\mathbf E}^0[\mathbf{L} ]<\infty$;
indeed, $\int_0^\infty v/l(v)\,\text{\rm d} v<\infty$.
However for the fast fading, slow noise case
the finiteness of the mean local delay of the typical node
depends on whether $W$ has finite
{\em exponential moments} of order $Tl(r)\mu$.
This is a rather strong assumption concerning the tail distribution
function of $W$. Often this moment is finite only for
some sufficiently small value of $Tl(r)\mu$.
To see the wireless contention phase transition in this model consider
the following example.
\begin{Exe}
\label{R:wireless-contention-phase-trans-bipolar}
Let us assume exponential noise with mean $1/\nu$.
Then ${\mathcal{L}}_W(-\xi)=\nu/(\nu-\xi)<\infty$ provided
$Tl(r)\mu<\nu$ and infinite for $Tl(r)\mu>\nu$.
This means that in the corresponding Poisson Bipolar MANET
with a Rayleigh fading, exponential noise,
the mean local delay
of the typical node is finite whenever $Tl(r)<\nu/\mu$ and infinite otherwise.
Here are a few incarnations of this phase transition:
\begin{itemize}
\item For fixed mean transmission power $\mu^{-1}$ (we recall that a typical
situation is that where fading has mean 1 and where $\mu^{-1}$ is actually
the effective transmission power) and mean
thermal noise $\nu^{-1}$, there is a threshold
on the distance $r$ between transmitter and receiver below which
mean local delay of the typical node is finite and above which they
are infinite;
\item For fixed mean thermal noise $\nu^{-1}$ and fixed distance $r$,
there is a threshold on mean transmission power $\mu^{-1}$ {\em above}
which the mean local delay of the typical node is finite and {\em
below} which it is infinite;
\item For fixed mean transmission power $\mu^{-1}$ and fixed distance $r$,
there is a threshold on mean thermal noise power $\nu^{-1}$ {\em below}
which the mean local delay of the typical node
is finite and {\em above} which it is infinite.
\end{itemize}
The fact that all transmissions contend for the shared wireless channel may lead
to infinite mean local delays of the typical node
if the system is stressed by either of the
phenomena listed above: too distant links, a too high thermal noise or a too
transmission power.
\end{Exe}
\begin{Remark}{\em (RESTART)}\label{R.restart}
There is a direct interpretation of the
local delay of the typical node in terms of the so called {\em RESTART
algorithm}
\index{RESTART algorithm}
assume a file of random size $B$ is to be transmitted over
an error prone channel. Let $\{A_n\}_{n\ge1}$ be the sequence
of channel inter-failure times.
If $A_1>B$ (resp. $A_1\le B$), the transmission succeeds
(resp. fails) at the first attempt.
If the transmission fails at the first attempt, one has
to restart the whole file transmission in the second attempt and so on. Let
$$ N= \inf \{n\ge 1 \ s.t.\ A_n>B\}$$
be the first attempt where the file is successfully transmitted.
In the classical RESTART scheme, the sequence $\{A_n\}_{n>0}$
is assumed to be i.i.d. and independent of $B$.
It can then be proved (see \cite{Asmussen}) that when $B$
has infinite support and $A_n$ is light tailed (say exponential), then $N$
is heavy tailed. This observation comes as a surprise because one
can get heavy tails (including infinite first moments) in situations
where $B$ and $A_n$ are both light tailed.
Consider the fast fading, slow noise case (and ignore the
interference for simplicity)
Then the local delay of the typical node can be seen as an instance of this
algorithm with the following identification:
$A_n=F_0^0(n)e_0(n)$ and $B=TW l(r)$.
In the next section,
we will see other incarnation of the above RESTART algorithm
in the nearest-receiver models with
deterministic $W$, where the role of the unboundedness of $B$ is
played by the distance to the receiver;
cf.~Remark~\ref{R.restart-cont}.
The above interpretation in terms of the RESTART scheme formally shows
that the local delay of the typical node is heavy tailed
and thus it is not surprising that in certain cases its mean is infinite.
However, the physical phenomena at hand are quite different here and
in the classical RESTART context, so
let us now comment on what exactly this heavy tailedness means in our
MANET context.
Let $B_0(R)$ be the ball centered at $0$ of radius $R$ and let
$\mathbf{L}_i$ be the number of time slots required by the node $X_i$ to
transmit a packet.
The ergodic interpretation of the Palm probability
implies that, for all $m$,
$${\mathbf P}^0\{\,\mathbf{L}_0>m\,\} =\lim_{R\to \infty} \frac 1 {\Phi(B_0(R))}
\sum_{X_i\in B_0(R)}1\hspace{-0.30em}{\mbox{I}}(\mathbf{L}_i>m),$$
where the last limit is in the almost sure sense.
The fact that the distribution of $\mathbf{L}_0$ is heavy tailed under
${\mathbf P}^0$ means that the (discrete) law ${\mathbf P}^0\{\,\mathbf{L}_0>m\,\}$
has no exponential moments. In view of the above ergodic
interpretation, this is equivalent to saying that the asymptotic
fraction of MANET modes which experience a local delay
of more than $m$ time slots decreases slowly with $m$ (more slowly than any
exponential function).
Finally, we remark that the fact that the {\em mean local delay of
the typical node
is infinite does not imply that the mean throughput of the typical node
is null}. The last quantity boils down to the probability
of success of node $X_0$ under ${\mathbf P}^0$, i.e., to $p_c$ and,
as already mentioned in Example~\ref{exe.basic}, it is positive for all
the considered models.
\end{Remark}
\subsection{Nearest Receiver Models}
In order to analyze mean local delays in these more complex receiver
models will study separately the impact of the thermal noise and
of the interference.
\subsubsection{Noise Limited Networks}
\label{sss:delay-noise-limitted-NN}
Consider first the IPNR and MNN models under the assumption that
the interference is perfectly canceled
(and that only noise has to be taken into account).
In what follows we consider the fast noise scenario.
\paragraph{IPNR model}
The following result follows from
(\ref{e.delay-INR-main}) with ${\mathcal{D}}_I(s)=1$
and with ${\mathcal{D}}_W$ given in
Proposition~\ref{p.delay-bipolar-fast-fading}:
\begin{Cor}
In the IPNR model with fast Rayleigh fading and fast noise,
if interference is perfectly canceled, then
\begin{equation*}
{\mathbf E}^0[\mathbf{L}] = 2\pi \lambda_0 \int_0^\infty
\frac {r \exp(-\pi \lambda_0 r^2)} {p {\cal L}_W(\mu l(r)T)} \text{\rm d} r.
\end{equation*}
\end{Cor}
Hence, for the simplified path loss function~(\ref{simpl.att})
${\mathbf E}^0[\mathbf{L}]<\infty$ whenever
\begin{equation}
\label{eq:oppsufcondnoisenndir}
{\cal L}_W(\xi) \ge
\eta \exp\left\{-\pi \lambda_0 \left(\frac{\xi}{ \mu T A^{\beta}}\right)^{2/\beta}\right\}
\left(\xi^{2(1+\epsilon)/\beta}\right),
\ \xi\to \infty,
\end{equation}
for some positive constants $\epsilon$ and $\eta$,
and whenever some natural local integrability conditions also hold.
This condition requires that there be a sufficient probability mass of $W$
in the neighborhood of 0. For instance,
under any of the path-loss models suggested in Section~\ref{ss.OPL}, this holds true for a thermal
noise with a rational Laplace transform (e.g. Rayleigh).
The condition~(\ref{eq:oppsufcondnoisenndir}) is sharp in the sense
that when ${\cal L}_W(\xi)$ is
asymptotically smaller that the expression in the right-hand-side
of~(\ref{eq:oppsufcondnoisenndir}) with $(1+\epsilon)$ replaced by
$(1-\epsilon)$
for some positive constants $\epsilon$ and $\eta$,
then ${\mathbf E}^0[\mathbf{L}]=\infty$. This is the case, e.g. when $W$ is a
positive constant.
\paragraph{MNN model}
In the MNN model (with $0<p<1$) and the simplified path loss
function~(\ref{simpl.att}), similar arguments show that
the same threshold as above holds with
\begin{equation}
\label{eq:oppsufcondnoisennds}
{\cal L}_W(\xi) \ge
\eta \exp\left\{-\pi \lambda \left(\frac{\xi}{ \mu T A^{\beta}}\right)^{2/\beta}\right\}
\left(\xi^{2(1+\epsilon)/\beta}\right),
\ \xi\to \infty,
\end{equation}
implying that ${\mathbf E}^0[\mathbf{L}]<\infty$ and a similar converse statement.
\subsubsection{Interference Limited Networks}
\label{sss:delay-interf-limitted-NN}
\index{network!interference limited}
In this section we assume and $W\equiv 0$.
\paragraph{IPNR model}
In the IPNR case with the simplified path loss
function~(\ref{simpl.att}),
using the fact that
\begin{equation}\label{e.PINR-interf}
2 \pi \int_0^\infty \frac{pT l(r)}{l(v)+ (1-p)T l(r)} v\, \text{\rm d} v = p
(1-p)^{\frac 2 \beta-1} T^{\frac 2 \beta} K(\beta) r^2,
\end{equation}
with
\begin{equation}
\label{eq:kdebeta}
K(\beta)=\frac{2\pi\Gamma(2/\beta)\Gamma(1-2/\beta)}{\beta}=
\frac{2\pi^2}{\beta\sin(2\pi/\beta)}\, .
\end{equation}
Using the above observations we get the following result
from~(\ref{e.delay-INR-main})
and~(\ref{eq:ch9basicld}):
\begin{Cor}\label{c:PINR-delay-interflim}
In the IPNR model with $W=0$, fast Rayleigh fading and the
path loss function~(\ref{simpl.att}), we have
\begin{equation*}
{\mathbf E}^0[\mathbf{L}] = 2\pi \lambda_0 \frac 1 p
\int_0^\infty r \exp\left(-\pi \lambda_0 r^2 + \lambda \theta(p,T,\beta) r^2\right) \text{\rm d} r\, ,
\end{equation*}
with
\begin{equation}\label{e:theta-delay}
\theta(p,T,\beta)= \frac p {(1-p)^{1 -\frac 2 \beta}}
T^{\frac 2 \beta} K(\beta) \,.
\end{equation}
\end{Cor}
Notice that $\theta(p,T,\beta)$ is increasing in $p$ and in $T$.
We hence get the following {\em incarnation of the wireless contention phase transition}:
\begin{itemize}
\item If $p\ne 0$ and $\lambda_0 \pi > \lambda \theta(p,T,\beta)$, then
\begin{eqnarray*}
{\mathbf E}^0[\mathbf{L}]&=&
\frac 1 p
\frac{\pi \lambda_0 }{\pi \lambda_0- \lambda \theta(p,T,\beta)} \\
&=&\frac 1 p
\frac{\lambda_0 }{\lambda_0- \lambda
\frac 2 \beta \Gamma(\frac 2 \beta)\Gamma(1-\frac 2 \beta)
p(1-p)^{2/\beta-1} T^{2/\beta}}\\
&<& \infty\, .
\end{eqnarray*}
\item If either $\lambda_0 \pi < \lambda \theta(p,T,\beta)$ or $p=0$,
then ${\mathbf E}^0[\mathbf{L}] =\infty$.
\end{itemize}
\begin{Remark}{\em (Pole of the path-loss function)}
The above phase transition is not linked to the pole of the
simplified path loss function~(\ref{simpl.att}) used in the analysis.
To show this, one can consider e.g. $l(u)=(\max(1,u))^4$ and
evaluate explicitly the integral in the left-hand-side
of~(\ref{e.PINR-interf})
(we skip the details due to the lack of space) and conclude that
that the mean delay is finite if $p\ne 0$ and
$$ \lambda_0> \lambda \hat \theta= \lambda \frac{\pi}{2}\frac{p}{\sqrt{1-p}} \sqrt{T} $$
and infinite if either $p=0$ or $ \lambda_0< \lambda \hat \theta$.
\end{Remark}
\begin{Remark}
Here are a few comments on this phase transition.%
\index{phase transition!wireless contention}
\begin{itemize}
\item The fact that $p=0$ ought to be avoided for having ${\mathbf E}^0[\mathbf{L}] <\infty$
is clear;
\item The fact that intensity of potential transmitters
$\lambda_0$ cannot be arbitrarily small when the other
parameters are fixed is clear too as this implies that:
(i)~at any given time slot,
the transmitters compete for too small set of receivers;
(ii)~targeted receivers are too far away from their transmitter.
\item For $T$ and $\beta$ fixed, stability (${\mathbf E}^0[\mathbf{L}]<\infty$) requires that
receivers outnumber potential transmitters by a factor which grows like
$p(1-p)^{2/\beta-1}$ when $p$ varies; if this condition is not satisfied,
this drives the system to instability because
some receivers have too persistent interferers nearby (for instance, if $p=1$,
a receiver may be very close from a persistent transmitter which will
most often succeed, forbidding (or making less likely) the success of
any other transmitter which has the very same receiver).
\end{itemize}
\end{Remark}
\paragraph{MNN model}
Fix $a,r\ge0$. For the path loss function~(\ref{simpl.att}) we have
\begin{eqnarray*}
\lefteqn{
\int_{ar}^\infty \frac{pT l(r)}{l(v)+ (1-p)T l(r)} v \,\text{\rm d} v}\\
&=&\frac 1 2 r^2 p (1-p)^{\frac 2 {\beta} -1} T^{\frac 2
{\beta}}H(a,T(1-p),\beta/2)\, ,
\end{eqnarray*}
with
\begin{equation}
H(a,w,b)
= \int_{a^2 w^{-1/b}}^\infty \frac 1 {1+u^{b}} \,\text{\rm d} u.
\end{equation}
Let
\begin{equation}
J(w,b) = \int_{\theta=-\pi/2}^{\pi/2}
H(2\cos(\theta),w,b) \,\text{\rm d} \theta\, .
\end{equation}
From~(\ref{e.delay-MNN-main}) and~(\ref{eq:ch9basiclddsm}),
we then get the same type of phase transitions as for the IPNR
model above:
\begin{itemize}
\item If $p\ne 0$
\begin{equation}
\frac p{(1-p)^{1- \frac 2 \beta}}
T^{\frac 2 {\beta}} \left( \frac{K(\beta)} {2} + J(T(1-p),\frac \beta 2)\right)
< \pi,\end{equation}
then ${\mathbf E}^0[\mathbf{L}] <\infty$;
\item If
\begin{equation}
\label{eqalo:dsminstab}
\frac p{(1-p)^{1- \frac 2 \beta}}
T^{\frac 2 {\beta}} \left(\frac{K(\beta)} {2} + J(T(1-p),\frac \beta 2)\right)
> \pi,
\end{equation}
then ${\mathbf E}^0[\mathbf{L}] =\infty$.
\end{itemize}
We can use the bounds of Remark~\ref{remal:dsmbounds} to get the
following and simpler conditions:
\begin{itemize}
\item If $p\ne 0$ and
$\theta (p,T,\beta) < \pi$
then ${\mathbf E}^0[\mathbf{L}] < \infty$.
\item If
$\theta (p,T,\beta) > 2\pi\,$
then ${\mathbf E}^0[\mathbf{L}] = \infty$.
\end{itemize}
\begin{Remark}{\em (RESTART, cont.)}\label{R.restart-cont}
We continue the analogy with the RESTART algorithm described in Remark~\ref{R.restart}.
\index{RESTART algorithm}
Consider the fast Rayleigh fading, with slow, constant noise $W=Const$
in the context of one of the nearest receiver models.
Then the local delay time of a packet can be seen as an instance of this
algorithm with the following identification:
$A_n=F^0_0(n)e_0(n)$ and $B=TW l({\cal D})$, where ${\cal D}$ is the (random) distance between
the node where the packet is located and the target receiver.
The support of $ l({\cal D})$ is unbounded (for instance, in the
Poisson receiver
model, for all the considered path-loss models, the density of ${\cal D}$ at $r>0$ is
$\exp(-\lambda_0\pi r^2) r$ for $r$ large).
The interference limited case can be seen as an extension of the
RESTART algorithm where the file size varies over time. More
precisely, the model corresponding to e.g. MNN is that where at attempt $n$,
the file size is $B_n=f(\Phi,C_n)$, where $\Phi$ is the Poisson p.p. and
$\{C_n\}_{n>0}$ is an independent i.i.d. sequence (here $C_n$
is the set of fading variables and MAC decisions at time $n$).
\end{Remark}
\section{Finite Mean Delays and Diversity}
\label{s.FiniteMean}
As we saw in the last subsection, the existence
of big void regions as found in Poisson configurations leads
to the surprising property in the nearest receiver models, that the
mean (time) local delay is finite everywhere
but may have an infinite spatial average in rather classical scenarios.
We describe below a few ways of getting finite mean spatial average of
the mean local delay
in fast fading scenarios. All the proposed methods rely on an
increase of {\em diversity}: more variability in fading, more
receivers, more mobility, more flexible (adaptive) coding schemes.
\index{diversity}
\subsection{Heavy Tailed Fading}
\paragraph{Weibull Fading}
Assume the path loss function~(\ref{simpl.att}) and deterministic $W>0$,
Recall from the discussion after
condition~(\ref{eq:oppsufcondnoisenndir})
(cf. also~(\ref{eq:oppsufcondnoisennds}) that in this case
the mean local delay of the typical node
is infinite in the IPNR and MNN model (due to the
noise constraint)
if one has the (fast) Rayleigh fading.
However if we assume that $F$ is Weibull of shape parameter $k$ i.e.
${\mathbf P}[F>x]=\exp(-(x/c)^k)$,
for some $c$, with $c$ and $k$ positive constants, then
the condition $k<2/\beta$ is sufficient
to have ${\mathbf E}^0[\mathbf{L}] <\infty$ in the noise limited scenario. Indeed then
${\mathbf P}(F>l(r)WT)= \exp\left\{-(l(r)TW/c)^k\right\}\ge
\exp\left\{-(TW/c)^k (Ar)^{2-\epsilon}\right\}$,
for $r\ge 1/A$, and some $\epsilon>0$.
Therefore the finiteness of ${\mathbf E}^0[\mathbf{L}]$ (with canceled
interference)
follows from the fact that
the integral
$$
\int_{1/A}^\infty r \exp\left\{-\pi \lambda_0 r^2 +
(TW/c)^k (Ar)^{2-\epsilon} \right\} \,\text{\rm d} r
$$
is finite.
\paragraph{Lognormal Fading}
Assume now
$F$ is lognormal with parameters $(\mu,\sigma)$,%
\index{fading!lognormal}
that is $\log (F)$ is ${\cal N}(\mu,\sigma^2)$ (Gaussian of mean $\mu$
and variance $\sigma^2$) and that
$W$ is constant. Using
$$
{\mathbf P}(F>x)
\sim
\frac 1 {(\log(x)-\mu)/\sigma}
\exp\left(-(\log(x)-\mu)^2/2\sigma^2\right),
$$
when $x\to \infty$ and
$$ {\mathbf P}(F>x) \ge \frac {(\log(x)-\mu)/\sigma} {1 +(\log(x)-\mu)^2/\sigma^2}
\exp\left(-(\log(x)-\mu)^2/2\sigma^2\right)$$
for all $x>0$
one can show that ${\mathbf E}^0[\mathbf{L}]<\infty$ in the noise-limited scenario
for IPNR and MNN model
with the path loss function~(\ref{simpl.att})
(we skip the details).
Let us conclude that a fading with heavier tails may be useful {\em
in the noise limited scenario} in that the
mean delay of the typical node may be infinite for the Rayleigh case and finite
for this heavier tailed case.
\subsubsection{Networks with an Additional Periodic Infrastructure}
\index{network!with periodic infrastructure}
The second line of thoughts is based on the idea that extra receiver should be added
to fill in big void regions.
We assume again fast Rayleigh fading
in conjunction with the
``Poisson + periodic'' independent receiver model.
In this receiver model we assume that the pattern of potential
receiver consists of Poisson p.p. and
an additional periodic infrastructure.
Since there is a receiver at distance at most, say, $\kappa$
from every point,
the closest receiver from the origin is at a distance at most $\kappa$ and
$$ {\mathbf E}^0[\mathbf{L}] = \int_0^\kappa {\mathcal{D}}_W(\mu Tl(r))\,
{\mathcal{D}}_I^{INR}(Tl(r))\, D(\text{\rm d} r)\,,$$
where ${\mathcal{D}}_W$, ${\mathcal{D}}_I^{INR}$ are as in
Proposition~\ref{p:delay-NN-main}
and $D(\cdot)$ is the distribution function of the
distance from the origin to the nearest receiver in
this model.
This latter integral is obviously finite.
Notice that periodicity is not required here. The only important property
is that each location of the plane has a node at a distance which is
upper bounded by a constant.
\subsubsection{High Mobility Networks}
\label{secalo:HMN}
It was already mentioned in Example~\ref{exe.basic}
that if one can assume that the whole network is independently re-sampled
at each time slot (including node locations
$\Phi$, which is {\em not} our default option) --- an assumption which
can be justified when there is a high mobility of nodes
---
then ${\mathbf E}^0[\mathbf{L} ]=1/p_c<\infty$ provided $p_c>0$.
This observation can be refined in at least two ways:
\begin{itemize}
\item Assume the IPNR model, with fixed potential receives,
and high mobility of MANETS nodes, i.e., with $\Phi=\Phi(n)$
i.i.d. re-sampled at each $n\ge1$. Assume also fast noise and
fading.
Then one can easily argue that
\vbox{\begin{eqnarray*}
\lefteqn{{\mathbf E}^0[\mathbf{L} ]}\\
&&\hspace{-3em} =
2\pi \lambda_0 \int_{r>0} r \exp(-\pi \lambda_0 r^2)
\frac{1}{p{\cal L}_W(\mu l(r)T) {\cal L}_{I^1}(\mu l(r)T)} \text{\rm d} r.
\end{eqnarray*}}
The finiteness of the last integral can be assessed using arguments
similar to those given above.
\item Assume now the IPNR model, with i.i.d. potential
receives $\Phi_0(n)$ and static MANET $\Phi$.
Assume also fast noise and fast fading.
We found no closed form expression for the mean local delay of the
typical node in this
case, however using some convexity arguments it can be shown that
it is smaller that the mean local delay of the typical node
in the original IPNR model
\end{itemize}
{\medskip\noindent \bf Remark:~\nolinebreak}
An important remark is in order. In the examples considered in this
section, we perform (at least some part of) the space average
{\em together} with the time average to get the mean local delay.
This operation, which makes sense in the case of high mobility (of
potential receivers, MANET nodes) more easily leads to a finite
mean local delay of the typical node.
In contrast, in the previous sections (case of static $\Phi$ and
$\Phi_0$)
we perform the time average first
and then the space average, and we get a different result, which can
for instance
be infinite.
\subsubsection{Adaptive Coding and Shannon Local Delay}
One may argue that if the mean delays are infinite in
the previously considered models, it is primarily because of the
{\em coverage logic},
where one transmits full packets at
time slots when the receiver is covered at the
required SINR and where one wastes all the other time slots.
This results in a RESTART mechanism
(cf. Remark~\ref{R.restart} and~\ref{R.restart}),
which in turn explains why we have heavy tails and infinite means.
Adaptive coding
offers the possibility of breaking the coverage/RESTART logic:
it gives up with minimal requirements on SINR
and it hence provides some non-null throughput at each time slot,
where this throughput depends on the current value of the SINR,
e.g. via Shannon's formula as briefly described in what follows.
Let ${\mathcal{T}}_0=\log(1+\text{SINR}_0)$ be the bit rate obtained
by the typical node $X_0$ under, Palm probability ${\mathbf P}^0$, at time
slot 0, where $\text{SINR}_0=\text{SINR}_0(0)$ is given by~(\ref{eq:SINR}).
It is natural to define the {\em Shannon local delay} of the typical
node $X_0$ as
$$\mathbf{L}^{Sh}=\mathbf{L}^{Sh}_0=
\frac{1}{p{\mathbf E}^0[{\mathcal{T}}_0\,|\,{\mathcal{S}}]}\,,$$
namely as the inverse of the time average of ${\mathcal{T}}_0$
given all the static elements (cf. Section~\ref{s.TheLD}).
This definition is the direct analogue of that of the
local delay in the packet model.
Observe that
\begin{eqnarray}\nonumber
\lefteqn{{\mathbf E}^0[\log (1+\text{SINR}_0)\,|\,{\mathcal{S}}]}\\
&&=\int_0^\infty {\mathbf P}^0\{\,\log(1+\text{SINR}_0)>t\,|\,{\mathcal{S}}\,\}
\,\text{\rm d} t \nonumber\\
&&=\int_0^\infty {\mathbf P}^0\{\,\text{SINR}_0>e^t-1\,|\,{\mathcal{S}}\}\,\text{\rm d} t \nonumber\\
&&=\int_0^\infty \pi_c(e^t-1|{\mathcal{S}})\text{\rm d} t\, \label{e:Tviap}
\end{eqnarray}
where $\pi_c(v|{\mathcal{S}})$ is as defined in (\ref{chalopic})
and where we made the dependence on $T=v$ explicit.
Consequently, we obtain
\begin{equation}\label{ss.TheLDSh}
{\mathbf E}^0[\mathbf{L}^{Sh}]={\mathbf E}^0\biggl[\frac1{\int_0^\infty\pi_c(v\,|\,{\mathcal{S}})/(v+1)\,\text{\rm d}
v}\biggr]\,.
\end{equation}
We now show two examples where ${\mathbf E}^0[\mathbf{L}]=\infty$ but
${\mathbf E}^0[\mathbf{L}^{Sh}]<\infty$.
\paragraph{Bipolar Receiver model}
Consider the slow noise, fast Rayleigh fading scenario.
We saw in Remark~\ref{R:wireless-contention-phase-trans-bipolar} that in this case,
for Poisson Bipolar, noise limited networks,
a necessary condition for ${\mathbf E}^0[\mathbf{L}]<\infty$ is that the noise $W$
has finite exponential moment ${\mathbf E}[e^{\{WTl(r)\mu}]<\infty$.
For the mean Shannon local delay of the typical node
we have
$${\mathbf E}^0[\mathbf{L}^{Sh}]=
{\mathbf E}\biggl[\frac{W}
{\int_0^\infty e^{-vl(r)\mu}/(v/W+1)\,\text{\rm d} v}\biggr]\,,
$$
in the noise limited case.
It is easy to see that the last expression is finite provided
${\mathbf E}[W]<\infty$ (which is much less constraining than the finiteness of
the exponential moment).
\paragraph{IPNR model}
Consider now the IPNR model, with fast Rayleigh fading.
Consider the interference limited case and with the
path loss function~(\ref{simpl.att})
It follows from the discussion after
Corollary~\ref{c:PINR-delay-interflim},
that if $\lambda_0\pi<\lambda\theta(p,T,\beta)$,
where $\theta(\cdot)$ is given by~(\ref{e:theta-delay}), then
${\mathbf E}^0[\mathbf{L}]=\infty$.
For the Shannon delay, in the interference limited case, we have
{\small
\begin{eqnarray*}
\lefteqn{{\mathbf E}^0[\mathbf{L}^{Sh}]=2\pi\lambda_0\frac{1}{p}\int_0^\infty
re^{-\lambda_0\pi r^2}}\\
&&\hspace{-2em}
\times{\mathbf E}^0\biggl[\biggl(\int_0^\infty
\frac{1}{v+r^\beta}\exp\Bigl\{\sum_{X_i\not=X_0}\log
{\mathcal{L}}_{eF}\bigl(\mu v/|X_i|^\beta\bigr)
\Bigr\}\,\text{\rm d} v
\biggr)^{-1}\biggr]\,\text{\rm d} r\,.
\end{eqnarray*}}
Using the inequalities
{\small \begin{eqnarray*}
\int_0^\infty\frac{\exp\{\dots\}}{v+r^\beta}\,\text{\rm d} v
&\ge&\frac{1}{2r^\beta}\int_0^{r^\beta}\exp\{\dots\}\,\text{\rm d} v
+ \int_{r^\beta}^\infty\frac{\exp\{\dots\}}{2v}\,\text{\rm d} v\\
&\ge&
\min\Bigl(\frac{1}{2r^\beta},1\Bigr)
\int_0^\infty\min\Bigl(\frac1{2v},1\Bigr)\exp\{\dots\}\,\text{\rm d} v\\
&\ge&
\min\Bigl(\frac{1}{2r^\beta},1\Bigr)
\int_0^{1/2}\exp\{\dots\}\,\text{\rm d} v\,.
\end{eqnarray*}}
Note that $\int_0^\infty 2r/(\min(r^{-\beta},2))e^{-\lambda_0\pi r^2}\,\text{\rm d}
r<\infty$. Using Jensen's inequality, we get that for all $X_i$
$$\log {\mathcal{L}}_{eF}\bigl(\mu v/|X_i|^\beta\bigr)\ge
-{\mathbf E}[eF]\mu v/|X_i|^\beta=-p v/|X_i|^\beta$$ for $|X_i|>\rho$,
where $\rho>0$ is some fixed constant. From this and from the inequality
${\mathcal{L}}_{eF}\ge 1-p$ for $|X_i|\le\rho$,
we conclude that ${\mathbf E}^0[\mathbf{L}^{Sh}]<\infty$ provided
$$
{\mathbf E}^0\biggl[\frac{\exp\Bigl\{-\log(1-p)\Phi(\{X_i:|X_i|\le\rho\})\Bigr\}}
{\int_0^{1/2}\exp\Bigl\{-pv\sum_{|X_i|>\rho}|X_i|^{-\beta}
\Bigr\}\,\text{\rm d} v}\biggr]<\infty\,.
$$
Using the independence property of the Poisson p.p., the fact that the
Poisson variable $\Phi(\{X_i:|X_i|\le\rho\})$ has finite
exponential moments, it remains to prove that
{\small
$$
{\mathbf E}^0\biggl[\biggl(\int_0^{1/2}\!\!\!
e^{-(pv\sum_{|X_i|>\rho}|X_i|^{-\beta})}\,\text{\rm d} v\biggr)^{-1}\biggr]
={\mathbf E}^0\Bigl[\frac{pJ}{1-e^{-pJ/2}}\Bigr]<\infty\,,
$$}
where $J=\sum_{|X_i|>\rho}|X_i|^{-\beta}$.
Note that for $J$ small, the expression under the expectation is close
to~2, whereas for $J$ bounded away from 0, we have
\begin{eqnarray*}
{\mathbf E}^0\Bigl[\frac{pJ}{1-e^{-pJ/2}}1\hspace{-0.30em}{\mbox{I}}(J>\epsilon)\Bigr]
&\le& (1-e^{-p\epsilon/2})p{\mathbf E}^0[J]\\
&=&(1-e^{-p\epsilon/2})2p\pi\lambda\int_\rho^\infty t^{1-\beta}\,\text{\rm d}
t,
\end{eqnarray*}
which is finite since $\beta>2$. Note that the last inequality
is essentially (modulo the problem of the pole of the
simplified path loss function~(\ref{simpl.att}) at 0 equivalent to the
finiteness of the mean of the
shot-noise.
\section{Conclusion}
\setcounter{equation}{0}
\label{s.conclusion}
In the present paper, we introduced a space-time scenario
for describing the dynamics of a MANET using Spatial Aloha.
This was used to analyze the law of the time to transmit a typical packet
from a typical node to its next-hop node in such networks.
This analysis was shown to lead to non trivial observations on the spatial
variability of the local delays in such MANETs, when assuming that the time
scale of the physical layer and the MAC layer is much smaller than
that of mobility.
In this case, {\em the local delay of the typical node
has heavy tails and infinite mean values}
in most standard scenarios, which however
{\em does {\em not} imply that the mean throughput of the typical node
is null}.
In addition, a new kind of phase transition,
related to the mean local delay of the typical node (being
the spatial, large-population, average of mean delays experienced by
individual nodes) was identified for the
interference limited case; closed form expressions were also given for the
thresholds separating the two phases in some computational cases.
Various ways of guaranteeing that the network is in the phase where
the spatial average of the mean delays is finite were discussed, some
based on an increase of
the variability (very high mobility, heavy tailed fading), some based on
bounding the distance to the next-hop. It was also shown that adaptive coding
offers fundamentally different performance compared to the coverage/outage
scheme, allowing for finite spatial mean local delays in cases
when the non-adaptive coverage/RESTART scheme gives infinite values.
The discussion on how to reduce the mean value of local delays (and in
particular how to move from an infinite to a finite mean value)
opens many interesting research directions which are left for a companion
paper: for instance, it would be useful to understand what classes of
moderate mobility lead to such a decrease. The same question is natural
for fading scenarios, or point processes representing additional receivers.
\vspace{-2ex}
\singlespacing
\bibliographystyle{plainnat}
{\footnotesize
|
\section{Introduction}
In this paper we are interested in curves $(\Gamma_t)_{t>0}$ in $\hbox{\bf R}^2$
which are half lines with an end point attached at the origin. These
lines are assumed to move with normal velocity
\begin{equation}\label{eq:law}
V_n= c+ \kappa
\end{equation}
where $\kappa$ is the curvature of the line and {\color{black} $c
\in\hbox{\bf R}$} is a given constant. We will see that this problem reduces
to the study of the following quasi-linear parabolic equation in
non-divergence form
\begin{equation}\label{eq::4a}
\displaystyle{r\bar{u}_t=c\sqrt{1 + r^2\bar{u}_r^2}
+ \bar{u}_r \left(\frac{2+ r^2 \bar{u}_r^2}{1+ r^2 \bar{u}_r^2}\right)
+ \frac{r\bar{u}_{rr}}{1 + r^2\bar{u}_r^2}},
\qquad t>0,r>0
\end{equation}
This paper is devoted to the proof of a comparison principle in the
class of sub-linear weak (viscosity) solutions and to the study of the
associated Cauchy problem.
\subsection{Motivations and known results}
\paragraph{Continuum mechanics.} From the viewpoint of
applications, the question of defining the motion of spirals in a two
dimensional space is motivated by the seminal paper of Burton, Cabrera
and Frank \cite{bcf51} where the growth of crystals with the vapor is
studied. When a screw dislocation line reaches the boundary of the
material, atoms are adsorbed on the surface in such a way that a
spiral is generated; moreover, under appropriate physical assumptions,
these authors prove that the geometric law governing the dynamics of
the growth of the spiral is precisely given by \eqref{eq:law} where
$-c$ denotes a critical value of the curvature. We mention that there
is an extensive literature in physics dealing with crystal growth in
spiral patterns.
\paragraph{Different mathematical approaches.} First and foremost,
defining geometric flows by studying non-linear parabolic equations is
an important topic both in analysis and geometry. Giving references in
such a general framework is out of the scope of this paper. As far as
the motion of spirals is concerned, the study of the dynamics of
spirals have been attracting a lot of attention for more than ten
years. Different methods have been proposed and developed in order to
define solutions of the geometric law~\eqref{eq:law}. {\color{black} A
brief list is given here.} A phase-field approach was first proposed
in \cite{kp} and the reader is also referred to
\cite{on00,on03}. Other approaches have been used; for instance,
``self-similar'' spirals are constructed in \cite{ishimura98} by
studying an ordinary differential equation. In \cite{gik02}, spirals
moving in (compact) annuli with homogeneous Neumann boundary condition
are constructed. From a technical point of view, the classical
parabolic theory is used to construct smooth solutions of the
associated partial differential equation; in particular, gradient
estimates are derived. We point out that in \cite{gik02}, the
geometric law is anisotropic, and is thus more general than
\eqref{eq:law}. In \cite{smereka00,ohtsuka03}, the geometric flow is
studied by using the level-set approach \cite{os88,cgg91,es91}. As in
\cite{gik02}, the author of \cite{ohtsuka03} considers spirals that
typically move into a (compact) annulus and reaches the boundary
perpendicularly. \medskip
The starting point of this paper is the following fact: to the best of our
knowledge, no geometric flows were constructed to describe the dynamics
of spirals by mean curvature by taking into account both the
singularity of the pinned point and the unboundedness of the domain.
\paragraph{The equation of interest.}
We would like next to explain with more details the main aims of our
study. By parametrizing spirals, we will see (\textit{cf.}
Subsection~\ref{subsec:level-set}) that the geometric
law~\eqref{eq:law} is translated into the quasi-linear parabolic
equation \eqref{eq::4a}. We note that the coefficients are unbounded
(they explode linearly with respect to $r$) and that the equation is
singular: indeed, as $r \to 0$, either $r u_t \to 0$ or first order
terms explode. Moreover, initial data are also unbounded. In such a
framework, we would like to achieve: uniqueness of weak (viscosity)
solutions for the Cauchy problem for a large class of initial data, to
construct a unique steady state (i.e. a solution of the form $\lambda
t + \bar\varphi(r)$), and finally to show the convergence of general
solutions of the Cauchy problem to the steady state as time goes to
infinity. This paper is mainly concerned with proving a uniqueness
result and constructing a weak (viscosity) solution; the study of
large time asymptotic will be achieved in \cite{fr2}.
\paragraph{Classical parabolic theory.}
Classical parabolic theory \cite{fried,LSU} could help us to construct
solutions but there are major difficulties to overcome. For instance,
Giga, Ishimura and Kohsaka \cite{gik02} studied a generalization of
\eqref{eq::4a} in domains of the form $R_{a,b}= {\color{black} \{ a <
r < b \}}$ with $a>0$ and $b>0$, with Neumann boundary conditions at
$r=a,b$. Roughly speaking, we can say that our goal is to see what
happens when $a \to 0$ and $b \to \infty$. First, we mentioned above
that the equation is not (uniformly) parabolic in the whole domain
$R_{0,\infty}={\color{black} \left\{0<r<+\infty\right\}}$. Second, in
such analysis, the key step is to obtain gradient
estimates. Unfortunately, the estimates from \cite{gik02} in the case
of \eqref{eq::4a} explode as $a$ goes to $0$. Third, once a solution
is constructed, it is natural to study uniqueness but {\em even in the
setting of classical solutions} there are substantial difficulties.
To conclude, classical parabolic theory can be useful in order to get
existence results, keeping in mind that getting gradient estimates for
\eqref{eq::4a} is not at all easy, but such techniques will not help
in proving uniqueness.
Recently, several authors studied uniqueness of quasilinear equations
with unbounded coefficients (see for instance \cite{BBBL03,ck}) by
using viscosity solution techniques for instance. But unfortunately,
Eq.~\eqref{eq::4a} does not satisfy the assumptions of these papers.
\paragraph{Main new ideas.}
New ideas are thus necessary to handle these difficulties, both for
existence and uniqueness. As far as uniqueness is concerned, one has
to figure out what is the relevant boundary condition at $r=0$.
We remark that solutions of \eqref{eq::4a} satisfy at least formally a
Neumann boundary condition at the origin
\begin{equation}\label{neum}
0=c + 2 \bar{u}_r \quad \mbox{for}\quad r=0.
\end{equation}
In some sense, we thus can say that the boundary condition is embedded
into the equation. Second, taking advantage of the fact that the
Neumann condition is compatible with the comparison principle,
viscosity solution techniques (also used in \cite{BBBL03}) permit us
to get uniqueness even if the equation is degenerate and also in a
very large class of weak (sub- and super-) solutions.
But there are remaining difficulties to be overcome. First, the
Boundary Condition~\eqref{neum} is only true asymptotically (as $r\to
0$) and the fact that it is embedded into the equation makes it
difficult to use. We will overcome this difficulty by making a proper
change of variables (namely $x=\ln r$, see below for further details)
{\color{black} and proving a comparison principle (whose proof is
rather involved; in particular many new arguments are needed in
compare with the classical case) in this framework}. Second,
classical viscosity solution techniques for parabolic equations do not
apply directly to \eqref{eq::4a} because of polar coordinates. More
precisely, the equation do not satisfy the fundamental structure
conditions as presented in \cite[{Eq.~(3.14)}]{cil92} when polar
coordinates are used. But the mean curvature equation has been
extensively studied in Cartesian coordinates \cite{es91,cgg91}. Hence
this set of coordinates should be used, at least far from the origin.
\paragraph{Perron's method and smooth solutions.}
We hope we convinced the reader that it is really useful, if not
mandatory, to use viscosity solution techniques to prove uniqueness.
It turns out that it can also be used to construct solutions by using
Perron's method \cite{ishii}. This technique requires to construct
appropriate barriers and we do so for a large class of initial
data. The next step is to prove that these weak solutions are smooth
if additional growth assumptions on derivatives of initial data are
imposed; we get such a result by deriving non-standard gradient
estimates (with viscosity solution techniques too).
We would like also to shed some light on the fact that this notion of
solution is also very useful when studying large time asymptotic (and
more generally to pass to the limit in such non-linear
equations). Indeed, convergence can be proved by using the
half-relaxed limit techniques if one can prove a comparison
principle. See \cite{fr2} for more details.
\subsection{The geometric formulation}
\label{subsec:level-set}
In this section, we make precise the way spirals are
defined. We will first define them as parametrized curves.
\paragraph{Parametrization of spirals.}
We look for interfaces $\Gamma$ parametrized as follows:
$\Gamma = \{ r e^{ -i \bar u(r)} : r \ge 0 \} {\color{black} \ \subset {\mathbb C}}$ for some
function $\bar u :[0,+\infty) \to \hbox{\bf R}$. If now the spiral
moves, i.e. evolves with a time variable $t >0$, then the function
$\bar u$ also depends on $t>0$.
\begin{defi}[Spirals]
A moving \emph{spiral} is a family of curves $(\Gamma_t)_{t>0}$ of
the following form
\begin{equation}\label{spiral:param}
\Gamma_t = \{ r e^{i\theta} : r >0, \theta \in \hbox{\bf R}, \; \theta + \bar u (t,r) = 0 \}
\end{equation}
for some function $\bar u : [0,+\infty) \times [0,+\infty) \to \hbox{\bf R}$.
This curve is oriented by choosing the normal vector field equal to
$(-i + r \partial_r \bar{u} (t,r))e^{-i\bar{u}(t,r)}$.
\end{defi}
\begin{figure}[ht]
\centering\epsfig{figure=F1.eps,width=90mm}
\caption{Motion of the spiral}\label{f1}
\end{figure}
With the previous definition in hand, the geometric law~\eqref{eq:law}
implies that $\bar u$ satisfies \eqref{eq::4a} with the initial condition
\begin{equation}\label{eq::4b}
\bar{u}(0,r)=\bar{u}_0(r) \quad \mbox{for}\quad r\in (0,+\infty) \, .
\end{equation}
\paragraph{Link with the level-set approach.}
In view of \eqref{spiral:param}, we see that our approach is closely
related to the level-set one. We recall that the level-set approach
was introduced in \cite{os88,es91,cgg91}; in particular, it permits to
construct an interface moving by mean curvature type motion, that is to say
satisfying the geometric law~\eqref{eq:law}. It consists in defining
the interface $\Gamma_t$ as the $0$-level set of a function $\tilde{U}
(t,\cdot)$ and in remarking that the geometric law is verified only if
$\tilde{U}$ satisfies a non-linear evolution equation of parabolic
type. In an informal way, we can say that the quasi-linear evolution
equation~\eqref{eq::4a} is a "graph" equation associated with the
classical mean curvature equation (MCE), but written in polar
coordinates.
More precisely, if ${\color{black} \tilde{U} (t,X)=\theta+ \bar u (t,r)}$
with $X {\color{black} =(r\cos\theta,r\sin\theta)}\in \hbox{\bf R}^2$, then $\bar u$ will satisfy
\eqref{eq::4a} as long as $\tilde{U}$ solves {\color{black} the following level-set equation}
\begin{equation}\label{eq::rm2}
\tilde{U}_t = c |D_X\tilde{U}|
+ \widehat{D_X\tilde{U}}^\perp \cdot D^2_{XX}\tilde{U} \; \widehat{D_X\tilde{U}}^\perp
{\color{black} \quad \mbox{for}\quad X\not= 0}
\end{equation}
(where $\hat{p}=p/|p|$ and $p^\perp =(-p_2,p_1)$ for $p=(p_1,p_2)\in\hbox{\bf R}^2$).
{\color{black} Notice that the angle $\theta$ is multivalued, i.e. only defined modulo $2\pi$.
Such an approach is for instance systematically developed in \cite{ohtsuka03}.}
\subsection{Main results}
\paragraph{Comparison principle.}
{\color{black} Our first main result is a comparison principle}: it says that all
sub-solutions lie below all super-solutions, provided they are ordered
at initial time.
\begin{theo}[Comparison principle for
\eqref{eq::4a}] \label{thm comp spirales sous lin en r}
Assume that {\color{black} $\bar{u}_0:(0,+\infty)\to \hbox{\bf R}$} is a {\color{black} globally} Lipschitz continuous function.
Consider a sub-solution $\bar u$ and a super-solution $\bar v$ of
{\color{black} \eqref{eq::4a},\eqref{eq::4b} (in the sense of Definition \ref{defi::1})}
such that there exist $C_1>0$ {\color{black} and
for all $t \in [0,T)$} and $r>0$,
{\color{black} \begin{equation}\label{eq:uetv}
\bar u(t,r) -\bar u_0(r)\le C_1 \quad \text{and} \quad
\bar v(t,r) -\bar u_0(r) \ge -C_1.
\end{equation}}
If $\bar u (0,r) \le \bar u_0 (r) \le \bar v(0,r)$ for all $r \ge 0$,
then $\bar u \le \bar v$ in {\color{black} $[0,T) \times (0,+\infty)$.}
\end{theo}
\begin{rem}
The growth of the sub-solution $u$ and the super-solution $v$ is made precise
by assuming Condition~\eqref{eq:uetv}. Such a condition is motivated by
the large time asymptotic study carried out in \cite{fr2}; indeed, we construct
in \cite{fr2} a global solution of the form $\lambda t + \bar u_0(r)$.
\end{rem}
The proof of Theorem \ref{thm comp spirales sous lin en r} is rather
involved and we will first state and prove a comparison principle in
the set of bounded functions for a larger class of equations (see
Theorem~\ref{thm comp spirales}). We do so in order to exhibit
the structure of the equation that makes the proof work. We then
turn to the proof of Theorem \ref{thm comp spirales sous lin en r}.
Both proofs are based on the doubling of variable method, which
consists in regularizing the sub- and super-solutions. Obviously,
this is a difficulty here because one end point of the curve is
attached at the origin and the doubling of variables at the origin is
not well defined. To overcome this difficulty, we work with
logarithmic coordinates $x=\ln r$ for $r$ close to $0$. But then the
equation becomes
{\color{black} $$u_t = c e^{-x} \sqrt{1 + u_x^2} + e^{-2x}u_x
+e^{-2x} \frac{u_{xx}}{1 + u_x^2}$$}
We then apply the doubling of variables in the $x$ coordinates. There
is a persistence of the difficulty, because we have now to bound terms
like
$$
A:= ce^{-x}\sqrt{1+u_x^2} - ce^{-y}\sqrt{1+v_y^2}
$$
that can blow up as $x,y \to -\infty$. We are lucky enough to be able
to show roughly speaking that $A$ can be controlled by {\color{black}
the doubling of variable of the term $e^{-2x}u_x$ which appears to
be the main term (in a certain sense) as $x$ goes to $-\infty$.}
In view of the study from \cite{fr2}, $\bar u_0$ has to be chosen
sub-linear in Cartesian coordinates and thus so are the sub- and
super-solutions to be compared. The second difficulty arises when
passing to logarithmic coordinates for large $r$'s; indeed, the
sub-solution and the super-solution then grow exponentially in $x=\ln
r$ at infinity and we did not manage to adapt the previous reasoning
in this setting. There is for instance a similar difficulty when
dealing with the mean curvature equatio. Indeed, in this framework,
for super-linear initial data, the uniqueness of the solution is not
known in full generality (see \cite{BBBL03,ck}). In other words, the
change of variables do not seem to work far from the origin. We thus
have to stick to Cartesian coordinates for large $r$'s {\color{black}
(using a level-set formulation)} and see the equation in different
coordinates when $r$ is either small or large (see
Section~\ref{sec:sublinear}).
\paragraph{Existence theorem.}
In order to get an existence theorem, we have to restrict the growth
of derivatives of the initial condition. We make the following
assumptions: the initial condition is globally Lipschitz continuous
and its mean curvature is bounded. We recall that the mean curvature
of a spiral parametrized by $\bar u$ is defined by
$$
\kappa_{\bar u}(r)=u_r\left(\frac {2+(r
\bar u_r)^2}{(1+(r\bar u_r)^2)^{\frac32}}\right) + \frac{r \bar
u_{rr}}{(1+(r\bar u_r)^2)^{\frac 32}}.
$$
We can now state our second main result.
\begin{theo}[The general Cauchy problem]\label{th:cauchywc}
Consider $\bar{u}_0 \in {\color{black} W^{2,\infty}_{loc}(0,+\infty)}$. Assume that $\bar{u}_0$
is {\color{black} globally Lipschitz continuous and that $\kappa_{\bar u_0}\in L^\infty(0,+\infty)$.}
Then there exists a unique solution $u$ of
\eqref{eq::4a},\eqref{eq::4b} {\color{black}on $[0,+\infty)\times (0,+\infty)$ (in the sense of Definition \ref{defi::1})}
such that for all $T>0$, there exists
{\color{black} $\bar C_T >0$} such that for all $t\in {\color{black} [0,T)}$ and $r>0$,
\begin{equation}\label{baru:baru0sst}
|\bar u(t,r) -\bar u_0(r)|\le {\color{black} \bar C_T} .
\end{equation}
Moreover, $\bar u$ is Lipschitz continuous with respect to space and
$\frac 12$-H\"older continuous with respect to time. More precisely,
there exists a constant $C$ depending only on $|(\bar u_0)_r|_\infty$ and
$|\kappa_{\bar u_0}|_\infty$ such that
$$
|\bar u (t, r+\rho)-\bar u(t,r)|\le C|\rho|
$$
and
\begin{equation}\label{eq::rm1}
|\bar u (t+h,r)-\bar u(t,r)|\le C\sqrt{|h|}.
\end{equation}
\end{theo}
{\color{black}
\begin{rem}
Notice that Theorem \ref{th:cauchywc} allows us to consider an
initial data $\bar u_0$ which does not satisfy the compatibility
condition (\ref{neum}), like for instance $\bar u_0\equiv 0$ with
$c=1$. Notice also that we do not know if the solution constructed
in Theorem \ref{th:cauchywc} is smooth (i.e. belongs to
$C^\infty((0,+\infty)^2)$).
\end{rem}
To get such a result, we first construct smooth solutions requiring that
the compatibility condition (\ref{neum}) is satisfied by the
initial datum, like in the following result.}
\begin{theo}[Existence and uniqueness of smooth solutions for the Cauchy problem]\label{th:cauchy}
Assume that {\color{black} $\bar{u}_0\in W^{2,\infty}_{loc}(0,+\infty)$ with
$$(\bar{u}_0)_r \in W^{1,\infty}(0,+\infty)\quad \mbox{or}\quad \kappa_{\bar u_0}\in L^\infty(0,+\infty)$$}
and
that it satisfies the following compatibility condition for some $r_0>0$:
\begin{equation}\label{cond:compatibilite}
|c+\kappa_{\bar u_0}|\le C r\quad {\rm for }\quad 0\le r\le {\color{black} r_0}.
\end{equation}
Then there exists a unique continuous function: $\bar u :
[0,+\infty)\times [0,+\infty)$ which is $C^\infty$ in {\color{black} $(0,+\infty)
\times (0,+\infty)$}, which satisfies \eqref{eq::4a},\eqref{eq::4b}
{\color{black} (in the sense of Definition \ref{defi::1})},
and such that there exists $\bar C >0$ such that
{\color{black}
$$|\bar u(t,r+\rho) -\bar u(t,r)|\le \bar C |\rho|$$
and
\begin{equation}\label{baru:baru0}
|\bar u(t+h,r) -\bar u(t,r)|\le \bar C |h|.
\end{equation}}
\end{theo}
{\color{black} \begin{rem} Condition \eqref{cond:compatibilite} allows
us also to improve the H\"{o}lder estimate (\ref{eq::rm1}) and to
replace it by the Lipschitz estimate (\ref{baru:baru0}). With the
help of this Lipschitz estimate (\ref{baru:baru0}), we can
conclude that the solution constructed in Theorem \ref{th:cauchy}
is smooth. Notice also that our space-time Lipschitz estimates on
the solution allow us to conclude that $\bar u(t,\cdot)$ satisfies
(\ref{cond:compatibilite}) with the constant $C$ replaced by some
possible higher constant. This implies in particular that $\bar
u(t,\cdot)$ satisfies the compatibility condition (\ref{neum}) for
all time $t\ge 0$.
\end{rem}}
{\color{black}
\paragraph{Open questions.}$\mbox{ }$\\
\noindent {\bf A. Weaker conditions on the initial data }\\
It would be interesting to investigate the existence/non-existence and
uniqueness/non-uniqueness of solutions when we allow the initial data
$\bar u_0$ to be less than globally Lipschitz. For instance what
happens when the initial data describes an infinite spiral close to
the origin $r=0$, with either $\bar u_0(0^+)=+\infty$ or $\bar
u_0(0^+)=-\infty$? On the other hand, what happens if the growth of
$\bar u_0$ is super-linear
as $r$ goes to $+\infty$?\\
\noindent {\bf B. More general shapes than spiral }\\
One of our main limitation to study only the evolution of spirals in
this paper is that we were not able to prove a comparison principle in
the case of the general level-set equation (\ref{eq::rm2}). The
difficulty is the fact that the gradient of the level-set function
$\tilde{U}$ may degenerate exactly at the origin where the curve is
attached. The fact that a spiral-like solution is a graph
$\theta=-\bar u(t,r)$ prevents the vanishing of the gradient of
$\tilde{U}$ at the origin $r=0$. If now we consider more general
curves attached at the origin, it would be interesting to study the
existence and uniqueness/non-uniqueness of solutions with general
initial data, like the curves on Figure \ref{F4}.
\begin{figure}[ht]
\centering\epsfig{figure=F4.eps,width=70mm}
\caption{Examples of non spiral initial data}\label{F4}
\end{figure}
}
\paragraph{Organization of the article.} In Section~\ref{prel}, we
recall the definition of viscosity solutions for the quasi-linear
evolution equation of interest in this paper. The change of variables
that will be used in the proof of the comparison principle is also
introduced. In Section~\ref{sec:bounded}, we give the proof of
Theorem~\ref{thm comp spirales sous lin en r} {\color{black} in the case of bounded solutions}. The
proof in the general case is given in Section~\ref{sec:sublinear}. In
Section~\ref{sec:classical}, a classical solution is constructed under
an additional compatibility condition on the initial datum (see
Theorem~\ref{th:cauchy}). First, we construct a viscosity solution by
Perron's method (Subsection~\ref{subsec:perron}); second, we derive
gradient estimates (Subsection~\ref{subsec:gradestim}); third, we
explain how to prove that the viscosity solution is in fact a
classical one (Subsection~\ref{subsec:class}). The construction of the
solution without compatibility assumption (Theorem~\ref{th:cauchywc})
is made in Section~\ref{sec:withoutcompatibility}. Finally, proofs of
technical lemmas are gathered in Appendix~\ref{app:a}.
\paragraph{Notation.} If $a$ is a real number, $a_+$ denotes $\max(0,a)$ and $a_-$ denotes $\max(0,-a)$
If $p=(p_1,p_2) \in \hbox{\bf R}^2$, $p \neq 0$, then $\hat p$ denotes $p / |p|$
and $p^\perp$ denotes $(-p_2,p_1)$.
\section{Preliminaries}
\label{prel}
\subsection{Viscosity solutions for the main equation}
In view of \eqref{eq::4a}, it is convenient to introduce the
following notation
\begin{equation}\label{eq:barF}
\bar{F}(r,q,Y)
= c \sqrt{1 + r^2q^2} + q \left(\frac{2+ r^2 q^2}{1+ r^2 q^2}\right)
+ \frac{r Y}{1 + r^2 q^2} .
\end{equation}
We first recall the notion of viscosity solution for an
equation such as \eqref{eq::4a}.
\begin{defi}[Viscosity solutions for \eqref{eq::4a},\eqref{eq::4b}] \label{defi::1} ~ \\
{\color{black} Let $T\in (0,+\infty]$.}
A lower semi-continuous (resp. upper semi-continuous) function {\color{black} $u:
[0,T) \times (0,+\infty) \to \hbox{\bf R}$} is a (viscosity)
\emph{super-solution} (resp. \emph{sub-solution}) of
\eqref{eq::4a},\eqref{eq::4b} {\color{black} on $[0,T)\times (0,+\infty)$} if for any $C^2$ test
function $\phi$ such that $u-\phi$ reaches a local minimum
(resp. maximum) at {\color{black} $(t,r) \in [0,T) \times (0,+\infty)$}, we
have
\begin{itemize}
\item[(i)] If {\color{black} $t> 0$}:
\begin{equation*}
r\phi_t \ge \bar F(r,\phi_r,\phi_{rr})\quad
\bigg( \text{resp. } r\phi_t \le \bar F(r,\phi_r,\phi_{rr}) \bigg).
\end{equation*}
\item[(ii)] If $t=0$:
$$u(0,r)\ge \bar u_0(r)\quad \bigg({\rm resp.}\; u(0,r)\le \bar u_0(r)\bigg).$$
\end{itemize}
A continuous function {\color{black} $u: [0,T) \times (0,+\infty) \to \hbox{\bf R}$} is a
(viscosity) \emph{solution} of \eqref{eq::4a},\eqref{eq::4b} {\color{black} on $[0,T)\times (0,+\infty)$} if it is
both a super-solution and a sub-solution.
\end{defi}
\begin{rem}
We do not impose any condition at $r=0$; in other words,
it is not necessary to impose a condition on the whole parabolic boundary
of the domain. This is due to the ``singularity'' of our equation at $r=0$.
\end{rem}
Since we only deal with this weak notion of solution,
(sub-/super-)solutions will always refer to (sub-/super-)solutions in
the viscosity sense. \medskip
When constructing solutions by Perron's method, it is necessary to use
the following classical discontinuous stability result. The reader is
referred to \cite{cgg91} for a proof.
\begin{pro}[Discontinuous stability] \label{pro:stability}
Consider a family $(u_\alpha)_{\alpha \in \mathcal{A}}$ of sub-solutions of
\eqref{eq::4a},\eqref{eq::4b} which is uniformly bounded from above. Then
the upper semi-continuous envelope of $\sup_{\alpha \in \mathcal{A}}
u_\alpha$ is a sub-solution of \eqref{eq::4a},\eqref{eq::4b}.
\end{pro}
\subsection{A change of unknown function}
\label{subsec:change}
We will make use of the following change of unknown function:
$u(t,x)=\bar{u}(t,r)$ with $x=\ln r$ satisfies for all $t>0$ and $x \in \hbox{\bf R}$
\begin{equation}\label{eq:polar-spir}
u_t = c e^{-x} \sqrt{1 + u_x^2} + e^{-2x}u_x
+e^{-2x} \frac{u_{xx}}{1 + u_x^2}
\end{equation}
submitted to the initial condition: for all $x \in \hbox{\bf R}$,
\begin{equation}\label{eq:ci}
u(0,x)=u_0(x)
\end{equation}
where $u_0(x) = \bar{u}_0 (e^x)$.
Eq.~\eqref{eq:polar-spir} can be rewritten $u_t = F(x,u_x,u_{xx})$
with
\begin{equation}\label{defi:F}
F(x,p,X) = c e^{-x} \sqrt{1 + p^2} + e^{-2x}p
+e^{-2x} \frac{X}{1 + p^2} .
\end{equation}
Remark that functions $F$ and $\bar{F}$ are related by the following formula
\begin{equation}\label{eq FF'}
F(x,u_x,u_{xx}) = \frac1r \bar{F} (r, \bar{u}_r, \bar{u}_{rr}) \, .
\end{equation}
Since the function $\ln$ is increasing and maps $(0,+\infty)$ onto
$\hbox{\bf R}$, we have the following elementary lemma which will be used
repeatedly throughout the paper.
\begin{lem}[Change of variables]\label{lem:equiv}
A function $\bar u$ is a solution of \eqref{eq::4a},\eqref{eq::4b}
if and only if the corresponding function $u$ is a solution of
\eqref{eq:polar-spir}-\eqref{eq:ci} with $u_0 (x) =
\bar{u}_0(e^x)$.
\end{lem}
The reader is referred to \cite{cil92} (for instance) for a proof of
such a result.
When proving the comparison principle in the general case, we will also have
to use Cartesian coordinates. From a technical point of view, the following lemma
is needed.
\begin{lem}[Coming back to the Cartesian coordinates]\label{lem:cartesien}
We consider a sub-solution $u$ (resp. super-solution $v$) of
\eqref{eq:polar-spir}-\eqref{eq:ci} and we define the function $\tilde U$
(resp. $\tilde V$) $: (0,+\infty) \times \hbox{\bf R}^2 \to \hbox{\bf R}$ by
$$\tilde
U(t,X)=\theta(X)+u(t,x(X)) \quad \text{\rm (resp. }
\quad \tilde V(t,Y)=\theta(Y)+u(t,x(Y)) \text{)}
$$
where $(\theta(Z),x(Z))$ is defined such that
$Z=e^{x(Z)+i\theta(Z)}\ne 0$.
Then $\tilde U$ (resp. $\tilde V$) is sub-solution (resp. super-solution) of
\begin{equation}\label{eq:1000}
\left\{\begin{array}{l}
\displaystyle{w_t = c |Dw| + \frac {Dw^\perp}{|D w|} D^2 w \frac {Dw^\perp}{|D w|}}\\
\\
w(0,x)=\theta(X)+\bar u_0(x(X)).
\end{array}\right.
\end{equation}
\end{lem}
\begin{rem}
In Lemma \ref{lem:cartesien}, for $Z\not=0$, the angle $\theta(Z)$
is only defined modulo $2\pi$, but is locally uniquely defined by
continuity. Then $D\theta, D^2\theta$ are always uniquely defined.
\end{rem}
\section{A comparison principle for bounded solutions}\label{sec:bounded}
As explained in the introduction, we first prove a comparison
principle for \eqref{eq::4a} in the class of bounded weak (viscosity)
solutions. In {\color{black} comparison} with classical comparison results for geometric
equations (see for instance \cite{es91,cgg91,sato94,gs91}), the
difficulty is to handle the singularity at the origin ($r=0$).
In order to clarify why a comparison principle holds true
for such a singular equation, we consider the following generalized
case
\begin{equation}\label{eq:gen-r}
\bar{u}_t = \frac{\bar{b}(\bar{u}_r,r \bar{u}_r)}{r}
+ \sigma^2 (r \bar{u}_r) \bar{u}_{rr}
\end{equation}
which can be written, with $x = \ln r$,
\begin{equation}\label{eq:gen-x}
u_t = e^{-x} b(e^{-x} u_x,u_x) + e^{-2x} \sigma^2 (u_x) u_{xx}
\end{equation}
where $b(q,p)= \bar{b} (q,p) - \sigma^2 (p)q$.
\paragraph{Assumption on $\mathbf (b,\sigma)$.}
\begin{itemize}
\item $\sigma \in W^{1,\infty} (\hbox{\bf R})$;
\item There exists $\delta_1, \delta_2, \delta _3, \delta_4>0$ such that
\begin{itemize}
\item
for all $q \in \hbox{\bf R}$ and $p_1,p_2 \in \hbox{\bf R}$,
$$
|b(q,p_1)- b(q,p_2)| \le \delta_1 |p_1 -p_2|;
$$
\item
for all $p \in \hbox{\bf R}$ and $q_1 \le q_2$,
$$
\delta_2 (q_2 - q_1) \le b(q_2,p) - b(q_1,p);
$$
$$|b(q_1,p)-b(q_2,p)|\le \delta _3 |q_1-q_2|;
$$
\item for all $p\in \hbox{\bf R}$
$$|b(0,p)|\le \delta_4 \sqrt {1+|p|^2}$$
\item
{\color{black} we have $\|\sigma\|_\infty^2 < 2 \delta_2$.}
\end{itemize}
\end{itemize}
In our special case, $\sigma (p) = (1+p^2)^{-\frac12}$ and
$b(q,p) = c \sqrt{1+p^2} + q $, {\color{black} and the assumption on $(b,\sigma)$ is satisfied.}
\begin{theo}[Comparison principle for
\eqref{eq:gen-r}-\eqref{eq:ci}] \label{thm comp spirales}
Assume that $u_0:(0,+\infty)\to \hbox{\bf R}$ is Lipschitz continuous. Consider a bounded
sub-solution $u$ and a bounded super-solution $v$ of
\eqref{eq:gen-r},\eqref{eq:ci} in the sense of Definition \ref{defi::1} with $\bar F$
given by the right hand side of (\ref{eq:gen-r}).
Then $u \le v$ in $(0,+\infty) \times
\hbox{\bf R}$.
\end{theo}
{\color{black} \begin{rem}
For radial solutions of the heat equation $u_t = \Delta u$ on $\hbox{\bf R}^n\backslash \left\{0\right\}$, we get $b(q,p)=(n-1)q$ and $\sigma(p)=1$. Therefore
the assumption on $(b,\sigma)$ is satisfied if and only if $1<2(n-2)$.
Notice that in particular for $n=2$ this assumption is not satisfied.
\end{rem}}
\begin{proof}[Proof of Theorem \ref{thm comp spirales}]
We classically fix $T>0$ and argue by contradiction by assuming that
$$
M = \sup_{0<t<T,x \in \hbox{\bf R}} (u(t,x) - v(t,x)) >0.
$$
\begin{lem}[Penalization] \label{lem:penal}
For $\alpha, \varepsilon, \eta>0$ small enough, and any {\color{black} $K\ge 0$,}
the supremum
$$
M_{\varepsilon, \alpha} = \sup_{0< t <T,x,y \in \hbox{\bf R}} \left\{ u(t,x) - v(t,y) -
e^{Kt} \frac{(x-y)^2}{2 \varepsilon} - \frac{\eta}{T-t} - \alpha
\frac{x^2}2 \right\}
$$
is attained at $(t,x,y)$ with $t>0$, $M_{\varepsilon,\alpha} > M/3>0$,
$|x-y| \le C_0 \sqrt{\varepsilon}$ and $\alpha |x| \le C_0 \sqrt{\alpha}$
for some $C_0>0$ only depending on $\|u\|_\infty$ and $\|v\|_\infty$.
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lem:penal}]
The fact that $M>0$ means that there exist $t^*>0$ and $x^* \in \hbox{\bf R}$ such that
$$
u(t^*,x^*) - v(t^*,x^*) \ge M/2> 0 .
$$
Since $u$ and $v$ are bounded functions, $M_{\varepsilon,\alpha}$ is attained at
a point $(t,x,y)$. By optimality of $(t,x,y)$, we have in particular
\begin{multline*}
u(t,x) - v(t,y) - e^{Kt} \frac{(x-y)^2}{2 \varepsilon} - \frac{\eta}{T-t}
- \alpha \frac{x^2}2 \\
\ge u(t^*,x^*) - v(t^*,x^*) - \frac{\eta}{T-t^*} - \alpha \frac{(x^*)^2}2
\ge M/3
\end{multline*}
for $\alpha$ and $\eta$ small enough (only depending on $M$). In
particular,
$$
\frac{(x-y)^2}{2 \varepsilon} + \alpha \frac{x^2}2 \le \|u\|_\infty + \|v\|_\infty \, .
$$
Hence, there exists a constant $C_0$ only depending on $\|u\|_\infty$ and $\|v\|_\infty$ such that
\begin{equation}\label{estim:penal}
|x-y| \le C_0 \sqrt{\varepsilon} \quad \text{ and } \quad \alpha |x| \le C_0 \sqrt{\alpha} \, .
\end{equation}
Assume that $t=0$. In this case, we use the fact
that $u_0$ is Lipschitz continuous and \eqref{estim:penal} in order
to get
\begin{eqnarray*}
\frac{M}3 &\le& u_0 (x) - u_0 (y)\le \|D u_0\|_\infty |x-y|\le C_0 \|D
u_0\|_\infty\sqrt\varepsilon
\end{eqnarray*}
which is absurd for $\varepsilon$ small enough (depending only on $M, C_0$ and $ \|D u_0\|_\infty$).
Hence $t>0$ and the proof of the lemma is complete.
\end{proof}
In the remaining of the proof, $\varepsilon$ is fixed (even if we will choose
it small enough) and $\alpha$ goes to $0$ (even if it is not necessary
to pass to the limit). In view of the previous discussion, we can
assume that, for $\varepsilon$ small enough, we have $t>0$ for all $\alpha>0$
small enough (independent on $\varepsilon$). We thus can write two viscosity
inequalities. It is then classical to use Jensen-Ishii's Lemma and combine
viscosity inequalities in order to get the following result.
\begin{lem}[{\color{black}Consequence of viscosity inequalities}] \label{lem:ji}
\begin{multline} \label{eq:ji}
\frac{\eta}{(T-t)^2} + K e^{Kt} \frac{(x-y)^2}{2\varepsilon} \le
e^{-x} b(e^{-x} (p+\alpha x), p+\alpha x) - e^{-y} b (e^{-y} p, p) \\
+ e^{-2x} \|\sigma\|_\infty^2 \alpha + \frac{e^{Kt}}{\varepsilon}
(e^{-x} \sigma (p+\alpha x) - e^{-y} \sigma (p))^2
\end{multline}
where $p = e^{Kt}\frac{(x-y)}\varepsilon$.
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lem:ji}]
Jensen-Ishii 's Lemma \cite{cil92} implies that for
all $\gamma_1 >0$, there exist four real numbers $a,b,A,B$ such that
\begin{eqnarray}
\label{eq:visc1} a &\le& e^{-x} b(e^{-x} (p+\alpha x), p+\alpha x)
+ e^{-2x} \sigma^2 (p+\alpha x) (A+\alpha) \\
\label{eq:visc2}
b &\ge& e^{-y} b (e^{-y} p, p) + e^{-2y} \sigma^2 (p) B
\end{eqnarray}
Moreover $a,b$ satisfy the following inequality
\begin{equation}
\label{eq:timederivative}
a-b \ge \frac{\eta}{(T-t)^2} + K e^{Kt} \frac{(x-y)^2}{2\varepsilon}
\end{equation}
(it is in fact an equality)
and for any $\gamma_1>0$ small enough, there exist two real numbers
$A,B$ satisfying the following matrix inequality
$$
\left[\begin{array}{cc} A & 0 \\ 0 & -B \end{array}
\right] \le \frac{e^{Kt}}{\varepsilon} (1+{\gamma_1}) \left[\begin{array}{cc} 1 &
-1 \\ -1 & 1 \end{array} \right] \, .
$$
This matrix inequality implies
\begin{equation}\label{matrix ineq}
A \xi_1^2 \le B \xi_2^2 + \frac{e^{Kt}}{\varepsilon}(1+{\gamma_1}) (\xi_1 -\xi_2)^2
\end{equation}
for all $\xi_1,\xi_2 \in \hbox{\bf R}$. Use this inequality with $\xi_1 = e^{-x} \sigma (p+\alpha x)$
and $\xi_2 = e^{-y} \sigma (p)$ and let $\gamma_1 \to 0$ in order to get the desired inequality.
\end{proof}
We can next make appear which error terms depend on $\alpha$ and
which ones depend on $\varepsilon$. For all $\gamma_2>0$,
\begin{equation} \label{eq:1}
\frac{\eta}{T^2} + K e^{Kt} \frac{(x-y)^2}{2\varepsilon} \le T_\alpha + T_\varepsilon
\end{equation}
where
\begin{multline*}
T_\alpha = e^{-x} b(e^{-x} (p+\alpha x), p+\alpha x) - e^{-x} b (e^{-x} p, p) \\
+ \|\sigma\|_\infty^2 e^{-2x} \alpha + (1 + \gamma_2^{-1})
\frac{e^{Kt-2x}}{\varepsilon} \|\sigma'\|_\infty^2 (\alpha x)^2
\end{multline*}
and $T_\varepsilon = T_\varepsilon^1 + T_\varepsilon^2$ with
$$
T_\varepsilon^1 = e^{-x} b(e^{-x} p, p )- e^{-y} b(e^{-y}p,p) \quad \text{and} \quad
T_\varepsilon^2= (1+\gamma_2)\|\sigma\|^2_\infty \frac{e^{Kt-2x}}{\varepsilon}
(1 - e^{x-y})^2 .
$$
It remains to estimate $T_\varepsilon$.
\begin{lem}[Estimate for $T_\varepsilon$]\label{lem:estim1}
For all $\gamma_3 >0$,
\begin{equation} \label{eq:3} T_\varepsilon \le
((C - 2 \delta_2)e^{-2x} + \gamma_3 ) e^{Kt}
\frac{(x-y)^2}\varepsilon + \gamma_3 \varepsilon
\end{equation}
where {\color{black} $C = \frac{\delta_4^2}{4\gamma_3} + (1+\gamma_2)
\|\sigma\|^2_\infty + o_\varepsilon (1)$}.
\end{lem}
\begin{proof}
Through a Taylor expansion, we obtain
\begin{align*}
T_\varepsilon^1 = & -e^{-y_\theta} b( e^{-y_\theta}p,p)(x-y) - e^{-2y_\theta} \partial_q b (e^{-y_\theta}p,p)p(x-y) \\
\le & - \varepsilon e^{-Kt} b(e^{-y_\theta}p,p) e^{-y_\theta} p - \delta_2 e^{Kt-2y_\theta}\frac{(x-y)^2}{\varepsilon}
\end{align*}
where $y_\theta = \theta y + (1-\theta) x$ for some $\theta \in [0,1]$,
{\color{black} and we have used the fact that for all $q\in \hbox{\bf R}$,
$$q (b(q,p)-b(0,p)) \ge \delta_2 q^2$$
We} get for any $\gamma_3 >0$
\begin{align*}
T_\varepsilon^1 \le & {\color{black} \delta_4} e^{-y_\theta} \frac{|x-y|}{\sqrt{\varepsilon}} \sqrt{\varepsilon}\sqrt{1+p^2}
- 2\delta_2 e^{Kt-2y_\theta} \frac{(x-y)^2}\varepsilon \\
\le & \frac{\delta_4^2}{4\gamma_3} (1+o_\varepsilon (1))e^{Kt-2x} \frac{(x-y)^2}\varepsilon + e^{-Kt} \gamma_3 \varepsilon + \gamma_3
e^{Kt} \frac{(x-y)^2}\varepsilon \\
& - 2\delta_2 (1+o_\varepsilon (1))e^{Kt-2x} \frac{(x-y)^2}\varepsilon,
\end{align*}
{\color{black}where we have used that $y_\theta = x + o_\varepsilon (1)$.
Now, since $y = x + O(\sqrt{\varepsilon})$, we also have
$$
T_\varepsilon^2 = (1+\gamma_2)(1+o_\varepsilon (1))
\|\sigma\|^2_\infty \frac{e^{Kt-2x}}{\varepsilon} (x-y)^2
$$}
and we can conclude.
\end{proof}
Combining \eqref{eq:1} and \eqref{eq:3} finally
yields
$$
\frac{\eta}{T^2} + \left( \frac{K}2 - \gamma_3 + (2 \delta_2-C)
e^{-2x} \right) e^{Kt} \frac{(x-y)^2}{\varepsilon} \le T_\alpha + \gamma_3\varepsilon.
$$
It suffices now to choose $\gamma_i$, $i=1,2,3$, such that $C \le 2 {\color{black} \delta_2}$ and
then choose $K = 2 \gamma_3$ and $\varepsilon$ small enough to get: $\frac{\eta}{2 T^2} \le T_\alpha$.
The following lemma
permits to estimate $T_\alpha$.
\begin{lem}[Estimate for $T_\alpha$]\label{lem:estim2}
For $D>0$ large enough, we have
\begin{equation} \label{eq:2}
\frac{T_\alpha}D \le e^{-2x} {\color{black} (-\alpha x_-+ \alpha x_+)}
+ e^{-x} |\alpha x|+ e^{-2x} \alpha + e^{Kt-2x}\frac{\sqrt{\alpha}}{\varepsilon}|\alpha x|.
\end{equation}
\end{lem}
\begin{proof}
Using assumptions on $b$ immediately yields
$$
T_\alpha \le e^{-2x} {\color{black} (-\delta_2 \alpha x_-+ \delta_3 \alpha x_+)}
+\delta_1 e^{-x} |\alpha x|+
\|\sigma\|_\infty^2 e^{-2x} \alpha + (1 + \gamma_2^{-1})
\frac{e^{Kt-2x}\sqrt{\alpha}}{\varepsilon} \|\sigma'\|_\infty^2 C_0 |\alpha x|
$$
where we used that $|\alpha x| \le C_0 \sqrt{\alpha}$.
\end{proof}
We next consider $a>0$ such that for all $x \le -a$, we have
{\color{black} $$
-|x|e^{-2x} + 2|x|e^{-x} + e^{-2x} \le 0.
$$}
We now distinguish cases.
{\bf Case 1: $\mathbf{x_n \le -a}$ for some $\mathbf{\alpha_n \to 0}$.} We choose $n$ large enough
so that $e^{Kt} \frac{\sqrt{\alpha_n}}\varepsilon \le {\color{black} 1}$ and we get
{\color{black} $$
\frac{\eta}{2T^2} \le D \alpha (-|x|e^{-2x} + 2|x|e^{-x} + e^{-2x} ) \le 0
$$}
which implies $\eta \le 0$. {\color{black} Contradiction.}
{\bf Case 2: $\mathbf{x \ge -a}$ for all $\alpha$ small enough.} We use \eqref{estim:penal} and get
$$
\frac{\eta}{{\color{black} 2}T^2} \le D e^{2a} (2 \alpha |x| + \alpha + e^{Kt} \frac{\sqrt{\alpha}}\varepsilon |\alpha x|)
\le D ( 2 C_0 \sqrt{\alpha} + \alpha + e^{Kt} \frac{C_0 \alpha}{\varepsilon})
$$
and we let $\alpha \to 0$ to get $\eta \le 0$ in this case too.
The proof is now complete.
\end{proof}
\section{Comparison principle for sub-linear solutions}
\label{sec:sublinear}
\begin{proof}[Proof of Theorem~\ref{thm comp spirales sous lin en r}]
Thanks to the change of unknown function described in Subsection~\ref{subsec:change},
we can consider the functions $u$ and $v$ defined on $(0,+\infty) \times \hbox{\bf R}$ which
are sub- and super-solutions of \eqref{eq:polar-spir}. We can either prove that
$\bar u \le \bar v$ in $(0,+\infty) \times (0,+\infty)$ or that $u \le v$ in
$(0,+\infty) \times \hbox{\bf R}$.
For $\theta\in \hbox{\bf R}$, we define
$$
U(t,x,\theta)=\theta+u(t,x)\quad{\rm and}\quad
V(t,x,\theta)=\theta+v(t,x).
$$
Note that $U$ and $V$ are respectively sub and super-solution of
\begin{equation}\label{eq:1010}
W_t(t,x,\theta)=c e^{-x} |D W| + e^{-2x} DW\cdot e_1 + e^{-2x} \frac
{DW^\perp}{|DW|}D^2W \frac {DW^\perp}{|DW|}.
\end{equation}
We fix $T>0$ and we argue by contradiction by assuming that
$$M=\sup_{t \in [0,T], x, \theta\in \hbox{\bf R}}
\{U(t,x,\theta)-V(t,x,\theta)\}>0.$$
In order to use the doubling variable technique, we need
a smooth interpolation function $\Psi$ between polar coordinates
for small $r$'s and Cartesian coordinates for large $r$'s.
Precisely, we choose $\Psi$ as follows.
\begin{lem}[Interpolation between logarithmic and Cartesian
coordinates]\label{lem:psi}
There exists a smooth ($C^\infty$) function $\psi:\hbox{\bf R} ^2\mapsto \hbox{\bf R}^3$
such that
\begin{equation}
\label{eq:ass-psi}
\left\{\begin{array}{rll}
\psi(x,\theta+2\pi)&=\psi(x,\theta)& \\
\psi(x,\theta)&=(x,e^{i\theta}) &\quad{\rm if} \quad x\le 0\\
\psi(x,\theta)&=(0, e^xe^{i\theta})&\quad {\rm if} \quad x\ge 1
\end{array}\right.
\end{equation}
and such that there exists two constants $\delta_0 >0$ and $m_\psi>0$ such
that
for $x \le 1$ and $\theta \in \hbox{\bf R}$,
\begin{equation}\label{psi:add}
{\color{black} \mbox{if}\quad \psi(x,\theta) = (a,b) \text{ then } |b| \le e}
\end{equation}
and such that for all $x,y,\sigma,\theta$, if
$|\psi(x,\theta)-\psi(y,\sigma)| \le \delta_0$ and $|\theta -\sigma | \le
\frac\pi2$, then
\begin{eqnarray}\label{psi:grad-inf}
|\psi(x,\theta)-\psi(y,\sigma)| &\ge& m_\psi
|(x,\theta)-(y,\sigma)|, \\
\label{psi:grad-inf-bis}
\left| D\psi(x,\theta)^T {\color{black} \odot}\
(\psi(x,\theta)-\psi(y,\sigma)) \right| &\ge& m_\psi
|(x,\theta)-(y,\sigma)|
\end{eqnarray}
{\color{black} where $\odot$ is the tensor contraction defined for a $p$-tensor $A=(A_{i_1,\dots, i_p})$ and a
$q$-tensor $B=(B_{j_1,\dots, j_q})$ by}
$$(A\odot B)_{i_1,\dots,i_{p-1},j_2,\dots, j_{q}}=\sum_{k}
A_{i_1,\dots,i_{p-1},k}B_{k,j_2,\dots, j_q}.$$
\end{lem}
The proof of this lemma is given in Appendix~\ref{app:a}.
\paragraph{Penalization.} We consider the following approximation of $M$
\begin{multline}\label{eq:Mea}
M_{\varepsilon,\alpha}=\sup_{t \in [0,T],x,\theta,y,\sigma \in \hbox{\bf R}} \bigg\{
U(t,x,\theta)-V(t,y,\sigma) \\ -e^{Kt}\frac{|\psi(x,\theta)-
\psi(y,\sigma)|^2}{2\varepsilon}-\frac 1 \varepsilon \left(|\theta-\sigma|-\frac \pi 3\right)_+^2
-\frac \alpha 2 \left|\psi(x,\theta)\right|^2 -\frac
\eta {T-t}\bigg\}
\end{multline}
where $\varepsilon,\alpha, \eta$ are small parameters and $K\ge
0$ is a large constant to be fixed later. For $\alpha$ and $\eta$ small
enough we remark that $M_{\varepsilon,\alpha}\ge \frac M 2>0$. In order to prove
that the maximum $M_{\varepsilon,\alpha}$ is attained, we need the following lemma
whose proof is postponed until Appendix~\ref{app:a}.
\begin{lem}[A priori estimates]\label{lem:est}
There exists a constant $C_2>0$ such that the following estimate holds
true for any $x,y,\theta,\sigma \in\hbox{\bf R}$
\begin{eqnarray*}
|u_0(x)-u_0(y)|&\le& C_2 +e^{Kt}\frac
{|\psi(x,\theta)-\psi(y,\sigma)|^2}{4\varepsilon}\\
|\theta-\sigma|&\le& C_2+\frac 1 {2\varepsilon}
\left(|\theta-\sigma|-\frac \pi 3\right)^2_+.
\end{eqnarray*}
\end{lem}
Using this lemma, we then deduce that
\begin{multline}\label{eq:est1}
U(t,x,\theta)-V(t,y,\sigma) -e^{Kt}\frac
{|\psi(x,\theta)-\psi(y,\sigma)|^2}{2\varepsilon}-\frac 1
\varepsilon\left(|\theta-\sigma|-\frac \pi 3\right)^2_+\\
\le {\color{black} u(t,x)-u_0(x)-v(t,y)
+u_0(y)}+|\theta-\sigma|+|u_0(x)-u_0(y)|\\
-e^{Kt}\frac
{|\psi(x,\theta)-\psi(y,\sigma)|^2}{2\varepsilon}-\frac 1
\varepsilon\left(|\theta-\sigma|-\frac \pi 3\right)^2_+\\
\le 2 C_1 + 2 C_2 -e^{Kt}\frac
{|\psi(x,\theta)-\psi(y,\sigma)|^2}{4\varepsilon}-\frac 1
{2\varepsilon}\left(|\theta-\sigma|-\frac \pi 3\right)^2_+.
\end{multline}
Using the $2\pi$-periodicity of $\psi$, the maximum is achieved for
$\theta\in [0,2\pi]$. Then, using the previous estimate and the fact
that $-\alpha(\psi(x,\theta))^2\to -\infty$ as $|x|\to\infty$, we deduce
that the maximum is reached at some point that we still denote
$(t,x,\theta,y,\sigma)$.
\paragraph{Penalization estimates.}
Using Estimate~\eqref{eq:est1} and the fact that $M_{\varepsilon,\alpha}\ge
0$, we deduce that there exists a constant $C_0=4(C_1+C_2)$ such that
\begin{equation}\label{eq:est2}
\alpha (\psi(x,\theta))^2+\frac 1
\varepsilon\left(|\theta-\sigma|-\frac \pi3 \right)^2_++e^{Kt}\frac
{|\psi(x,\theta)-\psi(y,\sigma)|^2}{2\varepsilon}\le C_0 .
\end{equation}
On the one hand, an immediate consequence of this estimate is that
$$
|\theta-\sigma|\le \frac \pi 2
$$
for $\varepsilon$ small enough. On the other hand, we deduce from
\eqref{eq:est2} and \eqref{psi:grad-inf}
$$
m_\psi \frac{|\theta-\sigma|^2 + |x-y|^2}{2\varepsilon}\le C_0.
$$
Hence, we have $|\theta - \sigma| \le \frac{\pi}4$ for $\varepsilon$ small enough
so that the constraint $|\theta -\sigma| \le \frac \pi 3$ is not
saturated. We can also choose $\varepsilon$ small enough so that
$$
|x-y| \le \frac12.
$$
In the sequel of the proof, we will also need a better estimate on the
term $\alpha (\psi(x))^2$; precisely, we need to know that $\alpha
(\psi(x))^2\to 0$ as $\alpha\to 0$. Even if such a result is classical
(see \cite{cil92}), we give details for the reader's convenience. To
prove this, we introduce
\begin{align*}
M_{\varepsilon,0}= \sup_{t \in [0,T], x,\theta, y, \sigma \in \hbox{\bf R}} \bigg\{&
U(t,x,\theta)-V(t,y,\sigma)-e^{Kt}\frac{|\psi(x,\theta)-
\psi(y,\sigma)|^2}{2\varepsilon}\\
&-\frac 1 \varepsilon \left(|\theta-\sigma|-\frac \pi 3\right)^2_+ -\frac \eta
{T-t}\bigg\}
\end{align*}
which is finite thanks to \eqref{eq:est1}.
We remark that $M_{\varepsilon,\alpha}\le M_{\varepsilon,0}$ and that $M_{\varepsilon,\alpha}$ is
non-decreasing when $\alpha$ decreases to zero. We then deduce that
there exists $L$ such that $M_{\varepsilon,\alpha}\to L$ as $\alpha\to 0$. A simple
computation then gives that
$$
\frac \alpha 4 (\psi(x,\theta))^2\le M_{\varepsilon,\frac \alpha 2}-
M_{\varepsilon,\alpha}\to 0\quad {\rm as}\quad \alpha \to 0
$$
and then
\begin{equation}\label{eq:est3}
\frac \alpha 2 (\psi(x,\theta))^2\to 0\quad {\rm as}\quad \alpha \to 0.
\end{equation}
\paragraph{Initial condition.} We now prove the following lemma.
\begin{lem} [Avoiding $t=0$]\label{lem:ci}
For $\varepsilon$ small enough, we have $t>0$ for all $\alpha>0$ small enough.
\end{lem}
\begin{proof}
We argue by contradiction. Assume that $t=0$. We then
distinguish two cases.
If the corresponding $x$ and $y$ are small ($x\le 2$ and
$y\le 2$) then, since $u_0$ is Lipschitz continuous and
\eqref{psi:grad-inf} holds true, there exists a constant $L_0>0$ such
that
\begin{align*}
0<\frac M 2\le M_{\varepsilon,\alpha}\le &
U(0,x,\theta)-V(0,y,\sigma)-\frac{|\psi(x,\theta)-\psi(y,\sigma)|^2}{2\varepsilon} \\
\le & L_0|(x,\theta)-(y,\sigma)| - m_\psi\frac
{|(x,\theta)-(y,\sigma)|^2}{2\varepsilon} \\
\le & \frac {L_0^2}{2m_\psi} \varepsilon
\end{align*}
which is absurd for $\varepsilon$ small enough.
The other case corresponds to large $x$ and $y$ ($x\ge 1$ and
$y\ge 1$). In this case, since $\bar u_0$ is Lipschitz continuous,
we know that there exists a constant $L_1>0$
such that
$$
0\le \frac M 2 \le M_{\varepsilon, \alpha}\le |U(0,x,\theta)-V(0,y,
\sigma)|\le |\theta-\sigma| + L_1 |e^{x}-e^{y}|.
$$
Using the fact
$$
|\theta-\sigma|+ L_1 |e^{x}-e^{y}|\le \left(\frac 1 {m_\psi} +
L_1\right) |\psi(x,\theta)-\psi(y,\sigma)|\le \left(\frac 1
{m_\psi} + L_1\right)\sqrt{ 2 C_0} \sqrt \varepsilon
$$
we get a contradiction for $\varepsilon$ small enough.
\end{proof}
Thanks to Lemma~\ref{lem:ci}, we will now write two viscosity
inequalities, combine them and exhibit a contradiction. We recall that
we have to distinguish cases in order to determine properly in which
coordinates viscosity inequalities must be written (see the
Introduction).
\paragraph{Case 1: There exists $\alpha_n\to 0$ such that $x\ge \frac 32$
and $y\ge \frac 32$.}
We set $X=e^{x+i\theta}$ and $Y=e^{y+i\sigma}$. Consider $\tilde U$
and $\tilde V$ defined in Lemma~\ref{lem:cartesien}. Remark that,
even if $\theta(X)$ is defined modulo $2\pi$, the quantity
$\theta(X)-\theta(Y)$ is well defined (for $|X|, |Y|\ge e$ and
$|X-Y|\le \frac 12$) and thus so is $\tilde U(t,X)- \tilde V(t, Y)$.
Recall also that $\tilde{U}, \tilde{V}$ are respectively sub and
super-solutions of the following equation
$$w_t = c|Dw| + \widehat{Dw}^\perp\cdot D^2 w \cdot \widehat{Dw}^\perp$$
Moreover, using the explicit form of $\psi$, we get that
$$
M_{\varepsilon,\alpha}=\sup_{t \in [0,T], X,Y \in \hbox{\bf R}^2 \backslash B_1(0)}
\left\{\tilde U(t,X)-\tilde V(t,Y) - \frac
{e^{Kt}}{2\varepsilon} |X-Y|^2 - \frac \alpha 2 |X|^2 - \frac \eta{T-t}\right\}.
$$
Moreover, $-|D_X \tilde U|\le - \frac 1 {|X|}$ (in the viscosity
sense). We set
$$p= \frac{X-Y} \varepsilon e^{Kt}.
$$
We now use the
Jensen-Ishii Lemma \cite{cil92} in order to get four real {\color{black} numbers} $a,
b, A, B$ such that
\begin{eqnarray*}
a&\le& c|p+\alpha X| + \frac {(p+\alpha X)^\perp}{|p+\alpha X|}(A+ \alpha I) \frac
{(p+\alpha X)^\perp}{|p+\alpha X|}, \\
b&\ge& c|p| + \frac {p^\perp}{|p|}B \frac
{p^\perp}{|p|}.
\end{eqnarray*}
Moreover, $p$ satisfies the following estimate
\begin{equation}\label{eq:p}
|p+\alpha X|\ge \frac 1 {|X|},\quad |p|\ge \frac 1 {|Y|},
\end{equation}
$a, b$ satisfy the following equality
$$a-b=\frac \eta {(T-t)^2} + K e^{Kt} \frac {|X-Y|^2}{2\varepsilon}$$
and $A,B$ satisfy the following matrix inequality
$$\left[\begin{array} {cc}
A&0\\
0&-B
\end{array}\right]
\le \frac {2 e^{Kt}}{\varepsilon}
\left[\begin{array} {cc}
I&-I\\
-I&I
\end{array}\right].
$$
This matrix inequality implies
\begin{equation}\label{eq:matrix}
A\xi_1^2\le B\xi_2^2 + \frac {2 e^{Kt}}{\varepsilon} |\xi_1-\xi_2|^2
\end{equation}
for all $\xi_1,\xi_2\in \hbox{\bf R}^2$. Subtracting the two viscosity
inequalities, we then get
\begin{align*}
\frac \eta {T^2}\le &
c|p+\alpha X|-c|p| +\alpha + \frac {(p+\alpha X)^\perp}{|p+\alpha X|}A\frac
{(p+\alpha X)^\perp}{|p+\alpha X|}- \frac {p^\perp}{|p|}B \frac
{p^\perp}{|p|}\\
\le& \alpha |c| |X| +\alpha+ \frac {2 e^{Kt}}{\varepsilon} \left( \frac {p+\alpha X}{|p+\alpha
X|}-\frac {p}{|p|}\right)^2\\
\le & |c| \sqrt{C_0}\sqrt \alpha +\alpha+\frac {2 e^{Kt}}{\varepsilon}\left(2\left(\frac
{\alpha X}{\frac 1 {|X|}}\right)^2 + 2 \left(\frac {p}{|p|} \frac {{\color{black} |\alpha
X|}}{|p+|\alpha X||}\right)^2\right)\\
\le & |c| \sqrt{C_0}\sqrt \alpha +\alpha+\frac {8 e^{Kt}}{\varepsilon}\left({\alpha
|X|^2}\right)^2
\end{align*}
where we have used successively \eqref{eq:matrix}, \eqref{eq:est2} and
\eqref{eq:p}. Recalling, by \eqref{eq:est3} that $\alpha |X|^2=o_\alpha(1)$,
we get a contradiction for $\alpha$ small enough.
\paragraph{Case 2: There exists $\alpha_n\to 0$ such that $x\le -\frac 12$ and
$y\le -\frac 12$.}
Using the explicit
form of $\psi$ and the fact that $U(t,x,\theta)=\theta+u(t,x)$ and
$V(t,y,\sigma)=\sigma+v(t,y)$ with $u$ and $v$ respectively sub and
super-solution of \eqref{eq:polar-spir}, we remark that
$$
M_{\varepsilon,\alpha}=\sup_{t',x',y'}\{u(t',x')-v(t',y')-e^{Kt'}\frac{|\psi(x',\theta)-
\psi(y',\sigma)|^2}{2\varepsilon}-\frac \alpha 2 |x'|^2 -\frac \eta {T-t'}
+\theta-\sigma -\frac \alpha 2\}.
$$
Moreover, the maximum is reached at $(t,x,y)$, where we recall that
$(t,x,\theta,y,\sigma)$ is the point of maximum in \eqref{eq:Mea}.
Using the Jensen-Ishii Lemma \cite{cil92}, we then deduce the
existence, for all $\gamma_1 >0$, of four real numbers $a,b,A,B$ such
that
\begin{eqnarray*}
a &\le& ce^{-x} \sqrt{1 + (p+\alpha x)^2} + e^{-2x} (p+\alpha x )
+ e^{-2x} \frac{A + \alpha }{1 + (p+\alpha x )^2} \\
b &\ge& ce^{-y} \sqrt{1 + p^2} + e^{-2y} p + e^{-2y} \frac{B}{1+p^2}
\end{eqnarray*}
where
$$p=\frac{x-y}\varepsilon e^{Kt}.$$
These inequalities are exactly \eqref{eq:visc1} and \eqref{eq:visc2}.
Moreover $a,b$ satisfy the following inequality
$$
a-b = \frac{\eta}{(T-t)^2} + K e^{Kt}
\frac{|\psi(x,\theta)-\psi(y,\sigma)|^2}{2\varepsilon} \ge
\frac{\eta}{(T-t)^2} + K e^{Kt}
\frac{|x-y|^2}{2\varepsilon}
$$
and we obtain \eqref{eq:timederivative}. Moreover, $A,B$ satisfy the
following matrix inequality
$$
\left[\begin{array}{cc} A & 0 \\ 0 & -B \end{array} \right] \le
\frac{e^{Kt}}{\varepsilon} (1+{\gamma_1}) \left[\begin{array}{cc} 1 & -1 \\
-1 & 1 \end{array} \right]
$$
which implies \eqref{matrix ineq}. On one hand, from \eqref{eq:visc1}, \eqref{eq:visc2},
\eqref{eq:timederivative} and \eqref{matrix ineq},
we can derive \eqref{eq:ji}. On the other hand, \eqref{eq:est2}, the
fact that $x\le 0$, $y\le 0$ and Lemma~\ref{lem:psi} imply
\eqref{estim:penal} (with a different constant). We thus can apply Lemmas
\ref{lem:estim1}, \ref{lem:estim2} and deduce the desired
contradiction.
\paragraph{Case 3: There exists $\alpha_n\to 0$ such that $-1\le x, y \le 2$.}
Since $\psi\in C^\infty $, there then exists $M_\psi>0$ (only
depending on the function $\psi$) such that for all $x \in [-1,2]$ and
$\theta \in [-\pi,3 \pi]$,
\begin{equation}\label{eq:mpsi}
|\psi(x,\theta)| + |D \psi(x,\theta)|+|D^2 \psi(x,\theta)|
+|D^3\psi(x,\theta)| \le M_\psi.
\end{equation}
For simplicity of notation, we denote $(x,\theta)$ by $\bar x$ and
$(y,\sigma)$ by $\bar y$. We next define
$$
p_{\bar x}=\frac {e^{K t }}\varepsilon D\psi(\bar x)^T \odot
(\psi(\bar x)-\psi(\bar y))\quad{\rm and}\quad p_{\bar y}=\frac {e^{K
t }}\varepsilon D \psi(\bar y)^T \odot (\psi(\bar x)-\psi(\bar y)).
$$
We have $p_{\bar{x}}, p_{\bar{y}} \in\hbox{\bf R}^2$ and we set $(e_1,e_2)$ a
basis of $\hbox{\bf R}^2$.
\begin{lem}[Combining viscosity inequalities for $\alpha=0$]
We have for $\alpha=0$
\begin{equation}\label{eq:301}
\frac \eta{T^2}+ Km_\psi^2 e^{Kt}\frac {|\bar x-\bar y|^2}{2\varepsilon}\le ce^{-x}
|p_{\bar x}|- ce^{-y} |p_{\bar y}| + e^{-2x} p_{\bar x}\cdot e_1 -
e^{-2y} p_{\bar y}\cdot e_1 + \frac {2e^{Kt}}\varepsilon ({\cal I}_1+{\cal I}_2)
\end{equation}
where
\begin{eqnarray*}
{\cal I}_1&=&(\psi(\bar x)-\psi(\bar y))\odot \bigg( D^2\psi(\bar x) e^{-x} \widehat
{p_{\bar x}^\perp}\cdot e^{-x} \widehat {p_{\bar x}^\perp}
-D^2\psi(\bar y) e^{-y} \widehat{p_{\bar y}^\perp}\cdot e^{-y} \widehat {p_y^\perp} \bigg) \\
{\cal I}_2&=&\left|D\psi(\bar x)e^{-x} \widehat {p_{\bar x}^\perp} -D\psi(\bar y)e^{-y}
\widehat {p_{\bar y}^\perp} \right|^2.
\end{eqnarray*}
\end{lem}
\begin{proof}
Recall that $U$ and $V$ are respectively sub and
super-solution of \eqref{eq:1010} and use the Jensen-Ishii Lemma \cite{cil92} in order to deduce that there exist
two real numbers $a,b$ and two $2 \times 2$ real matrices $A,B$ such that
\begin{eqnarray*}
a &\le& {\color{black} c}e^{-x} |\tilde p_{\bar x}| + e^{-2x} \tilde p_{\bar x}\cdot e_1 \\
&&+ e^{-2 x} \frac {\tilde p_{\bar x}^\perp}{|\tilde p_{\bar x}|} \left(A +
\alpha(\psi(\bar x)\odot D^2\psi(\bar x)+ D\psi(\bar x)^T \odot D\psi(\bar x)
)\right)\frac {\tilde p_{\bar x}^\perp}{|\tilde p_{\bar x}|} \\
b &\ge& {\color{black} c}e^{-y} |p_{\bar y}| + e^{-2y} p_{\bar y}\cdot e_1 +
e^{-2y} \frac {p_{\bar y}^\perp}{|p_{\bar y}|} B \frac {p_{\bar
y}^\perp}{| p_{\bar y}|}
\end{eqnarray*}
where
$$\tilde p_{\bar x}=p_{\bar x}+\alpha D\psi(\bar x)^T\odot
\psi(\bar x).
$$
Remark that , since $D_\theta U=D_\theta V=1$, there exists
$\delta_0>0$ such that
$$\tilde p_{\bar x}\ge \delta_0>0\quad{\rm and}\quad p_{\bar y}\ge \delta_0>0.$$
Moreover $a,b$ satisfy the following equality
$$
a-b = \frac{\eta}{(T-t)^2} + K e^{Kt}
\frac{|\psi(\bar x)-\psi(\bar y)|^2}{2\varepsilon}
$$
and $A,B$ satisfy the following matrix inequality
\begin{multline*}
\left[\begin{array}{cc} A & 0 \\ 0 & -B \end{array} \right] \le
\frac{2 e^{Kt}}{\varepsilon}\bigg\{ \left[\begin{array}{cc} (\psi(\bar
x)-\psi(\bar y))\odot D^2\psi(\bar x) & 0 \\ 0 & -(\psi(\bar
x)-\psi(\bar y))\odot D^2\psi(\bar
y) \end{array} \right]\\
+\left[\begin{array}{cc} D\psi(\bar x)^T \odot D\psi(\bar x) &
-D\psi(\bar y)^T\odot D\psi(\bar x) \\ -D\psi(\bar y)^T\odot
D\psi(\bar x) & D\psi(\bar y)^T\odot D\psi(\bar y) \end{array}
\right]\bigg\} \, .
\end{multline*}
This implies
\begin{multline}\label{eq:ineqmatrix}
A\xi\cdot \xi\le B\zeta\cdot \zeta+ \frac {2e^{Kt}}\varepsilon
\bigg\{(\psi(\bar x)-\psi(\bar y))\odot D^2\psi(\bar x) \xi\cdot \xi-
(\psi(\bar x)-\psi(\bar y))\odot D^2\psi(\bar y) \zeta\cdot \zeta\nonumber\\
+\left|D\psi(\bar x)\xi-D\psi(\bar y)\zeta\right|^2\bigg\}
\end{multline}
for all $\xi,\zeta\in \hbox{\bf R}^2$. Combining the two viscosity inequalities
and using the fact that $|\psi(\bar x)-\psi(\bar y)|\ge m_\psi|\bar{x}-\bar{y}|$,
we obtain
\begin{align*}
\frac \eta{T^2}+ Km_\psi^2 e^{Kt}\frac {|\bar x-\bar y|^2}{2\varepsilon}\le & {\color{black} c}e^{-x} |\tilde
p_{\bar x}|- {\color{black} c}e^{-y} |p_{\bar y}| + e^{-2x} \tilde p_{\bar x}\cdot
e_1 - e^{-2y} p_{\bar y}\cdot e_1 \\
+& \alpha e^{-2 x} . \widehat {\tilde p_{\bar x}^\perp}
\left(\psi(\bar x)\odot D^2\psi(\bar x)+ D\psi(\bar x)^T
D\psi(\bar x) \right)\widehat {\tilde p_{\bar x}^\perp} \\
+& \frac {2e^{Kt}}\varepsilon (\tilde {\cal I}_1 + \tilde {\cal I}_2)
\end{align*}
where $\tilde {\cal I}_1$ and $\tilde {\cal I}_2$ are defined respectively as
${\cal I}_1$ and ${\cal I}_2$ with $p_{\bar x}$ replaced by $\tilde p_{\bar x}$.
Remarking that there exists a constant $C>0$ such that
\begin{eqnarray*} {\color{black} c}e^{-x} |\tilde p_{\bar x}| &+& e^{-2x} \tilde p_{\bar
x}\cdot e_1+\left|\alpha e^{-2 x} \frac {\tilde p_{\bar
x}^\perp}{|\tilde p_{\bar x}|}\left(\psi(\bar x)\odot
D^2\psi(\bar x)+ D\psi(\bar x)^T D\psi(\bar x) \right)\frac
{\tilde p_{\bar x}^\perp}{|\tilde p_{\bar x}|}\right|\\
&& \le {\color{black} c}e^{-x} | p_{\bar x}| + e^{-2x} p_{\bar
x}\cdot e_1+C\alpha \left( |D^2\psi(\bar x)|^2 + |D\psi(\bar x)|^2 + |\psi(\bar
x)|^2\right)\\
&& \le {\color{black} c}e^{-x} | p_{\bar x}| + e^{-2x} p_{\bar
x}\cdot e_1+3M_\psi^2 C \alpha
\end{eqnarray*}
and
\begin{align*}
|\tilde {\cal I}_1-{\cal I}_1|+|\tilde {\cal I}_2 -{\cal I}_2|\le& C \left|\widehat {\tilde
p_{\bar x}^\perp}-\widehat {p_{\bar x}^\perp}\right|\\
\le &C \left|\frac {\tilde p_{\bar x}-p_{\bar x}}{|\tilde p_{\bar x}|}\right|+|p_{\bar x}|\left|\frac
1{|\tilde p_{\bar x}|}-\frac 1{|p_{\bar x}|}\right|\\
\le &C \left|\frac {\tilde p_{\bar x}-p_{\bar x}}{\delta_0}\right|+\left|\frac
{|p_{\bar x}|-|\tilde p_{\bar x}|}{\delta_0}\right|\\
\le &2C \left|\frac {\tilde p_{\bar x}-p_{\bar x}}{\delta_0}\right|\\
\le&\frac{2C^2\alpha}{\delta_0}
\end{align*}
and sending $\alpha\to 0$ (recall that $\bar{x},\bar{y}$ lie in a compact domain), we get \eqref{eq:301}.
\end{proof}
\begin{lem}[Estimate on ${\mathcal I}_1$]\label{lem:tech1}
There exists a constant $\overline C_1$ such that
\begin{equation}\label{estim:i1}
|{\cal I}_1|\le \overline C_1 |x-y|^2 \end{equation}
\end{lem}
\begin{proof}
In order to prove \eqref{estim:i1}, we write
\begin{align*}
\frac{|{\cal I}_1|}{|\psi(\bar x)-\psi(\bar y))|} \le &
|(D^2\psi(\bar x)- D^2\psi(\bar y)) e^{-x} \widehat {p_{\bar x}^\perp}\cdot e^{-x} \widehat {p_{\bar x}^\perp}|\\
+&| D^2\psi(\bar y)
(e^{-x}-e^{-y}) \widehat {p_{\bar x}^\perp}\cdot e^{-x} \widehat{p_{\bar x}^\perp}|\\
+& |D^2\psi(\bar y) e^{-y} \left(\widehat {p_{\bar
x}^\perp}-\widehat{p_y^\perp}\right)\cdot e^{-x} \widehat
{p_{\bar x}^\perp}|\\
+& |D^2\psi(\bar y)
e^{-y} \widehat{p_y^\perp} \cdot (e^{-x}-e^{-y}) \widehat {p_{\bar x}^\perp}\\
+& |D^2\psi(\bar y) e^{-y} \widehat{p_{\bar y}^\perp}\cdot e^{-y}
\left(\widehat {p_{\bar x}^\perp}-\widehat{p_y^\perp}\right)|.
\end{align*}
Thanks to \eqref{eq:mpsi} and $\max(|x|,|y|) \le 2$, we have
\begin{eqnarray*}
|D^2\psi(\bar x)- D^2\psi(\bar y)| &\le& M_\psi |\bar x-\bar y|, \\
|e^{-x}-e^{-y}| &\le& e^2 |\bar x-\bar y|.
\end{eqnarray*}
We also have the following important estimate
\begin{align*}
\left|\widehat {p_{\bar x}^\perp}-\widehat {p_{\bar y}^\perp}\right|\le&
\left|\frac {p_{\bar x}-p_{\bar y}}{|p_{\bar x}|}\right|+|p_{\bar y}|\left|\frac
1{|p_{\bar x}|}-\frac 1{|p_{\bar y}|}\right|\\
\le & \left|\frac {p_{\bar x}-p_{\bar y}}{|p_{\bar x}|}\right|+\left|\frac
{|p_{\bar y}|-|p_{\bar x}|}{|p_{\bar x}|}\right|\\
\le & 2 \left|\frac {p_{\bar x}-p_{\bar y}}{|p_{\bar x}|}\right|\\
\le& 2\frac{\frac {e^{Kt}}\varepsilon|D\psi(\bar x)-D\psi(\bar y)||\psi(\bar
x)-\psi(\bar y)|}{\frac {e^{Kt}}\varepsilon m_\psi |\bar x-\bar y|}\\
\le & \frac{2 M_{\psi}^2}{m_\psi}|\bar x-\bar y|
\end{align*}
where we have used the fact that $|p_{\bar x}|\ge \frac {e^{Kt}}\varepsilon
m_\psi|\bar x -\bar y|$ (see \eqref{psi:grad-inf-bis}). This finally gives
that there exists a constant $\overline C_1$ (depending on $m_\psi$
and $M_\psi$) such that \eqref{estim:i1} holds true.
\end{proof}
Using the fact that $|p_{\bar x}|,|p_{\bar y}|\le C \frac {e^{Kt}}\varepsilon |\bar x-\bar
y|$, we can prove in a similar way the following lemma.
\begin{lem}[Remaining estimates]\label{lem:tech2}
There exist three positive constants $\overline C_2,\overline C_3$ and $\overline C_4$ such that
\begin{eqnarray*}
|{\cal I}_2| &\le & \overline C_2 |\bar x-\bar y|^2, \\
ce^{-x} |p_{\bar x}|- ce^{-y} |p_{\bar y}|&\le &\overline C_3 \frac
{e^{Kt}}\varepsilon |\bar x-\bar y|^2 ,\\
e^{-2x} p_{\bar x}\cdot e_1 - e^{-2y} p_{\bar y}\cdot e_1 &\le&
\overline C_4 \frac {e^{Kt}}\varepsilon |\bar x-\bar y|^2.
\end{eqnarray*}
\end{lem}
Use now Lemmas~\ref{lem:tech1} and \ref{lem:tech2} in
order to derive from \eqref{eq:301} the following inequality
$$
\frac \eta{T^2}+ Km_\psi e^{Kt}\frac {|\bar x-\bar
y|^2}{2\varepsilon}\le\overline C \frac {e^{Kt}}\varepsilon |\bar x-\bar y|^2
$$
with $\overline C=\overline C_1+\overline C_2+\overline C_3+\overline C_4$.
Choosing $K\ge \frac {2 \overline C}{m_\psi}$, we get a contradiction.
\end{proof}
\section{Construction of a classical solution}
\label{sec:classical}
{\color{black} In this section, our main goal is to prove Theorem \ref{th:cauchy}
which claims the existence and uniqueness of classical solutions
under suitable assumptions on the initial data $\bar u_0$.
Notice that assumptions \eqref{cond:compatibilite} on the initial data
imply in particular that
\begin{equation}\label{cond:compatibilite1}
c+2(\bar{u}_0)_r(0)=0 .\end{equation}}
To prove Theorem \ref{th:cauchy}, we first construct a unique weak (viscosity)
solution. We then prove gradient estimates from which it is not
difficult to derive that the weak (viscosity) solution is in fact
smooth; in particular, it thus satisfies the equation in a classical
sense.
\subsection{Barriers and Perron's method}\label{subsec:perron}
Before constructing solutions of \eqref{eq::4a} submitted to the
initial condition~\eqref{eq::4b}, we first construct appropriate
barrier functions.
\begin{pro}[Barriers for the Cauchy problem] \label{prop:barriers}
Assume that {\color{black} $\bar{u}_0\in W^{2,\infty}_{loc}(0,+\infty)$ and
$$(\bar{u}_0)_r \in W^{1,\infty} (0,+\infty)\quad \mbox{or}\quad \kappa_{\bar u_0}\in L^\infty(0,+\infty)$$}
with $\bar{u}_0$
such that \eqref{cond:compatibilite} holds true. {\color{black} Then} there exists a constant
$\bar{C}>0$ such that $\bar{u}^\pm (t,r)=\bar{u}_0 (r) \pm \bar{C}t$ are
respectively a super- and a sub-solution of
\eqref{eq::4a},\eqref{eq::4b}.
\end{pro}
\begin{proof}
It is enough to prove that the following quantity is finite
$$
\bar{C} = \sup_{r \ge 0} \frac1r \left| \bar{F}(r,(\bar{u}_0)_r (r), (\bar{u}_0)_{rr}(r))
\right| = \max (\bar{C}_1,\bar{C}_2)
$$
with
$$
\bar{C}_1 = \sup_{r \in [0,{\color{black} r_0}]} \frac{| \bar{F}(r,(\bar{u}_0)_r (r),
(\bar{u}_0)_{rr}(r))|}r, \quad
\bar{C}_2 = \sup_{r \in [{\color{black} r_0},+\infty)} \frac{| \bar{F}(r,(\bar{u}_0)_r (r) (x),
(\bar{u}_0)_{rr}(r))|}r.
$$
On one hand, thanks to \eqref{cond:compatibilite} and the Lipschitz
regularity of $\bar u_0$, we have $\bar{C}_1$ is finite. On the other
hand, thanks to Lipschitz regularity and $(\bar u_0)_r \in
W^{1,\infty}$ {\color{black} or $\kappa_{\bar u_0}\in L^\infty$}, $\bar{C}_2$ is also finite. The proof is now complete.
\end{proof}
We now construct a viscosity solution for
\eqref{eq::4a},\eqref{eq::4b}; this is very classical with the results
we have in hand, namely the strong comparison principle and the
existence of barriers. However, we give a precise statement and a
sketch of proof for the sake of completeness.
\begin{pro}[Existence by Perron's method]\label{pro:cauchy}
Assume that $\bar u_0\in C(0,+\infty)$ and that there
exists
{\color{black} $$u^+(t,r):=\bar u_0(r)+f(t) \quad \left(\mbox{resp.}\ u^-(t,r):=\bar u_0(r)-f(t)\right)$$
for some continuous function $f$ satisfying $f(0)=0$, which are respectively a
super- and a sub-solution of \eqref{eq::4a},\eqref{eq::4b}.} Then,
there exists a {\color{black} (continuous)} viscosity solution $\bar{u}$ of
\eqref{eq::4a},\eqref{eq::4b} such that {\color{black} \eqref{baru:baru0sst}} holds true for
some constant {\color{black} $\bar C_T$} depending on {\color{black} $f$}.
{\color{black} Moreover $\bar u$ is the unique viscosity solution of
\eqref{eq::4a},\eqref{eq::4b} such that \eqref{baru:baru0sst} holds
true.}
\end{pro}
\begin{proof}
In view of Lemma~\ref{lem:equiv}, it is enough to construct a
solution $u$ of \eqref{eq:polar-spir} satisfying \eqref{eq:ci} with
$u_0(x) = \bar{u}_0 ({\color{black} e^x})$.
Consider the set
$$
\mathcal{S} = \{ v : (0,+\infty) \times \hbox{\bf R} \to \hbox{\bf R}, \text{ sub-solution
of } \eqref{eq:polar-spir} \text{ s.t. } v \le u^+ \} \, .
$$
Remark that it is not empty since $u^- \in \mathcal{S}$ (where
$u^\pm(t,x)=\bar u^\pm(t,r)$ with $x=\ln r$). We now
consider the upper envelope $u$ of $(t,r)\mapsto \sup_{v\in {\mathcal
S}} v(t,r)$. By Proposition~\ref{pro:stability}, it is a
sub-solution of \eqref{eq:polar-spir}. The following lemma derives
from the general theory of viscosity solutions as presented in
\cite{cil92} for instance.
\begin{lem}\label{lem:bump}
The lower envelope $u_*$ of $u$ is a super-solution of
\eqref{eq:polar-spir}.
\end{lem}
We recall that the proof of this lemma proceeds by contradiction
and consists in constructing a so-called bump function around the
point the function $u_*$ is not a super-solution of the equation.
The contradiction comes from the maximality of $u$ in $\mathcal{S}$.
Since for all $v \in \mathcal{S}$,
$$
u_0 (x) - f(t) \le v \le u_0 (x) +f(t) ,
$$
with $f(0)=0$
we conclude that
$$
u_0(x) = u_* (0,x) = u (0,x) \, .
$$
If $\bar u$ satisfies \eqref{baru:baru0sst}, we use the comparison principle and get $u \le u_*$ in
$(0,T) \times \hbox{\bf R}$ for all $T>0$. Since $u_* \le u$ by construction, we deduce
that $u=u_*$ is a solution of \eqref{eq:polar-spir}.
The comparison principle also ensures that the solution we constructed
is unique. The proof of Proposition~\ref{pro:cauchy} is now complete.
\end{proof}
\subsection{Gradient estimates} \label{subsec:gradestim}
In this subsection, we derive gradient estimates for a viscosity
solution $\bar{u}$ of \eqref{eq::4a} satisfying \eqref{baru:baru0sst}.
\begin{pro}[Lipschitz estimates] \label{prop:gradestim} Consider a {\color{black} globally} Lipschitz continuous
function $\bar{u}_0$. We denote by $L_0>0$ and $L_1 >0$ such that
for all $r>0$,
$$
-L_0 \le (\bar{u}_0)_r (r) \le L_1.
$$
Let $\bar u$ be a viscosity solution $\bar{u}$ of
\eqref{eq::4a},\eqref{eq::4b} satisfying \eqref{baru:baru0sst}. Then $\bar
u$ is also Lipschitz continuous {\color{black} in space}: $\forall t>0$, $\forall r \ge
0$,
{\color{black} \begin{equation}\label{estim:gradient}
\left\{\begin{array}{ll}
-\max(1,L_0) \le \bar{u}_r (t,r)\le L_1 & \quad \mbox{if}\quad c\ge 0\\
\\
-L_0 \le \bar{u}_r (t,r)\le \max(1,L_1) & \quad \mbox{if}\quad c\le 0
\end{array}\right.
\end{equation}}
Moreover, if {\color{black} $\bar{u}_0\in W^{2,\infty}_{loc}(0,+\infty)$ with
$$(\bar{u}_0)_r\in W^{1,\infty}(0,+\infty)\quad \mbox{and}\quad \kappa_{\bar u_0}\in L^\infty(0,+\infty)$$}
and
\eqref{cond:compatibilite} holds true, then {\color{black} $\bar u$} is $\bar{C}$-Lipschitz
continuous with respect to $t$ for all $r >0$ where $\bar{C}$ denotes
the constant appearing in Proposition~\ref{prop:barriers}.
\end{pro}
\begin{proof} ~ \\
{\color{black} \noindent {\bf Step 1: gradient estimates}}\\
Proving \eqref{estim:gradient} {\color{black} for $c\ge 0$} is equivalent to prove that the
solution $u$ of \eqref{eq:polar-spir} satisfies the following gradient
estimate: $\forall t>0$, $\forall x \in \hbox{\bf R}$,
\begin{equation}\label{estim:gradient-x}
-\bar L_0e^x \le u_x (t,x)\le L_1 e^x
\end{equation}
where $\bar L_0=\max(1,L_0)$.
We will prove each inequality separately.
Since $\bar{u}$ is sublinear,
there exists $C_u>0$ such that for all $x \in \hbox{\bf R}$
$$
|u(t,x)| \le C_u (1 + e^x) .
$$
Eq.~\eqref{estim:gradient} is equivalent to prove
\begin{align*}
{\color{black} M^0} &= \sup_{t \in (0,T), x \le y \in \hbox{\bf R}} \left\{ u (t,x)+\bar L_0 e^x
- u (t,y) - \bar L_0 e^y \right\} \le 0 \\
{\color{black} M^1} &= \sup_{t \in (0,T), x \ge y \in \hbox{\bf R}} \left\{ u (t,x)-L_1 e^x
- u (t,y) + L_1 e^y \right\} \le 0.
\end{align*}
We first prove that ${\color{black} M^0} \le 0$. We argue by contradiction by assuming
that ${\color{black} M^0}>0$ and we exhibit a contradiction. The following supremum
$$
{\color{black} M_\alpha^0}= \sup_{t \in (0,T), x \le y \in \hbox{\bf R}} \left\{ u(t,x) - u(t,y) +
\bar L_0 e^x-\bar L_0 e^y -\frac
\alpha 2 x^2 - \frac \alpha 2 y^2
- \frac\eta{T-t} \right\}
$$
is also positive for $\alpha$ and $\eta$ small enough.
Using the fact that, by assumption on $\bar u_0$,
\begin{equation}\label{eq:borne1}
u(t,x)-u(t,y)+\bar L_0e^x-\bar L_0e^y\le
u(t,x)- u_0(x)+ u_0(x)- u_0(y)+ u_0(y)-u(t,y)+\bar L_0
e^x-\bar L_0 e^y\le 2C_1
\end{equation}
and the fact
that $-\frac \alpha 2 x^2-\frac \alpha 2 y^2\to -\infty $ as $x\to \pm \infty$
or $y\to \pm \infty$, we deduce that the supremum is achieved at a point
$(t,x,y)$ such that $t \in (0,T)$ and $x > y$.
Moreover, we deduce using \eqref{eq:borne1} and the fact that $M_\alpha>0$,
that there exists a constant $C_0:=4C_1$ such
that $x$ and $y$ satisfy the following inequality
$$
\alpha x^2+ \alpha y^2\le C_0.
$$
Thanks to Jensen-Ishii's Lemma (see e.g. \cite{cil92}), we conclude that
there exist $a,b,X,Y \in \hbox{\bf R}$ such that
\begin{eqnarray*}
a &\le& c e^{-x} \sqrt{1+ (-\bar L_0 e^x+\alpha x)^2} + e^{-2x} (-\bar
L_0e^x+\alpha x) + e^{-2x} \frac{X-\bar L_0e^x+ \alpha }
{1+ (-\bar L_0e^x+\alpha x)^2}, \\
b &\ge& c e^{-y} \sqrt{1+ (\bar L_0e^y+\alpha y)^2} - e^{-2y} (\bar
L_0e^y+\alpha y) + e^{-2y} \frac{Y-\bar L_0e^y-\alpha} {1+ (\bar L_0e^x+\alpha
y)^2}, \\
a-b &=& \frac{\eta}{(T-t)^2}, \qquad
\left[ \begin{array}{ll} X & 0 \\ 0 & -Y \end{array} \right] \le 0 .
\end{eqnarray*}
Subtracting the viscosity inequalities and using the last line yield
\begin{align*}
\frac{\eta}{T^2} \le& c e^{-x} \sqrt{1+ (\bar L_0e^x-\alpha x)^2}-c e^{-y} \sqrt{1+
(\bar L_0e^y+ \alpha y)^2}+ \alpha e^{-2x} (x+1)+ \alpha e^{-2y}
(y+1) \\
&-\bar L_0e^{-x}+\bar L_0e^{-y}-\frac {\bar L_0 e^{-x}} {1+
(\bar L_0e^x-\alpha x)^2}+ \frac{\bar L_0e^y} {1+ (\bar L_0e^x+\alpha y)^2}
\end{align*}
Using the fact that the functions $z\mapsto \sqrt{1+z^2}$ and $z\mapsto
\frac 1 {1+z^2}$ are $1$-Lipschitz, we deduce that
\begin{align*}
\frac{\eta}{T^2} \le& ce^{-x} \sqrt{1+ \bar L_0^2 e^{2x}}-ce^{-y} \sqrt{1+
\bar L_0^2 e^{2y}}+ e^{-x} \alpha(({\color{black} |c|}+\bar L_0)|x|+x e^{-x}+e^{-x})\\
&+ \alpha e^{-y}
(({\color{black} |c|}+\bar L_0)|y|+ y e^{-y} +e^{-y})
-\bar L_0e^{-x}+\bar L_0e^{-y}-\frac {\bar L_0 e^{-x}} {1+
(\bar L_0e^x)^2}+ \frac{\bar L_0e^y} {1+ (\bar L_0e^x)^2}
\end{align*}
Remarking that the function $z\mapsto e^{-z} (({\color{black} |c|}+\bar L_0)|z|+z
e^{-z}+e^{-z})$ is bounded from above by a constant $C_3$, we have
\begin{equation}\label{eq:g}
\frac{\eta}{T^2} \le 2 C_3 \alpha+ g(x)-g(y)
\end{equation}
where
$$g(x)=e^{-x} c \sqrt{1+ \bar L_0^2e^{2x}} -\bar L_0e^{-x}-\frac {\bar L_0
e^{-x}} {1+ \bar L_0^2e^{2x}}.
$$
{\color{black} \noindent {\bf Case A: $c\ge 0$}}\\
We now rewrite $g$ in the following way
\begin{align*}
g(x)=&c\sqrt{e^{-2x}+\bar L_0^2}-\bar L_0 e^{-x} -\frac {\bar L_0}{e^x+\bar
L_0^2 e^{3x}}\\
=&\frac{c^2 e^{-2x}(1-\bar L_0^2)-\bar L_0^2e^{-2x}}{c\sqrt{e^{-2x}+\bar
L_0^2}+\bar L_0 e^{-x}}-\frac {\bar L_0}{e^x+\bar
L_0^2 e^{3x}}\\
=&\frac{c^2(1-\bar L_0^2)}{c\sqrt{e^{2x}+\bar L_0^2e^{4x}}+\bar L_0
e^{x}}-\frac{\bar L_0^2}{\sqrt{e^{2x}+\bar L_0^2e^{4x}}+\bar L_0
e^{x}}-\frac {\bar L_0}{e^x+\bar
L_0^2 e^{3x}}
\end{align*}
and use the fact that $\bar L_0\ge 1$ to deduce that $g$ is
non-decreasing. Hence, we finally get
$$
\frac{\eta}{T^2} \le 2 C_3\alpha
$$
which is absurd for $\alpha$ small enough.
\bigskip
In order to prove that ${\color{black} M^1} \le 0$, we proceed as before and we obtain
\eqref{eq:g} where
{\color{black} $$g(x)=c \sqrt{e^{-2x}+ L_1^2} +L_1 e^{-x}+ \frac {L_1} {e^{x}+ L_1^2 e^{3x}}.
$$}
Remarking that $g$ is decreasing permits us to conclude in this case.\\
{\color{black} \noindent {\bf Case B: $c\le 0$}\\
We simply notice that the equation is not changed if we change $(w,c)$ in $(-w,-c)$.
}
\bigskip
{\color{black} \noindent {\bf Step 2: Lipschitz in time estimates}}\\
It remains to prove that $\bar{u}$ is $\bar{C}$-Lipschitz continuous
with respect to $t$ under the additional compatibility
condition~\eqref{cond:compatibilite}. To do so, we fix $h>0$ and we
consider the following functions:
$$
\bar{u}_h (t,r) = \bar{u} (t+h,r) - \bar{C} h \quad \text{ and }
\bar{u}^h (t,r) = \bar{u}(t+h,r) + \bar{C}h.
$$
Remark that $\bar{u}_h$ and $\bar{u}^h$ satisfy \eqref{eq::4a}. Moreover,
Proposition~\ref{prop:barriers} implies that
$$
\bar{u}_h(0,r) \le \bar{u}_0 (r) \le \bar{u}^h (0,r).
$$
Thanks to the comparison principle, we conclude that $\bar{u}_h \le
\bar{u} \le \bar{u}^h$ in {\color{black} $[0,+\infty)\times (0,+\infty)$}; since $h$ is arbitrary, we
thus conclude that $\bar{u}$ is $\bar{C}$-Lipschitz continuous with
respect to $t$. The proof of Proposition~\ref{prop:gradestim} is now
complete.
\end{proof}
\subsection{Proof of Theorem~\ref{th:cauchy}} \label{subsec:class}
It is now easy to derive Theorem~\ref{th:cauchy} from
Propositions~\ref{pro:cauchy} and \ref{prop:gradestim}.
\begin{proof}[Proof of Theorem~\ref{th:cauchy}]
Consider the viscosity solution $\bar{u}$ given by
Proposition~\ref{pro:cauchy}
{\color{black} with $f(t)=\bar C t$ where the constant $\bar C$
is given in the barrier presented in Proposition \ref{prop:barriers}.}
This function is
continuous. Moreover, thanks to Proposition~\ref{prop:gradestim},
$\bar{u}_t$ and $\bar{u}_r$ are bounded in the viscosity sense;
hence $u$ is Lipschitz continuous. In particular, there exists a set
$\tilde{N} \subset {\color{black} (0,+\infty)} \times (0,+\infty)$ of null measure such
that for all $(t,r) \notin \tilde{N}$, $\bar{u}$ is differentiable at
$(t,r)$.
Thanks to the equation
\begin{equation}\label{eq::vsd}
\displaystyle{\bar{u}_t-a(r,\bar{u}_r)\bar{u}_{rr}=f(r,\bar{u}_r)}
\quad \mbox{for}\quad (t,r)\in (0,+\infty)\times (0,+\infty)
\end{equation}
with
$$
a(r,\bar{u}_r)=\frac{1}{1 + r^2\bar{u}_r^2}, \quad
f(r,\bar{u}_r)=\frac{1}{r}\left\{c \sqrt{1 + r^2\bar{u}_r^2} + \bar{u}_r
\left(\frac{2+ r^2 \bar{u}_r^2}{1+ r^2 \bar{u}_r^2}\right)\right\}
$$
we also have that $\bar{u}_{rr}$ is locally bounded in the viscosity
sense. This implies that $\bar{u}$ is locally $C^{1,1}$ with respect
to $r$, and in particular, we derive from Alexandrov's theorem
\cite[p.~242]{eg92} that for all $t>0$ there exists a set $N_t \subset
[0,+\infty)$ of null measure such that for all $r \notin N_t$,
$\bar{u}(t,\cdot)$ is twice differentiable with respect to $r$,
i.e. there exist $p, A \in \hbox{\bf R}$ such that for $\rho$ in a neighborhood
of $r$, we have
\begin{equation}\label{eq:dl2}
\bar{u} (t,\rho) = \bar{u} (t,r) + p (\rho-r) + \frac12 A (\rho-r)^2 + o ((\rho-r)^2).
\end{equation}
From $\tilde{N}$ and $\{ N_t \}_{t >0}$, we can construct a set $N
\subset (0;+\infty) \times (0;+\infty)$ of null measure such that for
all $(t,r) \notin N$, $\bar{u}$ is differentiable with respect to time
and space at $(t,r)$ and there exists $A \in \hbox{\bf R}$ such that
\eqref{eq:dl2} holds true. We conclude that
$$
\bar{u} (s,\rho) = \bar{u} (t,r) + \partial_t \bar{u} (t,r) (s-t)
+ \partial_r \bar{u}(t,r) (\rho-r) + \frac12 A (\rho-r)^2 + o ((\rho-r)^2) + o (s-t).
$$
In particular, \eqref{eq::vsd} holds true for $(t,r) \notin N$.
We deduce from the previous discussion that $\bar{u}_t - \bar{u}_{rr}
=\tilde{f} \in L^\infty_{loc}$ holds true almost everywhere, and thus
in the sense of distributions. From the standard interior estimates
for parabolic equations, we get that $\bar{u} \in W^{2,1;p}_{loc}$ for
any $1<p<+\infty$. Then from the Sobolev embedding (see Lemma 3.3 in
\cite{LSU}), we get that for $p>3$, and $\alpha=1-3/p$, we have
$\bar{u}_r \in C^{\alpha,\alpha/2}_{loc}$.
We now use that \eqref{eq::vsd} holds almost everywhere.
Therefore we can apply the standard interior Schauder theory (in
H\"{o}lder spaces) for parabolic equations. This shows that $\bar
u\in C^{2+\alpha,1+\alpha/2}_{loc}$. Bootstrapping, we finally get
that $\bar u\in C^\infty_{loc}$, which ends the proof of the theorem.
\end{proof}
\section{Construction of a {\color{black} general} weak (viscosity) solution}
\label{sec:withoutcompatibility}
{\color{black} The main goal of this section is to prove Theorem \ref{th:cauchywc}.
We start with general barriers, H\"{o}lder estimates in time
and finally an approximation argument.}
\begin{pro}[Barriers for the Cauchy problem without the Compatibility
Condition] \label{prop:barriers2} Let $\bar{u}_0 \in {\color{black} W^{2,\infty}_{loc}(0,+\infty)}$
be such that there exists $C_0$ such that
\begin{equation}\label{eq:hypu0}
|(\bar u_0)_r|\le C_0 \quad{\rm and}\quad|\kappa_{\bar u_0}|\le C_0.
\end{equation}
Then, there exists a constant $\bar{C}>0$ (depending only on $C_0$)
such that for any function $B:[0,T]\to \hbox{\bf R}$ with $B(0)=0$
and $B'\ge \bar C(1+\bar Ct)$, $\bar{u}^\pm
(t,r)=\bar{u}_0 (r) \pm \frac{\bar C t}r \pm B(t)$ are respectively a
super- and a sub-solution of \eqref{eq::4a},\eqref{eq::4b}.
\end{pro}
\begin{proof}
We only do the proof for the super-solution since it is similar
(and even simpler) for the sub-solution. {\color{black} We also do the proof only in the case $c\ge 0$,
noticing that the equation is unchanged if we replace $(w,c)$ with $(-w,-c)$.}
It is convenient to write $A$ for $\bar C t$ and do the
computations with this function.
Since $|\kappa_{\bar u_0}|\le C_0$, we have
$$\left|\frac {r(\bar u_0)_{rr}}{(1+(r(\bar u_0)_r)^2)^{\frac 32}} + (\bar u_0)_r \left(\frac
{2+(r(\bar u_0)_r)^2}{(1+(r(\bar u_0)_r)^2)^{\frac 32}}\right)\right|\le C_0.
$$
{\color{black} Since $|u_r|\le C_0$}, there exists $c_1>0$ such that
$$\left|r(\bar u_0)_{rr}\right|\le c_1(1+(r(\bar u_0)_r)^2)^{\frac 32}.$$
We then have
\begin{align*}
\bar F(r,\bar u^+_r, \bar u^+_{rr})=&c\sqrt{1+(r \bar u^+_r)^2} +
\bar u_r
\left(\frac {2+ (r\bar u_r^+)^2}{1+(r\bar u_r^+)^2}\right)+\frac {r
\bar u_{rr}^+}{1+(r\bar u_r^+)^2}\\
=&c\sqrt{1+\left(\frac {-A}r+r(\bar u_0)_r\right)^2} +
\left(\frac{-A}{r^2}+(\bar u_0)_r\right)\left(\frac {2+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}{1+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}\right)\\
&+ \frac{\frac {2A}{r^2}+r(\bar u_0)_{rr}}{1+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}\\
\le &c( 1+ \frac A r + r|(\bar u_0)_r|) + (\bar u_0)_r \left(\frac {2+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}{1+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}\right) \\
&-\frac A {r^2} \left(\frac {\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}{1+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}\right) + c_1 \frac
{(1+(r(\bar u_0)_r)^2)^{\frac 32}}{1+\left(\frac {-A}r+r(\bar u_0)_r\right)^2}.
\end{align*}
Using \eqref{eq:hypu0}, we can write
$$(\bar u_0)_r \frac {2+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}{1+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}\le 2C_0$$
we get
\begin{align*}
\bar F(r,\bar u^+_r, \bar u^+_{rr})\le& c(1+ \frac A r +C_0r) + 2 C_0 -\frac
A {r^2} \left(\frac {\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}{1+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}\right) + c_1 \frac
{(1+(r(\bar u_0)_r)^2)^{\frac 32}}{1+\left(\frac {-A}r+r(\bar u_0)_r\right)^2}\\
\le& c(1+ \frac A r +C_0r) + 2C_0 -\frac
A {r^2} \left(\frac {\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}{1+\left(\frac
{-A}r+r(\bar u_0)_r\right)^2}\right) + c_1 \frac
{(1+r|(\bar u_0)_r|)^{3}}{1+\left(\frac {-A}r+r(\bar u_0)_r\right)^2}.
\end{align*}
We now set $\rho$ such that $r(\bar u_0)_r=\rho\frac A r$ and distinguish two
cases:
\medskip
\noindent{\bf Case 1: $\frac 1 2 <\rho <2$.} In this case,
\begin{align*}
\bar F(r,\bar u^+_r, \bar u^+_{rr})\le& c(1+ 2 C_0r +C_0r) +2C_0 + c_1\left(
1+r|(\bar u_0)_r|\right)^3\\
\le&c+ 3 c C_0 r+2C_0+ 4 c_1+ 4c_1r^3|(\bar u_0)_r|^3\\
\le&c+ 3 c C_0 r+2C_0+ 4 c_1+ 4c_1 \rho \frac A r r^2C_0^2 \\
\le&(c+2C_0 + 4 c_1)+ r( 3 cC_0 + 8c_1 C_0^2 A) \\
\end{align*}
where for the second line, we have used the fact that for $a,b\ge 0$,
$(a+b)^3\le 4 (a^3+b^3)$. On the other hand, we have $r\bar u_t^+=A'+rB'$.
Choosing $\bar C\ge \max (c+2C_0 + 4 c_1, 3 C_0+ 8c_1 C_0^2)$ we
get the desired result in this case.
\bigskip
\noindent{\bf Case 2: $\rho\le \frac 1 2$ or $\rho \ge2$.} In this case
$$
\frac {(1+r|(\bar u_0)_r|)^{3}}{1+\left(\frac {-A}r+r(\bar u_0)_r\right)^2}\le\frac{
4+4 r^3|(\bar u_0)_r |^3}{1+(\rho-1)^2\frac {A^2}{r^2}} \le 4 + 4 \frac
{\rho^2r|(\bar u_0)_r|}{(\rho-1)^2}\le 4+ 16 C_0 r.
$$
Then
$$
\bar F(r,\bar u^+_r, \bar u^+_{rr})\le c+ c\frac A r +2C_0 + 4c_1+
c C_0r +16c_1C_0r -\frac A {r^2} \frac {(\rho-1)^2 \left(\frac
{A}r\right)^2}{1+(\rho-1)^2\left(\frac {A}r\right)^2}
$$
We distinguish two sub-cases:
\medskip
\noindent{\sc Subcase 2.1: $\frac A r \le 2$}.
In this sub-case, we get
$$
\bar F(r,\bar u^+_r, \bar u^+_{rr})\le(3c+ 2C_0+4c_1)+r(cC_0+16c_1C_0)
$$
and we obtain the desired result taking $\bar C\ge\max(3c+
2C_0+4c_1,cC_0+16c_1C_0)$.
\medskip
\noindent{\sc Subcase 2.2: $\frac A r \ge 2$}.
In this subcase, $|\rho-1|\frac A r\ge 1$ and
$$\frac {(\rho-1)^2 \left(\frac
{A}r\right)^2}{1+(\rho-1)^2\left(\frac
{A}r\right)^2}\ge \frac 1 2 $$
and thus
\begin{align*}
\bar F(r,\bar u^+_r, \bar u^+_{rr})\le& (c+2 C_0 + 4c_1) + A(\frac c
r- \frac 1 {2r^2}) +cC_0r+16c_1C_0r \\
\le &(c+2C_0+ 4c_1)+ (dA +cC_0+16c_1C_0)r
\end{align*}
where for the last line we have used the fact that we can find
$d >0$ (only depending on $c$) such that $\frac c r-
\frac {1}{2r^2} \le d r$ for all $r >0$. We finally get the desired
result taking $\bar C\ge\max(c+2C_0+ 4c_1,cC_0+16c_1C_0,d)$. The
proof is now complete.
\end{proof}
\begin{pro}[Time H\"older estimate -- (I)]\label{pro:holder1}
Let $\bar u_0\in {\color{black} W^{2,\infty}_{loc}(0,+\infty)}$ satisfying
\eqref{eq:hypu0}. Let $\bar u$ be a solution of
\eqref{eq::4a},\eqref{eq::4b} satisfying \eqref{baru:baru0sst}. If
$\bar u$ is $L_0$-Lipschitz continuous with respect to the variable
$r$, then there exists a constant $C$, depending only on $C_0$ and
$L_0$ such that
$$
|\bar u(t,r)-\bar u_0(r)|\le C\sqrt t + B(t)
$$
where $B$ is defined in Proposition \ref{prop:barriers2}.
\end{pro}
\begin{rem}\label{rem:holder}
Let us note that in Proposition \ref{prop:barriers2}, we can choose
$B(t)=\bar C t (1+\frac{\bar C}2 t)$. Hence, we deduce from
Proposition \ref{pro:holder1} that there exists
$C>0$ such that for all $t \in [0,1]$,
\begin{equation}\label{eq::rm3}
|\bar u(t,r)-\bar u_0(r)|\le C\sqrt t.
\end{equation}
\end{rem}
\begin{proof}
Let $r_0>0$. Using Proposition \ref{prop:barriers2} and the
comparison principle, we deduce that there exists a
constant $\bar C$ and a function $B$ such that
$$\left| \bar u(t,r_0)-\bar u_0(r_0)\right|\le \bar C
\frac t{r_0} +B(t).$$
Since $\bar u$ is $L_0$-Lipschitz continuous in $r$, we also have
$$
|\bar u(t,0)-\bar u(t,r_0)|\le L_0r_0\quad{\rm and}\quad
|\bar u_0(0)-\bar u_0(r_0)|\le C_0 r_0.
$$
Combining the previous inequalities, we get that
$$
|\bar u(t,0)-\bar u_0(0)|\le (L_0+C_0) r_0+ \bar C \frac t {r_0} + B(t).
$$
Taking the minimum over $r_0$ in the right hand side, we get that
$$
|\bar u(t,0)-\bar u_0(0)|\le C_1 \sqrt t + B(t)
$$
with $C_1:=2\sqrt{(C_0+L_0)\bar C}$.
We finally deduce that
$$|\bar u(t,r)-\bar u_0(r)|\le \min\left\{\bar C\frac t r + B(t), C_1
\sqrt t + B(t) +(L_0+C_0) r\right\}.$$
The desired result is obtained by remarking that, if $r\le \sqrt t$,
then $C_1 \sqrt t + B(t) +(L_0+C_0) r\le (C_1+L_0+C_0)\sqrt t+ B(t)$, while
if $r\ge \sqrt t$, then $\bar C\frac t r + B(t)\le \bar C \sqrt t+ B(t)$.
\end{proof}
The next proposition asserts that the previous proposition is still
true if we do not assume that $\bar u$ is Lipschitz continuous with
respect to $r$.
\begin{pro}[{\color{black} Existence and time} H\"older estimate -- (II)]\label{pro:holder2}
Let $\bar u_0\in {\color{black} W^{2,\infty}_{loc}(0,+\infty)}$ satisfying
\eqref{eq:hypu0}. {\color{black} Then there exists a solution $\bar u$ of
\eqref{eq::4a},\eqref{eq::4b} satisfying \eqref{baru:baru0sst}. Moreover
there exists a constant $C$, depending only on $C_0$ such
that}
$$|\bar u(t,r)-\bar u_0(r)|\le C\sqrt t + B(t)$$
where $B$ is defined in Proposition \ref{prop:barriers2},
{\color{black}and there exists a constant $L_0$ (only depending on $C_0$) such that
$$|\bar u(t,r+\rho)-\bar u(t,r)|\le L_0|\rho|.$$}
\end{pro}
\begin{proof}
The initial datum is approximated with a sequence of initial data
satisfying \eqref{eq:hypu0} and the compatibility
condition~\eqref{cond:compatibilite}; passing to the limit will
give the desired result.
We can assume without loss of generality that $C_0 \ge \frac{c}2$.
Then we consider
$$
\bar u_0^\varepsilon=\Psi_\varepsilon U_0+(1-\Psi_\varepsilon)\bar u_0
$$
where $U_0\in C^\infty$ is such that
\begin{equation}\label{eq::rm5}
U_0(0)=\bar u_0(0), \quad (U_0)_r(0)=-\frac c2,\quad |(U_0)_r|\le C_0,\quad
r|(U_0)_{rr}|\le C_0 \quad {\rm for}\quad r\le 2
\end{equation}
and
$$
\Psi_\varepsilon(r)=\Psi_1\left(\frac r \varepsilon\right)
$$
where the non-increasing function $\Psi_1\in C^\infty$ satisfies
$$
\Psi_1=\left\{\begin{array}{lll}
1&{\rm if}& r\le 1,\\
0&{\rm if}& r\ge 2.
\end{array}\right.
$$
\begin{claim}\label{claim}
The initial condition $\bar u_0^\varepsilon$ satisfies the compatibility
condition~\eqref{cond:compatibilite} and \eqref{eq:hypu0} for some
constant $C_0$ which does not depend on $\varepsilon$.
\end{claim}
Let $u^\varepsilon$ denote the unique solution of \eqref{eq::4a} with initial
condition $\bar u_0^\varepsilon$ given by Proposition~\ref{pro:cauchy},
{\color{black} using the barrier (Proposition \ref{prop:barriers2}) provided by the Claim \ref{claim}.}
In
particular, $u^\varepsilon$ satisfies \eqref{baru:baru0sst} for some constant $\bar
C^\varepsilon$ depending on $\varepsilon$. Using Proposition~\ref{prop:gradestim}, we
deduce that $\bar u^\varepsilon$ is $L_0$-Lipschitz continuous with
$L_0:=\max(1,C_0)$. Then Proposition~\ref{pro:holder1} can be applied
to obtain the existence of a constant $C$ (depending only on $C_0$,
{\color{black} because $L_0$ now depends on $C_0$}) such
that for all $\varepsilon$
$$|\bar u^\varepsilon(t,r)-\bar u_0^\varepsilon(r)|\le C \sqrt t + B(t).$$
Taking $\varepsilon\to 0$ and using the stability of the solution and the
uniqueness of \eqref{eq::4a},\eqref{eq::4b}, we finally deduce the
desired result.
\end{proof}
We now prove the claim.
\begin{proof}[Proof of Claim~\ref{claim}.]
We have
$$(\bar u_0^\varepsilon)_r=(\Psi_\varepsilon)_r (U_0-\bar u_0)+\Psi_\varepsilon(U_0)_r+(1-\Psi_\varepsilon)(u_0)_r.
$$
Hence, since $(\Psi_\varepsilon)_r(0)=0$ and $\Psi_\varepsilon(0)=1$, we get
$$(\bar u_0^\varepsilon)_r(0)=(U_0)_r(0)=-\frac c2$$
which means that $\bar u_0^\varepsilon$ satisfies \eqref{cond:compatibilite1}.
{\color{black} Using the fact that
$\bar u_0^\varepsilon\in W^{2,\infty}_{loc}$
and (\ref{eq::rm5}), we get \eqref{cond:compatibilite}.}
Since $U_0(0)=\bar u_0(0)$ and $U_0$ and $\bar u_0$ are $C_0$-Lipschitz
continuous, we have
$$
|U_0(r)-\bar u_0(r)|\le 2 C_0 r.
$$
Let $c_1$ denote $ \sup_{\rho\ge 0} \rho
|(\Psi_1)_r(\rho)|<+\infty$. We then have
$$
|(\Psi_\varepsilon)_r (U_0-\bar u_0)|\le 2 C_0 \frac r \varepsilon
\left|(\Psi_1)_r\left(\frac r\varepsilon\right)\right| \le 2C_0 c_1.
$$
Hence
$$
|(\bar u_0^\varepsilon)_r|\le 2C_0(c_1+1).
$$
Let us now obtain an estimate on $\kappa_{\bar u_0^\varepsilon}$. Using the previous
bound, we only have to estimate
$$
\frac
{r(\bar u_0^\varepsilon)_{rr}}{\left(1+(r(\bar u_0^\varepsilon)_r)^2\right)^{\frac 32}}.
$$
If $r >2$, then $\bar u_0^\varepsilon=\bar u_0$ and the estimate follows from \eqref{eq:hypu0}.
If $r\le 2$, it is enough to estimate $r(\bar u_0^\varepsilon)_{rr}$. We have
$$
r(\bar u_0^\varepsilon)_{rr}=r(\Psi_\varepsilon)_{rr}(U_0-\bar u_0)+2r(\Psi_\varepsilon)_r((U_0)_r-(\bar u_0)_r)+r\Psi_\varepsilon
(U_0)_{rr}+r(1-\Psi_\varepsilon)(\bar u_0)_{rr}.
$$
Moreover there exists a constant $c_2$ (depending only on $C_0$) such
that for all $r \le 2$, $r|(\bar u_0)_{rr}|\le c_2$. Let $c_3$ denote
$\sup_{\rho\ge 0} \rho ^2|(\Psi_1)_{rr}(\rho)|<+\infty$. We then
have
$$
r|(\Psi_\varepsilon)_{rr} (U_0-\bar u_0)|\le 2 C_0 \frac {r^2} {\varepsilon^2}
\left|(\Psi_1)_{rr}\left(\frac r\varepsilon\right)\right| \le 2C_0 c_3.
$$
We finally deduce that for $r\le 2$,
$$
|r(\bar u_0^\varepsilon)_{rr}|\le 2C_0c_3+4C_0c_1+C_0+c_2
$$
which proves that $\bar u_0^\varepsilon$ satisfies \eqref{eq:hypu0} with a constant
$\bar C_0=2C_0c_3+4C_0c_1+C_0+c_2$ depending only on $C_0$.
\end{proof}
We now turn to the proof of Theorem \ref{th:cauchywc}.
\begin{proof}[Proof of Theorem \ref{th:cauchywc}]
{\color{black} The existence of $\bar u$ and its Lipschitz
continuity with respect to $r$ follows from
Proposition~\ref{pro:holder2}.
The uniqueness (and continuity) of $\bar u$
follows from the comparison principle
(Theorem~\ref{thm comp spirales sous lin en r}).}
Let us now prove that $\bar u$ is $\frac 12$-H\"older continuous
with respect to time. By Remark \ref{rem:holder}, there exists a
constant $C$ such that for $h\le 1$
$$|\bar u(h,r)-\bar u_0(r)|\le C \sqrt h.$$
{\color{black} with $C$ given in (\ref{eq::rm3}).
Proceeding as in Step 2 of the proof of Proposition \ref{prop:gradestim}, we get for $0\le h\le 1$:
$$\bar u(t+h, r)-\bar u(t,r)\le C\sqrt h.$$
The reverse inequality is obtained in the same way.
This implies (\ref{eq::rm1}). The proof is now complete.}
\end{proof}
|
\section{Introduction}
Due to limited computing power of standard serial computers, parallel
computing has become indispensable for investigating complex problems
in all fields of science. A frequently encountered question is how to
transform an existing serial scientific code into a new form that is
executable on a parallel computing platform. Although portable
parallel programming standards, such as MPI and OpenMP, have greatly
simplified the programming work, the task of parallelization may still
be quite complicated for domain scientists. This is because inserting
MPI calls or OpenMP directives directly into an existing serial code
often requires extensive code rewrite as well as detailed knowledge of
and experience with parallel programming.
The hope for non-specialists in parallel computing is that many
scientific applications possess high-level parallelism. That is, the
entire computational work can be decomposed into a set of individual
(and often collaborative) computational tasks, each of coarse grain,
and can be performed by an existing piece of serial code. Depending
on the specific application, the decomposition can be achieved by
identifying a set of different parameter combinations, or (fully or
almost) independent computations, or different data groups, or
different geometric subdomains. For a given type of decomposition,
the parallelization induced programming components, such as work
partitioning, domain partitioning, communication, load balancing, and
global administration, are often generic and independent of specific
applications. These generic components can thus be implemented as
reusable parallelization libraries once and for all. This is what we
exemplify in the present paper.
It is clear that a user-friendly parallelization approach relies on at
least two factors: (1) The existing serial code should be extensively
reused; (2) The programming effort by the end user must be limited.
To achieve these goals we suggest to use Python to wrap up pieces of
existing serial code (possibly written in other languages), and
implement the parallelization tasks in separate and generic Python
functions.
Python~\cite{python} is an extremely expressive and flexible
programming language at its core. The language has been extended with
numerous numerical and visualization modules such as
NumPy~\cite{NumPy} and SciPy~\cite{scipy}. The two requirements of a
user-friendly parallelization mentioned above are actually well met by
Python. First of all, Python is good at inter-operating with other
languages, especially Fortran, C, and C++, which are heavily used in
scientific codes. Using wrapper tools such as F2PY~\cite{f2py}, it is
easy to wrap up an existing piece of code in Fortran and C and provide
it with a Pythonic appearance.
Moreover, among its many strong features, Python is extremely flexible
with handling functions. Python functions accept both positional
arguments and keyword arguments. The syntax of a variable set of
positional and keyword arguments (known as ``\emp{(*args,**kwargs)}''
to Python programmers) allows writing libraries routines that work
with any type of user-defined functions. That is, the syntax makes it
possible to call a Python function without revealing the exact number
of arguments.
It is also straightforward to pass functions as input arguments to a
Python function and/or return a function as output. A callable class
object in Python can be used as if it were a stand-alone
function. Such a construction, or alternatively a closure (known from
functional programming), can be used to create functions that carry a
state represented through an arbitrarily complex data structure. The
result is that one can express the flow of a scientific code as a
Python program containing a set of calls to user-defined Python
functions. These user-defined functions can be ordinary functions or
classes that wrap pieces of the underlying scientific code. This is
what we call a \emph{function-centric} representation of the
scientific code. With such a function-centric approach, we can build a
general framework in Python for almost automatic parallelization of
the program flow in the original scientific code. Later examples will
convey this idea in detail.
Performance of the resulting parallel application will closely follow
the performance of the serial application, because the overhead of the
parallelization layer in Python is just due to a small piece of extra
code, as we assume the main computational work to take place in the
Python functions that call up pieces of the original scientific code.
In the parallelization layer, good performance can be ensured by using
efficient array modules in Python (such as \emp{numpy}~\cite{NumPy})
together with light-weight MPI wrappers (such as
\emp{pypar}~\cite{Pypar}). For examples of writing efficient Python
code segments for some standard serial and parallel numerical
computations, we refer the reader to Cai et al.~\cite{Cai2005SP}.
\paragraph{Related Work.}
In C++, generic programming via templates and object-oriented
programming has been applied to parallelizing serial scientific codes.
Two examples can be found in~\cite{SC.5.Cai.2002.a}
and~\cite{Cirak2008}, where the former uses C++ class hierarchies to
enable easy implementation of additive Schwarz preconditioners, and
the latter uses C++ templates extensively to parallelize finite
element codes. Many scientific computing frameworks have also adopted
advanced programming to incorporate parallelism behind the scene. In
these frameworks (see, e.g.,
\cite{petsc-web-page,ParFUM,Sierra,Trilinos,UG}) the users can write
parallel applications in a style quite similar to serial programming,
without being exposed to many parallelizing details. Likewise are
frameworks that are specially designed to allow coupling of different
serial and parallel components, such as Cactus~\cite{Cactus} and
MpCCI~\cite{MpCCI}. The Python programming language, however, has not
been widely used to parallelize existing serial codes. The Star-P
system~\cite{Star-P} provides the user with a programming environment
where most of the parallelism is kept behind the
scene. Hinsen~\cite{Hinsen07} has combined Python with BSP to enable
high-level parallel programming. In addition, quite a number of Python
MPI wrappers exist, such as \emp{pyMPI}~\cite{pyMPI},
\emp{pypar}~\cite{Pypar}, \emp{MYMPI}~\cite{myMPI},
\emp{mpi4py}~\cite{mpi4python,mpi4python2}, and
\emp{Scientific.MPI}~\cite{Scientific}. Efforts in incorporating
parallelism via language extensions of Python can be found
in~\cite{Benson07,cp-stackless-python,python-coarray}.
The contribution of the present paper is to show by examples that a
function-centric approach using Python may ease the task of parallel
scientific programming. This result is primarily due to Python's
flexibility in function handling and function arguments. As a result,
generic tasks that arise in connection with parallelization can often
be programmed as a collection of simple and widely applicable Python
functions, which are ready to be used by non-specialists to
parallelize their existing serial codes.
This paper contains three examples with different algorithmic
structures. A wide range of problems in science can be attacked by
extending and adapting the program code in these examples. Moreover,
readers whose problems are not covered by the examples will hopefully
from these examples understand how we solve programming problems by
identifying the principal, often simplified, underlying algorithmic
structure; then creating generic code to reflect the structure; and
finally applying the generic code to a specific, detailed case. Our
approach is much inspired by the success of mathematics in problem
solving, i.e., detecting the problem's principal structure and
devising a generic solution makes complicated problems tractable.
With Python as tool, we demonstrate how this strategy carries over to
parallelization of scientific codes.
The remainder of the paper is organized as follows. We give in
Section~\ref{sect:motive} a simple but motivating example, explaining
the principles of splitting a problem into a set of function calls
that can easily be parallelized. Generic parallelization of three
common types of real scientific applications are then demonstrated in
Section~\ref{sect:highlevel}. Afterwards, Section~\ref{sect:results}
reports the computational efficiency of the suggested parallelization
approach applied to specific cases in the three classes of scientific
problems. Some concluding remarks are given in
Section~\ref{sect:conclusion}.
\section{A Motivating Simple Example}
\label{sect:motive}
\subsection{Serial Version}
\label{sect:motive:serial}
Suppose we want to carry out a parameter analysis that involves a
large number of evaluations of a multi-variable mathematical function
$f(a_1,\ldots,a_q)$. The Python implementation of $f$ may use $p$
positional arguments and $k$ keyword arguments such that the total
$p+k$ arguments contain at least the variables $a_1,\ldots,a_q$ (i.e.,
$q\leq p+k$). As a very simple example, consider the parabola
$f(x,a,b,c)=ax^2+bx+c$ with the following Python implementation ($q=4,
p=1, k=3$): {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def func(x, a=0, b=0, c=1):
return a*x**2+b*x+c
\end{verbatimtab}
\end{quote}
}
\noindent
Suppose we want to evaluate \emp{func} for a particular set of input
parameters chosen from a large search space, where $x$, $a$, $b$, and
$c$ vary in specified intervals. The complete problem can be
decomposed into three main steps: (1) initialize a set of arguments to
\emp{func}; (2) evaluate \emp{func} for each entry in the set of
arguments; (3) process the set of function return values from all the
\emp{func} calls.
Step (1) calls a user-defined function \emp{initialize} which returns
a list of 2-tuples, where each 2-tuple holds the positional and
keyword arguments (as a tuple and a dictionary) for a specific call to
\emp{func}. Step (2) iterates over the list from step (1) and feed the
positional and keyword arguments into \emp{func}. The returned value
(tuple) is stored in a result list. Finally, step (3) processes the
result list in a user-defined function \emp{finalize} which takes this
list as input argument.
A generic Python function that implements the three-step parameter
analysis can be as follows: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def solve_problem(initialize, func, finalize):
input_args = initialize()
output = [func(*args, **kwargs) for args, kwargs in input_args]
finalize(output)
\end{verbatimtab}
\end{quote}
} Note that the use of {\em list comprehension} in the above code has
given a very compact implementation of the \emp{for}-loop for going
through all the evaluations of \emp{func}. The \emp{initialize},
\emp{func}, and \emp{finalize} functions are passed to
\emp{solve\_problem} as input arguments. These three user-defined
functions are independent of \emp{solve\_problem}.
As an example, assume that \emp{x} is a set of $n$ uniformly
distributed coordinates in $[0,L]$, and we vary $a$ and $b$ in
$[-1,1]$ each with $m$ values, while $c$ is fixed at the value 5. For
each combination of $a$ and $b$, we call \emp{func} with the vector
\emp{x} as a positional argument and the $a$, $b$, $c$ values as
keyword arguments, and store the evaluation results of \emp{func} in a
list named \emp{output}. The objective of the computations is to
extract the $a$ and $b$ values for which \emp{func} gives a negative
value for one or several of the coordinates $x\in [0,L]$. For this
very simple example, the concrete implementation of the
\emp{initialize} and \emp{finalize} functions can be put inside a
class named \emp{Parabola} as follows: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
class Parabola:
def __init__(self, m, n, L):
self.m, self.n, self.L = m, n, L
def initialize(self):
x = numpy.linspace(0, self.L, self.n)
a_values = numpy.linspace(-1, 1, self.m)
b_values = numpy.linspace(-1, 1, self.m)
c = 5
self.input_args = []
for a in a_values:
for b in b_values:
func_args = ([x], {'a': a, 'b': b, 'c': c})
self.input_args.append(func_args)
return self.input_args
def func(self, x, a=0, b=0, c=1):
return a*x**2+b*x+c
def finalize(self, output_list):
self.ab = []
for input, result in zip(self.input_args, output_list):
if min(result) < 0:
self.ab.append((input[0][1], input[0][2]))
\end{verbatimtab}
\end{quote}
}
Now, to find the combinations of $a$ and $b$ values that make
$ax^2+bx+c<9$, we can write the following two lines of code (assuming
$m=100$, $n=50$, and $L=10$): {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
problem = Parabola(100, 50, 10)
solve_problem(problem.initialize, problem.func, problem.finalize)
\end{verbatimtab}
\end{quote}
}
\noindent Note that the desired combinations of $a$ and $b$ values
will be stored in the list \emp{problem.ab}. Also note that we have
placed \emp{func} inside class \emp{Parabola}, to have all pieces of
the problem in one place, but having \emp{func} as stand-alone
function or a class method is a matter of taste.
Despite the great mathematical simplicity of this example, the
structure of the \emp{solve\_problem} function is directly applicable
to a wide range of much more advanced problems. Although
\emp{initialize} and \emp{finalize} are Python functions with very
simple arguments (none and a list, respectively), this is not a
limitation of their applicability. For example, the \emp{initialize}
step in our simple example needs values for $m$, $n$, and $L$, the $a$
and $b$ interval and so on, which can not be specified in the generic
\emp{solve\_problem} function. To overcome this limitation, the
information of $m$, $n$, and $L$ can be hard-coded (not recommended),
or transferred to \emp{initialize} through global variables (not
recommended in general) or carried with \emp{initialize} as a state,
either as class attributes or as a surrounding scope in a closure. We
have chosen the class approach, i.e., class attributes store
user-dependent data structures such that the \emp{initialize} and
\emp{finalize} methods can have the simple input argument structure
demanded by the generic \emp{solve\_problem} function. Alternatively,
a closure as follows can be used instead of a class (this construct
requires some knowledge of Python's scoping rules): {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def initialize_wrapper(m, n, L):
def initialize(self):
x = numpy.linspace(0, L, n)
a_values = numpy.linspace(-1, 1, m)
...
return input_args
return initialize
\end{verbatimtab}
\end{quote}
} Now, the returned \emp{initialize} function will carry with it the
values of $m$, $n$, and $L$ in the surrounding scope. The choice
between the class approach and the closure approach, or using global
variables in a straightforward global \emp{initialize} function, is up
to the programmer. The important point here is that \emp{initialize}
must often do a lot, and the input information to \emp{initialize}
must be handled by some Python construction. Similar comments apply to
\emp{finalize}.
\subsection{Parallel Version}
\label{sect:motive:parallel}
Let us say that we want to utilize several processors to share the
work of all the \emp{func} evaluations, i.e., the \emp{for}-loop in
the generic \emp{solve\_problem} function. This can clearly be
achieved by a task-parallel approach, where each evaluation of
\emp{func} is an independent task. The main idea of parallelization
is to split up the \emp{for}-loop into a set of shorter
\emp{for}-loops, each assigned to a different processor. In other
words, we need to split up the \emp{input\_args} list into a set of
sub-lists for the different processors. Note that this partitioning
work is generic, independent of both the \emp{func} function and the
actual arguments in the \emp{input\_args} list. Assuming homogeneous
processors and that all the function evaluations are equally
expensive, we can divide the \emp{input\_args} list into
\emp{num\_procs} (number of processors) sub-lists of equal length. In
case \emp{input\_args} is not divisible by \emp{num\_procs}, we adjust
the length of some sub-lists by 1:
{
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def simple_partitioning(length, num_procs):
sublengths = [length/num_procs]*num_procs
for i in range(length
sublengths[i] += 1
return sublengths
def get_subproblem_input_args(input_args, my_rank, num_procs):
sub_ns = simple_partitioning(len(input_args), num_procs)
my_offset = sum(sub_ns[:my_rank])
my_input_args = input_args[my_offset:my_offset+sub_ns[my_rank]]
return my_input_args
\end{verbatimtab}
\end{quote}
} Using the above generic \emp{get\_subproblem\_input\_args} function,
each processor gets its portion of the global \emp{input\_args} list,
and a shorter \emp{for}-loop can be executed there. Note that the
syntax of Python lists and \emp{numpy} arrays has made the function
very compact.
The next step of parallelization is to collect the function evaluation
results from all the processors into a single global \emp{output}
list. Finally, we let \emp{finalize(output)} run only on the master
processor (assuming that this work does not require parallelization).
For the purpose of collecting outputs from all the processors, the
following generic Python function can be used:
{
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def collect_subproblem_output_args(my_output_args, my_rank,
num_procs, send_func, recv_func):
if my_rank == 0: # master process?
output_args = my_output_args
for i in range(1, num_procs):
output_args += recv_func(i)
return output_args
else:
send_func(my_output_args, 0)
return None
\end{verbatimtab}
\end{quote}
}
The last two input arguments to the above function deserve some
attention. Both \emp{send\_func} and \emp{recv\_func} are functions
themselves. In the case of using the \emp{pypar} wrapper of MPI
commands, we may simply pass \emp{pypar.send} as the \emp{send\_func}
input argument and \emp{pypar.receive} as \emp{recv\_func}. Moreover,
switching to another MPI module is transparent with regard to the
generic function named \emp{collect\_subproblem\_output\_args}. It
should also be noted that most Python MPI modules are considerably
more user-friendly than the original MPI commands in C/Fortran. This
is because (1) the use of keyword arguments greatly simplifies the
syntax, and (2) any picklable (marshalable) Python data type can be
communicated directly.
Now that we have implemented the generic functions
\emp{get\_subproblem\_input\_args} and
\emp{collect\_subproblem\_output\_args}, we can write a minimalistic
parallel solver as follows: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def parallel_solve_problem(initialize, func, finalize,
my_rank, num_procs, send, recv):
input_args = initialize()
my_input_args = get_subproblem_input_args(input_args,
my_rank, num_procs)
my_output = [func(*args, **kwargs) \
for args, kwargs in my_input_args]
output = collect_subproblem_output_args(my_output, my_rank,
num_procs, send, recv)
if my_rank == 0:
finalize(output)
\end{verbatimtab}
\end{quote}
}We remark that the above function is generic in the sense that it is
independent of the actual implementation of \emp{initialize},
\emp{func}, and \emp{finalize}, as well as the Python MPI module being
used. All problems that can be composed from independent function
calls can (at least in principle) be parallelized by the shown small
pieces of Python code.
As a specific example of using this parallel solver, we may address
the problem of evaluating the parabolic function (\emp{func} and class
\emp{Parabola}) for a large number of parameters. Using the
\emp{pypar} MPI module and having the problem-dependent code in a
module named \emp{Parabola} and the general function-centric tools in
a module named \emp{function\_centric}, the program becomes as
follows:
{
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
from Parabola import func, Parabola
from function_centric import parallel_solve_problem
problem = Parabola(m=100, n=50, L=10)
import pypar
my_rank = pypar.rank()
num_procs = pypar.size()
parallel_solve_problem(problem.initialize,
func,
problem.finalize,
my_rank, num_procs,
pypar.send, pypar.receive)
pypar.finalize()
\end{verbatimtab}
\end{quote}
} To the reader, it should be obvious from this generic example how to
parallelize other independent function calls by the described
function-centric approach.
\section{Function-Centric Parallelization}
\label{sect:highlevel}
We have shown how to parallelize a serial program that is decomposable
into three parts: \emp{initialize}, calls to \emp{func} (i.e., a set
of independent tasks), and \emp{finalize}. In this section, we
describe how the function-centric parallelization is helpful for three
important classes of scientific applications: Markov chain Monte Carlo
simulations, dynamic population Monte Carlo simulations, and solution
of partial differential equations. We use Python to program a set of
simple and generic parallelization functions.
\subsection{Parallel Markov chain Monte Carlo Simulations}
\label{sect:3.1}
The standard Markov chain Monte Carlo algorithms are embarrassingly
parallel and have exactly the {\em same} algorithmic structure as the
example of parameter analysis in Section~\ref{sect:motive}. This
means that the functions \emp{initialize}, \emp{func}, and
\emp{finalize} can easily be adapted to Monte Carlo problems. More
specifically, the \emp{initialize} function prepares the set of random
samples and other input parameters. Some parametric model is computed
by the \emp{func} function, whereas \emp{finalize} collects the data
returned from all the \emp{func} calls and prepares for further
statistical analysis.
Function-centric parallelization of Markov chain Monte Carlo
applications closely follows the example in Section~\ref{sect:motive}.
We can reuse the three generic functions named
\emp{get\_subproblem\_input\_args},
\emp{collect\_subproblem\_output\_args}, and
\emp{parallel\_solve\_problem}, assuming that all the \emp{func}
evaluations are equally costly and all the processors are equally
powerful so there is no need for more sophisticated load balancing.
In Section~\ref{sect:pMCMC}, we will look at a real-life Markov chain
problem from political science (Appendix~\ref{sect:voting} gives its
mathematical description).
\subsection{Population Monte Carlo with Dynamic Load Balancing}
\label{sect:3.2}
A more advanced branch of Monte Carlo algorithms is population Monte
Carlo, see~\cite{iba-2000-16}. Here, a group of walkers, also called
the {\em population}, is used to represent a high-dimensional vector
and the computation is carried out by a random walk in the state
space. During the computation some of these walkers may be duplicated
or deleted according to some acceptance/rejection criteria, i.e., the
population is dynamic in time. Population Monte Carlo algorithms have
been proven useful in a number of fields, spanning from polymer
science to statistical sciences, statistical physics, and quantum
physics.
Unlike the examples so far, where the computational tasks were totally
independent and of static size, population Monte Carlo algorithms may
be viewed as an iteration in time where we repeatedly do some work on
a dynamic population, including moving the walkers of the population
and adjusting the population size, which in a parallel context calls
for dynamic load balancing.
\subsubsection{Serial Implementation}
A serial implementation of the time integration function can be as follows:
{
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def time_integration(initialize, do_timestep, finalize):
walkers, timesteps = initialize()
output = []
for step in range(timesteps):
old_walkers_len = len(walkers)
output.append(do_timestep(walkers))
walkers.finalize_timestep(old_walkers_len, len(walkers))
finalize(output)
\end{verbatimtab}
\end{quote}
}
The input arguments to the generic \emp{time\_integration} function
are three functions: \emp{initialize}, \emp{do\_timestep}, and
\emp{finalize}. This resembles the three-step structure discussed in
Section~\ref{sect:motive}. The \emp{do\_timestep} function can have a
unified implementation for all the variants of population Monte Carlo
algorithms. The other two input functions are typically programmed as
methods of a class that implements a particular algorithm (such as
diffusion Monte Carlo in Section~\ref{sect:pdmc}). Here, the
\emp{initialize} method sets up a population object \emp{walkers} (to
be explained below) and determines the number of time steps the
walkers are to be propagated. The \emp{finalize} method can, e.g.,
store the output for later analysis.
The purpose of the \emp{do\_timestep} function is to implement the
work for one time step, including propagating the walkers and
adjusting the population. An implementation that is applicable for
all population Monte Carlo algorithms may have the following form:
\begin{quote}
{
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def do_timestep(walkers):
walkers.move()
for walker in range(len(walkers)):
if walkers.get_marker(walker) == 0:
walkers.delete(walker)
elif walkers.get_marker(walker) > 1:
walkers.append(walker, walkers.get_marker(walker)-1)
return walkers.sample_observables()
\end{verbatimtab}
}
\end{quote}
The above implementation of \emp{time\_integration} and
\emp{do\_timestep} assumes that \emp{walkers} is an object of a class,
say with name \emp{Walkers}, that has a certain number of methods. Of
course, the flexibility of Python allows that the concrete
implementation of class \emp{Walkers} be made afterwards, unlike C++
and Java that require class \emp{Walkers} be written before
implementing \emp{time\_integration} and \emp{do\_timestep}. Here, we
expect class \emp{Walkers} to provide a generic implementation of a
group of walkers, with supporting methods for manipulating the
population. The most important methods of class \emp{Walkers} are as
follows:
\begin{itemize}
\item \emp{move()} carries out the work of moving each walker of the
population randomly according to some rule or distribution function.
\item \emp{get\_marker(walker)} returns the number of copies belonging
to a walker with index \emp{walker}, where 0 means the walker should
be deleted, 2 or more means that clones should be created.
\item \emp{append(walker, nchilds)} and \emp{delete(walker)} carry out
the actual cloning and removal of a walker with index \emp{walker}.
\item \emp{sample\_observables()} returns the observables at a given
time step, e.g., an estimate of the system energy.
\item \emp{finalize\_timestep(old\_size, new\_size)} does some
internal book keeping at the end of each time step, such as
adjusting some internal variables. It takes as input the total
number of walkers before and after the walker population has been
adjusted by the \emp{do\_timestep} function.
\item \emp{\_\_len\_\_} is one of Python's special class methods and
is in our case meant to return the number of walkers. A call
\emp{len(walkers)} yields the same result as
\emp{walkers.\_\_len\_\_()}.
\end{itemize}
For a real application, such as the diffusion Monte Carlo algorithm
(see Section~\ref{sect:pdmc} and Appendix~\ref{sect:dmc}), the
concrete implementation of the methods should reflect the desired
numerical algorithm. For example, the \emp{move} method of diffusion
Monte Carlo uses diffusion and branching as the rule to randomly move
each walker, and the \emp{finalize\_timestep} method adjusts the
branching ratio.
\subsubsection{Parallelization}
Parallelism in population Monte Carlo algorithms arises naturally from
dividing the walkers among the processors. Therefore, a parallel
version of the \emp{time\_integration} function may be as follows: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def parallel_time_integration(initialize, do_timestep, finalize,
my_rank, num_procs, send, recv, all_gather):
my_walkers, timesteps = initialize(my_rank, num_procs)
old_walkers_len = sum(all_gather(numpy.array([len(my_walkers)])))
my_output = []
for step in range(timesteps):
# do what is required at this time step and measure CPU time
t_start = time.time()
results = do_timestep(my_walkers)
my_output.append(results)
task_time = time.time() - t_start
# redistribute walkers and get walker size per process
num_walkers_per_proc = dynamic_load_balancing(\
my_walkers, task_time, my_rank, num_procs,\
send, recv, all_gather)
# finalize task for this time step
new_walkers_len = sum(num_walkers_per_proc)
my_walkers.finalize_timestep(old_walkers_len, new_walkers_len)
old_walkers_len = new_walkers_len
my_output = collect_subproblem_output_args(my_output, my_rank,
num_procs, send, recv)
if my_rank == 0:
finalize(my_output)
\end{verbatimtab}
\end{quote}
}
In comparison with its serial counterpart, the
\emp{parallel\_time\_integration} function has a few noticeable
changes. First, the input arguments have been extended with five new
arguments. The two integers \emp{my\_rank} and \emp{num\_procs} are,
as before, meant for identifying the individual processors and finding
the total number of processors. The other three new input arguments
are MPI communication wrapper functions: \emp{send}, \emp{recv}, and
\emp{all\_gather}, which can be provided by any of the Python wrapper
modules of MPI. The only exception is that \emp{pypar} does not
directly provide the \emp{all\_gather} function, but we can easily
program it as follows: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def all_gather (input_array):
array_gathered_tmp = pypar.gather (input_array, 0)
array_gathered = pypar.broadcast (array_gathered_tmp, 0)
return array_gathered
\end{verbatimtab}
\end{quote}
}
Second, we note that the \emp{initialize} function is slightly
different from the serial case, now accepting \emp{my\_rank} and
\emp{num\_procs} as input. This is because initial division of the
walkers is assumed to be carried out here, giving rise to
\emp{my\_walkers} on each processor. Third, a new function
\emp{dynamic\_load\_balancing} is called during each time step. This
function will be explained below in detail. Fourth, unlike that the
serial counterpart could simply pass the size of its walkers to
\emp{finalize\_timestep}, the parallel implementation needs to collect
the global population size before calling \emp{finalize\_timestep}.
We remark that each local population knows its own size, but not the
global population size. For this purpose, the
\emp{dynamic\_load\_balancing} function returns the individual local
population sizes as a \emp{numpy} array. Last, the
\emp{collect\_subproblem\_output\_args} function from
Section~\ref{sect:motive:parallel} is used to assemble all the
individual results onto the master processor before calling the
\emp{finalize} function.
As mentioned before, parallelization of population Monte Carlo
algorithms has to take into account that the total number of walkers
changes with time. Dynamic re-distribution of the walkers is
therefore needed to avoid work load imbalance. The generic
\emp{dynamic\_load\_balancing} function is designed for this purpose,
where we evaluate the amount of work for each processor and, if the
work distribution is too skew, we move the excess walkers from a busy
processor to a less busy one. The function first checks the
distribution of local population sizes. If the difference between the
smallest number of walkers and the largest number of walkers exceeds
some predefined threshold, \emp{dynamic\_load\_balancing} finds a
better population distribution and redistributes the walkers: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def dynamic_load_balancing(walkers, task_time, my_rank, num_procs, \
send, recv, all_gather):
walkers_per_proc = all_gather(numpy.array([len(walkers)]))
if imbalance_rate(walkers_per_proc) > walkers.threshold_factor:
timing_list = all_gather(numpy.array([task_time]))
rebalanced_work = find_optimal_workload(timing_list,
walkers_per_proc)
walkers = redistribute_work(walkers,
walkers_per_proc,
rebalanced_work,
my_rank, num_procs, send, recv)
return walkers_per_proc
\end{verbatimtab}
\end{quote}
}
Two helper functions \emp{find\_optimal\_workload} and
\emp{redistribute\_work} are used in the above implementation. Here,
\emp{find\_optimal\_workload} finds the optimal distribution of work,
based on how much time each local population has used. The
\emp{redistribute\_work} function carries out the re-shuffling of
walkers. A straightforward (but not optimal) implementation of these
functions goes as follows: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def find_optimal_workload(timing_list, current_work_per_proc):
total_work = sum(current_work_per_proc)
C = total_work/sum(1./timing_list)
tmp_rebalanced_work = C/timing_list
rebalanced_work = numpy.array(tmp_rebalanced_work.tolist(),'i')
remainders = tmp_rebalanced_work-rebalanced_work
while sum(rebalanced_work) < total_work:
maxarg = numpy.argmax(remainders)
rebalanced_work[maxarg] += 1
remainders[maxarg] = 0
return rebalanced_work
def redistribute_work(my_walkers, work_per_proc, rebalanced_work,
my_rank, num_procs, send, recv):
difference = work_per_proc-rebalanced_work
diff_sort = numpy.argsort(difference)
prev_rank_min = diff_sort[0]
while sum(abs(difference)) != 0:
diff_sort = numpy.argsort(difference)
rank_max = diff_sort[-1]
rank_min = diff_sort[0]
if rank_min == prev_rank_min and rank_max != diff_sort[1]:
rank_min = diff_sort[1]
if my_rank==rank_max:
send(my_walkers.cut_slice(rebalanced_work[my_rank]),\
int(rank_min))
elif my_rank==rank_min:
my_walkers.paste_slice(recv(int(rank_max)))
difference[rank_min] += difference[rank_max]
difference[rank_max] = 0
prev_rank_min = rank_min
return my_walkers
\end{verbatimtab}
\end{quote}
}
Careful readers will notice that two particular methods,
\emp{my\_walkers.cut\_slice} and \emp{my\_walkers.paste\_slices},
provide the capability of migrating the work load between processors
in the \emp{redistribute\_work} function. These two methods have to be
programmed in class \emp{Walkers}, like the other needed methods
described earlier: \emp{move}, \emp{get\_marker}, \emp{append},
\emp{delete}, and so on. The \emp{cut\_slice} method takes away
excess work from a local population and the \emp{paste\_slice} method
inserts additional work into a local population. Note that the input
argument to the \emp{cut\_slice} method is an index threshold meaning
that local walkers with indices larger than that are to be taken away.
The returned slice from \emp{cut\_slice} is a picklable Python object
that can be sent and received through MPI calls.
The generic \emp{redistribute\_work} function deserves a few more
words. Among its input arguments is the ideal work distribution,
\emp{rebalanced\_work}, which is calculated by
\emp{find\_optimal\_workload}. The \emp{redistribute\_work} function
first calculates the difference between the current distribution,
\emp{work\_per\_proc}, and the ideal distribution. It then
iteratively moves walkers from the processor with the most work to the
processor with the least work until the difference is evened out.
This load balancing scheme is in fact independent of population Monte
Carlo algorithms. As long as you have an algorithm repeatedly doing a
task over time and where the amount of work in the task varies over
time, this scheme can be reused. The only requirement is that an
application-specific implementation of class \emp{Walkers}, in terms
of method names and functionality, should match with
\emp{dynamic\_load\_balancing} and \emp{redistribute\_work}. It
should be noted that the given implementation of the latter function
is not optimal.
The algorithm of diffusion Monte Carlo, described in
Appendix~\ref{sect:dmc}, is a typical example of a population Monte
Carlo algorithm. The implementation is described in Section
\ref{sect:pdmc} and Appendix~\ref{sect:dmc}.
\subsection{Parallel Additive Schwarz Iterations}
\label{sect:3.3}
From the perspective of communication between processors,
parallelization of the Monte Carlo algorithms is relatively
easy. Parallel Markov chain Monte Carlo algorithms only require
communication in the very beginning and end, whereas parallel
population Monte Carlo algorithms only require communication at the
end of each time step. Actually, our function-centric approach to
parallelization can allow more frequent communication. To show the
versatility of function-centric parallelization, we apply it to an
implicit method for solving partial differential equations (PDEs)
where communication is frequent between processors.
More specifically, many PDEs can be solved by an iterative process
called {\em domain decomposition}. The idea is to divide the global
domain, in which the PDEs are to be solved, into $n$ overlapping
subdomains. The PDEs can then be solved in parallel on the $n$
subdomains. However, the correct boundary condition at the
\emph{internal} subdomain boundaries are not known, thus leading to an
iterative approach where one applies boundary conditions from the last
iteration, solves for the $n$ subdomain problems again, and repeats
the process until convergence of the subdomain solutions (see
e.g.~\cite{Chan:1994:DOA,Smith:1996:DPM}). This algorithm is commonly
called additive Schwarz iteration and can successfully be applied to
many important classes of PDEs
\cite{Cai2005AWR,Cai.2005.5,Langtangen:01b}. The great advantage of
the algorithm, especially from a software point of view, is that the
PDE solver for the global problem can be reused for each subdomain
problem. Some additional code is needed for communicating the
solutions at the internal boundaries between the subdomains. This code
can be implemented in a generic fashion in Python, as we explain
later.
Let us first explain the additive Schwarz algorithm for solving PDEs
in more detail. We consider some stationary PDE defined on a global
domain $\Omega$:
\begin{eqnarray}
\mathcal L(u)&=&f, \quad x\in\Omega,
\label{eqn:global-system}
\end{eqnarray}
subject to some boundary condition involving $u$ and/or its
derivatives. Dividing $\Omega$ into a set of overlapping subdomains
$\{\Omega_s\}_{s=1}^P$, we have the restriction of
(\ref{eqn:global-system}) onto $\Omega_s$, for all $s$, as
\begin{eqnarray}
\mathcal L(u) &=&f, \quad x\in\Omega_s.
\label{eqn:sub-system}
\end{eqnarray}
The additive Schwarz method finds the global solution $u$ by an
iterative process that generates a series of approximations $u_0$,
$u_1$, $u_2$ and so on. During iteration $k$, each subdomain computes
an improved local solution $u_{s,k}$ by locally solving
(\ref{eqn:sub-system}) for $u=u_{s,k}$ with $u_{s,k}=u_{k-1}$ as (an
artificial) boundary condition on $\Omega_s$'s non-physical internal
boundary that borders with neighboring subdomains. All the subdomains
can {\em concurrently} carry out the local solution of
(\ref{eqn:sub-system}) within iteration $k$, thus giving rise to
parallelism. At the end of iteration $k$, neighboring subdomains
exchange the latest local solutions in the overlapping regions to
(logically) form the global field $u_k$. The subdomain problems
(\ref{eqn:sub-system}) are of the same type as the global problem
(\ref{eqn:global-system}), which implies the possibility of reusing an
existing serial code that was originally implemented for
(\ref{eqn:global-system}). The additional code for exchange of local
solutions among neighbors can be implemented by generic communication
operations, independently of specific PDEs.
A generic implementation of parallel additive Schwarz iteration
algorithm can be realized as the following Python function: {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def additive_Schwarz_iterations(subdomain_solve, communicate,
set_BC, max_iter, threshold, solution,
convergence_test=simple_convergence_test):
iter = 0; not_converged = True # init
while not_converged and iter < max_iter:
iter += 1
solution_prev = solution.copy()
set_BC(solution)
solution = subdomain_solve()
communicate(solution)
not_converged = not convergence_test(\
solution, solution_prev, threshold)
\end{verbatimtab}
\end{quote}
}
In the above function, \emp{max\_iter} represents the maximum number
of additive Schwarz iterations allowed, and \emp{subdomain\_solve} is
a function that solves the subdomain problem of form
(\ref{eqn:sub-system}) and returns an object \emp{solution}, which is
typically a \emp{numpy} array containing the latest subdomain solution
$u_{s,k}$ on a processor (subdomain). However, \emp{solution} may
well be a more complex object, say holding a collection of scalar
fields over computational grids, provided that (i) the object has a
\emp{copy} method, (ii) \emp{convergence\_test} and \emp{communicate}
can work with this object type, and (iii) \emp{subdomain\_solve}
returns such an object. This flexibility in choosing \emp{solution}
reflects the major dynamic power of Python and provides yet another
illustration of the generality of the examples in this paper.
Given an existing serial code, for example in a language like Fortran
or C/C++, the \emp{subdomain\_solve} function is easily defined by
wrapping up an appropriate piece of the serial code as a Python class
(since \emp{subdomain\_solve} does not take any arguments, the
function needs a state with data structures, conveniently implemented
as class attributes as explained in Section~\ref{sect:motive:serial}).
The \emp{communicate} argument is a function for exchanging the latest
local solutions among the subdomains. After the call, the
\emp{solution} object is updated with recently computed values from
the neighboring subdomains, and contents of \emp{solution} have been
sent to the neighbors. The \emp{communicate} function is problem
independent and can be provided by some library. In our
implementation, the implementation is entirely in Python to take
advantage of easy programming of parallel communication in Python.
The \emp{set\_BC} argument is a function for setting boundary
conditions on a subdomain's internal boundary. This function depends
on the actual serial code and is naturally implemented as part of the
class that provides the \emp{subdomain\_solve} function.
The \emp{convergence\_test} function is assumed to perform an
appropriate convergence test. The default generic implementation can
test
$$\max_{1\le s\le P}\frac{\|u_{s,k}-u_{s,k-1}\|^2}{\|u_{s,k}\|^2}
$$ against a prescribed threshold value. An implementation reads
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def simple_convergence_test(solution, solution_prev, threshold=1E-3):
diff = solution - solution_prev
loc_rel_change = vdot(diff,diff)/vdot(solution,solution)
glob_rel_change = all_reduce(loc_rel_change, MAX)
return glob_rel_change < threshold
\end{verbatimtab}
\end{quote}
We remark that \emp{all\_reduce} is a wrapper of the MPI
\emp{MPI\_Allreduce} command and \emp{vdot} computes the inner product
of two \emp{numpy} arrays.
Unlike the three-component structure described in
Sections~\ref{sect:3.1} and \ref{sect:3.2}, the main ingredients for
parallel additive Schwarz iterations are the functions of
\emp{subdomain\_solve}, \emp{communicate}, \emp{set\_BC}, and
\emp{convergence\_test}. In other words, it is not natural to divide
the work of solving a PDE into \emp{initialize}, \emp{func}, and
\emp{finalize}. Nevertheless, function-centric parallelization is
also here user-friendly and gives a straightforward implementation of
\emp{additive\_Schwarz\_iterations} as above. The
\emp{convergence\_test} function shown above is clearly generic, and
so is the \emp{communicate} function in the sense that it does not
depend on the PDE. Both functions can be reused for different PDEs.
The other two functions are PDE dependent, however,
\emp{subdomain\_solve} normally wraps an existing serial code, while
the implementation of \emp{set\_BC} is typically very simple.
\section{Applications and Numerical Experiments}
\label{sect:results}
In this section we will address three real research projects involving
the three classes of algorithms covered in
Section~\ref{sect:highlevel}. The projects have utilized our
function-centric approach to parallelizing existing codes. That is,
we had some software in Fortran, C++, and R performing the basic
computations needed in the projects. The serial software was wrapped
in Python, adapted to components such as \emp{initialize}, \emp{func},
\emp{do\_timestep}, \emp{finalize}, \emp{subdomain\_solve},
\emp{communicate}, \emp{set\_BC}. Parallelization was then carried
out as explained in previous sections. An important issue to be
reported is the parallel efficiency obtained by performing the
parallelization in a Python layer that is separate from the underlying
serial scientific codes.
The Python enabled parallel codes have been tested on a Linux cluster
of 3.4 GHz Itanium2 processors, which are interconnected through
1Gbits ethernet. The purpose is to show that the function-centric
parallelization approach is easy to use and that satisfactory parallel
performance is achievable.
\subsection{Parallel Markov Chain Monte Carlo Simulations}
\label{sect:pMCMC}
The first case is from political science and concerns estimating
legislators' ideal points by the Markov chain Monte Carlo (MCMC)
method. For a detailed description of the mathematical problem and
the numerical method, we refer the reader to
Appendix~\ref{sect:voting}. This application fits into the setup in
Section~\ref{sect:3.1}. The statistical engine is provided by the
PSCL library~\cite{Jackman2006} in R~\cite{R}, for which there exists
a Python wrapper.
To use the function-centric parallelization described in
Section~\ref{sect:3.1}, we have written a Python class named
\emp{PIPE}. In addition to the constructor of the class (i.e., the
\emp{\_\_init\_\_} method), there are three methods as follows:
\begin{itemize}
\item \emp{initialize} sets up the functionality of the PSCL library
through the Python wrapper of R (named \emp{rpy}), and prepares the
input argument list needed for \emp{func}.
\item \emp{func} carries out the computation of each task by invoking
appropriate functions available through \emp{rpy} (in short,
\emp{func} is a Python wrapper to the R function \emp{ideal} from
the PSCL library).
\item \emp{finalize} summarizes the output and generates an array in R
format.
\end{itemize}
The resulting parallel Python program is now as short as {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
from function_centric import parallel_solve_problem
import pypar
my_rank = pypar.rank()
num_procs = pypar.size()
from pypipe import PIPE
problem = PIPE("EP1.RData", "rcvs", "NULL", "NULL")
parallel_solve_problem(problem.initialize, problem.func, problem.finalize,
my_rank, num_procs, pypar.send, pypar.receive)
pypar.finalize()
\end{verbatimtab}
\end{quote}
}
The practical importance of a parallel MCMC code is that large and
computationally intensive simulations are now easily doable. More
specifically, data from the European Parliament between 1979 and
2004~\cite{Hix2005} are used for simulation. During the five year
legislative terms, the European Parliament expanded the size of the
membership as well as the number of votes taken. (This trend has
continued since 2004.) It is hence increasingly computationally
intensive to estimate the ideal point model without reducing the
length of the Markov chain. We examined the parallel performance by
comparing the computing time for each of the five legislatures,
running the parallelized code on 32 CPUs. The results are reported in
Table~\ref{tab:results}. When comparing the results, the reader should
note that we have not made any attempts to optimize the \emp{ideal}
code (called by our \emp{func} function) for the purpose of
parallelization. This makes it straightforward to switch to new
versions of the \emp{ideal} function. We ran 100,000 MCMC
iterations. The parallel efficiency was about 90\%.
\begin{table}[ht]
\caption{\label{tab:results}Speedup results associated with voting analysis.}
\begin{indented}
\item[]\begin{tabular}{@{}llllll}
\br
\textbf{Legislature}& Votes& Members& 1 CPU &32 CPUs& Efficiency\\
\mr
1979 - 1984 & 810 & 548& 287m 32.560s & 10m 13.318s & 87.91$\%$\\
1984 - 1989 & 1853& 637 & 783m 59.059s& 26m 58.702s& 91.06$\%$\\
1989 - 1994 & 2475 & 597 & 1006m 59.258s &33m 26.140s & 94.11$\%$\\
1994 - 1999 & 3603 & 721 & 1905m 0.930s&66m 0.068s&90.20$\%$\\
1999 - 2004 & 5639 & 696&2898m 45.224s & 102m 7.786s &88.70$\%$\\
\br
\end{tabular}
\end{indented}
\end{table}
\subsection{Parallel Diffusion Monte Carlo Simulations}
\label{sect:pdmc}
As an example of population Monte Carlo methods, we will now look at
parallel Diffusion Monte Carlo (DMC) computations (see
Appendix~\ref{sect:dmc} for a detailed numerical description), which
is used here to simulate Bose-Einstein condensation. We recall from
Section~\ref{sect:3.2} that dynamic load balancing is needed in
connection with the parallelization, and can be provided by the
generic \emp{dynamic\_load\_balancing} function. To utilize the
parallel time integration function \emp{parallel\_time\_integration}
from Section~\ref{sect:3.2}, we need to program a parallel version of
the \emp{initialize} function. The \emp{do\_timestep} function from
Section~\ref{sect:3.2} can be used as is. {
\begin{quote}
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
def initialize(my_rank, num_procs):
nwalkers = 1000
nspacedim = 3
stepsize = 0.1
timesteps = 200
walkers_per_proc = simple_partitioning(nwalkers, num_procs)
my_walkers = Walkers(walkers_per_proc[my_rank], nspacedim, stepsize)
my_walkers.threshold_factor = 1.1
return my_walkers, timesteps
\end{verbatimtab}
\end{quote}
} This \emp{initialize} function is quite similar to its serial
counter part. As noted in Section~\ref{sect:3.2}, it takes as input
\emp{my\_rank} and \emp{num\_procs}. The \emp{simple\_partitioning}
function is called to partition the walker population. A
\emp{my\_walkers} object is assigned to each processor, and a
threshold factor is prescribed to determine when load balancing is
needed.
Together with the \emp{parallel\_time\_integration} function from
Section~\ref{sect:3.2}, the above \emp{initialize} function is the
minimum programming effort needed to parallelize a serial population
Monte Carlo code. For the particular case of our parallel Diffusion
Monte Carlo implementation, we also need to know the global number of
walkers in every timestep to be able to estimate its observables
globally. Moreover, the load balancing scheme needs the time usage of
each processor during each time step.
A class with name \emp{Walkers} needs to be implemented to match with
the implementations of \emp{parallel\_time\_integration},
\emp{dynamic\_load\_balancing}, and the above \emp{initialize}
function. The essential work is to provide a set of methods with
already decided names (see Section~\ref{sect:3.2}), such as
\emp{move}, \emp{append}, \emp{delete}, \emp{finalize\_timestep},
\emp{cut\_slice}, and \emp{paste\_slice}. A concrete example of the
\emp{Walkers} class is described with more details in
Appendix~\ref{sect:dmc}.
We report in Table~\ref{tab:results2} the timing results of a number
of parallel DMC computations. The parallel efficiency was about 85\%.
We increased the total number of walkers when more processors were
used in the simulation, such that the number of walkers assigned to
each processor remained as 200. Such a use of parallel computers for
DMC simulations mimics the everlasting wish of quantum physicists to
do larger computations as soon as more computing resource becomes
available. Note that in this scaled scalability test, good speedup
performance is indicated by an almost constant time usage independent
of the number of processors.
\begin{table}[ht]
\caption{\label{tab:results2}Timing results of the parallel DMC
simulations where each processor is constantly assigned with 200
walkers, all moved in 5000 time steps.}
\begin{indented}
\item[]\begin{tabular}{@{}lll}
\br
CPUs&Time&Efficiency\\
\mr
1&37m10.389s &N/A\\
5&42m32.359s &87.39\%\\
10&42m00.734s &88.48\%\\
20&42m29.945s &87.47\%\\
30&42m33.895s &87.33\%\\
40&43m30.092s &85.45\%\\
50&43m39.159s &85.16\%\\
\br
\end{tabular}
\end{indented}
\end{table}
\subsection{Parallel Boussinesq Simulations}
Simulating the propagation of ocean waves is the target of the our
third and final concrete case. The reader is referred to
Appendix~\ref{sect:add_sch} for the mathematical model and the
numerical method. The involved equations can be solved in parallel by
the additive Schwarz algorithm of Section~\ref{sect:3.3}.
Our starting point for parallelization is a 25 years old legacy
Fortran 77 code consisting of a set of subroutines. More specifically,
the most important subroutines are \emp{KONTIT} and \emp{BERIT}, which
target the two semi-discretized equations (\ref{eqn:dist-continu}) and
(\ref{eqn:dist-bernoulli}) of the mathematical model (see
Appendix~\ref{sect:add_sch}). These two Fortran 77 subroutines
contain intricate algorithms with nested layers of do-loops, which are
considered to be very difficult to parallelize by directly inserting
MPI calls in the Fortran code. Performing the parallelization outside
the Fortran code is therefore much more convenient. Using the proposed
framework in the present paper the parallelization is a technically
quite straightforward task.
The subdomain solver consists of calls to the subroutines \emp{KONTIT}
and \emp{BERIT}. The implementation of the Python function
\emp{subdomain\_solve} (see Section~\ref{sect:3.3}) requires a Python
interface to \emp{KONTIT} and \emp{BERIT}, which can easily be
produced by the F2PY software. Since a subdomain solver needs to set
artificial boundary conditions at non-physical boundaries, we have
programmed two light-weight wrapper subroutines in Fortran,
\emp{WKONTIT} and \emp{WBERIT}, which handles the boundary conditions
before invoking \emp{KONTIT} and \emp{BERIT}. We then apply F2PY to
make \emp{WKONTIT} and \emp{WBERIT} callable from Python. Since the
Fortran subroutines have lots of input data in long argument lists and
\emp{subdomain\_solve} takes no arguments, we have created a class
where the Fortran input variables are stored as class attributes:
\begin{quote}
{
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
import f77 # extension module for the Fortran code
class SubdomainSolver:
def __init__(self, ...):
# set input arguments to the Fortran subroutines as class attributes
# (nbit,F,YW,H,QY,WRK,dx,dy,dt,kit,ik,gg,alpha,eps,
# lower_x_neigh,upper_x_neigh,lower_y_neigh,upper_y_neigh)
def continuity(self):
self.Y, self.nbit = f77.WKONTIT(\
self.F, self.Y, self.YW, self.H, self.QY, self.WRK,
self.dx, self.dy, self.dt, self.kit, self.ik,
self.gg, self.alpha, self.eps, self.nbit,
self.lower_x_neigh, self.upper_x_neigh,
self.lower_y_neigh, self.upper_y_neigh)
def bernoulli(self):
# similar to the continuity method
sd = SubdomainSolver(...)
subdomain_solve1 = sd.continuity
subdomain_solve2 = sd.bernoulli
\end{verbatimtab}
}
\end{quote}
Note that since there are two PDEs (\ref{eqn:dist-continu}) and
(\ref{eqn:dist-bernoulli}), we have created two functions:
\emp{subdomain\_solve1} and \emp{subdomain\_solve2}. The main
computation of the resulting parallel program is in the following
while loop:
\begin{quote}
{
\renewcommand{\baselinestretch}{0.85}\footnotesize
\begin{verbatimtab}[8]
t = 0
while t < t_stop:
t = t+dt
additive_Schwarz_iterations(subdomain_solve1, communicate,
set_BC, 10, 1E-3, sd.Y)
additive_Schwarz_iterations(subdomain_solve2, communicate,
set_BC, 10, 1E-3, sd.F)
\end{verbatimtab}
}
\end{quote}
The \emp{additive\_Schwarz\_iterations} function from
Section~\ref{sect:3.3} can be placed in a reusable module. The
\emp{communicate} function is borrowed from a Python library for mesh
partitioning and inter-subdomain communication. The \emp{set\_BC}
function actually does not do anything for this particular
application.
Speedup results are reported in Table~\ref{tab:results3}, for which
the global solution mesh was fixed at $1000\times1000$, and the number
of time steps was 40. The results show that we can handle a quite
complicated mathematical problem in a black-box Fortran code with our
suggested simple framework and obtain a remarkable good speedup, with
just a trivial extension of the Fortran code.
\begin{table}[ht]
\caption{\label{tab:results3}The speedup results of the Python enabled
parallel Boussinesq simulations.}
\begin{indented}
\item[]\begin{tabular}{@{}llll}
\br
CPUs & Time & Speedup & Efficiency\\
\mr
1 & 166.66s & N/A & N/A\\
2 & 83.61s & 1.99 & 99.67\%\\
4 & 44.45s & 3.75 & 93.73\%\\
8 & 20.16s & 8.27 & 103.33\%\\
16 & 11.43s & 14.58 & 91.13\%\\
\br
\end{tabular}
\end{indented}
\end{table}
\section{Conclusion}
\label{sect:conclusion}
We have shown how serial scientific codes written in various common
languages, including Fortran, C, C++, and Python, can be parallelized
in a separate, small software unit written in Python. The advantage
of such an approach is twofold. First, the existing, often
complicated, scientific high-performance code remains (almost)
unchanged. Second, the parallel algorithm and its inter-processor
communication are conveniently implemented in high-level Python code.
This approach to parallelization has been implemented in a software
framework where the programmer needs to implement a few Python
functions for carrying out the key steps in the solution approach. For
example, our first application involves doing a set of independent
tasks in parallel, where a small Python framework deals with the
parallelism and demands the user to supply three functions:
\emp{initialize} for preparing input data to the mathematical model,
\emp{func} for calling up the serial scientific code, and
\emp{finalize} for processing the computational results. Some more
functions must be supplied in more complicated problems where the
algorithm evolves in time, with a need for dynamic load balancing and
more parallel communication.
Our simple software frameworks outlined in this paper are applicable
to many different scientific areas, and we have described some common
classes of problems: parameter investigation of a mathematical model,
standard Monte Carlo simulation, Monte Carlo simulation with need for
dynamic load balancing, and numerical solution of partial differential
equations. In each of these cases, we have outlined fairly detailed
Python code such that most technical details of the parallel
implementations are documented. This may ease the migration of the
ideas to new classes of problems beyond the scope of this paper.
In particular, the shown frameworks have been used to parallelize
three real scientific problems taken from our research. The problems
concern Markov Chain Monte Carlo models for voting behavior in
political science, Diffusion Monte Carlo methods for Bose-Einstein
condensation in quantum mechanics, and additive Schwarz and finite
difference methods for simulating ocean waves by a system of partial
differential equations. The results of our investigations of the
parallel efficiency are very encouraging: In all these real science
problems, parallelizing serial codes in the proposed Python framework
gives almost optimal speedup results, showing that there arises no
significant loss due to using Python and performing the
parallelization ``outside'' the serial codes.
As a conclusion, we believe that the ideas and code samples from this
paper can simplify parallelization of serial codes greatly, without
significant loss of computational efficiency. This is good news for
scientists who are non-experts in parallel programming but want to
parallelize their serial codes with as small efforts as possible.
\newpage
|
\section{Introduction}
A (non) magnetic impurity coupled to a spin-$1/2$ Heisenberg chain is
a prototype system that exemplifies ``Kondo"-type effects in a
correlated system. Starting with the proposal of
Kane-Fisher,\cite{kf} a weak link in a repulsive (attractive)
Luttinger liquid was shown to lead to an insulating (transmitting)
ground state. The cutting or healing of spin chains by a variety of
(non) magnetic defects has also been
established\cite{affleck,eggert,sorensen,laflorencie} as well as the
effect of a magnetic impurity on the ground state of the anisotropic
easy-plane Heisenberg chain.\cite{furusaki} Generically, a weak link
or coupling to a magnetic impurity in a Heisenberg antiferromagnetic
chain leads to a ground state corresponding to two open chains. In the
exceptional case of two adjacent links or a ferromagnetic (attractive in the
fermionic language) easy axis anisotropy a healing of the defect is
conjectured.\cite{furusaki} This screening effect is characterized by
a Kondo-like temperature and screening
length.\cite{eggert,sorensen} These phenomena have so far mostly been
studied either as they are reflected on ground state properties,
e.g. finite size gaps, entanglement or, somewhat indirectly, as a
temperature dependent induced staggered susceptibility.\cite{eggert}
In this work we use an exceptional physical probe for the study of
these effects, namely, the thermal transport in the spin-$1/2$
Heisenberg chain that is truly singular. Although the Heisenberg
model describes a strongly correlated system, the thermal conductivity
is purely ballistic as the energy current commutes with the
Hamiltonian,\cite{znp} a result that is related to the integrability
of this model.\cite{higherspin} Thus the only scattering present is
due to the defect and thus its frequency/temperature/coupling strength
dependence can be isolated and clearly analyzed. In this context it
was already found that a single potential impurity renders the thermal
transport incoherent\cite{static} with the frequency dependence of the
thermal conductivity well described by a Lorentzian, at least for
a weak impurity. This is in sharp contrast to the case of a
non-interacting system where in spite of the impurity the
transport remains coherent described within the Landauer formalism by
a finite transmission coefficient through the impurity. Thus a single static
impurity materializes the many-body character of scattering states.
Besides its theoretical interest, the effect of (non)magnetic impurities on
the thermal transport of quasi-one dimensional materials as SrCuO$_2$,
Sr$_2$CuO$_3$ and the ladder compound La$_5$Ca$_9$Cu$_{24}$O$_{41}$ has
recently become possible to explore experimentally.\cite{hess}
In this work we use numerical diagonalization techniques -
(full) exact diagonalization (ED), the Finite-Temperature Lanczos
method (FTLM)\cite{ftlm} and the Microcanonical Lanczos method
(MCLM)\cite{mclm} - to study the thermal transport in the
Heisenberg chain model either coupled to a magnetic impurity
or perturbed by single and double weak links. These state of
the art techniques are crucial in the attempt to look for the subtle
low temperature many-body effects associated with Kondo screening.
\section{Model}
We consider the one dimensional anisotropic spin-$1/2$ Heisenberg model
in the presence of a magnetic impurity out of the chain or weak links,
\begin{eqnarray}
H&=&\sum_{l=0}^{L-1} J_{l,l+1} h_{l,l+1}+J'
(s^x_0S^x+s^y_0S^y+\Delta' s^z_0S^z),
\nonumber\\
h_{l,l+1}&=&s^x_ls^x_{l+1}+s^y_ls^y_{l+1}+\Delta
s^z_ls^z_{l+1},
\label{ham}
\end{eqnarray}
where $s^{\alpha}, \alpha=x,y,z$ are spin-$1/2$ operators, $J_{l,l+1}>0 $
the in-chain magnetic exchange coupling that we take antiferromagnetic,
$J'$ the chain-impurity coupling,
$\Delta, \Delta'$ anisotropy parameters and {\bf S} a spin-S
magnetic-impurity operator ($\hbar=1$).
In this work we mostly consider a spin-$1/2$ impurity. We
assume periodic boundary conditions, ${\bf s}_{L}={\bf s}_{0}$,
and uniform couplings $J_{l,l+1}=J$, except in
the study of weak links (see below).
We vary the anisotropy parameters $\Delta, \Delta'$,
with $\Delta=\Delta'$, in order to
look for the (healing) cutting of the chain effects mentioned above.
In our study, based on standard linear response theory,
the frequency $\omega$ dependence of the real part
of the thermal conductivity (regular component) is given by
\begin{equation}
\kappa(\omega)=-\frac{\beta}{\omega}\chi''(\omega),~~
\chi(\omega)=
\frac{i}{L}\int_0^{+\infty} dt e^{i\omega t}
\langle[j^{\epsilon}(t),j^{\epsilon}]\rangle,
\end{equation}
where $\beta=1/T$, $T$ is the temperature and $k_B=1$.
We determine the energy current from the hydrodynamic ($q\rightarrow 0$)
limit of the energy continuity equation
$\partial H_q/\partial t\sim q j^{\epsilon}$
with $H_q=\sum_l e^{iql} h_{l,l+1}$ as,
\begin{eqnarray}
j^{\epsilon}&=&
\sum_{l=0}^{L-1} J_{l-1,l}J_{l,l+1}\: {\bf s}_l\cdot( {\bf s}_{l+1}\times{\bf
s}_{l-1})\nonumber\\ &+&\frac{JJ'}{2}{\bf s}_0\cdot( {\bf
S}\times{\bf s}_{L-1}+ {\bf s}_1\times{\bf S}),
\label{je}
\end{eqnarray}
\noindent
showing for simplicity the case $\Delta=1$ ($\Delta\ne 1$ is obtained by
$s_l^z \rightarrow \Delta s_l^z$ in the cross-product terms). When
$J'=0$ and all $J_{l,l+1}=J$ the energy current commutes with the Hamiltonian,
the transport
is purely ballistic and the thermal conductivity consists of only a
$\delta(\omega)$-peak proportional to the thermal Drude weight.
\section{High temperature limit}
Starting from the high temperature ($\beta\rightarrow 0$) limit
we can obtain a first impression on the behavior of the frequency
dependence of $\kappa(\omega)$ from the 0th and 2nd moments,
$\mu_n=\int d\omega \omega^n \kappa(\omega)$
which are equal to (for the isotropic point, $\Delta=1$),
\begin{eqnarray}
\mu_0&=&\text{const.}\times\frac{6}{T^2}
\big(J^2+\frac{2}{L}\mathcal{B} ^2\big),\quad
\big(\text{const.}=\pi\frac{J^2}{64}\big)\label{mu}\\
\mu_2&=& \text{const.}\times \frac{\mathcal{B}^2}{LT^2}
\left(39J^2-12JJ'+3J'^2+36\mathcal{B}^2\right)\:,\nonumber
\end{eqnarray}
\noindent
where ${\cal B}^2=(J'^{\:2}/3)S(S+1)$ is the characteristic impurity
spin dependence. One could expect the 2nd moment to reflect the
width of $\kappa(\omega)$ and thus to be related to the inverse
scattering time $1/\tau$. We note that for this impurity problem
an assumption of a Gaussian form
$\kappa(\omega)=\kappa_{dc}\exp^{-(\omega\tau)^2}$ would imply from
the $L$ dependence of $\mu_{0,2}$ that $\kappa_{dc}=\kappa(0)$ would
scale as $\sqrt{L}$ and $1/\tau \sim 1/\sqrt{L}$. This is,
however, incorrect as is also evident from the disagreement with higher
moments, $n>2$, which behave all as $\mu_n \propto 1/L$. For
weak-coupling cases, such as a single impurity weakly coupled to the host
chain, we should therefore
rather expect a Lorentzian-like frequency dependence with a
static $\kappa(0) \propto L$ and a characteristic frequency width
$1/\tau \propto 1/L$.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{l_scale.pdf}
\caption{Frequency-dependent thermal conductivity in the high-$T$
limit scaled as $\kappa(\omega L)/L$ for ($\Delta=1$):
(a) weak coupling $J'=0.5J$, (b) strong coupling
$J'=2J$ (curves are normalized to unity).}
\label{l_scale}
\end{figure}
In Fig. \ref{l_scale} we show the frequency dependence of the
thermal conductivity, normalized and appropriately scaled with system size.
Note that in the high-$T$ ($\beta
\rightarrow 0$) limit the relevant (but still nontrivial) quantity
is $T^2 \kappa(\omega)$ which is implicitely extracted by the normalization.
We thus present results of the normalized $\kappa(\omega
L)/L$ for a weak, $J'=0.5J$ and strong, $J'=2J$ coupling case
respectively. The data up to $L=16$ were obtained by full ED
while for $L=18-22$ the MCLM was used.\cite{mclm} The $\delta-$peaks at
the excitation frequencies are binned in windows $\delta\omega=0.01$,
which also gives the frequency resolution of the spectra. For
$J'=0.5J$ we find a simple Lorentzian form while in the strong
coupling case the behavior is nonmonotonic with a maximum at a finite
frequency $O(1/L)$. In both cases the proposed $L$ scaling is indeed
realized.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{s_scale.pdf}
\caption{Frequency dependence of the normalized thermal conductivity
$\kappa(\omega L)/L$ in the high-$T$ limit for a variety of
impurity spin values $S=1/2,1,3/2,2$ and for: (a)
$J'/J=0.5,0.3,0.22,0.18$ corresponding to the weak coupling
${\cal B}^2=(J'^{\:2}/3)S(S+1)\simeq0.06$, (b)
$J'/J=1.5,0.92,0.67,0.53$ corresponding to the stronger coupling ${\cal
B}^2\simeq 0.57$.}
\label{s_scale}
\end{figure}
As for the scaling with impurity spin $S$ suggested by the
proportionality of the 2nd moment to ${\cal B}^2=(J'^{\:2}/3)S(S+1)$
we show in Fig. \ref{s_scale} MCLM results for $\kappa(\omega
L)/L$ for a series of S-values and couplings $J'$ so that the effective
perturbation strength ${\cal B}^2$ retains its value. We
find indeed that at both weak as well as strong coupling the scaling
is well obeyed, giving a wider applicability to our results.
They can be applied to a range of impurity spin values becoming directly
relevant in the interpretation of experiments.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{w_scale}
\caption{Memory function $\tilde N''(\omega)$ for a strong coupling
$J'=2J$ and for various lattice sizes $L=12-24$, using both ED and
FTLM. Inset: the scaled function $\tilde N''(\omega L)$ is shown at
low frequencies.}
\label{w_scale}
\end{figure}
Now let us address the generic $L\rightarrow \infty$ behavior. We can
discuss it by considering the memory function $N(\omega)$ representation
defined via the general complex function $\bar{\kappa}(\omega)$,
\begin{eqnarray}
\bar{\kappa}(\omega)&=&i\beta\frac{\chi_0} {\omega+N(\omega)},~~~
\chi_0= \chi(\omega\rightarrow 0).
\end{eqnarray}
where the real $\kappa(\omega)=\bar{\kappa}'(\omega)$ and
$N''(\omega) \sim 1/\tau$ plays the role of the (frequency
dependent) thermal-current relaxation rate. The lowest moments $\mu_n$
can be evaluated (in principle) exactly in the high-$T$
limit\cite{ins} on a finite size lattice of $L$ sites. Involving
only local quantities, at least for $0<n<L/2$, they should behave as
$\mu_n = \tilde \mu_n/L$ whereby $\tilde \mu_n$ is size independent
for $n<L/2$. It is plausible that also higher moments, $n>L/2$, behave
as $\mu_n \propto 1/L$.
If $\tilde \mu_n$ for $n>L/2$ would be also size independent,
then this would imply the scaling
$N(\omega)=\frac{1}{L}\tilde N(\omega)$, with a universal (size
independent) $\tilde N(\omega)$. Consequently
\begin{equation}
\bar{\kappa}(\omega)=\frac{i\beta\chi_0 L}{(\omega L) + \tilde N (\omega)},
\end{equation}
with the real part $\kappa(\omega)$ for $L\rightarrow \infty$ and
$\omega \to 0$ obeying
the Lorentzian scaling relation,
\begin{equation}
\frac{\kappa(\omega L)}{L}=\frac{\beta\chi_0 \tilde N''(\omega\rightarrow 0)}
{(\omega L)^2 + \tilde N''(\omega\rightarrow 0)^2},
\end{equation}
provided that $N''(\omega\rightarrow 0)$ is finite. This is, however,
clearly not what we observe in Fig. \ref{l_scale}, where from the
non-Lorentzian shape we must conclude that the memory function also scales
as $\tilde N(\omega L)$ and thus,
\begin{equation}
\frac{\kappa(\omega L)}{L}=\frac{\beta\chi_0 \tilde N''(\omega L)}
{(\omega L + \tilde N' (\omega L) )^2 + \tilde N''(\omega L)^2}.
\end{equation}
This is not in contradiction with the moments argument, since the
higher moments, $n>L/2$, determine the low frequency behavior.
So we can argue that
at high frequencies $\tilde N(\omega)$ scales as $\omega$ while at low
frequencies as $\omega L$. This scenario is indeed verified in
Fig. \ref{w_scale} at the low/high frequency regimes, where
$N(\omega)$ is extracted from the $\kappa(\omega)$ data.
The FTLM method is used for lattice sizes $L\ge16$ with
$M_L=500$ Lanczos steps and smoothed with an additional frequency
broadening $\delta\omega=0.03$.
On the other hand, we can also explain the observed general $\kappa(\omega
L)/L$ scaling with the similarity to a noninteracting system - with
an impurity. In the latter case, the characteristic scaling $L \omega$
is signature of ``free'' oscillations in the system.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{j_scale}
\caption{Impurity coupling $J'$ dependence of scaled $\tilde N''(L
\omega)/J'^2$ and the comparison with the perturbative
result. Results are obtained for $\Delta=1$ and $L=16$ via ED.}
\label{j_scale}
\end{figure}
To study the crossover from weak to strong coupling regime we show in
Fig. \ref{j_scale} the evolution of the relaxation-rate function
$\tilde N''(\omega L)$ with impurity coupling $J'$ along with a
perturbative evaluation $\tilde N_0''(\omega L)$ using the eigestates
of the Hamiltonian without the impurity.\cite{gw} It is interesting
that the memory function shows an increasingly pronounced structure
with minima at approximately the same frequencies, multiples of
$2\pi/L$ {\it independently of $J'$} and which are not present in the
perturbative calculation. In particular the characteristic frequency
of the minima decreases as the anisotropy parameter $\Delta$ decreases
and thus it apparently related to the velocity of elementary excitations
(spinons) in the system. We can conjecture that this peak structure
is due to a resonant mode, created by
multiple forward/backward scattering on the impurity,
characteristic for the noninteracting system.
It is remarkable that this happens even in this
high temperature limit. This effect has already been seen in
integrable systems where a perturbation seems to affect the totality
of the energy spectrum.\cite{ins}
Now the picture is clear, $\tilde N''(\omega)$
increases as $J'^2$, scales as $\omega L$ at low frequencies and at
the same time develops a structure that dominates the behavior of
$\kappa(\omega L)$ turning the Lorentzian weak-coupling shape to a
nontrivial one at strong coupling.
\section{Weak links - finite $T$}
Next we examine the behavior of the thermal conductivity
$\kappa(\omega)$ as we lower the temperature, starting with the
influence of static weak exchange links.
Kane-Fisher\cite{kf} for a Luttinger liquid and
Eggert and Affleck\cite{affleck} (EA) for the isotropic
spin-$1/2$ Heisenberg chain, proposed that a weak link
leads to an open chain (cutting) in the low energy
limit. In contrast, a defect of two adjacent weak links is ``healed"
leading to a uniform chain at $T=0$.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{1wl}
\caption{Frequency dependence of: (a) the normalized thermal
conductivity $\kappa(\omega L)/L$, (b) the extracted memory function
$\tilde N''(\omega L)$, for a chain of $L=22$ sites with one weak link
$\tilde J=0.7J$ and various $T/J=0.3 - 2.0$ . (c) Temperature
dependence of $\kappa_{dc}(T)/L$.}
\label{wl1}
\end{figure}
To analyze this effect we consider a chain with only one weak link, that
is one altered bond with coupling e.g. $J_{0,1}=\tilde J$ in an
otherwise uniform chain ($J'=0$, there is no spin impurity).
The characteristic Kane-Fisher temperature is given in the weak coupling
limit by $T_{KF}\sim (J-\tilde J)^2/J$.
In Fig. \ref{wl1}a we show the corresponding $\kappa(\omega L)/L$ for
$\tilde J=0.7J$ and a series of temperatures. The data are obtained
using the FTLM method for a chain of $L=22$ spins, by $M_L=2000$
Lanczos steps and smoothed by an additional frequency broadening
$\delta \omega = 0.007$. From Fig.~\ref{wl1}a
we notice that $\kappa(\omega L)/L$
develops a strongly nonmonotonic frequency dependence by lowering the
temperature, with a maximum at a finite frequency that suggests a flow
to the strong coupling limit similar to the one discussed before by
increasing $J'$. In Fig.~\ref{wl1}b, the extracted $\tilde N''(\omega L)$
for various $T$ is presented, with the development of a
characteristic structure that explains the nonmonotonic behavior of
$\kappa(\omega)$. The increasing value of $\tilde N''(0) \sim 1/\tau$
with decreasing temperature indeed corresponds to the effect of
``cutting" of the chain.
Nonmonotonic is also the frequency dependence of
$\kappa(\omega L)/L$ for the case of two adjacent equal
weaker links, $J_{L-1,0}=J_{0,1}=\tilde J= 0.7 J$, as shown in
Fig. \ref{wl2}a. However, in this case we observe in Fig. \ref{wl2}b
the opposite behavior of $\tilde N''(\omega)$. Namely ``healing" of the
double defect deduced by the decreasing
$\tilde N''(0)$ as the temperature is lowered in agreement with theoretical
prediction.\cite{affleck} We should note that both cutting/healing are
low frequency effects at frequencies $\omega L\:\: O(1)$.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{2wl}
\caption{Frequency dependence of: (a) the normalized thermal conductivity
$\kappa(\omega L)/L$,
(b) the extracted memory function $\tilde N''(\omega L)$
for a chain of $L=22$ sites with two adjacent weak links $\tilde J=0.7 J$
and various $T/J=0.3 - 2.0$. (c) Temperature dependence of $\kappa_{dc}(T)/L$.}
\label{wl2}
\end{figure}
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{wl12n0}
\caption{Temperature dependence of $\tilde N''(0)$ for $\tilde{J}=0.5,0.7J$
showing cutting/healing behavior for one and two weak links.}
\label{wl12n0}
\end{figure}
To summarize the observed behavior we show in Fig. \ref{wl12n0},
the $T$-dependence of the relaxation rate $\tilde N''(0)$
for two different couplings $\tilde{J}/J=0.5,0.7$,
for one and two weak links, respectively. The presented results confirm the
existence of the cutting behavior at low $T$ for a single link, as well
as the healing by lowering $T$ for two adjacent and equal links.
As expected, both effects appear only at low $T/J < 1$ while the dependence
of the characteristic $T_{KF}$ on $\tilde J/J$ is less pronounced.
\section{Spin coupled to the chain - finite $T$}
Finally we can study the effect of lowering the temperature on the
scattering by a magnetic impurity.
According to EA it leads to cutting the chain at $T=0$
irrespective of the sign of $J'$.
This proposal was extended by Furusaki and Hikihara\cite{furusaki}
to the anisotropic spin chain $-1 < \Delta \le 1$ where they
furthermore proposed that for $-1 < \Delta < 0$ (attractive case in
the fermionic language) there is ``healing" of the impurity, in analogy
to the case of two adjacent weak links.
In the Kondo problem the characteristic temperature in the
weak coupling limit is given by $T_K\sim v \exp(-c/J')$ with
$c$ being a constant, $v$ the velocity
of spin excitations and $J'$ the Kondo coupling. In the case of a spin-$1/2$
chain it was shown\cite{frahm} that the exponential dependence is
replaced by $T_K\sim \exp (-\pi \sqrt {1/J' -(S'+1/2)^2})$
and a next-nearest neighbor coupling
$J_2\simeq 0.2412$ is needed to recover the traditional Kondo case.
We should note that in the model studied the impurity spin is attached
only at the end of the chain - in contrast to our model - but plausibly
the behavior is qualitatively similar. To get a qualitative idea of
orders of magnitude for our problem \cite{laflorencie} for
$J'=0.3J$, $T_K\sim 0.014$, $\xi_K\sim 40$,
for $J'=0.6J$, $T_K\sim 0.388$, $\xi_K\sim 4$ and $J'=J$, $\xi_K=0.65$.
As in our study we are limited to $T \ge 0.4$ in order to see a ``Kondo"
crossover we must consider a coupling
$J' \ge 0.5J$ and thus we are in the relatively strong coupling regime,
with typical screening length of the order $\xi_K \sim 1$.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{cutheal.pdf}
\caption{Frequency dependent normalized thermal conductivity
$\kappa(\omega L)/L$ for strong coupling $J'=2J$, $\Delta=\pm0.5$ and
three $T/J=50,2,0.4$. Inset: $T$-dependence of $N''(0)$ for $\Delta=\pm0.5$
and $\Delta=+1$.}
\label{cutheal}
\end{figure}
In Fig. \ref{cutheal} we show $\kappa(\omega L)/L$ for a chain of $L=22$ sites
at strong coupling $J'=2J$ and two representative cases $\Delta=\pm 0.5$
as we lower the temperature.
Indeed we find at low frequencies the gradual development of the corresponding
``cutting/healing" behavior which we exemplify in the inset by $\tilde N''(0)$
as a function of temperature both for $\Delta=\pm 0.5$ and the
most typical isotropic case $\Delta=+1.0$.
It is remarkable that the tendency to increase-decrease the scattering time
is already evident from high $T$, presumably due to the local
character of the effect because of the strong $J'$ coupling.
We note in passing that the $\omega L$ scaling is found
not just at high $T$ but rather at all $T$ (not shown).
Next in Fig. \ref{memjp} we show $\tilde N''(0)$ as a function of $T$
for a series of increasing $J'$ couplings. The ``cutting" effect for
the repulsive case $\Delta=+0.5$ is present for all values of $J'$
with no easily distiguishable ``Kondo" temperature. We are always
dealing with screening lengths well less than the system size where presumably
no subtle many-body effects come into play. On the other hand,
in the attractive case
$\Delta=-0.5$, we do not observe ``healing" for the weakest coupling
$J'=+0.5$ where the screening length is expected to be several lattice sites.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{memjp}
\caption{$\tilde N''(0)$ vs. $T$ for the
repulsive (attractive) case $\Delta=+0.5 (-0.5)$ for different
$J'/J=0.5,1.0,1.5$.}
\label{memjp}
\end{figure}
Finally, in Fig. \ref{kdc} we summarize the $T$-dependence of $\kappa_{dc}/L$
for a variety of coupling strenghts $J'/J$ and
$\Delta=\pm 0.5$. The experimentally most interesting
case $\Delta=+1$ corresponding to isotropic antiferromagnetic
as well as ferromagnetic
impurity coupling is shown in Fig. \ref{kdciso}.
For $\Delta>0$ we observe in Fig. \ref{kdc}a and
Fig. \ref{kdciso} a continuous decrease of the $\kappa_{dc}$
with increasing $J'$. This can be explained
with the formation of a local singlet, at least for $T<J'$
which blocks the current through the impurity region.
On the other hand, the $\Delta<0$ case in Fig. \ref{kdc}b
reveals a saturation of $\kappa_{dc}$ with $J'$,
at least for intermediate large $J'$.
However, for severe perturbations ($J'\gg J$) the impurity cannot be
healed by the chain leading inevitably to a further decrease of the $\kappa_{dc}$.
\begin{figure}[ht]
\includegraphics[angle=0, width=.9\linewidth]{kdpm}
\caption{Temperature dependence of $\kappa_{dc}/L$ for a variety of
impurity couplings $J'$ and for: (a) repulsive $\Delta=+0.5$,
(b) attractive $\Delta=-0.5$.}
\label{kdc}
\end{figure}
\begin{figure}
\includegraphics[angle=0, width=.9\linewidth]{kfaf-iso}
\caption{Temperature dependence of $\kappa_{dc}/L$ for a variety of
impurity couplings $J'$, $\Delta=1$ and for: anti-ferromagnetic
couplings (top), ferromagnetic couplings (bottom).}
\label{kdciso}
\end{figure}
\section{Conclusions}
In conclusion, by analysing the unique behavior of the thermal conductivity
of the spin-1/2 Heisenberg model
several effects of the local static and dynamical impurities have been established:
\\(\textit{a})
A single local impurity, either static as the local field \cite{static} and
weak link, or dynamical as the spin coupled to the chain turn the
dissipationless thermal conductivity into an incoherent one.
Numerical results for the dynamical conductivity, best studied at high-$T$,
reveal that a single impurity in a system of $L$ sites shows a universal
scaling form $\kappa(\omega L)/L$ at least in the low-$\omega$ regime.
For weak perturbation, as weakly coupled spins outside
the chain, the scaling form is of the simple Lorentzian type.
On the contrary large local perturbation can lead to a nontrivial
form with the maximum response at $\omega >0$.
\\(\textit{b})
Furthermore, universal oscillations in the dynamical
relaxation rate $N''(\omega)$ become visible, from
the weak coupling regime already, with the period
$\omega \propto 1/L$ being a remnant of the impurity
multiple-scattering phenomena in a noninteracting system.
\\(\textit{c})
Our results confirm the existence of the Kondo-type effects of
impurities on lowering the temperature. In the case of weak links
and for the isotropic Heisenberg model cutting and healing
effects are observed at lower $T$ for a single weak link
and a pair of identical weaker links, respectively, in accordance with theoretical predictions.\cite{kf, eggert} In the case of a spin coupled to the chain the
cutting/healing effects at low $T$ depend on the sign of the anisotropy
$\Delta$. For ferromagnetic anisotropy ($\Delta<0$), the chain
screens the impurity and the system enters the weak coupling regime
as the temperature is decreased. The opposite behavior is obtained
for antiferromagnetic anisotropy ($\Delta>0$), where the system
flows to the strong coupling limit at lower temperatures.
\\(\textit{d})
Obtained data can be used to model the behavior observed in experiments
on materials with spin chains doped with magnetic and nonmagnetic
impurities.\cite{hess}
\section{acknowledgments}
This work was supported by the FP6-032980-2 NOVMAG project and
by the Slovenian Agency grant No. P1-0044.
|
\section*{Abstract}
Before contributing new knowledge, individuals must attain requisite
background knowledge or skills through schooling, training, practice,
and experience. Given limited time, individuals often choose either to
focus on few areas, where they build deep expertise, or to delve less
deeply and distribute their attention and efforts across several
areas. In this paper we measure the relationship between the
narrowness of focus and the quality of contribution across a range of
both traditional and recent knowledge sharing media, including
scholarly articles, patents, Wikipedia, and online question and answer
forums. Across all systems, we observe a small but significant
positive correlation between focus and quality.
\section*{Introduction}
Prior to making novel or useful contributions to scholarly work, an
individual must acquire pertinent knowledge. Since acquisition of such
knowledge often involves time-consuming study or extensive experience,
most individuals tend to build their expertise in a small range of
fields; and experiencing success in one subject area may, through
positive reinforcement and the ability to obtain more resources,
result in additional focus on that subject. On the other hand, there
are myriad successful individuals who dabble in many sciences,
inventors who invent diverse gadgets, and Wikipedia editors who edit
pages on seemingly disparate topics. One might argue that their
versatility is a reflection of superior abilities that can yield greater opportunities for
varied collaborations and cross-polination of ideas. In this paper, we present the first comprehensive, large-scale analysis of this relationship between
individual focus and performance across a broad range of
both traditional and modern knowledge collections.
Much previous research in this area has aimed to quantify the
benefit of interdisciplinarity among researchers at the \emph{group}
level. A study of scholarly articles in the UK, for example, found
that research articles whose coauthors are in different departments at the same
university receive more citations than those authored in a single
department, and those authored by individuals across different
universities yield even more citations on average~\cite{katz1997mcw}.
Multi-university collaborations that include a top tier-university
were found to produce the highest-impact research articles~\cite{jones2008mur}.
It has also been demonstrated that scholarly work covering a range of
fields -- and patents generated by larger teams of coauthors--tend to
have greater impact over time~\cite{wuchty2007idt}. Collaborations
between experienced researchers who have not previously collaborated
fare better than repeat collaborations~\cite{guimera2005team}. In the
area of nanotechnology authors who have a diverse set of collaborators
tend to write articles that have higher
impact~\cite{rafols08diversity}. Finally, diverse groups can,
depending on the type of task, outperform individual experts or even
groups of experts~\cite{page2007dpd}.
All of this work is evidence of a benefit in bringing together diverse
individuals. It does not demonstrate, however, whether diversity in
research focus is beneficial at the \emph{individual} level. One
exception is a study of political forecasting that established that ``foxes'', individuals who know many little things, tend to make better
predictions about future outcomes than ``hedgehogs'' who focus on one big thing ~\cite{tetlock2005expert}. Our work addresses
knowledge contribution in a much broader context than forecasting and,
more importantly, quantifies the relationship between individuals'
narrowness of focus and the corresponding quality of their contributions.
\section*{Results}
In order to cover a broad range of knowledge-generating activities, we study several collections of traditional scholarly publications in
addition to recent, Web-based media collections. The traditional
media we consider include patents and research articles. Our patent collection
consists of 5.5 million patents filed with the USPTO between 1976 and
2006. We consider two sources of research articles:
JSTOR and the American Physical Society (APS). The JSTOR corpus consists of 2 million articles from 1108
journals in the natural sciences, social sciences, and humanities, and
the APS dataset we consider covers over 200,000 research articles in
the single discipline of physics. We complement data from these traditional publication venues with data from two recent, online types of knowledge-sharing activity: Wikipedia and a collection of
question-answering forums. Wikipedia is a collaborative online effort
to document all of human knowledge in a systematic way into a popular
internet-based encyclopedia. The Question and Answer forums we study are
Yahoo Answers (English)~\cite{adamic2008YA}; Baidu Knows
(Chinese)~\cite{jiang09baidu}; and Naver Knowledge iN
(Korean)~\cite{nam09naver}. On each of the sites, millions of
questions are answered each month by individuals with a wide range of
expertise.
Online knowledge-sharing activity includes not just those who specialize in knowledge generation and dissemination, i.e. professional researchers and scholars, but also others who gained their expertise through study and experience. In addition to providing data on different types of
individuals, these datasets represent knowledge generation at
different scales. Authoring a research article or patent in
most cases involves weeks to years of research, culminating in a
significant new result worthy of publication. In contrast,
contributing a fact to Wikipedia or answering a question posed in an
online forum may involve little more than a simple recall of
previously attained knowledge -- and a few minutes of the contributor's
time.
In evaluating focus across such a broad range of activities, we aimed
to use a metric that captures three qualities: \emph{variety},
or how many different areas an individual contributes to;
\emph{balance}, or how evenly their efforts are distributed among
these areas; and \emph{similarity}, or how related those areas are~\cite{rafols08diversity}. We use the
Stirling measure
$\mathcal{F}$, which captures all three aspects~\cite{stirling2007gfa}:
\begin{equation}
\mathcal{F} := \sum_{i,j} s_{i j} p_{i} p_{j}, \nonumber
\end{equation}
\noindent where $p_{i}$ is the proportion of the individual's
contributions in category $i$ and $s_{ij}=n_{ij}/n_{j}$ is a measure of similarity between categories $i$ and $j$, inferred from the number of joint contributors $n_{ij}$ between two categories $i$ and $j$. This $F$ metric assigns a narrower (higher) focus value to an individual who contributes to fewer,
related areas than to someone who contributes
in many unrelated areas.
The categories across which focus is measured differ by the type of knowledge-sharing medium. An inventor's proportion $p_{c}$ of
contributions in subject $c$ is proportional to the number of times the class $c$ is assigned by the inventor or patent examiner to the inventor's patents. Articles in the APS dataset are classified according to the Physics and Astronomy Classification Scheme. For JSTOR articles, in the absence of a pre-defined category structure, we used unsupervised topic models on the full text of
authors' research articles~\cite{blei2003latent}. Wikipedia articles are situated within Wikipedia's category hierarchy, while
answers provided in Q\&A forums are sorted according to the hierarchy of categories of
the corresponding questions.
For each dataset, we sought a relevant, objective measure of
quality of a contribution and evaluated it in the context of its peers. For research
articles, we measured each article's citation count relative to those
of other articles in the same discipline and year~\cite{valderas2007}.
Likewise, patents' citation counts were compared with those of other
patents in the same patent classes and years. In doing this, we
control for discipline-specific factors that can impact a publication's citation count such as publication cycle length and number of publications in the discipline~\cite{stringer2008ejr,seglen1997ifj}.
For Wikipedia contributions, we consider the percentage of words an
author newly introduces to an article that survive subsequent
revisions~\cite{adler2007cdr}.
Finally, for Q\&A forums, we rely on the asker's rating of answers: a good
contributor should have their answer selected as best more often than
expected by chance.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.55\textwidth]{Figure1qvsf.pdf}
\end{center}
\caption{\textbf{Mean quality as a function of focus of contribution}.}
\label{fig:qvsf}
\end{figure}
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.55\textwidth]{Figure2fvsq.pdf}
\end{center}
\caption{\textbf{Mean focus as a function of quality of contribution}.}
\label{fig:fvsq}
\end{figure}
Using these measures of focus and quality, we find that focus is
weakly yet consistently positively correlated with quality across all types of
knowledge contribution systems, as summarized in
Table~\ref{tab:focusQQ}. The relationship between focus and quality is detailed further in Figure~\ref{fig:qvsf}, which shows the variation in
average quality for individuals grouped by their levels of focus. As
focus increases, so does average quality; but the trend levels off or
even reverses for extremely focused individuals. This is clarified by
plotting average focus at a given level of quality (see
Figure~\ref{fig:fvsq}). While high-quality contributors are more
narrowly focused than others on average, very poor contributors sometimes
also dwell in a single area. In Q\&A forums, we find further that
narrowly focused users with poor track records of giving best answers
tend to give answers that are significantly shorter than those of
other users.
\begin{table*}[!htpb]
\begin{center}
\caption{\bf{Pearson correlation between quantity, focus, and quality.} All correlations are significant at
$p < 0.001$. } \label{tab:focusQQ}
\begin{tabular}{l|l|l|l}\\
type & $\rho$(log(quantity),focus) & $\rho$(log(quantity),quality) & $\rho$(focus,quality) \\ \hline \hline
Research articles JSTOR & -0.055 & 0.058 & 0.112\\ \hline
Research articles APS & 0.302 & 0.130 & 0.173\\ \hline
Patents & 0.286 & 0.100 & 0.094\\ \hline
Wikipedia & -0.274 & 0.133 & 0.084\\ \hline
Yahoo Answers & 0.150 & 0.116 & 0.084\\
Baidu Knows &0.095 & 0.083 &0.111\\
Naver KnowledgeIn &0.066 &0.102 &0.169\\
\hline \hline
\end{tabular}
\end{center}
\end{table*}
We note that these datasets provide incomplete views of contributors'
activity. JSTOR archives over a thousand journals, but does not include many more. Inventors' patents prior to 1976 are not captured in our data. Likewise, we parsed
only a subset of the Wikipedia revision history; and while our Yahoo
Answers data set spanned the complete activity of a sample of users,
our Baidu and Naver data sets covered only several months each (see
supplementary materials).
Nevertheless, we believe our results to be robust. We expect that we would find equally strong or stronger correlation between focus and quality, if we had complete records of each individual's contributions. An indication of this is that the correlations between focus and quality strengthen for users for whom we observe a higher number of contributions. For example, the correlation between focus and impact
for inventors with 10-20 patents is just $0.079\pm0.003$, but for
inventors with 50 to 100 patents it increases to
$0.166\pm0.014$. Similarly, Wikipedia users editing between 10 and 20
pages display a correlation of $0.088\pm0.065$, while those editing
between 50 and 100 pages display a correlation of $0.243\pm0.038$.
We find our results to be robust in several other respects as well.
Aside from being consistent across a wide range of media and
performance metrics, our results hold when focus is measured at different levels of granularity, e.g. when using top-level patent, Wikipedia, and Q\&A categories as opposed to subcategories in those data sets, and when we
construct 250 as opposed to 100 topics in the JSTOR dataset. While the
distribution of focus shifts downward as we increase the granularity,
the correlations between focus and other variables remain
qualitatively similar.
\begin{table*}[!htpb]
\begin{center}
\caption{\bf{Pearson correlations between quality, focus, and quantity,
when self-citations are removed.} All correlations are significant at
$p < 0.001$. } \label{tab:focusQQnoself}
\begin{tabular}{l|l|l|l}\\
type &$\rho$(log(quantity),quality) & $\rho$(focus,quality) \\ \hline \hline
research articles JSTOR & 0.051 & 0.095\\ \hline
research articles APS & 0.059 & 0.036\\ \hline
patents & 0.059 & 0.073 \\ \hline
\end{tabular}
\end{center}
\end{table*}
We also find our results to be consistent,
though weaker for the paper and patent data sets, when self-citations, i.e. citations between two papers that share an author with the same last name, are removed. Table~\ref{tab:focusQQnoself} summarizes these findings. Self-citations may inflate the impact of
prolific and focused authors who have greater opportunity and
justification to cite their own work.
We also find
consistent results using alternative focus measures such as Shannon
Entropy, $-\sum_{i} p_{i} ln(p_{i})$. Entropy captures the balance and variety of contribution, but
not similarity, and is negatively correlated with focus. We find the
results to be qualitatively consistent to those obtained using the
Stirling measure. Table~\ref{tab:entropyQQ} and Figures~\ref{fig:QvsE} and \ref{fig:EvsQ} correspond to Table~\ref{tab:focusQQ} and
Figures~\ref{fig:qvsf} and \ref{fig:fvsq} respectively, but present results using contributor entropy rather than focus.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.55\textwidth]{Figure3qvse.pdf}
\end{center}
\caption{\textbf{Mean quality as a function of entropy of contribution.}}
\label{fig:QvsE}
\end{figure}
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.55\textwidth]{Figure4evsq.pdf}
\end{center}
\caption{\textbf{Mean entropy as a function of quality of contribution}.}
\label{fig:EvsQ}
\end{figure}
\begin{table*}[!htpb]
\caption{\bf{Person correlations between focus, entropy, quantity, and quality. } All correlations are significant at $p <
0.001$.} \label{tab:entropyQQ}
\begin{tabular}{l|c|c|c}\\
data set & $\rho$(entropy,focus) & $\rho$(log(quantity),entropy) & $\rho$(entropy,quality) \\ \hline \hline
research articles JSTOR & -0.606 & 0.082 & -0.218 \\ \hline
research articles APS & -0.601 & 0.083 & -0.116 \\ \hline
patents & -0.848 & 0.160 &-0.117\\ \hline
Wikipedia & -0.755 & 0.629 & -0.103\\ \hline
Yahoo Answers & -0.878 & -0.081 & -0.121\\
Baidu Knows & -0.918 & -0.034 &-0.108\\
Naver KnowledgeIn & -0.946 &-0.175 &-0.223\\
\hline \hline
\end{tabular}
\end{table*}
One remaining concern is that focus and quality are both correlated with a third variable that holds greater explanatory power. One such potential variable is that of quantity. Quantity itself is positively correlated with quality, revealing a possible link between contributor success and motivation or
resources. However, focus remains a significant factor in
the quality of contributions, even once quantity is accounted for (see Table~\ref{tab:regression}). We also note that quantity's correlation with focus varies by medium studied, as shown in Table~\ref{tab:focusQQ}.
The correlation is positive for patents
and Q\&A forums, but negative for research and Wikipedia
articles. Individuals generating many patents or answers tend to focus
their contributions more narrowly, but authors who write a greater
number of research and Wikipedia articles tend to make broader contributions.
\begin{table*}[!htpb]
\caption{\bf{Regression models using quantity and focus as predictors of quality}} \label{tab:regression}
\begin{tabular}{cccccccc}\\
variable & JSTOR & APS & patents & Wikipedia & YA & Baidu & Naver \\ \hline \hline
focus & 1.899$^{***}$ & 0.553$^{***}$ & 0.501$^{***}$ & 0.353$^{***}$ & 0.245$^{***}$ & 0.143$^{***}$ & 0.146$^{***}$ \\
& (0.085) & (0.040) & (0.009) & (0.033) & (0.046) & (0.005) & (0.015)\\
log(quantity) & 0.177$^{***}$ & 0.124$^{***}$ & 0.163$^{***}$ & $0.027^{***}$& 0.046$^{***}$ & 0.052$^{***}$ & 0.039$^{***}$ \\
& (0.014) & (0.008) & (0.003) & (0.002) & (0.005) & (0.002) & (0.008) \\
\hline
$R^{2}$ &0.017 & 0.025 & 0.015 & 0.023 & 0.022 & 0.023 & 0.022\\ \hline \hline
\end{tabular}
\end{table*}
Finally we examine whether individuals narrow or broaden their focus over time. For JSTOR research articles,
patents, and Q\&A forums, a majority of contributors narrow their focus over time (see
Table~\ref{tab:focuschange}). Wikipedia contributors and physicists,
on the other hand, do not appear to specialize further. In addition to
a change in focus, we also observe a slight change in quality. Across
most datasets, contributors tend to improve the quality over time; exceptions include Baidu
Knows, where the change in answer quality is not significant, and JSTOR, where there is a
statistically significant decline in contribution quality. One might speculate that a researcher's early success permits him or her to continue producing publications, but that the quality of those publications may fall due to factors such as moving from a primary contributor to a project management role.
\begin{table*}[!htpb]
\begin{center}
\caption{\bf{Change in focus from first half to second half of contributions.}}
\label{tab:focuschange}
\begin{tabular}{l|c|c|c|} \\
data set & \% who increased focus & av. change focus & av. change quality \\ \hline \hline
research articles JSTOR & 62.0\% & 0.024 & -0.394 \\ \hline
research articles APS & 44.0\% & -0.012 & -0.157 \\ \hline
patents & 79.5\% & 0.139 & 0.080 \\ \hline
Wikipedia & 49.5\% & not. sig. & 0.066 \\ \hline
Yahoo Answers & 69.8\% & 0.056 & 0.056 \\
Baidu Knows & 61.4\% & 0.040 & not. sig. \\
Naver KnowledgeIn & 68.8\%& 0.023 & 0.034 \\
\hline \hline
\end{tabular}
\end{center}
\end{table*}
\section*{Discussion}
In conclusion, we have quantified the value an individual's
focus in contributing knowledge through different media and across a
wide range of fields. Consistently we observe a slight but
significant correlation between an individual's degree of focus and
this individual's quality of contribution. The relationship persists
even when quantity of contributions is taken into account.
How should individuals invest their time? While our results do not demonstrate
causality, the overall trend appears to favor those who do a few
things and do them well. However, individuals who focus in a very
narrow field tend to contribute work that is on average less-well
recognized than that of their slightly less focused peers.
This work immediately suggests several areas for future research. It
would be useful to understand the benefit of narrowing one's focus in
the context of the research field and media type, in addition to the
quality and diversity of one's prior efforts. One could also examine whether highly focused authors of research articles or patents may be introducing marginal or
incremental contributions.
In addition, while these results have shed light on the value of focus
in the context of the individual, they say nothing about focus
in the context of a group. After all, several studies have
demonstrated the value of interdisciplinary collaborations in the
sciences, and we believe that large-scale online knowledge sharing
systems such as those discussed are successful precisely because they
bring together individuals of different backgrounds. This leaves open
the question of whether interdisciplinary collaborations between more
{\em focused}--yet \emph{collectively diverse}--individuals are more
fruitful.
\section*{Materials and Methods}
Table~\ref{tab:datalist} summarizes the seven data sets we used to study
focus and contribution. For each dataset we selected a threshold criterion for
the minimum level of activity needed for an individual to be included.
\begin{table*}[!htpb]
\begin{center}
\caption{\bf{Description of data.}}\label{tab:datalist}
\begin{tabular}{l|rrl} \hline
dataset & time span & \# individuals & threshold \# contributions\\ \hline \hline
JSTOR & 1668-2006 & 37,031 & 10\ articles\\
APS & 1977-2006 & 22,351 & 10\ articles\\ \hline
patents & 1976-2006 & 604,113 & 10 patents\\ \hline
Wikipedia & 2001-2006 & 7,129 & 40 edits\\ \hline
Yahoo Answers & 08/05-03/09 & 5,256 & 40 answers\\
Baidu Knows & 12/07-05/08 & 65,854 & 40 answers\\
Naver KnowledgeIn & 12/08-02/09 & 5,918 & 40 answers\\
\hline \hline
\end{tabular}
\end{center}
\end{table*}
\textbf{Research articles}. A snapshot of \textbf{JSTOR} data includes 2.0
million research articles with 6.6 million citations between them.
JSTOR spans over a century of publications in the natural and social
sciences, the arts, and humanities. For this data set, we needed to address name ambiguity.
For example there were 26,000 instances where a person with the last
name of Smith authored an article and 728 unique combinations of
initials appearing alongside ``Smith''. Identifying two different
individuals as being one and the same would tend to introduce data
points with low focus and an inflated number of articles. Since both
variables are related to quality, we sought to exclude such
instances. We excluded authors with $\sqrt{F_{L}*L_{F}} > 200$, where
$F_{L}$ is the number of first names or initials the inventor's last
name occurs with in the data set, and $L_{F}$ is the number of last
names the inventor's first name occurs with. We also collapsed matching names and
initials if there was only one matching first name/inital pair and
the last name occurred with fewer than 50 first names. This left us
with 37,031 authors with 10 or more publications, for whom we were
reasonably sure that they were uniquely identified.
Using Latent Dirichlet Allocation~\cite{blei2003latent}, we generate 100
topics over the entire corpus of research articles. Each document was
assigned a normalized score for each of the 100 topics, and the
pairwise topic similarity matrix $s$ was computed from cosines of
vector values across documents. An author's distribution across topics
was computed by averaging the topic vectors of all of the articles they
authored. For robustness, we repeated the analysis with 250 topics
instead of 100, and found quantitatively similar correlation between
focus and quality, although focus scores were lower due to the finer
granularity. The quality of an article is measured as the number
of times the article is cited, divided by the number of times other
articles in the same area and year are cited. Citations originate
within the dataset. By normalizing quality by area, we mitigate the
possible biases introduced by some areas being better represented in
the dataset than others.
Our database of \textbf{American Physical Society} publications included
Physical Review Letters, and Physical Review A-E journal articles. We
excluded Reviews of Modern Physics as we were considering the impact
of original research rather than review articles. The data set
contained 396,134 articles published between 1893 and 2006, with
3,112,706 citations between them. For our purposes, we were limited to
the 261,161 articles with PACS (Physics and Astronomy Classification
Scheme) codes associated with articles published after 1977. The PACS
hierarchy has 5 levels, and we performed our analysis at the level of
the 76 main categories, such as 42 (Optics) and the 859 2nd level
categories, e.g. 42.50 (Quantum Optics).
\textbf{ Patents}. The patent data set contains all 5,529,055 patents
filed between 1976 and 2006, in 468 top level categories. We construct
a similarity matrix for the 468 categories, reflecting the frequency
with which inventors in one category also file patents in
another. There are 3,643,520 patents citing 2,382,334 others, for a
total of 44,556,087 citations. We excluded inventors with
$\sqrt{F_{L}*L_{F}} > 150$. This makes it unlikely that we would
identify two separate individuals as being one. We measure an
inventor's impact according to a citation count normalized by the
average number of citations for other patents in the same year and
categories as those filed by the inventor.
\textbf {Q\&A forums}: We obtained snapshots of activity on Q\&A
forums with uniquely identified users posting answers to questions in
distinct categories. We perform our analysis at the subcategory level,
which gives us enough resolution to differentiate the question topics,
while supplying a sufficient number of observations in each
subcategory. We use best answers as a proxy for answer quality. The
best answer is selected by the user who posed the question. If this
user does not select a best answer, it may be selected via a vote by
other users. The quality metric we used was the $\gamma$ score, which
compares the number of answers the user gives that were selected as
best among others, relative to the expected number of best
answers. The expected number depends on the number of other answers
provided to each question. $\gamma = $ {\em (observed -
$b_{e})/b_{e}$}. $b_{e}$ the expected number of best answers is
simply given by $b_{e} = \sum_{k}1/a_{k}$, where $a_{k}$ is the total
number of users answering question $k$.
\textbf {Wikipedia}: Our Wikipedia dataset is a meta-history dump file
of the English Wikipedia generated on Nov. 4th, 2006. The
dump file has the entire revision history of about 1.5 million
encyclopedia pages, of which we parsed 100,000, or about
7\%. In order to verify that our sample is unbiased with
respect to topic distribution, we compare the category and
subcategory distributions of our sample to that of a larger corpus
of 1 million pages. The two distributions have a nearly
perfect correlation ($\rho > 0.96^{***}$).
Articles are a product of varying number of revisions, from several to
10,000 for single article. Revisions are contributed
by either registered or anonymous users. Since anonymous users'
revision histories are non-traceable, we only consider registered users
whose unique user names are associated with at least 40
revisions. We excluded Wikipedia administrators from our study because
they may perform a primarily editorial role. In like manner, to better filter the noise of measuring the
quality of words by the final version of articles, we only choose
pages in which ewer than 5\% of the revisions occurred in the
30 days prior to the data dump.
A Wikipedia contributor's focus and entropy were calculated from the
second-level categories of the pages they edit. Each Wikipedia article
belongs to one or more categories. We truncated each hierarchical category
to one of the roughly 500 second-level categories.
The quality of a contribution is measured in terms of $w_{new}$, the
number of new words added by a user to Wikipedia articles, such that
the words were not present in any previous revisions of those
articles. We found a high correlation between the number of new words
that survive 5 revisions, and the number $w_{surv}$ that survive to
the last revision of the article ($\rho > 0.97^{***}$), consistent
with previous analyses of edit persistence~\cite{panciera09wiki}. We
therefore constructed a simple metric by taking the proportion of new
words introduced by the user that are retained in the last version of
a sufficiently frequently edited article:
$w_{surv}/w_{new}$~\cite{adler2007cdr}.
\section*{Acknowledgments}
We thank IBM for providing the patent data, and JSTOR, APS, and Katy Borner
for providing the article citation data. We would also like to thank
Michael McQuaid, Jure Leskovec, Scott Page and Eytan Adar for helpful comments.
This research was supported by MURI award FA9550-08-1-0265 from the Air Force
Office of Scientific Research and NSF award IIS 0855352.
\bibliographystyle{plain}
|
\section{Introduction.}
The spherically symmetric LTB dust models \cite{LTB} are among the best known and most useful exact solutions of Einstein'e equations. Since they allow us to examine non--linear effects analytically, or at least in a tractable way, there is an extensive literature (see \cite{kras,kras2} for comprehensive reviews) using them, mostly as models of cosmological inhomogeneities \cite{KH1,KH2,KH3,KH4,ltbstuff}, but also as in other theoretical contexts, such as gravitational collapse and censorship of singularities \cite{sscoll,joshi} or quantum gravity \cite{quantum}. There is also a widespread literature \cite{LTB1,LTB2,LTBkolb,num3,num2,LTBfin} considering LTB models as tools to probe how the cosmic acceleration associated to recent observations can be accounted for inhomogeneities, without introducing an exotic source like dark energy. LTB models are also a standard choice \cite{LTBchin,LTBave1,LTBave2,LTBave3} to apply Buchert's scalar averaging formalism \cite{ave_review}, in which the effects of dark energy could be mimicked by ``back--reaction'' terms (see \cite{celerier} for a review of all this literature). In practically all articles the models are parametrized by a standard set of free functions and analytic solutions (implicit and parametric).
While the literature is vast and exhaustive, there is still room to explore further development on their theoretical properties (see for example \cite{wainwright} in this context). The present article deals with a theoretical problem that has not been, as far as we are aware, previously examined in the literature, namely: the asymptotic behavior of LTB dust models (through their covariant scalars) in the radial spacelike direction, which can be defined in a covariant manner in terms of spacelike geodesics (radial rays) whose tangent vectors are orthogonal to the 4--velocity and to the orbits of SO(3). In the remaining of this section we explain and summarize the contents of the article.
Basic background on LTB models is provided in section 3: the metric, field equations, classification in kinematic classes: parabolic, hyperbolic and elliptic, as well as a covariant time slicing that defines the space slices ${}^3{\cal{T}}[t]$ orthogonal to the 4--velocity field and marked by constant values of $t$. In section 3 we introduce an alternative set of quasi--local scalar variables \cite{suss02,suss08,sussQL,suss09,suss10}, leading to a ``fluid flow'' description of the dynamics of the models that is similar (and equivalent) to the ``1+3'' approach of Ellis, Bruni and coworkers \cite{ellisbruni89,1plus3}, which for LTB models (as with all spherically symmetric spacetimes) reduce to scalar equations \cite{LRS}. As we showed in \cite{sussQL,suss09,suss10}, these variables can also be understood in the framework of a perturbation formalism on a FLRW ``background'' defined by the quasi--local scalars, which satisfy FLRW dynamics, while their fluctuations are gauge invariant and covariant non--linear perturbations. The quasi--local variables and their fluctuations are potentially useful for a numeric approach to LTB models (see \cite{suss09} and sections XI--XIII of \cite{suss10}), but they are also very handy for analytic and qualitative work, since they lead to an initial value parametrization of the models in which all covariant scalars can be given by simple scaling laws that depend on the two metric functions (scaled to an initial ${}^3{\cal{T}}[t_i]$) and initial value functions.
In section 4 we discuss the definition and properties of the proper radial length, $\ell$, along the radial rays, which is the affine parameter of these geodesics. Since we need to explore how scalars behave as $\ell\to\infty$, we will only consider LTB models (admitting a symmetry center) in which this limit can be realized, which means``open'' models whose 3--dimensional space slices ${}^3{\cal{T}}[t]$ orthogonal to the 4--velocity are homeomorphic (topologically equivalent) to $\mathbb{R}^3$, as $\ell$ is always finite at all ${}^3{\cal{T}}[t]$ in ``closed'' elliptic models with slices ${}^3{\cal{T}}[t]$ homeomorphic to $\mathbb{S}^3$.
In order to probe the asymptotic behavior of covariant (local and quasi--local) scalars along the radial rays, we should (ideally) evaluate these scalars as functions of $\ell$, which without numeric work is practically an impossible task (even qualitatively), as $\ell$ is an integral of a metric function that can only be determined (at best) in implicit form in terms of another metric function. Instead, we provide the conditions so that this radial asymptotic regime can be examined in terms of the dependence of scalars on a well defined radial coordinate. Once this is done, the limit $r\to\infty$ will correspond to $\ell\to\infty$, and as a consequence, a covariant characterization of the asymptotic regime can be provided in terms of the radial coordinate and initial value functions.
Since we are considering LTB models that comply with basic regularity conditions (no shell crossings and scalars may only diverge at a central singularity), we define in section 5 a regular asymptotic regime in the radial direction by demanding that all scalars are smooth and finite as $r\to\infty$ (which now corresponds to $\ell\to\infty$). We also discuss (Lemmas 1 and 2) the relation between $\ell$ and $R=\sqrt{g_{\theta\theta}}$, which is another important invariant in spherically symmetric spacetimes \cite{hayward}. Since the quasi--local scalars satisfy less complicated scaling laws than local scalars, it is more practical to study their radial asymptotics first. Hence, we prove in Lemma 3 in section 6 that both types of scalars, the local and quasi--local, share the same asymptotic behavior (given a common set of assumptions on the initial value functions). We also choose a radial coordinate gauge (as there is a radial coordinate gauge freedom in the LTB metric).
In section 7 we consider the radial asymptotic behavior of the $r$--dependent initial value functions, which are the quasi--local scalars and their fluctuations evaluated at a fiducial ``initial'' slice ${}^3{\cal{T}}_i={}^3{\cal{T}}[t_i]$ (the $A_{qi}$ and $\delta^{(A)}_i$), and are necessarily restricted by regularity conditions (the Hellaby--Lake conditions \cite{ltbstuff,suss10,ltbstuff1}). We assume for these functions a uniform asymptotic convergence to specific (but not restrictive) asymptotic trial analytic functions of $r$: power law, logarithmic or exponential. However, before using these convergence forms to probe the scalars $A_q,\,A$ and the $\delta^{(A)}$ by means of the scaling laws derived in section 4, we introduce in section 8 the distinction between the ``asymptotic limit'', which is simply the limit of scalars as $r\to\infty$, and the notion of ``asymptotic state'', which we define as the set of these limits, together with suitable series expansions around them, evaluated for the metric functions and all covariant scalars and fluctuations.
We examine the asymptotic limits and states separately for parabolic (section 9), hyperbolic (section 10), elliptic (section 11) models and in section 12 for special models with a simultaneous big bang (initial central singularity) and maximal expansion. A summary of these asymptotic limits and states, as well as a brief discussion, are provided in section 13. Depending on the initial value functions, the scalars in open LTB models have as asymptotic limit, either a FLRW cosmology (with zero or negative spatial curvature) or a section of Minkowski spacetime given in ``non--standard'' coordinates that generalize those defining Milne spacetime (the particular solution ``[s2]'' of \cite{ltbstuff}). Within those LTB models whose asymptotic limit is a Minkowski section we can recognize asymptotic states that clearly identify various particular cases of LTB models: sections of Minkowski spacetime that include the Milne universe, self--similar LTB solutions (see pages 344--345 of \cite{kras2} and \cite{sscoll,selfsim}) or sections of the Schwarzschild--Kruskal spacetime given in terms of coordinates constructed with radial geodesics (Lema\^\i tre and Novikov coordinates, see page 332 of \cite{kras2}).
As we comment in section 13, LTB models asymptotic to a FLRW cosmology can be understood in the context of embedding a dust inhomogeneity in a dust FLRW background, without resorting to an artificial matching at a fixed comoving boundary. In other words, these models can be fully relativistic and less artificial representations of the Newtonian ``spherical collapse models'' used as toy models of structure formation (see \cite{suss09} for discussion on this point). This embedding into homogeneous cosmology is not compatible with LTB models asymptotic to Minkowski (in any of their asymptotic states). Instead, some of these models (specially the vacuum dominated hyperbolic ones) can be considered as toy models of a dust inhomogeneity surrounded by a large cosmic void. In this context, these configurations can approximate a spherical realization of the notion of ``finite infinity'' (or ``fi'') suggested by Ellis \cite{fiEllis}, and considered further by Wiltshire \cite{fiWilt}, as a description of an intermediate scale in which galactic cluster structures can be studied as (approximately) asymptotically flat configurations.
The article contains three appendices: Appendix A provides the analytic (parametric and implicit) solutions in the conventional variables, Appendix B discusses the possibility of considering local scalars (instead of quasi--local ones) as initial value functions, while Appendix C summarizes the various particular case spacetimes that follow from specializing the free parameters of LTB models (in our description: the initial value functions). As shown in sections 9--12, the radial asymptotic state of any open LTB model corresponds to a specific spacetime in this list.
\section{LTB models, kinematic classes and a fluid flow time slicing.}
LTB dust models in their conventional variables are characterized by the following metric and field equations $G^{ab}=\kappa\rho u^au^b$
\begin{equation} ds^2=-c^2dt^2+\frac{R'{}^2}{1+E}\,dr^2+R^2(d\theta^2+
\sin^2\theta d\varphi^2),\label{LTB1} \end{equation}
\begin{eqnarray} \dot R^2 &=& \frac{2M}{R} +E,\label{fieldeq1}\\
2M' &=& \kappa \rho \,R^2 \,R',\label{fieldeq2}\end{eqnarray}
where $\rho$ is the rest--mass density, $\kappa=8\pi G/c^2$,\, $E=E(r),\,M=M(r)$, while $\dot R=u^a\nabla_a R=\partial R/\partial (ct)$ and $R'=\partial R/\partial r$.
It is common usage in the literature (see \cite{kras,kras2,ltbstuff}) to classify the solutions of (\ref{fieldeq1}) in ``kinematic equivalence classes'' given by the sign of $E$, which determines the existence of a zero of $\dot R^2$. Since $E=E(r)$, the sign of this function can be, either the same in the full range of $r$, in which case we have LTB models of a given kinematic class, or it can change sign in specific ranges of $r$, defining LTB models with regions of different kinematic class (see \cite{ltbstuff}). These kinematic classes are
\numparts\begin{eqnarray} E = 0,\qquad \hbox{Parabolic models or regions}\label{par}\\
E \geq 0,\qquad \hbox{Hyperbolic models or regions}\label{hyp}\\
E \leq 0,\qquad \hbox{Elliptic models or regions}\label{ell} \end{eqnarray}\endnumparts
where the equal sign in (\ref{hyp}) and (\ref{ell}) holds only in a symmetry center. The solutions of the Friedman--like field equation (\ref{fieldeq1}) for each kinematic class are given in Appendix A. The case $M=0$ with $E>0$ arbitrary has been classified in \cite{ltbstuff} as the solution ``[s2]'' and is locally equivalent to Minkowski spacetime.
\subsection{A fluid flow time slicing and covariant scalars.}
The normal geodesic 4--velocity in (\ref{LTB1}) defines a natural time slicing in which the space slices are the 3--dimensional Riemannian hypersurfaces ${}^3{\cal{T}}[t]$, orthogonal to $u^a$, with metric $h_{ab}=u_au_b+g_{ab}$, and marked by arbitrary constant values of $t$.
Each spacelike slice is a warped product ${}^3{\cal{T}}[t]=\chi[t](r)\times_R \mathbb{S}^2(\theta,\varphi)$, where the warping function is $R(t,r)\geq 0$, the fibers are concentric 2--spheres $\mathbb{S}^2$ with surface area $4\pi R^2(t,r)$ (orbits of SO(3)), while the leaves $\chi[t](r)$ are ``radial rays'' or curves of the form $C(r)=[ct_0, r,\theta_0,\varphi_0]$, with $t_0,\theta_0,\varphi_0$ constants. The rays are orthogonal to the fibers and isometric to each other in any given ${}^3{\cal{T}}[t]$, and are also geodesics of the ${}^3{\cal{T}}[t]$ and spacelike geodesics of the LTB metric (\ref{LTB1}). Since we are assuming the existence of (at least) one symmetry center, then every $\chi[t](r)$ for $t$ constant is diffeomorphic to $\mathbb{R}^+=\{x\,|\,x\geq 0\}$. Hence, we will consider every LTB scalar function as equivalent, under the time slicing given by $u^a$, to a one parameter family of real valued functions $A[t]: \mathbb{R}^+\to \mathbb{R}$ so that $A[t](r)=A(t,r)$.
Besides $\rho$ given by (\ref{fieldeq2}), other covariant objects associated with LTB models are the expansion scalar, $\Theta$, the Ricci scalar of the space slices, ${}^3{\cal{R}}$, plus the shear and electric Weyl tensors, $\sigma_{ab},\,E_{ab}$
\begin{eqnarray} \fl\Theta &=& \tilde\nabla_au^a=\frac{2\dot R}{R}+\frac{\dot R'}{R'},\qquad
{}^3{\cal{R}} = -\frac{2(E\,R)'}{R^2R'},\label{ThetaRR}\\
\fl \sigma_{ab} &=& \tilde\nabla_{(a}u_{b)}-(\Theta/3)h_{ab}=\Sigma\,\Xi^{ab},\qquad
E^{ab}= u_cu_d C^{abcd}={\cal{E}}\,\Xi^{ab},\label{SigEE}\end{eqnarray}
where $h_{ab}=u_au_b-g_{ab}$,\, $\tilde\nabla_a = h_a^b\nabla_b$,\, and $C^{abcd}$ is the Weyl tensor, $\Xi^{ab}=h^{ab}-3\eta^a\eta^b$ with $\eta^a=\sqrt{h^{rr}}\delta^a_r$ being the unit tangent vector along the radial rays (orthogonal to $u^a$ and to the orbits of SO(3)). The scalars ${\cal{E}}$ and $\Sigma$ in (\ref{SigEE}) are
\begin{equation}\Sigma = \frac{1}{3}\left[\frac{\dot R}{R}-\frac{\dot R'}{R'}\right],\qquad
{\cal{E}} = -\frac{\kappa}{6}\,\rho+ \frac{M}{R^3}.\label{SigEE1}\end{equation}
Since LTB models (as all spherically symmetric spacetimes) are LRS (locally rotationally symmetric), they can be completely characterized by covariant scalars. Considering (\ref{fieldeq2}), (\ref{ThetaRR}), (\ref{SigEE}) and (\ref{SigEE1}), these are the local ``fluid flow'' scalars
\begin{equation} \{\rho,\,\Theta,\, {}^3{\cal{R}},\,\Sigma,\, {\cal{E}}\},\label{locscals}\end{equation}
whose evolution equations completely determine the dynamics of LTB models in the fluid flow or ``1+3'' approach \cite{ellisbruni89,1plus3}, and thus provide an alternative approach to that based on the analytic solutions of (\ref{fieldeq1}).
\section{Quasi--local scalars and their fluctuations.}
For every scalar function $A$ in LTB models we define its quasi--local dual $A_q$ as the family of real valued functions $A_q[t]: \mathbb{R}^+\to \mathbb{R}$ given by
\footnote{This section provides the minimal background material to make this article as self--contained as possible. The reader is advised to consult reference \cite{suss10} for details on the quasi--local scalar representation of LTB models.}
\footnote{Since it is clear that $t$ is a fixed arbitrary parameter we will omit henceforth the notation $[t]$ unless it is needed.}
\begin{equation} A_q=\frac{\int_0^{r}{A\,{\cal{F}}\,{\rm{d}}{\cal{V}}_p}}{\int_0^{r}{{\cal{F}}\,{\rm{d}}{\cal{V}}_p}}=\frac{\int_0^{r}{A R^2 R'\,{\rm{d}} x}}{\int_0^{r}{R^2 R'{\rm{d}} x}},\label{aveq_def}\end{equation}
where the integration is along arbirtary slices ${}^3{\cal{T}}[t]$,\,with ${\cal{F}}\equiv (1+E)^{1/2}$,\, ${\rm{d}}{\cal{V}}_p=\sqrt{h_{ab}}{\rm{d}} r{\rm{d}}\theta{\rm{d}}\varphi$, and we are using the notation $\int_0^r{... {\rm{d}} x}=\int_{x=0}^{x=r}{... {\rm{d}} x}$. The quasi--local scalars comply with the following properties
\numparts\begin{eqnarray}
A_q'=(A_q)' = \frac{3R'}{R}\,\left[A-A_q\,\right],\label{propq2}\\
A(r) - A_q(r) = \frac{1}{R^3(r)}\int_0^r{A' \,R^3 \,{\rm{d}} x}.\label{propq3}\end{eqnarray}\endnumparts
Given the pair of scalars $\{A,\,A_q\}$, we define the relative fluctuations as
\begin{equation} \delta^{(A)} \equiv \frac{A-A_q}{A_q}=\frac{A'_q/A_q}{3R'/R} =\frac{1}{A_q(r) R^3(r)}\int_0^r{A'\,R^3\,{\rm{d}} x},\label{Dadef}\end{equation}
where we used (\ref{propq2}) and (\ref{propq3}).
\subsection{Scaling laws and initial value functions.}
The quasi--local scalars lead in a natural manner to an initial value parametrization of LTB models, so that all quantities can be scaled in terms of their value at a fiducial (or ``initial'') slice ${}^3{\cal{T}}_i\equiv {}^3{\cal{T}}[t_i]$, where $t=t_i$ is arbitrary. Hence, the subindex ${}_i$ will denote henceforth ``initial value functions'', which will be understood to be scalar functions evaluated at $t=t_i$. This procedure suggests rephrasing the metric functions $R$ and $R'$ as dimensionless scale factors
\begin{equation} L = \frac{R}{R_i},\qquad
\Gamma = \frac{R'/R}{R'_i/R_i}=1+\frac{L'/L}{R'_i/R_i},\label{LGdef}\end{equation}
leading to scaling laws for all covariant scalars, which now become functions of the scale factors $L,\,\Gamma$ and initial value functions. Considering the definition (\ref{aveq_def}), the quasi--local duals of $\rho,\,\Theta$ and ${}^3{\cal{R}}$ depend only on $L$ and initial value functions
\begin{eqnarray} m_q = \frac{m_{qi}}{L^3} = \frac{M}{R^3},\label{mq}\\
k_q = \frac{k_{qi}}{L^2} =-\frac{E}{R^2},\label{kq}\\
{\cal{H}}_q^2 = \frac{\dot L^2}{L^2}=\frac{\dot R^2}{R^2}=2m_q-k_q=\frac{2m_{qi}-k_{qi}L}{L^3},\label{Hq}\end{eqnarray}
where, to simplify the notation, we have introduced (and will use henceforth) the definitions
\begin{equation}\fl 2m\equiv \frac{\kappa}{3}\,\rho,\quad 2m_q\equiv \frac{\kappa}{3}\,\rho_q,\qquad k\equiv \frac{{}^3{\cal{R}}}{6},\quad k_q\equiv \frac{{}^3{\cal{R}}_q}{6},\qquad {\cal{H}}\equiv\frac{\Theta}{3},\quad {\cal{H}}_q\equiv\frac{\Theta_q}{3}.\label{mkHdefs}\end{equation}
Scaling laws for the local scalars follow readily from (\ref{fieldeq2}) and (\ref{ThetaRR}) as:
\numparts\begin{eqnarray} m=\frac{m_{qi}}{L^3}\,[1+\delta^{(m)}] =\frac{m_i}{L^3\,\Gamma},\label{slaw1}\\
k=\frac{k_{qi}}{L^2}\,[1+\delta^{(k)}] = \frac{k_i}{L^2\,\Gamma}\,\left[1+\frac{\Gamma-1}{3\,(1+\delta_i^{(k)})}\right].\label{slaw2}\end{eqnarray}\endnumparts
Since local fluid flow scalars in (\ref{locscals}) are expressible in terms of $m_q,\,{\cal{H}}_q$ and $k_q$ and their fluctuations (\ref{Dadef})
\numparts\begin{eqnarray} \fl m=m_q\,\left[1+\delta^{(m)}\right],\quad {\cal{H}}={\cal{H}}_q\,\left[1+\delta^{(\HH)}\right],\quad k=k_q\,\left[1+\delta^{(k)}\right],\label{qltransf}\\ \fl \Sigma = -{\cal{H}}_q\,\delta^{(\HH)},\quad {\cal{E}}=-m_q\,
\delta^{(m)}.\label{qltransf2}\end{eqnarray}\endnumparts
we get, with the help from (\ref{Dadef}), (\ref{mq}), (\ref{kq}) and (\ref{Hq}), the scaling laws for the fluctuations
\numparts\begin{eqnarray} 1+\delta^{(m)}=\frac{1+\delta_i^{(m)}}{\Gamma},\label{slawDm}\\
\frac{2}{3}+\delta^{(k)} = \frac{2/3+\delta_i^{(k)}}{\Gamma},\label{slawDk}\end{eqnarray}\endnumparts
\begin{eqnarray} 2\delta^{(\HH)} &=& \frac{2m_q\,\delta^{(m)}-k_q\,\delta^{(k)}}{2m_q-k_q}=\frac{2m_{qi}\,\delta^{(m)}-k_{qi}\,L\,\delta^{(k)}}{2m_{qi}-k_{qi}\,L}\nonumber\\
&=& \frac{2m_{qi}\,[\delta_i^{(m)}+1-\Gamma]-k_{qi}\,L\,[\delta_i^{(k)}+\frac{2}{3}(1-\Gamma)]}{[2m_{qi}-k_{qi}L]\,\Gamma},\label{slawDh}\end{eqnarray}
which allow us to obtain scaling laws for the local expansion scalar, ${\cal{H}}$, and the scalar function associated with the shear tensor, $\Sigma$.
We note that the scalars $m_q,\,{\cal{H}}_q,\,k_q$ and their fluctuations are covariant objects, as $M,\,E,\,R,\,\dot R=u^a\nabla_a R$ are invariants in spherically spacetimes \cite{hayward}. Since the conventional variables $M$ and $E$ depend only on $r$, it is convenient to define them as initial value functions
\begin{equation}M = m_{qi} R_i^3,\qquad E = -k_{qi} R_i^2. \label{ME}\end{equation}
Given (\ref{LGdef}) and (\ref{ME}), the LTB metric (\ref{LTB1}) takes the form
\begin{equation} {\rm{d}} s^2=-c^2{\rm{d}} t^2+L^2\left[\frac{\Gamma^2\,{R'_i}^2\,{\rm{d}} r^2}{1-k_{qi}\,R_i^2}+R_i^2\left({\rm{d}}\theta^2+\sin^2\theta{\rm{d}} \phi^2\right)\right],\label{LTB2}\end{equation}
The Friedman--like equation (\ref{fieldeq1}) now takes the form (\ref{Hq}). Its solutions, are equivalent to those of (\ref{fieldeq1}) in Appendix A, but now expressed in terms of $L,\,m_{qi},\,k_{qi}$. These solutions will be given explicitly in sections 9, 10 and 11. We can compute from these solutions a scaling law for $\Gamma$.
\subsection{Curvature singularities.}
The scaling laws (\ref{mq})--(\ref{slawDh}) clearly indicate the existence of two possible curvature singularities whose coordinate locus is
\numparts\begin{eqnarray} L(t,r) &=& 0,\qquad \hbox{central singularity}\label{Lzero}\\
\Gamma(t,r) &=& 0 \qquad \hbox{shell crossing singularity}.\label{Gzero}\end{eqnarray}\endnumparts
so that for reasonable initial value functions (bounded and continuous), all scalars $A_q=m_q,\,k_q,\,{\cal{H}}_q$ diverge as $L\to 0$, whereas local scalars $A=m,\,k,\,{\cal{H}}$ can also diverge if $\Gamma\to 0$ (even if $L>0$). Notice that if $\Gamma>0$, then all scalars $A$ and $A_q$ only diverge at the central singularity $L=0$, which is an intrinsice feature of LTB models. However, if $\Gamma\to 0$ for $L>0$, then all the relative fluctuations $\delta^{(A)}$ diverge (with $A_q\ne 0$), so that local scalars $A$ diverge while their quasi--local duals $A_q$ remain bounded. This is an obviously unphysical effect of shell crossings that must be avoided. We will denote by ``regular LTB models'' all configurations for which shell crossing singularities are absent, thus complying with
\begin{equation} \Gamma > 0 \quad\forall\;\; (ct,r)\quad\hbox{such that}\quad L>0 \label{noshxG}\end{equation}
In order to test this regularity condition we need to compute $\Gamma$, which will be done separately for parabolic, hyperbolic and elliptic models in sections 9, 10 and 11.
\section{A well behaved proper radial length.}
In order to examine the radial dependence of scalar functions we need to define the ``radial direction'' in precise and covariant terms. There is no inherent covariant meaning in the radial coordinate. In fact, the metrics (\ref{LTB1}), (\ref{LTB2}) and $h_{ab}$ are invariant under arbitrary re--scalings $r=r(\bar r)$, indicating the existence of a coordinate gauge freedom that can always be used, either to simplify computations or to eliminate any initial value function by using it as radial coordinate. Yet, radial rays are geodesic curves whose affine parameter is radial proper length, and so the radial dependence of scalars at every individual ${}^3{\cal{T}}[t]$ acquires a covariant meaning by relating it to this parameter.
The proper radial length along an arbitrary ${}^3{\cal{T}}[t]$ can be defined as the function $\ell[t]:\mathbb{R}^+\to \mathbb{R}$ such that
\begin{eqnarray} \ell[t](r)=\int_{0}^{r}{\sqrt{g_{rr}}\,{\rm{d}} x}=\int_0^r{\frac{R'}{{\cal{F}}}\,{\rm{d}} x}=\int_0^r{L\,\Gamma\, \frac{R'_i}{{\cal{F}}}\,{\rm{d}} x},\label{elldef}\\
\hbox{where}\qquad {\cal{F}} \equiv \sqrt{1+E} =[1-k_{qi} R_i^2]^{1/2},\label{Fdef}\end{eqnarray}
so that $\ell[t](0)=0$ for all $t$. A well behaved proper length must necessarily satisfy $\ell(r)>0$ and $\ell\,'(r)>0$ for $r>0$ at all ${}^3{\cal{T}}[t]$, so that $r_2>r_1\,\Leftrightarrow\,\ell(r_2)>\ell(r_1)$ at all ${}^3{\cal{T}}[t]$. Considering (\ref{Lzero}) and (\ref{noshxG}), these requirements are satisfied for regular radial rays in regular LTB models if the following regularity condition among initial value function holds
\begin{equation} \hbox{sign}\,(R'_i(r)) = \hbox{sign}\,{\cal{F}}(r).\label{RirF}\end{equation}
Thus, if $R_i$ has a zero, it must be a common zero of ${\cal{F}}$ (and of the same order). While proper length is the natural parameter to characterize radial dependence in a covariant manner, it is not convenient to use it as a spacetime coordinate because in general:\, $\ell=\ell(t,r)$ (the same remark applies to another important scalar like $R$). Hence, for practical reasons we need to describe the radial dependence of scalars in terms of their radial coordinate. Since we can always use radial coordinate gauge freedom to fix $R_i$, (\ref{RirF}) can be understood as a consistency condition on the radial coordinate, so that it effectively mirrors the dependence on proper radial length along radial rays.
Since (\ref{elldef}) is valid for all ${}^3{\cal{T}}[t]$, then for regular LTB models ($\Gamma>0$) the sign of $R'$ is the same as the sign of $R'_i$ for all $r$. Therefore, as long as (\ref{RirF}) holds, the radial coordinate can be used to probe the radial profiles of scalars at an arbitrary ${}^3{\cal{T}}[t]$, as these profiles will be qualitatively analogous to those with respect to $\ell$:
\numparts\begin{eqnarray} R' = \frac{\partial R}{\partial \ell}\,\ell\,'=\frac{\partial R}{\partial \ell}\,\frac{R'}{{\cal{F}}}\quad\Rightarrow\quad \hbox{sign}\,(R') = \hbox{sign}\,(\partial R/\partial\ell),\label{Rrxi}\\
A' = \frac{\partial A}{\partial \ell}\,\ell\,'=\frac{\partial A}{\partial \ell}\,\frac{R'}{{\cal{F}}}\quad\Rightarrow\quad \hbox{sign}\,(A') = \hbox{sign}\,(\partial A/\partial\ell).\label{Arxi} \end{eqnarray}\endnumparts
While a zero of $A'$ necessarily corresponds to a zero of $\partial A/\partial\ell$ at any individual ${}^3{\cal{T}}[t]$, in time dependent scalars $A=A(t,r)$, a zero of $A'$ can arise (in general) at different values of $r$ (or $\ell$) for different ${}^3{\cal{T}}[t]$, or it can arise in some of the ${}^3{\cal{T}}[t]$ and not in others, all of this without violating (\ref{noshxG}), (\ref{RirF}) or (\ref{Rrxi})--(\ref{Arxi}). As we mentioned above, the case of $R$ is different: a zero of $R'$ must be common to a zero of ${\cal{F}}$, and so it necessarily occurs at a fixed $r$ and is common to all ${}^3{\cal{T}}$.
Notice that (\ref{RirF}) could be violated even if shell crossings are absent ((\ref{noshxG}) holds). In these situations there would be a surface layer at the comoving worldline marking the zero of $R'_i$ \cite{ltbstuff,suss02}. It is evident that such a surface layer also implies an ill--defined proper radial length.
\section{A regular asymptotic regime in the radial direction.}
A radial asymptotic regime is associated with convergence and behavior of metric functions and covariant scalars the limit $\ell\to\infty$ along radial rays in the space slices ${}^3{\cal{T}}(t)$. However, given the existence of (at least) one symmetry center, these slices can be homeomorphic to either $\mathbb{R}^3$ (``open'' models) or to $\mathbb{S}^3$ (``closed'' models), which must be elliptic because of (\ref{RirF}). Since $\ell$ is everywhere finite in closed models, then the limit $\ell\to\infty$ can only be realized in open models, which can be of all kinematic classes (hyperbolic, parabolic or elliptic with $-1<E<0$). We will only consider open models for the remaining of this article.
Assuming regular open models for which (\ref{noshxG}) and (\ref{RirF}) hold, we will consider the radial asymptotic regime as regular if the scalars $\{m,\,{\cal{H}},\,k\}$ and $\{m_q,\,{\cal{H}}_q,\,k_q\}$ converge to finite values as $\ell\to\infty$. If we adopt as a criterion to define a curvature singularity that curvature scalars diverge if a given coordinate locus is reached by geodesics in finite affine parameter values, and since $\ell$ is such a parameter, then there would be no singularity (technically speaking) if these scalars diverge as $\ell\to\infty$. Also, it is possible to conceive a situation in which $k,\,{\cal{H}}$ and their quasi--local duals diverge (even at finite $\ell$) while the densities $m,\,m_q$ remain bounded, so that a singularity does not arise because curvature scalars remain bounded (see \cite{ltbstuff} and Appendix A of \cite{suss10}). While these situations cannot be ruled out, we will exclude them from consideration and will, nevertheless, assume henceforth that all covariant scalars ($m,\,k,\,{\cal{H}}$ and their quasi--local duals) may only diverge at a central singularity (\ref{Lzero}), remaining bounded everywhere and also in the limit $\ell\to \infty$. \footnote{The relative fluctuations $\delta^{(A)}$ might diverge for finite $r$ under regular conditions, see Appendix A4 of \cite{suss10}}
\subsection{Relation between $\ell$ and $R$.}
Since $R'>0$ holds everywhere in open models and $R$ is an important invariant scalar involved in the definition of $\ell$, it is necessary to examine the relation between the limits $\ell\to\infty$ and $R\to\infty$. We prove now the following
\begin{quote}
\noindent \underline{Lemma 1}. Let $\ell(r)$ given by (\ref{elldef}) be the proper radial length along an arbitrary ${}^3{\cal{T}}[t]$ homeomorphic to $\mathbb{R}^3$ in a regular LTB model, the limit
\begin{equation}
\mathop {\lim }\limits_{\ell \to \infty } R(\ell ) = \infty ,\label{Lemma1}\end{equation}
holds for all parabolic and hyperbolic models, and for elliptic models in which ${\cal{F}}=\sqrt{1+E}$ converges to a nonzero constant as $\ell\to\infty$.\\
\noindent \underline{Proof}. As a consequence of (\ref{Rrxi}), we have
\begin{equation}\fl \qquad {\cal{F}}(r(\ell)) = \frac{\partial R(\ell)}{\partial \ell}\quad\Rightarrow\quad R(\ell)=\int_0^\ell{{\cal{F}}(\bar\ell) {\rm{d}}\bar\ell},\label{RintF}\end{equation}
which is valid at each ${}^3{\cal{T}}[t]$ separately. Since ${\cal{F}}=1$ for parabolic models and ${\cal{F}}\geq 1$ and ${\cal{F}}'\geq 0$ hold for all regular hyperbolic models, the result follows directly from (\ref{RintF}). For open elliptic models, ${\cal{F}}$ lies in the range $0<{\cal{F}}\leq 1$ with ${\cal{F}}(0)=1$. If ${\cal{F}}'\leq 0$ then as $\ell\to\infty$ we have ${\cal{F}}\to {\cal{F}}_0$ with $0\leq {\cal{F}}_0<1$, then (\ref{RintF}) implies
\begin{equation} {\cal{F}}_0\,\ell\leq R(\ell)\leq \ell,\label{F0}\end{equation}
and so the result follows. If ${\cal{F}}\to 1$ as $\ell\to\infty$, then ${\cal{F}}'$ must have a zero at some $r=r_{\rm{tv}}$ (which is a minimum of ${\cal{F}}$ because ${\cal{F}}(0)=1$), then ${\cal{F}}\geq {\cal{F}}(r_{\rm{tv}})$ holds for all $r\ner_{\rm{tv}}$ and ${\cal{F}}'>0$ for all $r>r_{\rm{tv}}$. In this case constraining bounds similar to (\ref{F0}) can be constructed from (\ref{RintF}) with ${\cal{F}}_0={\cal{F}}(r_{\rm{tv}})$ and the result follows. However, if ${\cal{F}}\to 0$ as $\ell\to\infty$, then (depending on how ${\cal{F}}$ converges to zero) $R$ might converge to a finite constant in this limit.\\
\noindent \underline{Corollary}. If the limit (\ref{Lemma1}) holds in one ${}^3{\cal{T}}[t]$, it must hold in all ${}^3{\cal{T}}[t]$. The proof is straightforward, since Lemma 1 is valid for arbitrary ${}^3{\cal{T}}[t]$. Also, the ${}^3{\cal{T}}[t]$ constitute a smooth foliation of LTB models and (if standard regularity holds) the rays in all ${}^3{\cal{T}}[t]$ are complete geodesics in the direction of increasing $\ell$.\\
\noindent \underline{Lemma 2:\, (converse of Lemma 1)}. The limit
\begin{equation}
\mathop {\lim }\limits_{R \to \infty } \ell(R ) = \infty \label{Lemma2}\end{equation}
holds for all parabolic and elliptic models, and for all hyperbolic models in which $k_q\to$ constant as $R\to\infty$.\\
\noindent \underline{Proof}. Since ${\cal{F}}$ is related to $k_q$ by (\ref{kq}) and (\ref{ME}), we can rewrite (\ref{elldef}) (and also (\ref{RintF})) along an arbitrary ${}^3{\cal{T}}[t]$ as
\begin{equation} \ell(R)=\int_0^R{\frac{{\rm{d}} R}{[1-k_q R^2]^{1/2}}},\label{xi_vs_k}\end{equation}
where we now consider $k_q=k_q(R)$ and we used the fact that ${\rm{d}} R=R' {\rm{d}} x$ is an exact differential because $t$ is kept constant as the integral in (\ref{elldef}) is evaluated. In parabolic models we have $k_q=0$ everywhere, so the result follows trivially from (\ref{xi_vs_k}). For open elliptic models, (\ref{xi_vs_k}) implies that $k_q R^2\leq 1$ must hold for all $R$, which is equivalent to $1/{\cal{F}}=[1-k_qR^2]^{-1/2}\geq 1$. Therefore, we have in general
\begin{equation} \ell(R) \geq \int_0^R{{\rm{d}} R}=R,\label{xi_vs_ke}\end{equation}
and so $\ell\to\infty$ if $R\to\infty$. For hyperbolic models, (\ref{xi_vs_k}) takes the form
\begin{equation} \ell(R)=\int_0^R{\frac{{\rm{d}} R}{[1+|k_q| R^2]^{1/2}}}.\label{xi_vs_kh}\end{equation}
It is evident then, that the integral in (\ref{xi_vs_kh}) converges only if $|k_q(R)|$ diverges, hence $\ell(R)$ will always diverge as $R\to\infty$ if $|k_q(R)|\to$ constant in this limit. Notice that this result is valid at individual (but arbitrary) ${}^3{\cal{T}}[t]$. In general, $|k_q(R)|$ would converge to a different constant in different ${}^3{\cal{T}}[t]$.
\end{quote}
\section{Radial asymptotics of local and quasi--local scalars.}
Bearing in mind that we are only considering open models for which $R'>0$ holds everywhere, then a convenient (and practical) way to choose the radial coordinate is
\begin{equation}R_i(r)= R_0\,r,\label{Rgauge}\end{equation}
where $R_0$ is a constant length scale. As a consequence of (\ref{Rgauge}), radial dependence becomes dependence on the initial value function $R_i$ and $R_0$ provides a characteristic length scale for the radial coordinate. Since (\ref{Rgauge}) complies with (\ref{RirF}), then $r\to\infty$ implies $\ell\to\infty$ in all open models (including elliptic models for which $R$ tends to a finite constnt in this limit). Also, as a consequence of the corollary of Lemma 1, $r\to\infty$ also implies $R\to \infty$ for all open models (save the above mentioned elliptic models). Unless specified otherwise, we will henceforth assume that the radial coordinate has been fixed by the gauge (\ref{Rgauge}).
Assuming the coordinate gauge (\ref{Rgauge}) and that (\ref{noshxG}) and (\ref{RirF}) hold, we examine now the relation between the radial asymptotic behavior of local and quasi--local scalars.
\begin{quote}
\noindent \underline{Lemma 3}. The following result holds in any given ${}^3{\cal{T}}[t]$
\begin{equation}
\mathop {\lim }\limits_{r \to \infty } A(r) = A_0 \quad \Leftrightarrow \quad \mathop {\lim }\limits_{r \to \infty } A_q (r) = A_0,\label{Lemma3}
\end{equation}
where $A_0$ is a constant.
\smallskip
\noindent \underline{Proof}. We consider the case when $A'\geq 0$ for sufficiently large values of $r$. The case when $A'\leq 0$ is analogous. If the limit of $A$ as $r\to\infty$ is $A_0$, then for all $\epsilon>0$ there exists $u(\epsilon)$ such that $A_0-\epsilon<A(r)<A_0$ holds for all $r>u(\epsilon)$. Constraining $A$ by means of this inequality in the definition of $A_q$ in (\ref{aveq_def}) leads immediately to $A_0-\epsilon<A_q(r)<A_0$, hence $A_q\to A_0$ as $r\to\infty$. The converse result follows from (\ref{propq2}).
\smallskip
\noindent The following results follow trivially from Lemma 3:
\smallskip
\noindent \underline{Corollary 1}:\, $A_q'\to 0$ and $A'\to 0$ both hold as $r\to\infty$.
\noindent \underline{Corollary 2}:\, $A_q\delta^{(A)}=A-A_q\to 0$ as $r\to\infty$.
\end{quote}
\noindent
Since we have assumed that (\ref{RirF}) holds, these results are valid as $\ell\to\infty$ along individual but arbitrary ${}^3{\cal{T}}[t]$, though the constant $A_0$ will be different at different ${}^3{\cal{T}}[t]$ and (in general) $A_0=A_0(t)$. These results are useful because it is easier to probe the asymptotics on the quasi--local scalars $A_q$ first, as they satisfy less complicated scaling laws (that do not involve $\Gamma$) and the analytic solutions are given in terms of $m_{qi}$ and $k_{qi}$. Once we have worked out the $A_q$, the asymptotic behavior of the $A$ follows from Lemma 3 and its corollaries.
\subsection{The relative fluctuations $\delta^{(A)}$.}
\noindent
Lemma 3 yields the following general result:
\begin{quote}
\noindent \underline{Corollary 3 of Lemma 3}:\, If $A\to A_0 \ne 0$ (or, equivalently, $A_q\to A_0 \ne 0$) as $r\to\infty$ in an arbitrary ${}^3{\cal{T}}[t]$, then $\delta^{(A)}\to 0$ in this limit, irrespective of how fast or slow $A$ and $A_q$ converge to $A_0$. The proof follows directly from Corollary 2 of Lemma 3.
\end{quote}
\noindent
However, if $A\to 0$ as $r\to\infty$, then this corollary no longer applies. The limit value of $\delta^{(A)}$ in this case is not (necessarily) zero, but depends on the specific asymptotic form in which $A$ and $A_q$ decay to zero.
\section{Asymptotics of initial value functions.}
The results proven so far are valid at individual (but arbitrary) ${}^3{\cal{T}}[t]$. However, it is very difficult to actually test them because we need to evaluate the involved scalars as functions of $r$ for arbitrary fixed $t$. With the exception of parabolic models, this is very hard because $L$ is, in general, given implicitly in the solutions of the Friedman--like equation (\ref{Hq}) (equivalent to (\ref{fieldeq1})). However, the coordinate gauge (\ref{Rgauge}) provides a simple relation between $R_i$ and $r$, and this allows us to obtain general analytic results on the asymptotic behavior of the initial value functions $A_i,\,A_{qi}$ and their fluctuations $\delta^{(A)}_i$. As we show in sections 9, 10, 11 and 12, these analytic expressions yield analytic forms for the asymptotic behavior of these variables for $t\ne t_i$.
\subsection{Asymptotics of the $\delta^{(A)}_i$}
In order to obtain the limit of the $\delta^{(A)}_i$ when $A,A_q\to A_0=0$, we need to make specific assumptions on the convergence of these scalars to zero.
\begin{quote}
\noindent\underline{Uniform asymptotic convergence}. Let $A_i$ and $\tilde A_i$ be smooth integrable scalar functions on ${}^3{\cal{T}}[t_i]$ both tending to zero as $r\to\infty$. The scalar $A_i$ uniformly converges asymptotically to $\tilde A_i$ if for every $\epsilon>0$ there exists a real positive number $y$, such that $|A_i-\tilde A_i|<\epsilon$ for all $r>y$. We denote this convergence by $A_i\sim \tilde A_i$. Notice that this definition is also applicable to the quasi--local scalars $A_{qi}$ converging asymptotically to a function $\tilde A_{qi}$.
\end{quote}
\noindent
We can assume a given asymptotic convergence either for $A_i$ or for $A_{qi}$. We show bellow how prescribing $A_{qi}$ we obtain $A_i$ and $\delta^{(A)}_i$. The alternative approach (prescribing $A_i$ to obtain $A_{qi}$ and $\delta^{(A)}_i$) is discussed in Appendix B.
It follows directly from the definition of uniform asymptotic convergence that if $A_{qi}\sim \tilde A_{qi}$ then $A'_{qi}\sim \tilde A'_{qi}$ holds for $r>y$. Therefore, considering (\ref{propq2}) and (\ref{Dadef}), if we assume $A_{qi}\sim \tilde A_{qi}$ we obtain
\numparts\begin{eqnarray} A_i\sim \tilde A_{qi} + \frac{\tilde A'_{qi}}{3R'_i/R_i}=\tilde A_{qi} +\frac{r\,\tilde A'_{qi}}{3},\label{Asim}\\
\delta^{(A)}_i \sim \frac{\tilde A'_{qi}/\tilde A_{qi}}{3R'_i/R_i}=\frac{\tilde r\,A'_{qi}}{3\,\tilde A_{qi}}.\label{Daasq}\end{eqnarray}\endnumparts
where we eliminated $R_i$ from (\ref{Rgauge}).
\subsection{Initial density}
While a non--negative $m_{qi}$ is an initial value function of LTB models of all kinematic classes (parabolic, hyperbolic or elliptic), the sign and regularity conditions associated with the quasi--local spatial curvature, $k_{qi}$, depends on the kinematic class. Hence, we examine in this subsection the admissible forms of asymptotic convergence for $m_{qi}$, leaving the discussion for the convergence of $k_{qi}$ for sections 10 and 11 that deal with the asymptotics of hyperbolic and open elliptic models.
For whatever form we might choose for the asymptotic convergence of $m_i$ or $m_{qi}$, we must be very careful that both of these densities remain non--negative, which (if standard regularity holds and from the scaling laws (\ref{mq}) and (\ref{slaw1})) implies that $m_q$ and $m$ are also non--negative at all ${}^3{\cal{T}}[t]$. Since $m_{qi}$ appears as initial value function in the analytic solutions in our parametrization of the models (see \cite{suss10}), we examine specific asymptotic convergence forms for it.
Considering (\ref{Asim}), the asymptotic convergence forms for $m_{qi}$ in which both $m_i$ and $m_{qi}$ remain non--negative everywhere are:
\numparts\begin{eqnarray}
\fl \hbox{Logarithmic:}\quad\quad m_{qi}\sim m_0/\ln\,r^{\alpha},\;\; (\alpha>0)\quad\qquad M\sim m_0R_0^3 r^3/\ln r^{\alpha}, \label{decmq2}\\
\fl \hbox{Power law:}\qquad m_{qi}\sim m_0 r^{-\alpha}\;\; (0\leq\alpha\leq 3),\qquad\;\; M\sim m_0R_0^3 r^{3-\alpha}.\label{decmq3}\end{eqnarray}\endnumparts
For the power law (\ref{decmq3}) with $\alpha=0$, we have $m_{qi}\sim m_0$, and so Lemma 3 yields $m_i\sim m_0$ and $\delta_i^{(m)}\sim 0$. For (\ref{decmq3}) with $\alpha>0$ and for (\ref{decmq2}) $m_{qi}\to 0$, and so Lemma 3 also implies $m_i\to 0$, but the asymptotic behavior of $m_i$ is not identical. We have from (\ref{Asim})--(\ref{Daasq}):
\numparts\begin{eqnarray}
\fl \hbox{Power law:}\qquad m_i\sim m_0\left(1-\frac{\alpha}{3}\right) r^{-\alpha},\qquad \delta_i^{(m)} \sim \tilde\delta_i^{(m)}=-\frac{\alpha}{3}.\label{PL_decq}\\
\fl \hbox{Logarithmic:}\qquad m_i\sim \frac{m_0(3\ln r-1)}{3\alpha\ln^2 r},\quad\quad \delta_i^{(m)}\sim \tilde\delta_i^{(m)}= -\frac{1}{3\ln r}\to 0.\label{Ldecq}
\end{eqnarray}\endnumparts
Notice that a power law with $\alpha>3$ in (\ref{decmq3}) implies that $m_i$ in (\ref{PL_decq}) becomes negative for sufficiently large $r$, though it remains close to zero and (from Lemma 3) it still tends to zero as $r\to\infty$. The same would happen for an exponential decay $m_{qi}\sim m_0\exp(-\alpha r)$. This undesirable behavior is consistent with the asymptotic forms for $M$ in (\ref{decmq2})--(\ref{decmq3}), since standard regularity requires this function to be monotonously increasing, thus admitting for $m_{qi}$ only a power law decay with $\alpha\leq 3$ or a slow logarithmic decay. The fact that $m_i<0$ happens for steeper decays of $m_{qi}$ follows because $m_i\leq m_{qi}$ holds for $r>y$ if the radial gradients are negative in this range, thus, if $m_{qi}$ becomes very close to zero in a very steep decay it may force $m_i$ to become negative (but close to zero).
While not all assumptions on the decay of $m_{qi}$ to zero yield a positive local density $m_i$, all assumptions on the decay of $m_i\to 0$ necessarily yield $m_i$ and $m_{qi}$ positive in the full range $r>y$. This is so because $m_{qi}\geq m_i\geq 0$ holds in this range (as $m'_i$ and $m'_{qi}$ are negative, see (\ref{propq2})). We discuss this issue in Appendix B.
\section{Asymptotic limits and asymptotic states.}
We explore now the asymptotic behavior of regular LTB models along radial rays of arbitrary ${}^3{\cal{T}}[t]$ ({\it{i.e.}} $t\ne t_i$ arbitrary and finite), and looking at parabolic, hyperbolic and elliptic models separately. This involves evaluating asymptotic series expansions of the metric functions and covariant scalars around their limit $r\to\infty$, under the assumption that the initial value functions $m_{qi}$ and $k_{qi}$ converge to admissible forms of uniform asymptotic convergence summarized in tables 1 and 2, and given explicitly by (\ref{decmq2})--(\ref{decmq3}) and (see section 10 and 11) (\ref{deckq2})--(\ref{deckq3}) and (\ref{deckq11})--(\ref{deckq22}). Since every LTB model can be completely determined by $m_{qi}$ and $k_{qi}$ (assuming the gauge (\ref{Rgauge})), once we assume admissible convergence forms for these initial value functions, we can find asymptotic convergence forms and asymptotic series for the metric functions $L,\,\Gamma$, the scalars $m_q,\,{\cal{H}}_q,\,k_q$, their fluctuations $\delta^{(m)},\,\delta^{(\HH)},\,\delta^{(k)}$, as well as local scalars $m,\,{\cal{H}},\,k$ and other auxiliary quantities ($M,\,E,\,t_{\textrm{\tiny{bb}}}$, etc). All this information completely characterizes an ``asymptotical state'' for every class of models based on a specific assumption on the convergence forms of $m_{qi}$ and $k_{qi}$ and the kinematic class (parabolic, hyperbolic and elliptic).
The resulting limits and expansions in the characteristic functions contained in the asymptotic states will be compared to the equivalent parameters in those spacetimes, listed in Appendix C, that are particular and limiting cases of LTB models, namely: dust FLRW cosmologies, Schwarzschild--Kruskal, Minkowski (including Milne) and self--similar LTB solutions. The radial asymptotics of specific classes of LTB models can be then characterized on the basis of this comparison, which we express below more precise terms:
\begin{quote}
Let $\{{\cal {M}},g\}$ be an LTB model manifold with metric $g$ given by (\ref{LTB2}), with $\{{\cal {M}}_{(0)},\bar g_{(0)}\}$ being the spacetime manifold and metric of any one of the particular cases listed in Appendix C. Let $X=\{L,\,\Gamma,\,A,\,A_q\}$, where $A,\,A_q$ denote the scalars $m,\,k,\,{\cal{H}}$ and $m_q,\,k_q,\,{\cal{H}}_q$ in $\{{\cal {M}},g\}$, with $X_{(0)}$ being the equivalent set of functions in $\{{\cal {M}}_{(0)},g_{(0)}\}$. \\
\noindent{\underline{Definition.}} We shall say that $\{{\cal {M}},g\}$ is {\it radially asymptotic} to $\{{\cal {M}}_{(0)},\bar g_{(0)}\}$, or that $\{{\cal {M}},g\}$ {\it converges asymptotically} to $\{{\cal {M}}_{(0)},\bar g_{(0)}\}$ in the radial direction, if the series expansion for every function in $X$ around its limit as $r\to\infty$ coincides with its equivalent function in $X_{(0)}$ up to leading terms.
\end{quote}
\noindent
Since it is clear that we are considering asymptotic behavior along radial rays, we will simply state that a given LTB model or class of models is ``asymptotic to'' or ``converges to'' a given particular case. Notice that it is sufficient to evaluate the limit as $r\to\infty$ for the quasi--local scalars $A_q$, as (from Lemma 3) this limit will be the same for the local scalars $A$.
However, the evaluation of the strict limit as $r\to\infty$ of the functions in $X$ is not sufficient to characterize the asymptotic behavior of a given class of LTB models, since the same limit can correspond to different asymptotic expansions and convergences. For example, the limit as $r\to\infty$ of the metric (\ref{LTB2}) of some LTB models can coincide with a Minkowski metric in spherical coordinates ($L,\Gamma \to 1$), but if we consider the expansions of $L$ and $\Gamma$ around this limit, the metric of these models could become equivalent (up to leading terms in the expansions) to the metric of an LTB self--similar solution or a section of Schwarzschild--Kruskal spacetime (whose metrics themselves have a Minkowski limit as $r\to\infty$). In order to distinguish these cases, we need to consider all metric functions and scalars in $X$, evaluate their asymptotic expansions (asymptotic state) and compare them (up to leading terms) with the functions and $X_{(0)}$. Notice that Another point that needs explaining is why we did not include the fluctuations in the set $X$ in the convergence criterion above. The reason is that $A\to 0$ (or equivalently $A_q\to 0$) as $r\to 0$ implies (in general) a nonzero limit for $\delta^{(A)}$, while the equivalent of $\delta^{(A)}$ in $X_{(0)}$ could be identically zero. This can happen when $X_{(0)}$ is associated with a Minkowskian particular case, for which $m=m_q=0$ holds identically, but then we have $m_q\to 0$ and $m\to 0$ in an LTB model converging to this particular case, and thus $\delta^{(m)}$ does not tend to zero. However, the asymptotic convergence is well characterized because the nonzero limit of $\delta^{(m)}$ is consistent with $m_q\to 0$ as $r\to\infty$. A summary of the asymptotic behavior for different classes of LTB models obtained in sections 9--12 is given in section 13.
\section{Radial asymptotics of parabolic models.}
For $k_{qi}=0$ the Friedmann--like equation (\ref{Hq}) yields a closed analytic expression for $L$
\begin{equation} L =\left[1+\frac{3}{2}\sqrt{2m_{qi}}\,c(t-t_i)\right]^{2/3},\label{par2}\end{equation}
which is equivalent to the parametric solution (\ref{par1}), and we are only considering expanding configurations ($L$ increases for $t>t_i$). The bang time, $t_{\textrm{\tiny{bb}}}$, follows by setting $L=0$ and $t=t_{\textrm{\tiny{bb}}}$ in (\ref{par2})
\begin{equation} ct_{\textrm{\tiny{bb}}} = ct_i-\frac{2}{3\sqrt{2m_{qi}}}=ct_i-\frac{2}{3{\cal{H}}_{qi}}.\label{tbbpar} \end{equation}
By differentiating (\ref{par2}) and from (\ref{LGdef}), we obtain
\begin{equation} \Gamma = 1+\delta_i^{(m)}-\frac{\delta_i^{(m)}}{L^{3/2}}.\label{Gp}\end{equation}
Since $k_{qi}=k_i=0$ and $R_i$ has been fixed by (\ref{Rgauge}), the only initial value functions are $m_{qi}$ and $\delta_i^{(m)}$. Necessary and sufficient conditions to fulfill (\ref{noshxG}) (absence of shell crossings) are given by the Hellaby--Lake conditions, which for parabolic models and in terms of our initial value functions takes the form \cite{suss10}
\begin{equation} -1\leq \delta_i^{(m)}\leq 0,\label{noshx_par}\end{equation}
which from (\ref{Dadef}) implies $m'_{qi}\leq 0$, so that $m_{qi}$ tends to zero or to a nonzero constant and its admissible asymptotic forms (preventing a negative $m_i$) are given by (\ref{decmq2})--(\ref{decmq3}).
Assuming an asymptotic convergence $m_{qi}\sim \tilde m_{qi}$ given by either one of (\ref{decmq2}) or (\ref{decmq3}) the asymptotic convergence for $L$ follows from (\ref{par2}) as
\begin{equation} L\sim\tilde L=\left[1+\frac{3}{2}\,\sqrt{2\tilde m_{qi}}\,c(t-t_i)\right]^{2/3},\label{asPar1a}\end{equation}
Considering (\ref{mq}), (\ref{Hq}), (\ref{ME}) and (\ref{tbbpar}) we get
\numparts\begin{eqnarray}
\fl {\cal{H}}_q\sim \tilde{\cal{H}}_q=\frac{\sqrt{2\tilde m_{qi}}}{\tilde L^{3/2}},\qquad m_q\sim \tilde m_q=\frac{\tilde m_{qi}}{\tilde L^3}.\label{asPar1b}\\
\fl M\sim\tilde M=\tilde m_{qi}\,R_0^3\,r^3,\qquad ct_{\textrm{\tiny{bb}}}\sim ct_i-\frac{2}{3\sqrt{2\tilde m_{qi}}}.\label{asPar1c}\end{eqnarray}\endnumparts
These asymptotic forms are valid for any regular parabolic model, while (from Lemma 3) the limits of $m,\,{\cal{H}}$ as $r\to \infty$ are the same as those of $m_q,\,{\cal{H}}_q$. Also, Lemmas 1 and 2 imply that $R\to\infty$ for all regular initial value functions. The asymptotic limits and states depend on the choice of $m_{qi}$. Figures 1a and 1b depict the domain in the $(ct,r)$ plane parabolic models with asymptotic limit to FLRW and Minkowski.
\subsection{Parabolic models asymptotic to spatially flat FLRW.}
If $m_{qi}\to m_0>0$, so that $\tilde m_{qi}= m_0$ (the power law form (\ref{decmq3}) with $\alpha=0$), then $\delta_i^{(m)}\to 0$ as $r\to \infty$ follows from the corollary 3 of Lemma 1. We have the asymptotic convergence forms
\numparts\begin{eqnarray} \fl L\sim \tilde L(t)=\left[1+\frac{3}{2}\,\sqrt{2m_0}\,c(t-t_i)\right]^{2/3},\qquad \Gamma\sim 1\label{asPar2a}\\
\fl {\cal{H}}_q\sim \tilde{\cal{H}}_q=\frac{\sqrt{2m_0}}{\tilde L^{3/2}(t)},\qquad m_q\sim \tilde m_q=\frac{m_0}{\tilde L^3(t)},\qquad \delta^{(m)}\sim 0,\label{asPar2c}\\
\fl M\sim \tilde M=m_0\,R_0^3\,r^3,\qquad ct_{\textrm{\tiny{bb}}}\sim c\tildet_{\textrm{\tiny{bb}}}=ct_i-\frac{2}{3\sqrt{2m_0}}=\hbox{const.}.\label{asPar2b}\label{asPar2c}\end{eqnarray}\endnumparts
so that
\begin{equation}\fl L\to \tilde L(t),\quad \Gamma\to 1,\quad m_q\to \tilde m_q(t),\quad {\cal{H}}_q\to \tilde {\cal{H}}_q(t),\quad \delta^{(m)}\to 0,\quad \delta^{(\HH)}\to 0.\end{equation}
Since $\delta_i^{(m)}$ and $\delta^{(m)}$ tend to zero and $\Gamma\to 1$, then $m\to m_q$ and ${\cal{H}}\to{\cal{H}}_q$ (which is consistent with Lemma 3). It is evident that the asymptotic state from these expansions is consistent with the corresponding parameters of a spatially flat dust FLRW model (see equations (\ref{FLRW})--(\ref{FLRW_L0})).
\subsection{Minkowski limit and its asymptotic states.}
If $m_{qi}\sim\tilde m_{qi}\to 0$ as $r\to \infty$, then (\ref{asPar1a}) and (\ref{Gp}) yield
\begin{equation} \fl L\sim \tilde L\approx 1+\sqrt{2\tilde m_{qi}}\,c(t-t_i),\qquad \Gamma\sim\tilde\Gamma\approx 1+\frac{3}{2}\tilde\delta_i^{(m)} \sqrt{2\tilde m_{qi}}\,c(t-t_i).\label{asPar3a}\end{equation}
where we have written ``$\approx$'' (instead of ``$\sim$'') because we used the approximation $(1+\epsilon)^{2/3}\approx 1+(2/3)\epsilon$, valid for $\epsilon\ll 1$. The power law decay of $m_{qi}$ is useful to illustrate the asymptotic convergence forms for the remaining quantities. Considering (\ref{decmq3}) with $0<\alpha\leq 3$, we obtain
\numparts\begin{eqnarray}
\fl {\cal{H}}_q\sim \tilde{\cal{H}}_q=\frac{\sqrt{2m_0}r^{-\alpha/2}}{\tilde L^{3/2}}\to 0,\qquad m_q\sim \tilde m_q=\frac{m_0\,r^{-\alpha}}{\tilde L^3}\to 0,\label{asPar3c}\\
\fl \delta_i^{(m)}\sim\tilde\delta_i^{(m)}=-\alpha/3,\qquad \delta^{(m)}\sim\tilde\delta^{(m)}= -\alpha/3,\qquad 2\tilde\delta^{(\HH)}\approx -\frac{\alpha}{3}.\label{asPar3d}\\
\fl M\sim\tilde M=m_0\,R_0^3\,r^{3-\alpha},\qquad ct_{\textrm{\tiny{bb}}}\sim ct_i-\frac{2r^{\alpha/2}}{3\sqrt{2m_0}}\to -\infty,\label{asPar3b}\end{eqnarray}\endnumparts
where we used (\ref{slawDm}) and (\ref{slawDh}). From Lemma 3 we have $m\to 0$ and ${\cal{H}}\to 0$, which agrees with (\ref{qltransf}). If we only consider the limit as $r\to\infty$, then $\tilde m_{qi}\to 0$ clearly implies a Minkowski limit, since $L\to 1$ and $\Gamma\to 1$ transforms the LTB metric (\ref{LTB2}) (with $k_{qi}=0$) into a Minkowski metric in spherical coordinates. The fact that $\delta^{(m)},\,\delta^{(\HH)}$ tend to a nonzero value is consistent with $m_q,\,{\cal{H}}_q$ tending to zero (as corollary 3 of Lemma 3 does not apply). If $m_{qi}$ decays logarithmically as in (\ref{decmq2}), we obtain the same Minkowski limit, but with fluctuations $\delta^{(A)}$ now also tending to zero.
\smallskip
\noindent
The asymptotic states of parabolic LTB models with $\tilde m_{qi}\to 0$ are:
\begin{itemize}
\item \underline{Asymptotic to spatially flat self similar solution.} If $\alpha=2$, then $\tilde L=\tilde L(\zeta)$, where $\zeta=c(t-t_i)/r$ is the self--similar variable in (\ref{SSvar}). It is evident that the metric and all functions asymptotically converge to their respective forms in the spatially flat self--similar solution given by (\ref{SS}), (\ref{SS_L0}), (\ref{SSvar})--(\ref{MEtbb42}) with $k_0=0$, which is the case $E=0$ in equation (2.29) of \cite{selfsim} (see also pages 344--345 of \cite{kras2} and \cite{sscoll}). Notice that parabolic LTB models with this asymptotic behavior are not self--similar solutions, they only converge to a self--similar solution as $r\to\infty$ along the ${}^3{\cal{T}}[t]$: the latter solution only follows if $m_{qi}=m_0\,r^{-2}$ holds exactly in all the domain of $r$.
\item \underline{Asymptotic to Schwarzschild--Kruskal.} If $\alpha=3$, then a comparison with (\ref{Schw}), (\ref{MEtbb3}) and (\ref{Schw_L0}) implies that we have an asymptotic convergence to Schwarzschild--Kruskal solution in spatially flat comoving coordinates (Lema\^\i tre coordinates, see page 332 of \cite{kras2}). This convergence is consistent with $M\sim m_0R_0^3=$ constant, as opposed to $M$ diverging for $\alpha<3$. Notice that $\delta^{(m)}\to -1$, hence $m=m_q(1+\delta^{(m)})$ converges to zero much faster than $m_q\sim m_0R_0^3/R^3$.
\item \underline{Asymptotic to Minkowski in generalized Milne coordinates.} If $\alpha\ne 2,3$, then comparison of $\tilde L$ in (\ref{asPar3a}) and (\ref{Mink_L}) shows an asymptotic state compatible with that of a locally Minkowski spacetime in coordinates that generalize Milne's. However, the form (\ref{asPar3a}) does not not follow from the particular solution with $M=m_{qi}=0$ in (\ref{s2case}). Rather, we have an asymptotic state compatible with a section of Minkowski, but one based on expansions around an asymptotic limit $m_{qi}=0$ in models (parabolic) for which $m_{qi}$ is not strictly zero but $k_{qi}=0$ holds everywhere. This Minkowsky asymptotic state occurs also for hyperbolic and elliptic models converging to a parabolic model (see sections 10 and 11).
\end{itemize}
\section{Radial asymptotics of hyperbolic models.}
Analytic solutions for hyperbolic models follow from solving the Friedman--like equation (\ref{Hq}) for $k_{qi}<0$. These solutions are equivalent to those of (\ref{fieldeq1}) given by (\ref{hypRt}) and (\ref{hypRt}) in parametric and implicit form. We will work with the implicit form given in terms of our variables by
\begin{equation} \phi = Z_h(x_iL),\label{phi_L_h}\end{equation}
where $Z_h$ is the function
\begin{equation} u \mapsto Z_h(u)=u ^{1/2} \left( {2 + u } \right)^{1/2} - \hbox{arccosh}(1 + u ).\label{hypZ1a} \end{equation}
and
\numparts\begin{eqnarray}\phi &\equiv& \phi(t,x_i,y_i) = y_i\, c(t-t_i)+Z_h(x_i),\label{phih}\\
x_i &=& \frac{|k_{qi}|}{m_{qi}},\qquad y_i=\frac{|k_{qi}|^{3/2}}{m_{qi}}.\label{xy}
\end{eqnarray}\endnumparts
The bang time emerges from setting $L=0$ in (\ref{phi_L_h}) as the following function of $m_{qi}$ and $|k_{qi}|$:
\begin{equation} ct_{\textrm{\tiny{bb}}} = ct_i-\frac{Z_h(x_i)}{y_i}.\label{tbbh} \end{equation}
The metric function $\Gamma$ follows from (\ref{phi_L_h}) by implicit derivation since it is related to $L'/L$ by (\ref{LGdef}). The result is
\begin{equation}\fl \Gamma = 1+3(\delta_i^{(m)}-\delta_i^{(k)})\left(1-\frac{{\cal{H}}_q}{{\cal{H}}_{qi}}\right)-3{\cal{H}}_q\,c(t-t_i)\,\left(\delta_i^{(m)}-\frac{3}{2}\delta_i^{(k)}\right),\label{Ghe}\end{equation}
where ${\cal{H}}_q$ and ${\cal{H}}_{qi}$ follow from (\ref{Hq}), while $c(t-t_i)$ is given by (\ref{phi_L_h}) and (\ref{phih}). The Hellaby--Lake conditions \cite{ltbstuff,suss10,ltbstuff1} to fulfill the condition (\ref{noshxG}) for absence of shell crossings are given in terms of initial value functions by \cite{suss10}
\begin{equation} ct_{\textrm{\tiny{bb}}}'\leq 0,\qquad \delta_i^{(k)}\geq -\frac{2}{3},\qquad \delta_i^{(m)}\geq -1,\label{noshxGh}\end{equation}
where
\begin{equation}\frac{ct_{\textrm{\tiny{bb}}}'}{3R'_i/R_i}
= \frac{\delta_i^{(m)}-\delta_i^{(k)}}{{\cal{H}}_{qi}}-c(t_i-t_{\textrm{\tiny{bb}}})\left(\delta_i^{(m)}-\frac{3}{2}\delta_i^{(k)}\right),\label{tbbr}\end{equation}
follows by differentiating (\ref{tbbh}).
\subsection{Initial negative curvature}
It is evident from (\ref{Dadef}), (\ref{ME}) that the profile and asymptotic convergence of the initial value function $k_{qi}<0$ in a hyperbolic model is restricted by the second regularity condition in (\ref{noshxGh}), which for a negative $k_{qi}$ necessarily implies that $E={\cal{F}}^2-1=-k_{qi} R_0^2 r^2$ must be monotonously increasing (and that $k_i<0$). In order to examine the case $k_{qi}<0$, we remark (see (\ref{Dadef})) that
\begin{equation} \hbox{sign}(\delta^{(k)})=\hbox{sign}(k'_q/k_q). \end{equation}
Hence, $\delta^{(k)}$ has the same sign in the combination $k_q\leq 0,\, k'_q\geq 0$ as in $k_q\geq 0,\, k'_q\leq 0$. As a consequence, if we assume that $k_{qi}< 0$ holds everywhere (with possibly $k_{qi}\to 0$ asymptotically), then we can examine negative spatial curvature by applying (\ref{PL_decq})--(\ref{Ldecq}) directly to $|k_{qi}|$. Considering all these points together with (\ref{ME}) and (\ref{Rgauge}), the asymptotic forms of $|k_{qi}|$ and $E$ compatible with standard regularity conditions and with a well defined asymptotic radial range ($r\to\infty \Rightarrow \xi_i\to\infty$) are
\numparts\begin{eqnarray}
\fl |k_{qi}|\sim k_0/\ln\,r,\qquad\qquad\quad\qquad \tilde\delta_i^{(k)}\to 0,\qquad E\sim k_0\,R_0^2\,r^2/\ln\,r, \label{deckq2}\\
\fl |k_{qi}|\sim k_0 r^{-\beta}\;\; (0\leq \beta\leq 2),\qquad\quad \tilde\delta_i^{(k)}\to -\frac{\beta}{3},\qquad E\sim k_0\,R_0^2\,r^{2-\beta}.\label{deckq3}\end{eqnarray}\endnumparts
where $k_0>0$ is a constant. Notice that $|k_{qi}|\sim k_0$ (or $\beta=2$) allows for both an increasing $k'_{qi}>0$ or decreasing $k'_{qi}<0$ behavior, though (from Lemma 3) $k'_{qi}\to 0$ as $r\to\infty$ and $|k_{qi}|\to k_0$. In both cases (\ref{deckq2}) and (\ref{deckq3}) we have $|k_{qi}|$ decaying to zero. The asymptotic forms of the local curvature $|k_i|$ are analogous to those of $|k_{qi}|$ given above, since the allowed decay forms are not steeper than $r^{-3}$ (see Appendix B).
\begin{table}
\begin{center}
\begin{tabular}{|c| c| c| c| c| c|}
\hline
\hline
{} &{} &{} &{Section 10.3} &{Section 10.4} &{Section 10.5}
\\
{$\tilde m_{qi}$} &{$|\tilde k_{qi}|$} &{$\tilde x_i$} &{$\tilde x_i\to 0$} &{$\tilde x_i\to\infty$} &{$\tilde x_i\to x_0$}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{PL} &{PL} &{$x_0\, r^{\alpha-\beta}$} &{$0\leq \alpha<\beta\leq 2$} &{$0\leq\beta<\alpha\leq 3$} &{$0\leq \alpha=\beta\leq 2$}
\\
{} &{} &{} &{} &{} &{}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{PL} &{LOG} &{$x_0\, r^\alpha/\ln r^\beta$} &{$\alpha= 0$} &{$0<\alpha\leq 3$} &{}
\\
{} &{} &{} &{$\beta>0$} &{$\beta>0$} &{}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{LOG} &{PL} &{$x_0\, r^{-\beta}\ln r^\alpha$} &{$0<\beta\leq 2$} &{$\beta= 0$} &{}
\\
{} &{} &{} &{$\alpha>0$} &{$\alpha>0$} &{}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{LOG} &{LOG} &{$x_0\,\alpha/\beta$} &{} &{} &{$\alpha,\beta > 0$}
\\
{} &{} &{} &{} &{} &{}
\\
\hline
\hline
\end{tabular}
\end{center}
\caption{{\bf{Combinations of the admissible asymptotic convergence forms for hyperbolic models}}. The table presents the asymptotic convergence of $\tilde x_i=|\tilde k_{qi}|/\tilde m_{qi}$ for the four possible combinations of the forms $\tilde m_{qi}$ and $|\tilde k_{qi}|$ in (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq2})--(\ref{deckq3}). The terms ``PL'' and ``LOG'' stand for ``power law'' and ``logarithmic'' decay forms, respectively given by (\ref{decmq2}), (\ref{deckq2}) and (\ref{decmq3}), (\ref{deckq3}). The behavior of $\tilde x_i$ examined in sections 10.3--10.5 follows from the restrictions on the parameters $\alpha,\,\beta$, hence the results of these sections can be readily applied to initial value functions having each asymptotic convergence form.}
\label{table1}
\end{table}
\subsection{Asymptotic approximations for the analytic solutions}
There is no closed explicit form for $L$ as (\ref{par2}), hence we must work with the implicit solutions (\ref{phi_L_h}), with the correspondence rule of the function $Z_h$ given by (\ref{hypZ1a}). Since the admissible forms in (\ref{deckq2})--(\ref{deckq3}) imply that $|k_{qi}|$ is bounded in the radial asymptotic regime, then the radial asymptotic behavior of $\phi=Z_h(x_iL)$ depends on the behavior of $x_i$ in this regime, which in turn depends on the choice of initial value functions $m_{qi}$ and $|k_{qi}|$. Since $Z_h$ is a monotonously increasing function of its arguments, then (\ref{phi_L_h}) implies
\numparts\begin{eqnarray}
\fl x_i\to 0\quad\quad\Rightarrow\quad\;\; Z_h(x_i)\to 0\quad \Rightarrow\quad \phi\to 0\quad \Rightarrow\quad Z_h(x_iL)\to 0,\label{ashpat1}\\
\fl x_i\to \infty\quad\;\;\Rightarrow\quad Z_h(x_i)\to \infty\quad \Rightarrow\quad\phi\to \infty\quad \Rightarrow\quad Z_h(x_iL)\to \infty,\label{ashpat2}\end{eqnarray}\endnumparts
while, another possibility is given by
\begin{equation}\fl x_i\to x_0=\hbox{const.},\qquad \phi\to |k_{qi}|^{1/2}\,x_0\,c(t-t_i) +Z_h(x_0)\to Z_h(x_0 L).\label{ashpat3}\end{equation}
The three patterns outlined by (\ref{ashpat1})--(\ref{ashpat2}) and (\ref{ashpat3}), and listed in the 4th, 5th and 6th columns of table \ref{table1}, provide all possible combinations of asymptotic behavior of regular hyperbolic models in terms of their initial conditions (through $x_i$). It is possible to apply appropriate approximations for each one of these cases in order to obtain $L$, which follows by inverting (\ref{phi_L_h})
\begin{equation} L = \frac{Z_h^{-1}(\phi)}{x_i}.\label{L_phi_h}\end{equation}
Given an asymptotic form for $L$ obtained in terms of admissible convergence forms of $m_{qi}$ and $|k_{qi}|$ in (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq2})--(\ref{deckq3}), we can obtain asymptotic convergence (or approximated) forms for the remaining quantities. We examine separately the asymptotic behavior associated with each of the three cases (\ref{ashpat1})--(\ref{ashpat2}) and (\ref{ashpat3}), each one listed in the fourth, fifth and sixth columns of table \ref{table1}. Figures 1a and 1b respectively depict the domain in the $(ct,r)$ plane of hyperbolic models with asymptotic limit to FLRW and Minkowski.
\subsection{Matter dominated asymptotics: convergence to parabolic models. }
If $m_{qi}$ dominates over $|k_{qi}|$ in the limit $r\to \infty$, then $ x_i\to 0$ and we have the situation described by (\ref{ashpat1}). Hence, we can consider the following approximation for $Z_h$
\begin{equation} Z_h(u)\approx \frac{\sqrt{2}}{3}\,u^{3/2}-\frac{\sqrt{2}}{20}\,u^{5/2}\quad\hbox{for}\quad u\approx 0,\label{apprhyp1}\end{equation}
where we use the sign ``$\approx$'' to distinguish this approximation from the approximations that follow from assumptions on uniform asymptotic convergence (``$\sim$'') on the initial value functions. Applying (\ref{apprhyp1}) (up to the leading term) to $Z_h(x_i)$ and $Z_h(x_i L)$ in (\ref{phi_L_h}), and considering that $m_{qi}\sim \tilde m_{qi}$ and $|k_{qi}|\sim |\tilde k_{qi}|$, yields
\begin{equation} L \sim \tilde L\approx \left[1+\frac{3}{2}\sqrt{2\tilde m_{qi}}\,c(t-t_i)\right]^{2/3},\label{asHyp11}\end{equation}
which coincides with the form of $\tilde L$ in (\ref{asPar1a}), though now $L$ does not (necessarily) converge uniformly to $\tilde L$, but approximates it up to a certain order related to the convergence of $x_i$ to zero (hence this approximation depends also on assumptions on $|k_{qi}|$, which is strictly zero in parabolic models). Nevertheless, (\ref{asHyp11}) indicates that hyperbolic models with initial conditions in which $x_i\to 0$ converge to a parabolic model. Though, we have now $|k_{qi}|\ne 0$, hence the asymptotic state might not define a general parabolic model and the forms of the involved quantities should not be identical to those of parabolic models in the previous section. In order to discuss these issues, we examine below the two types of asymptotic states associated with (\ref{asHyp11}) that emerge from the form of $\tilde m_{qi}$:
\begin{itemize}
\item \underline{Asymptotic to spatially flat FLRW.}
If $m_{qi}\sim m_0>0$ (the power law form (\ref{decmq3}) with $\alpha=0$, see top entry in the fourth column of table \ref{table1}), then $\delta_i^{(m)}\to 0$ and (\ref{asHyp11}) yields $L\approx \tilde L(t)$ with $\tilde L(t)$ given by (\ref{asPar2a}), which is the scale factor of a spatially flat FLRW dust model. The conventional variables $M,\,E$ and $t_{\textrm{\tiny{bb}}}$ have the asymptotic forms
\begin{equation}\fl M\sim m_0R_0^3r^3,\quad E\sim |\tilde k_{qi}|R_0 r^2,\quad ct_{\textrm{\tiny{bb}}} \sim ct_i-\frac{2}{3\sqrt{2m_0}}\to \hbox{const}. \label{asHyp12} \end{equation}
where $|\tilde k_{qi}|$ can take the power law form (\ref{deckq3}) with $\beta>0$ or the logarithmic decay (\ref{deckq2}). Since $\tilde L(t)$ is constant along the ${}^3{\cal{T}}[t]$ and $|k_{qi}|\to 0$, then the asymptotic forms of the remaining time--dependent scalars follow from series expansions of (\ref{mq}), (\ref{kq}), (\ref{Hq}), (\ref{slawDm}), (\ref{slawDk}), (\ref{slawDh}) and (\ref{Ghe}) around $|k_{qi}|\sim|\tilde k_{qi}|$, leading up to order $\tilde |k_{qi}|$ to
\numparts\begin{eqnarray} \fl m_q \approx \frac{m_0}{\tilde L^3},\quad |k_q|\approx \frac{|\tilde k_{qi}|}{\tilde L^2}\to 0,\quad {\cal{H}}_q\approx \frac{\sqrt{2m_0}}{\tilde L^{3/2}}\left(1+\frac{|\tilde k_{qi}|\tilde L}{4m_0}\right)\to \frac{\sqrt{2m_0}}{\tilde L^{3/2}},\label{asHyp13a}\\
\fl \Gamma\approx 1+ \frac{3\,(\tilde L^{5/2}-1)\,|\tilde k_{qi}|\tilde\delta_i^{(k)}}{4\,m_0\,\tilde L^{3/2}}\to 1,\quad \delta^{(m)}\approx -\frac{3(\tilde L^{5/2}-1)\,|\tilde k_{qi}|\tilde\delta_i^{(k)}}{4\,m_0\,\tilde L^{3/2}}\to 0,\label{asHyp13b}\\
\fl \delta^{(k)}\approx \tilde \delta_i^{(k)} -\frac{(2+3\tilde\delta_i^{(k)})\,(\tilde L^{5/2}-1)\,|\tilde k_{qi}|\tilde\delta_i^{(k)}}{4\,m_0\,\tilde L^{3/2}}\to \tilde \delta_i^{(k)},\label{asHyp13c}\\
\fl 2\delta^{(\HH)} \approx \frac{1}{2}\left(1-\frac{3(\tilde L^{5/2}-1)}{2\tilde L^{5/2}}\right)\,\frac{\tilde L\,|\tilde k_{qi}|\tilde\delta_i^{(k)} }{m_0}\to 0.\label{asHyp13d}\end{eqnarray}\endnumparts
where $|\tilde k_{qi}|,\,\,\tilde\delta_i^{(k)}$ can take any of the admissible asymptotic forms (\ref{deckq2})--(\ref{deckq3}). The fact that, as $r\to\infty$, we have $m_q,\,{\cal{H}}_q$ tending to time dependent forms, while $\Gamma\to 1$ and $\delta^{(k)}$ tends to a perturbative term, clearly shows that these hyperbolic models are asymptotic to the spatially flat FLRW dust cosmology.
\item \underline{Asymptotic limit to Minkowski.}
If $\tilde m_{qi}\to 0$ and $|\tilde k_{qi}|\to 0$, following either power law decays with $0<\alpha<\beta$ or any combination of power law and logarithmic decay, as given by (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq2})--(\ref{deckq3}), we have the cases listed in the fourth column of table \ref{table1} (excluding the case PL--PL with $\alpha=0$). For all these combinations of $\tilde m_{qi}\to 0$ and $|\tilde k_{qi}|\to 0$ (\ref{asHyp11}) becomes as $r\to\infty$
\begin{equation} \tilde L \approx 1+\sqrt{2\tilde m_{qi}}\,c(t-t_i)\to 1,\label{asHyp14a}\end{equation}
where we used $(1+\epsilon)^{2/3}\approx 1+2/3\epsilon$, which holds for $\epsilon\ll 1$. The conventional variables $M,\,E$ and $t_{\textrm{\tiny{bb}}}$ take the asymptotic forms
\begin{equation}\fl M\sim \tilde m_{qi}R_0^3r^3,\quad E\sim |\tilde k_{qi}|R_0 r^2,\quad ct_{\textrm{\tiny{bb}}} \sim ct_i-\frac{2}{3\sqrt{2\tilde m_{qi}}}\to -\infty.\label{asHyp14b} \end{equation}
Hyperbolic models with these initial value functions converge to parabolic models whose asymptotic limit is Minkowski (see previous section), but they do not converge to a general parabolic model of this type. This point can be better illustrated by considering a power law decay for both $\tilde m_{qi}$ and $|\tilde k_{qi}|$ as in (\ref{decmq3}) and (\ref{deckq3}) (see the case PL--PL in the top entry of fourth column of table \ref{table1} with $\alpha>0$). Since we must have $\alpha<\beta$ for $x_i\to 0$ to hold, then there cannot be a convergence to parabolic models for which $2<\alpha\leq 3$ holds because regularity conditions imply $\beta\leq 2$.
\smallskip
The scalars $m_q$ and $|k_q|$ and ${\cal{H}}_q$ take the forms
\numparts\begin{eqnarray} \fl m_q\approx \frac{\tilde m_{qi}}{\tilde L^3}\to 0,\quad |k_q|\approx \frac{|\tilde k_{qi}|}{\tilde L^2}\to 0,\quad \frac{|k_q|}{m_q}\approx \frac{|\tilde k_{qi}|}{\tilde m_{qi}}\to 0,\label{asHyp15a}\\
\fl {\cal{H}}_q\approx \frac{\sqrt{2\tilde m_{qi}}}{\tilde L}\,\left(1+\frac{\tilde x_i}{4}\right)\approx \sqrt{2\tilde m_{qi}}\,\left[1-\sqrt{2\tilde m_{qi}}c(t-t_i)+\frac{\tilde x_i}{4}\right] \to 0,\label{asHyp15b}
\end{eqnarray}\endnumparts
where $\tilde L$ is given by (\ref{asHyp14a}). Considering the expansions above and only terms linear in $\sqrt{\tilde m_{qi}}$ or $\tilde x_i$, and since $\sqrt{1+\tilde x_i/2}\approx 1+\tilde x_i/4$ holds for $x_i\ll 1$, we obtain up to leading order
\numparts\begin{eqnarray}
\fl \Gamma \approx 1+ \frac{3}{2}\sqrt{2\tilde m_{qi}}\,c(t-t_i)\,\tilde \delta_i^{(m)} \to 1,\label{asHyp16a}\\
\fl \delta^{(m)} \approx \tilde\delta_i^{(m)}- \frac{3}{2}\sqrt{2\tilde m_{qi}}\,c(t-t_i)\,\tilde \delta_i^{(m)}\,(1+\tilde\delta_i^{(m)})\to \tilde\delta_i^{(m)},\label{asHyp16b}\\
\fl \delta^{(k)} \approx \tilde\delta_i^{(k)} - \frac{1}{2}\sqrt{2\tilde m_{qi}}\,c(t-t_i)\,\tilde\delta_i^{(m)}\, (2+3\tilde\delta_i^{(k)}) \to \tilde\delta_i^{(k)},\label{asHyp16c}\\
\fl 2\delta^{(\HH)} \approx \tilde \delta_i^{(m)} - \frac{3}{2}\sqrt{2\tilde m_{qi}}c(t-t_i)\tilde \delta_i^{(m)} (1+\tilde\delta_i^{(m)})-(\tilde\delta_i^{(m)}-\tilde\delta_i^{(k)})\frac{\tilde x_i}{2}\to \tilde\delta_i^{(m)}.\label{asHyp16d}\end{eqnarray}\endnumparts
where the asymptotic convergence form of $\tilde\delta_i^{(m)},\,\tilde\delta_i^{(k)}$ depends on the choice of $\tilde m_{qi},\,|\tilde k_{qi}|$ in (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq2})--(\ref{deckq3}).
\end{itemize}
\subsection{Vacuum dominated asymptotics.}
Consider now the case when $|k_{qi}|$ dominates over $m_{qi}$ in the limit $r\to \infty$, so that $ x_i\to \infty$ and we have the situation described by (\ref{ashpat2}), with initial value functions listed in the fifth column of table \ref{table1}. We can consider then the following approximation for $Z_h$
\begin{equation} Z_h(u)\approx u-\ln(u)\approx u\quad\hbox{for}\quad u\gg 1,\label{apprhyp2}\end{equation}
where, as in the previous subsection, we distinguish this approximation from that following a given assumption of uniform asymptotic convergence on the initial value functions (hence the sign ``$\approx$''). Applying (\ref{apprhyp2}) (up to the leading term) to $Z_h(x_i)$ and $Z_h(x_i L)$ in (\ref{phi_L_h}), and considering that $m_{qi}\sim \tilde m_{qi}$ and $|k_{qi}|\sim |\tilde k_{qi}|$, yields
\begin{equation} L \sim \tilde L\approx 1+|\tilde k_{qi}|^{1/2}\,c(t-t_i),\label{asHyp21}\end{equation}
so that $L$ approximates the solution of the equation $\dot L^2 = |k_{qi}|$, which is the Friedman equation (\ref{Hq}) with $m_q=0$ and $k_q<0$, or equivalently, the solution of (\ref{fieldeq1}) in (\ref{s2case}) with $M=0$ and $E>0$. From (\ref{Mink})--(\ref{MEtbb2}) in Appendix C2, these LTB models are the solutions [s2] in \cite{ltbstuff}, and are locally Minkowskian (equivalent to sections of Minkowski spacetime parametrized by non--standard coordinates that generalize the Milne universe). Evidently, hyperbolic models characterized by initial value functions complying with $x_i\to \infty$ converge in the radial direction to these Minkowskian sections. Notice that $k_q<0$ does not imply that spacetime curvature is nonzero, as $k_q$ is the quasi--local dual to the spatial curvature ($k$) of the 3--dimensional slices ${}^3{\cal{T}}[t]$. The 3--dimensional scalar curvature of these hypersurfaces can be non--trivial, even in a vacuum flat Minkowski space. On the other hand, $m_q$ is associated with a source $m$ appearing in $T^{ab}$ and generating spacetime curvature. If $|k_{qi}|$ dominates over $m_{qi}$ in the asymptotic radial range, then the frame dependent kinematic effects of the spatial curvature play the major role in the behavior of all incumbent scalars in this range.
The spatial curvature dominated hyperbolic models lead to the following asymptotic states depending on the convergence of the initial value function $|k_{qi}|$:
\begin{itemize}
\item \underline{Asymptotic to a Milne Universe.}
If $|k_{qi}|\sim k_0>0$ (which corresponds to the power law form (\ref{deckq3}) with $\beta=0$), while $\tilde m_{qi}\to 0$ has any of the forms (\ref{decmq2})--(\ref{decmq3}) with $\alpha>0$, then (\ref{asHyp21}) becomes
\begin{equation} L\approx \tilde L(t) = 1+\sqrt{k_0}\,c(t-t_i),\label{asHyp22a}\end{equation}
while the conventional variables take the asymptotic form
\begin{equation}\fl M\sim \tilde m_{qi}R_0r^3,\qquad E\sim k_0 R_0^2 r^2,\qquad ct_{\textrm{\tiny{bb}}} \approx ct_i-\frac{1}{\sqrt{k_0}}\to\hbox{const.} \label{asHyp22b} \end{equation}
By comparing with (\ref{Milne})--(\ref{MEtbb2}), it is evident that hyperbolic models with these initial value functions converge to the locally Minkowskian Milne universe.
\smallskip
Bearing in mind that $\tilde L=\tilde L(t)$ and $m_{qi}\to 0$, then the asymptotic forms of the remaining time--dependent scalars follow from series expansions of (\ref{mq}), (\ref{kq}), (\ref{Hq}), (\ref{slawDm}), (\ref{slawDk}), (\ref{slawDh}) and (\ref{Ghe}) around $\tilde m_{qi}=0$. We obtain up to leading order in $\tilde m_{qi}$ the following forms similar to (\ref{asHyp13a})--(\ref{asHyp13d})
\numparts\begin{eqnarray} \fl m_q \approx \frac{\tilde m_{qi}}{\tilde L^3}\to 0,\quad |k_q|\approx \frac{k_0}{\tilde L^2},\quad {\cal{H}}_q\approx \frac{\sqrt{k_0}}{\tilde L}\left(1+\frac{\tilde m_{qi}}{k_0\,\tilde L}\right)\to \frac{\sqrt{k_0}}{\tilde L},\label{asHyp23a}\\
\Gamma\approx 1- \frac{3 c(t-t_i)\tilde\delta_i^{(m)}\,\tilde m_{qi}^2}{k_0^{3/2}\,[1+\sqrt{k_0} c(t-t_i)]^2}\to 1, \label{asHyp23b}\\
\delta^{(m)}\approx \tilde\delta_i^{(m)} + \frac{3 c(t-t_i)\tilde\delta_i^{(m)}\,(1+\tilde\delta_i^{(m)})\,\tilde m_{qi}^2}{k_0^{3/2}\,[1+\sqrt{k_0} c(t-t_i)]^2}\to \tilde\delta_i^{(m)},\label{asHyp23c}\\
\delta^{(k)}\approx \frac{2 c(t-t_i)\tilde\delta_i^{(m)}\,\tilde m_{qi}^2}{k_0^{3/2}\,[1+\sqrt{k_0} c(t-t_i)]^2}\to 0,\label{asHyp23d}\\
2\delta^{(\HH)} \approx \frac{2\tilde\delta_i^{(m)}\,\tilde m_{qi}}{k_0\,[1+\sqrt{k_0} c(t-t_i)]}\to 0.\label{asHyp23e}\end{eqnarray}\endnumparts
where $\tilde m_{qi},\,\,\tilde\delta_i^{(m)}$ can take any of the admissible asymptotic forms (\ref{decmq2})--(\ref{decmq3}) with $\alpha>0$.
\item \underline{Asymptotic to a generalized Milne universe.}
If $x_i$ diverges as $r\to\infty$, but both $\tilde m_{qi},\,|\tilde k_{qi}|$ given by (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq2})--(\ref{deckq3}) tend to zero in this limit, we have the case listed in the fifth column of table \ref{table1} (of course, excluding the Milne universe discussed above, hence $\alpha,\,\beta>0$ necessarily holds). As a consequence, $L$ takes the form (\ref{asHyp21}), which implies $L\sim\tilde L\to 1$ as $r\to\infty$. The functions $M$ and $E$ take the forms in (\ref{asHyp14b}) with the appropriate values of $\alpha,\,\beta>0$, while $t_{\textrm{\tiny{bb}}}$ takes the asymptotic form
\begin{equation} ct_{\textrm{\tiny{bb}}} \approx ct_i - \frac{1}{|\tilde k_{qi}|^{1/2}} \to -\infty,\label{asHyp31}\end{equation}
Hyperbolic models with these initial value functions converge to generalized versions of the Milne universe (case [s2] in \cite{ltbstuff}). The scalars $m_q$ and $k_q$ are given by the same expressions as in (\ref{asHyp15a}), but with $\tilde L$ given by (\ref{asHyp21}) and $\tilde m_{qi},\,|\tilde k_{qi}|$ complying with $\tilde x_i=|\tilde k_{qi}|/\tilde m_{qi}\to\infty$. The scalar ${\cal{H}}_q$ in (\ref{Hq}) takes the following asymptotic form
\numparts\begin{eqnarray}
\fl\tilde {\cal{H}}_q &\approx& \frac{|\tilde k_{qi}|^{1/2}}{\tilde L}\,\left(1+\frac{2}{\tilde x_i\,\tilde L}\right)^{1/2} \approx \frac{|\tilde k_{qi}|^{1/2}}{\tilde L}\,\left(1+\tilde\chi_i\right)\nonumber\\
\fl &\approx& |\tilde k_{qi}|^{1/2}\left[1- |\tilde k_{qi}|^{1/2}c(t-t_i)+\chi_i\right] \to 0,\label{asHyp32a}\\
\fl &\hbox{with}&\qquad \chi_i\equiv \frac{1}{\tilde x_i\,\tilde L}\approx \frac{1}{\tilde x_i}. \label{asHyp32b}\end{eqnarray}\endnumparts
Considering only terms linear in $|\tilde m_{qi}|^{1/2}$ and $\tilde \chi_i$, we obtain up to leading order
\numparts\begin{eqnarray}
\Gamma \approx 1+ \frac{3}{2}c(t-t_i)\,\tilde\delta_i^{(k)}\,|\tilde k_{qi}|^{1/2}\to 1, \label{asHyp33a}\\
\delta^{(m)} \approx \tilde\delta_i^{(m)} -\frac{3}{2}c(t-t_i)\,(1+\tilde\delta_i^{(m)})\,\tilde\delta_i^{(k)}\,|\tilde k_{qi}|^{1/2}\to \tilde\delta_i^{(m)},\label{asHyp33b}\\
\delta^{(k)} \approx \tilde\delta_i^{(k)} -\frac{3}{2}c(t-t_i)\,(2+3\tilde\delta_i^{(k)})\,\tilde\delta_i^{(k)}\,|\tilde k_{qi}|^{1/2}\to \tilde\delta_i^{(k)},\label{asHyp33c}\\
2\delta^{(\HH)} \approx \delta^{(k)} +2(\tilde\delta_i^{(m)}-\tilde\delta_i^{(k)})\,\tilde\chi_i.\label{asHyp33d}\end{eqnarray}\endnumparts
where $\tilde\delta_i^{(m)},\,\tilde\delta_i^{(k)}$ depend on the choice of $\tilde m_{qi},\,|\tilde k_{qi}|$ in (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq2})--(\ref{deckq3}). See figure 1b.
\item \underline{Asymptotic to Schwarzschild--Kruskal.} If we choose the power law decay for $\tilde m_{qi}$ in (\ref{decmq3}) with $\alpha=3$, then
\begin{equation}m_{qi}\sim m_0r^{-3},\qquad M\sim m_0R_0^3=\hbox{const.},\label{asHyp41}\end{equation}
then a comparison with (\ref{Schw}), (\ref{MEtbb3}) and (\ref{Schw_L0}) reveals an asymptotic convergence to Schwarzschild--Kruskal solution in coordinates given by geodesic observers with positive binding energy (see page 332 of \cite{kras2}). Evidently, the hypersurfaces ${}^3{\cal{T}}[t]$ of the Schwarzschild--Kruskal spacetime itself, in this specific time slicing, has an asymptotic limit to Minkowski in the radial direction. The asymptotic form for $|\tilde k_{qi}|$ can be any of the admissible forms (\ref{deckq2})--(\ref{deckq3}), but in order to illustrate the asymptotic behavior of the scalars we will also assume a power law decay $|k_{qi}|\sim |\tilde k_{qi}|=k_0\,r^{-\beta}$ with $\beta\leq 2$, so that
\numparts\begin{eqnarray}\fl\tilde\chi_i\approx \frac{m_0}{k_0}r^{\beta-3}\ll |\tilde k_{qi}|,\qquad \tilde \delta_i^{(m)} \sim -1+O(\tilde\chi_i),\quad \tilde\delta_i^{(k)}\sim -\frac{\beta}{3}+O(\tilde\chi_i),\label{asHyp42a}\\
\fl E\sim k_0 R_0^2 r^{2-\beta},\qquad ct_{\textrm{\tiny{bb}}}\approx ct_i -\frac{r^{\beta/2}}{\sqrt{k_0}}\to-\infty. \label{asHyp42b}
\end{eqnarray}\endnumparts
The asymptotic forms for the remaining scalars simply follow from specializing the parameters of (\ref{asHyp32a})--(\ref{asHyp33c}) to this case:
\numparts\begin{eqnarray} {\cal{H}}_q\approx \frac{\sqrt{k_0}}{r^{\beta/2}}\,\left[1-\frac{\sqrt{k_0}\,c(t-t_i)}{r^{\beta/2}}+\frac{m_0}{k_0\,r^{3-\beta}}\right]\to 0,\label{asHyp43a}\\
\Gamma \approx 1- \frac{\beta\,\sqrt{k_0}\,c(t-t_i)}{2\,r^{\beta/2}}+O(r^{-\beta})\to 1,\label{asHyp43b}\\
\delta^{(m)} \approx -1 +O(r^{-\beta})\to -1,\label{asHyp43c}\\
\delta^{(k)} \approx -\frac{\beta}{3}+ \frac{\beta\,(2-\beta)}{6}\frac{c(t-t_i)}{r^{\beta/2}}+O(r^{-\beta}) \to -\frac{\beta}{3},\label{asHyp43d}\\
\delta^{(\HH)} \approx \delta^{(k)} -3\left(1-\frac{\beta}{3}\right)\,\frac{2m_0}{k_0\,r^{3-\beta}}\to -\frac{\beta}{3}.\label{asHyp43e}\end{eqnarray}\endnumparts
Notice that (\ref{asHyp43c}) implies that local density, $m$, decays to zero much faster than $m_q$, as expected in a convergence to Schwarzschild--like vacuum conditions.
\end{itemize}
\subsection{Generic hyperbolic asymptotics.}
We consider now the case (\ref{ashpat3}) when $x_i\to x_0 = $ constant, listed in the sixth column of table \ref{table1}. We will assume the power law decay (\ref{decmq3}) and (\ref{deckq3}) with $\alpha=\beta$, so that density and spatial curvature decay at the same rate (the treatment of the logarithmic decay is analogous). This decay implies $m_{qi}\sim m_0\,r^{-\gamma}$ and $|k_{qi}|\sim k_0\,r^{-\gamma}$, where $0\leq \gamma\leq 2$ (regularity rules out $\gamma>2$). Hence, we obtain from (\ref{phi_L_h})
\begin{equation}
\fl \phi \sim y_0\,\xi+Z_h(x_0),\qquad x_i\sim x_0=\frac{k_0}{m_0},\qquad \xi \equiv\frac{c(t-t_i)}{r^{\gamma/2}},\label{asHyp51}\end{equation}
where $y_0=\sqrt{k_0}\,x_0$ and $Z_h$ is given by (\ref{hypZ1a}). Thus, following (\ref{phi_L_h}) and (\ref{hypZ1a}), we have $L\sim \tilde L$ so that (\ref{L_phi_h}) becomes
\numparts\begin{eqnarray} y_0\,\xi+Z_h(x_0) \sim Z_h(x_0\,\tilde L(\xi)),\label{asHyp52a}\\
\tilde L(\xi)=\frac{1}{x_0}\,Z_h^{-1}(y_0\,\xi+Z_h(x_0)).\label{asHyp52b} \end{eqnarray}\endnumparts
where we can write $\tilde L=\tilde L(\xi)$ because $Z_h(x_0)$ is a constant. We have then for $\gamma>0$
\begin{equation}\mathop {\lim }\limits_{r \to \infty } \tilde L = \mathop {\lim }\limits_{\xi \to 0} \tilde L(\xi ) = \frac{1}{{x_0 }}Z_h^{ - 1} (Z_h (x_0 )) = 1.\label{asHyp52c}
\end{equation}
As a consequence of (\ref{asHyp51}), (\ref{asHyp52a})--(\ref{asHyp52b}) and (\ref{asHyp52c}), and bearing in mind that we are considering $t$ constant, the asymptotic form of $\tilde L$ for very large $r$ is follows from a series expansion of $\tilde L(\xi)$ for $\xi\ll 1$ up to order $\xi$ ({\it{i.e.}} order $O(r^{-\gamma/2})$):
\begin{equation} \tilde L(\xi)\approx \tilde L(0)+L_{,\xi}(0)\,\xi = 1+\sqrt{2m_0+k_0}\,\,\xi,\label{asHyp52d}\end{equation}
where we used (\ref{asHyp52c}) and $L_{,\xi}(0)=[{\rm{d}} L/{\rm{d}}\xi]_{\xi=0}$ was computed by implicit derivation of (\ref{asHyp52a}).
The following asymptotic convergence forms follow from (\ref{mq}), (\ref{kq}), (\ref{Hq}), (\ref{ME}) and (\ref{tbbh}):
\numparts\begin{eqnarray} \fl {\cal{H}}_q\sim \frac{[2m_0+k_0\,\tilde L]^{1/2}}{r^{\gamma/2}\tilde L^{3/2}},\quad m_q\sim\frac{m_0}{r^{\gamma}\tilde L^3},\quad |k_q|\sim \frac{k_0}{r^{\gamma}\tilde L^2}.\label{asHyp53b}\\
\fl M\sim m_0R_0^3 r^{3-\gamma},\quad E\sim k_0R_0^2 r^{2-\gamma},\quad ct_{\textrm{\tiny{bb}}}\sim ct_i-\frac{Z_h(x_0)\,r^{\gamma/2}}{y_0},\label{asHyp53a} \end{eqnarray}\endnumparts
Since $\delta_i^{(m)},\,\delta_i^{(k)}\to -\gamma/3$ and considering (\ref{asHyp52d}), then (\ref{slawDm}), (\ref{slawDk}), (\ref{Ghe}) and (\ref{asHyp13b}) lead to the following approximations up to order $r^{-\gamma/2}$
\numparts\begin{eqnarray} \Gamma \approx 1-\frac{\gamma}{6}\,(2m_0+k_0)^{1/2}\,\xi,\label{asHyp54a}\\
\delta^{(m)} \approx -\frac{\gamma}{3}\,\left[1-\frac{3-\gamma}{6}\,(2m_0+k_0)^{1/2}\,\xi\right],\label{asHyp54b}\\
\delta^{(k)} \approx -\frac{\gamma}{3}\,\left[1-\frac{2-\gamma}{6}\,(2m_0+k_0)^{1/2}\,\xi\right],\label{asHyp54c}\\
2\delta^{(\HH)}\approx -\frac{\gamma}{3}\left[1-\frac{\gamma\,[2m_0(3-\gamma)+k_0(2-\gamma)]}{6\,(2m_0+k_0)^{1/2}}\,\xi\right].\label{asHyp54d}
\end{eqnarray}\endnumparts
The remaining scalars $m,\,k,\,{\cal{H}}$ are approximately equal to $m_q,\,k_q,\,{\cal{H}}_q$ times a factor $1-\gamma/3$. If instead of a power law form we consider a logarithmic as in (\ref{decmq2}) and (\ref{deckq2}), we obtain the same forms, except that $\xi=c(t-t_i)/\ln r$ and $\delta_i^{(m)},\,\delta_i^{(k)}\to 0$.
The asymptotic expressions (\ref{asHyp52d}), (\ref{asHyp53a})--(\ref{asHyp53b}) and (\ref{asHyp54a})--(\ref{asHyp54d}) hold for any regular hyperbolic model for which $x_i=|k_{qi}|/m_{qi}\to x_0=$ constant. Since the free parameter is $\gamma$, we can identify the following asymptotic states:
\begin{itemize}
\item \underline{Asymptotic to negatively curved FLRW.} If $\gamma=0$, then we have $m_{qi}\sim m_0$ and $|k_{qi}|\sim k_0$, as well as $\xi=c(t-t_i)$. Hence, $\tilde L,\,\tilde {\cal{H}}_q$ and $\tilde m_q$ depend only on $t$, while asymptotically $M\propto r^3$ and $E\propto r^2$ and $t_{\textrm{\tiny{bb}}}\to$ constant. Comparing with (\ref{FLRW}), (\ref{MEtbb1}) and (\ref{FLRW_L}), we can see that $L,\,M,\,E$ and $t_{\textrm{\tiny{bb}}}$ converge to their corresponding forms of a FLRW dust model with negative spatial curvature.
\item \underline{Asymptotic to generalized Milne.} If $0<\gamma\leq 2$ then $\xi\to 0$ as $r\to\infty$, thus $m_q,\,|k_q|$ and ${\cal{H}}_q$ tend to zero, while $t_{\textrm{\tiny{bb}}}\to-\infty$ and $\Gamma\to 1$. These limits and the expansions around it, together with (\ref{Mink_L}) clearly point out that models with these parameters are asymptotic to a section of Minkowski in generalized Milne coordinates.
\item \underline{Asymptotic to the self--similar solution with negative spatial curvature.} If $\gamma=2$, then $m_{qi},\,|k_{qi}|\,\propto r^{-2}$ and all quantities in (\ref{asHyp52d}), (\ref{asHyp53a})--(\ref{asHyp53b}) and (\ref{asHyp54a})--(\ref{asHyp54d}) converge to their respective forms in the self--similar solution with negative spatial curvature in (\ref{SS}) (this is the case $E>0$ in equation (2.29) of \cite{selfsim}).
\end{itemize}
\section{Radial asymptotics of open elliptic models.}
As with the hyperbolic models, we will use the implicit analytic solution of (\ref{Hq}) for $k_{qi}>0$, equivalent to (\ref{elltR}), which in terms of our variables takes the form
\begin{equation} \phi = \left\{ \begin{array}{l}
Z_e(x_iL) \qquad\qquad \hbox{expanding phase}\\
\\
2\pi-Z_e(x_iL) \qquad \hbox{collapsing phase}\\
\end{array} \right.\label{ellZ2}\end{equation}
where the expanding and collapsing phases correspond to ${\cal{H}}_q>0$ and ${\cal{H}}_q<0$, and $Z_e$ is given by
\begin{equation} u \mapsto Z_e(u)= \arccos(1 - u )-u ^{1/2} \left( {2 - u } \right)^{1/2}.\label{ellZ1a} \end{equation}
with
\numparts\begin{eqnarray}\phi &\equiv& \phi(t,x_i,y_i) = y_i\, c(t-t_i)+Z_e(x_i),\label{phie}\\
x_i &=& \frac{k_{qi}}{m_{qi}},\qquad y_i=\frac{k_{qi}^{3/2}}{m_{qi}},\label{xye}
\end{eqnarray}\endnumparts
The scale factor $L$ is restricted by $0<L\leq L_{\textrm{\tiny{max}}}=2/x_i=2m_{qi}/k_{qi}$. The times associated with the initial singularity, the maximal expansion ($L_{\textrm{\tiny{max}}}$) and the collapsing singularity are given by
\begin{equation}\fl ct_{\textrm{\tiny{bb}}} = ct_i-\frac{Z_e(x_i)}{y_i},\qquad ct_{\textrm{\tiny{max}}}=ct_{\textrm{\tiny{bb}}}+\frac{\pi}{y_i},\qquad ct_{\textrm{\tiny{coll}}}=ct_{\textrm{\tiny{bb}}}+\frac{2\pi}{y_i}.\label{tmc}\end{equation}
Notice that (in general) $t_{\textrm{\tiny{max}}}=t_{\textrm{\tiny{max}}}(r)$ and $t_{\textrm{\tiny{coll}}}=t_{\textrm{\tiny{coll}}}(r)$, so (like $t_{\textrm{\tiny{bb}}}(r)$) neither one is (in general) simultaneous (see section 12 for the cases with simultaneous $t_{\textrm{\tiny{bb}}}$ or $t_{\textrm{\tiny{max}}}$). For every comoving observer $r =$ const., the time evolution is contained in the range $t_{\textrm{\tiny{bb}}}(r)<t<t_{\textrm{\tiny{coll}}}(r)$.
The metric function $\Gamma$ takes the same form as (\ref{Ghe}), with ${\cal{H}}_q$ given by (\ref{Hq}) with $k_{qi}>0$ and $c(t-t_i)$ follows from (\ref{ellZ2}) and (\ref{phie}). The Hellaby--Lake conditions \cite{ltbstuff,suss10,ltbstuff1} to comply with (\ref{noshxG}) are
\begin{equation} \frac{ct_{\textrm{\tiny{bb}}}'}{3R'_i/R_i}\leq 0,\qquad \frac{ct_{\textrm{\tiny{coll}}}'}{3R'_i/R_i}\geq 0,\qquad \delta_i^{(m)}\geq -1,\label{noshxGe}\end{equation}
where $ct_{\textrm{\tiny{coll}}}'$ follows from differentiating $ct_{\textrm{\tiny{coll}}}$ in (\ref{tmc})
\begin{equation} \frac{ct_{\textrm{\tiny{coll}}}'}{3R'_i/R_i} =\left(\delta_i^{(m)}-\frac{3}{2}\delta_i^{(k)}\right)\,c(t_{\textrm{\tiny{coll}}}-t_{\textrm{\tiny{bb}}})+\frac{ct_{\textrm{\tiny{bb}}}'}{3R'_i/R_i},\label{tcollr}\end{equation}
while $ct_{\textrm{\tiny{bb}}}'$ has the same form as (\ref{tbbr}) with $ct_{\textrm{\tiny{bb}}}$ given by (\ref{tmc}).
\subsection{Initial positive curvature.}
The asymptotic forms for $k_{qi}$ compatible with standard regularity must comply with the constraint $1+E={\cal{F}}^2=1-k_{qi} R_0^2 r^2$, and thus $k_{qi}r^2$ cannot diverge as $r\to\infty$. Hence, the admissible asymptotic forms are
\numparts\begin{eqnarray}
\fl k_{qi}\sim \tilde k_{qi}=k_0 r^{-\beta}\;\; (\beta\geq 2),\qquad {\cal{F}}^2\sim 1-k_0 R_0^2 r^{2-\beta},\qquad \delta_i^{(k)}\to -\frac{\beta}{3},\label{deckq11}\\
\fl k_{qi}\sim \tilde k_{qi}=k_0\,\textrm{e}^{-\beta r}, \qquad\qquad\quad {\cal{F}}^2\sim 1-k_0 R_0^2 r^2\textrm{e}^{-\beta r},\quad \delta_i^{(k)}\to -\frac{\beta}{3}\,r. \label{deckq22}\end{eqnarray}\endnumparts
Notice that the case $k_{qi}\sim k_0$ ({\it{i.e.}} (\ref{deckq11}) with $\beta=0$) is not admissible as an asymptotic form for $r\to\infty$, hence $k_{qi}\to 0$ holds in this limit for all open elliptic models (though $m_{qi}\to m_0 =$ const. is allowed). We also remark that the logarithmic decay is also ruled out because $k_{qi}$ must decay at least as fast as $r^{-2}$. Also, from (\ref{slaw2}) and (\ref{slawDk}), if $\delta_i^{(k)}<-1$ then we have $k<0$ even if $k_q>0$ in all ${}^3{\cal{T}}[t]$. This situation occurs if $\beta>3$ in (\ref{deckq11}) and in (\ref{deckq22}) for which $\delta_i^{(k)}\to-\infty$ (see Appendix B of \cite{suss10}).
\subsection{Asymptotic approximations for the analytic solutions.}
As with hyperbolic models, there is no closed explicit form for $L$ and we must work with the implicit solutions (\ref{ellZ2})--(\ref{tmc}), with $Z_e$ given by (\ref{ellZ1a}).
As opposed to the hyperbolic case, we have now two branches, but $\phi=Z_e(x_iL)$ in (\ref{ellZ2}) is a monotonously increasing/decresing function of $x_iL$ in the expanding/colapsing branches, so it has a branched inverse
\begin{equation}x_i L(\phi ) = \left\{ \begin{array}{l}
Z_e^{ - 1} (\phi ) \\
Z_e^{ - 1} (2\pi - \phi ) \\
\end{array} \right..\label{L_phi_e}\end{equation}
However, from (\ref{ellZ1a}), (\ref{LTB2}) and (\ref{Rgauge}) for open elliptic models with $k_{qi}>0$ we must have
\begin{equation} \fl 0<x_i\leq 2\qquad\hbox{and}\qquad 1-k_{qi}r^2R_0^2>0\quad\Rightarrow\quad k_{qi}\,r^2<1/R_0^2,\label{ell_restr}\end{equation}
so that $x_i$ is bounded and $k_{qi}r^2$ cannot be an increasing function as $r\to\infty$ (which is consistent with (\ref{deckq11})--(\ref{deckq22})). Hence (as opposed to the hyperbolic case) $\phi$ must remain bounded.
As a direct consequence of (\ref{deckq11})--(\ref{deckq22}) and (\ref{ell_restr}), the asymptotic convergence $k_{qi}\sim k_0=$ constant is not allowed. Therefore, open elliptic LTB models are incompatible with a positively curved FLRW asymptotic limit. This is not surprising, since these FLRW dust models are necessarily closed (the ${}^3{\cal{T}}[t]$ homeomorphic to $\mathbb{S}^3$).
As with the hyperbolic models, the asymptotic behavior of open elliptic models depends on the initial value functions through $x_i$. If we take into consideration the restrictions (\ref{deckq11})--(\ref{deckq22}) and (\ref{ell_restr}), then (\ref{ellZ2}) and (\ref{L_phi_e}) allow for the following limiting regimes
\begin{equation} x_i\to 0,\qquad x_i\to x_0<2,\qquad x_i\to 2,\label{asepat}\end{equation}
which will be discussed in detail, respectively, in sections 11.3 to 11.5. The combinations of the admissible asymptotic convergence forms compatible with (\ref{ell_restr}) that correspond to each of these cases are listed in 4th, 5th and 6th columns of table \ref{table2}. For reasons that will be explained in the following subsections, the collapsing branch will be needed only in the case $x_i\to 2$ in (\ref{asepat}) (see figure 2).
\begin{table}
\begin{center}
\begin{tabular}{|c| c| c| c| c| c|}
\hline
\hline
{} &{} &{} &{Section 11.3} &{Section 11.4} &{Section 11.5}
\\
{$\tilde m_{qi}$} &{$\tilde k_{qi}$} &{$\tilde x_i$} &{$\tilde x_i\to 0$} &{$\tilde x_i\to x_0<2$} &{$\tilde x_i\to x_0=2$}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{PL} &{PL} &{$x_0\, r^{\alpha-\beta}$} &{$\beta>\alpha$} &{$2\leq\beta=\alpha\leq 3$} &{$2\leq\beta=\alpha\leq 3$}
\\
{} &{} &{} &{$0<\alpha\leq 3$} &{} &{$k_0=2m_0$}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{PL} &{EXP} &{$x_0\, r^\alpha\textrm{e}^{-\beta r}$} &{$\beta>0$} &{} &{}
\\
{} &{} &{} &{$0<\alpha\leq 3$} &{} &{}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{LOG} &{PL} &{$x_0\, r^{-\beta}\ln r^\alpha$} &{$\beta\geq 2$} &{} &{}
\\
{} &{} &{} &{$\alpha>0$} &{} &{}
\\
\hline
\hline
{} &{} &{} &{} &{} &{}
\\
{LOG} &{EXP} &{$x_0\, \textrm{e}^{-\beta r}\ln r^\alpha$} &{$\beta>0$} &{} &{}
\\
{} &{} &{} &{$\alpha>0$} &{} &{}
\\
\hline
\hline
\end{tabular}
\end{center}
\caption{{\bf{Combinations of the admissible asymptotic convergence forms for open elliptic models}}. The table presents the asymptotic convergence of $\tilde x_i=\tilde k_{qi}/\tilde m_{qi}$ for the four possible combinations of the forms $\tilde m_{qi}$ and $\tilde k_{qi}$ in (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq11})--(\ref{deckq22}). The terms ``PL'' and ``EXP'' stand for ``power law'' and ``exponential'' decay forms, respectively given by (\ref{decmq3}), (\ref{deckq11}) and (\ref{decmq2}), (\ref{deckq22}). The behavior of $\tilde x_i$ examined in sections 11.3--11.5 follows from the restrictions on the parameters $\alpha,\,\beta$, hence the results of these sections can be readily applied to initial value functions having each asymptotic convergence form.}
\label{table2}
\end{table}
\subsection{Convergence to parabolic models.}
Since $x_i\leq 2$ there is no possibility for $k_{qi}$ to dominate $m_{qi}$ in the limit $r\to \infty$, though the opposite situation is possible and leads to the limit $ x_i\to 0$ corresponding to the first one of the cases in (\ref{asepat}) and listed in the fourth column of table \ref{table2}. As we show further ahead, the collapsing phase of (\ref{ellZ2}) will not be needed (see figure 2), while $Z_e$ is monotonously increasing in the expanding phase. We have then
\begin{equation}
\fl x_i\to 0\quad\quad\Rightarrow\quad\;\; Z_e(x_i)\to 0\quad \Rightarrow\quad \phi\to 0\quad \Rightarrow\quad Z_e(x_iL)\to 0.\label{asepat1}\end{equation}
Hence, we can approximate $Z_e$ in (\ref{ellZ1a}) by
\begin{equation} Z_e(u)\approx \frac{\sqrt{2}}{3}\,u^{3/2}+\frac{\sqrt{2}}{20}\,u^{5/2}\quad\hbox{for}\quad u\approx 0,\label{apprell1}\end{equation}
which, up to the leading term, coincides with (\ref{apprhyp1}). As in the density dominated hyperbolic case, we assume $m_{qi}\sim \tilde m_{qi}$ and $k_{qi}\sim \tilde k_{qi}$, and apply (\ref{apprell1}) to $Z_e(x_i)$ and $Z_e(x_i L)$ in the expanding branch of (\ref{ellZ2}), leading exactly to the same form for $L\approx \tilde L$ given by (\ref{asHyp11})
\begin{equation} L \sim \tilde L\approx \left[1+\frac{3}{2}\sqrt{2\tilde m_{qi}}\,c(t-t_i)\right]^{2/3},\label{asEll11}\end{equation}
which indicates that open elliptic models with initial conditions in which $x_i\to 0$ are asymptotic to a parabolic model.
While open elliptic and hyperbolic models with $x_i\to 0$ have the same radial asymptotic behavior, their time evolution is radically different, as it is restricted by the initial and collapsing singularity (see figure 2). Considering (\ref{apprell1}) and the fact that we are assuming $x_i\to 0$ and $\tilde k_{qi}\to 0$ for all choices in (\ref{deckq11})--(\ref{deckq22}), the maximal expansion and collapse times $ct_{\textrm{\tiny{max}}},\,ct_{\textrm{\tiny{coll}}}$ given by (\ref{tmc}) take the asymptotic forms
\numparts\begin{eqnarray} ct_{\textrm{\tiny{max}}} =ct_i+\frac{\pi-Z_e(x_i)}{k_{qi}^{1/2}\,x_i}\approx ct_i-\frac{\sqrt{2}}{3}\,\tilde x_i^{1/2}+\frac{\pi}{\tilde k_{qi}^{1/2}\,\tilde x_i}\to\infty,\label{ctmax1}\\
ct_{\textrm{\tiny{coll}}} =ct_i+\frac{2\pi-Z_e(x_i)}{k_{qi}^{1/2}\,\tilde x_i}\approx ct_i-\frac{\sqrt{2}}{3}\,\tilde x_i^{1/2}+\frac{2\pi}{\tilde k_{qi}^{1/2}\,\tilde x_i}\to\infty.\label{ctcoll1}\end{eqnarray}\endnumparts
where $\tilde m_{qi}$ and $\tilde k_{qi}$ take the admissible forms compatible with $\tilde x_i\to 0$. Since the locus of maximal expansion marked by $ct_{\textrm{\tiny{max}}}(r)$ corresponds to a maximum $L=L_{\textrm{\tiny{max}}}=2/x_i$, we have $L_{\textrm{\tiny{max}}}\to\infty$ as $r\to\infty$ and so: $R_{\textrm{\tiny{max}}}=R_0rL_{\textrm{\tiny{max}}}\to \infty$ in this limit.
Since the collapsing phase corresponds to $ct>ct_{\textrm{\tiny{max}}}(r)$, the fact that $ct=ct_{\textrm{\tiny{max}}}(r)$ diverges as $r\to\infty$ implies that the radial asymptotic range for all hypersurfaces ${}^3{\cal{T}}[t]$ occurs within the expanding phase, which justifies the fact that we did not need to consider the collapsing phase of (\ref{ellZ2}) for the study of the radial asymptotics in this case. This is clearly illustrated by figures 2a and 2b. Since open elliptic models have the same asymptotic behavior as in hyperbolic models in which $x_i\to 0$, the same asymptotic limits arise:
\begin{itemize}
\item \underline{Asymptotic to spatially flat FLRW.}
If $m_{qi}\sim m_0>0$ (the power law form (\ref{decmq3}) with $\alpha=0$, see top entry in the fourth column of table 2), then (\ref{asEll11}) yields $L\approx \tilde L(t)$ with $\tilde L(t)$ given by (\ref{asPar2a}), which is the scale factor of a spatially flat FLRW dust model. The conventional variables $M,\,E$ and $t_{\textrm{\tiny{bb}}}$ have the asymptotic forms as in (\ref{asHyp12}) (with $-\tilde k_{qi}$ instead of $|\tilde k_{qi}|$), while the maximal expansion and collapse times, $ct_{\textrm{\tiny{max}}}$ and $ct_{\textrm{\tiny{coll}}}$, are given by (\ref{ctmax1}) and (\ref{ctcoll1}) under the specialization (\ref{decmq3}) with $\alpha=0$, so that $x_i\to 0$. The asymptotic forms of the remaining time--dependent scalars are readily computed as in the hyperbolic case. The result is exactly the same forms as in (\ref{asHyp13a})--(\ref{asHyp13d}).
\item \underline{Asymptotic to Minkowski.}
If $\tilde m_{qi}\to 0$ and $\tilde k_{qi}\to 0$, following either power law decays with $0<\alpha<\beta$ or any combination of power law and exponenetial decay, as given by (\ref{decmq2})--(\ref{decmq3}) and (\ref{deckq11})--(\ref{deckq22}), we have the cases listed in the fourth column of table 2 (excluding the case PL--PL with $\alpha=0$ which corresponds to the spatially flat FLRW case). For all admissible combinations of initial value functions (\ref{asEll11}) becomes as $r\to\infty$
\begin{equation} \tilde L \approx 1+\sqrt{2\tilde m_{qi}}\,c(t-t_i)\to 1,\label{asEll14a}\end{equation}
while the conventional variables $M,\,E$ and $ct_{\textrm{\tiny{bb}}}$ take the same asymptotic forms as (\ref{asHyp14b}), while $ct_{\textrm{\tiny{max}}}$ and $ct_{\textrm{\tiny{coll}}}$ follow from (\ref{ctmax1})--(\ref{ctcoll1}), and the scalars $m_q$ and $k_q$ have the same forms as in (\ref{asHyp15a}). Considering only terms linear in $\sqrt{\tilde m_{qi}}$ or $\tilde x_i$, we obtain up to leading order
\begin{equation}\fl {\cal{H}}_q\approx \frac{\sqrt{2\tilde m_{qi}}}{\tilde L}\,\left(1-\frac{\tilde x_i}{4}\right)\approx \sqrt{2\tilde m_{qi}}\,\left[1-\sqrt{2\tilde m_{qi}}c(t-t_i)-\frac{\tilde x_i}{4}\right] \to 0,\label{asHyp15b}\end{equation}
The remaining quantities have the same asymptotic forms as (\ref{asHyp16a})--(\ref{asHyp16d}), but with $\tilde k_{qi}$ and $\tilde\delta_i^{(k)}$ now given by (\ref{deckq11})--(\ref{deckq22}).
\end{itemize}
\subsection{Generic elliptic asymptotics.}
We examine now the second case in (\ref{asepat2}) listed in the fifth column of table \ref{table2}:
\begin{equation}\fl x_i\to x_0=\hbox{const.}<2,\qquad \phi\to k_{qi}^{1/2}\,x_0\,c(t-t_i) +Z_e(x_0)\to Z_e(x_0L),\label{asepat2}\end{equation}
From table 2, the only combination compatible with $x_i\to x_0$ is that in which both, $\tilde m_{qi}$ and $\tilde k_{qi}$, have the power law forms (\ref{decmq3}) and (\ref{deckq11}) with
same exponent:
\begin{equation} \fl \tilde m_{qi}=m_0r^{-\gamma},\quad \tilde k_{qi}=k_0r^{-\gamma},\quad2\leq \gamma\leq 3,\quad x_0=\frac{k_0}{m_0}<2.\label{initvalsx0}\end{equation}
where the restriction on $\gamma$ follows from the possible common exponents in (\ref{decmq3}) and (\ref{deckq11}). Considering that $Z_e(x_0)<\pi$ for all $x_0<2$ (we examine the case $x_0=2$ in the following subsection), the maximal expansion and collapse times for these forms of $\tilde m_{qi}$ and $\tilde k_{qi}$ take the following asymptotic forms
\numparts \begin{eqnarray} ct_{\textrm{\tiny{max}}} \sim ct_i+\frac{\pi-Z_e(x_0)}{\sqrt{k_0}\,x_0}\,r^{\gamma/2}\to\infty,\label{ctmax2}\\
ct_{\textrm{\tiny{coll}}} \sim ct_i+\frac{2\pi-Z_e(x_0)}{\sqrt{k_0}\,x_0}\,r^{\gamma/2}\to\infty.\label{ctcoll2} \end{eqnarray}\endnumparts
As a consequence, only the expanding phase in (\ref{ellZ2}) is needed to examine the radial asymptotics in this case (just as with the case $x_i\to 0$). If $x_i\to x_0$, then as $r\to\infty$ the scale factor $L$ at maximal expansion tends to a finite value:\, $L=L_{\textrm{\tiny{max}}}=2/x_i\to 2/x_0$, though we have $R_{\textrm{\tiny{max}}}=R_0rL_{\textrm{\tiny{max}}}\to (2R_0/x_0)\,r\to \infty$ in this limit. Since $2/x_0>1$ for $x_0<2$ and $L=1$ at the initial hypersurface ${}^3{\cal{T}}[t_i]$, then this hypersurface necessarily lies in the expanding phase: $t_i<t_{\textrm{\tiny{max}}}$ (see figure 2b).
Since we are only considering the expanding phase in (\ref{ellZ2}), the equation $\phi=Z_e(x_0\tilde L)$ has exactly the same form as in (\ref{asHyp51}) and (\ref{asHyp52a})--(\ref{asHyp52b}), with $Z_e$ instead of $Z_h$. Thus, (\ref{asHyp52c}) holds and we have $L\to 1$ as $r\to\infty$, while the asymptotic expansion (\ref{asHyp52d}) takes now the form
\begin{equation}\tilde L \approx 1+\sqrt{2m_0-k_0}\,\,\xi,\qquad \xi=\frac{c(t-t_i)}{r^{\gamma/2}}.\label{asEll52d}\end{equation}
The asymptotic convergence forms for $M,\,E,\,t_{\textrm{\tiny{bb}}},\,m_q,\,k_q$ and ${\cal{H}}_q$ follow from (\ref{mq}), (\ref{Hq}), (\ref{ME}) and (\ref{tmc}), and are very similar to those in (\ref{asHyp53a})--(\ref{asHyp53b}):
\numparts\begin{eqnarray} \fl M\sim m_0R_0^3 r^{3-\gamma},\quad E\sim -k_0R_0^2 r^{2-\gamma},\quad ct_{\textrm{\tiny{bb}}}\sim ct_i-\frac{Z_e(x_0)\,r^{\gamma/2}}{y_0}\to-\infty,\label{asEll53a}\\
\fl {\cal{H}}_q\sim \frac{[2m_0-k_0\,\tilde L]^{1/2}}{r^{\gamma/2}\tilde L^{3/2}}\to 0,\quad m_q\sim\frac{m_0}{r^{\gamma}\tilde L^3}\to 0,\quad k_q\sim \frac{k_0}{r^{\gamma}\tilde L^2}\to 0.\label{asEll53b}\end{eqnarray}\endnumparts
where $y_0=\sqrt{k_0}x_0$ and $\tilde L$ is given by (\ref{asEll52d}).
Since $\delta_i^{(m)},\,\delta_i^{(k)}\to -\gamma/3$ and considering (\ref{asEll52d}), then (\ref{slawDm}), (\ref{slawDk}), (\ref{Ghe}) and (\ref{asHyp13b}) lead to the same approximations (up to order $r^{-\gamma/2}$) as (\ref{asHyp54a})--(\ref{asHyp54d}), but with $k_0$ replaced by $-k_0$. Since $\gamma=0$ is not possible, then all open elliptic models complying with (\ref{initvalsx0}) have a Minkowski asymptotic limit. The following asymptotic states emerge:
\begin{itemize}
\item \underline{Asymptotic to the self--similar solution with positive spatial curvature.} If $\gamma=2$, then $m_{qi},\,k_{qi}\,\propto r^{-2}$ and $M\propto\,r$ and $E\sim -k_0R_0^2$ hold. Hence, there is asymptotic convergence to the self--similar solution with positive spatial curvature in (\ref{SS}) with self--similar variable $\xi=c(t-t_i)/r=\zeta$ (this is the case $E<0$ in equation (2.29) of \cite{selfsim}).
\item \underline{Asymptotic to Schwarzschild--Kruskal.} If $\alpha=3$, then $M\sim m_0R_0^3=$ constant. Comparison with (\ref{Schw}), (\ref{MEtbb3}) and (\ref{Schw_L0}) reveals an asymptotic convergence to Schwarzschild--Kruskal solution in coordinates given by geodesic observers with negative binding energy (Novikov coordinates, see page 332 of \cite{kras2}).
\end{itemize}
\subsection{The case $x_0=2$.}
The particular case $x_i\sim x_0=2$ is characterized by the same initial value functions in (\ref{initvalsx0}) with $k_0=2m_0$ (sixth column of table \ref{table2}). Equation (\ref{ellZ2}) can be rewritten as
\begin{equation} Z_e(2\tilde L(\xi)) \sim \pi \pm 2\sqrt{2m_0}\,\xi,\qquad \xi=\frac{c(t-t_i)}{r^{\gamma/2}}\label{asepat3}\end{equation}
where the $+$ and $-$ signs respectively correspond to the expanding and collapsing phase, and we used the fact that $Z_e(2)=\pi$.
There is a fundamental difference (in comparison with the case $x_0<2$) in the locus of the maximal expansion time, $ct_{\textrm{\tiny{max}}}$. Considering that both $m_{qi}$ and $k_{qi}$ decay as $r^{-\gamma}$, then we can write in general $x_i\sim 2+O(r^{-\gamma})$. Hence, we have for $x_i\approx 2$
\begin{equation}\pi - Z_e(x_i)\approx 2\sqrt{2}r^{-\gamma/2}\,\left[1+\frac{1}{6}\,r^{-\gamma/2}+O(r^{-\gamma})\right],\label{x02exp}\end{equation}
which, from (\ref{tmc}), leads to
\numparts\begin{eqnarray} ct_{\textrm{\tiny{max}}} \sim ct_i +\frac{[\pi-Z_e(x_i)]\,r^{\gamma/2}}{2\sqrt{2m_0}}\approx ct_i+\frac{1}{\sqrt{m_0}},\label{ctmax02}\\
ct_{\textrm{\tiny{coll}}} \sim ct_i +\frac{[2\pi-Z_e(x_i)]\,r^{\gamma/2}}{2\sqrt{2m_0}}\approx ct_i+\frac{\pi\,r^{\gamma/2}}{2\sqrt{2m_0}}\to\infty,\label{ctcoll02}\\
ct_{\textrm{\tiny{bb}}} \sim ct_i -\frac{Z_e(x_i)\,r^{\gamma/2}}{2\sqrt{2m_0}}\approx ct_i-\frac{\pi\,r^{\gamma/2}}{2\sqrt{2m_0}}\to -\infty,\label{ctbb02}
\end{eqnarray}\endnumparts
so that $t_{\textrm{\tiny{bb}}}$ and $t_{\textrm{\tiny{coll}}}$ tend to curves that are symmetric with respect to the constant asymptotic line of $t_{\textrm{\tiny{max}}}$ (see figure 2c). Therefore, as a consequence of (\ref{ctmax02})--(\ref{ctbb02}), there exist an extended collapsing region for open elliptic models for which $x_i\sim x_0=2$. As opposed to the cases $x_i\to 0$ and $x_i\to x_0<2$, there is now a radial asymptotic range for the hypersurfaces ${}^3{\cal{T}}[t]$ in both the expanding phase and the collapsing phase $ct_{\textrm{\tiny{max}}}<ct<ct_{\textrm{\tiny{coll}}}$. In fact, models in which $x_0=2$ are asymptotic to the models with a simultaneous $t_{\textrm{\tiny{max}}}$ that will be examined in the following section.
If $x_i\to 2$, then:\, $L=L_{\textrm{\tiny{max}}}=2/x_i\to 1$ as $r\to\infty$, but $R_{\textrm{\tiny{max}}}=R_0rL_{\textrm{\tiny{max}}}\to R_0\,r=R_i\to \infty$ in this limit. Notice that (\ref{ctmax02}) implies that the locus of the maximal expansion itself lies in the expanding phase. The asymptotic limit of $L$ follows by inverting (\ref{asepat3}):
\begin{equation}2\tilde L \sim Z_e^{-1}(\pi\pm 2\sqrt{2m_0}\,\xi).\label{L_phi_x02}\end{equation}
Since $\xi\to 0$ as $r\to\infty$ and $\pi=Z_e(2)$, then (\ref{L_phi_x02}) implies $L\sim \tilde L\to 1$ in this limit (for both the expanding and collapsing phases). While this is the same limit as in the case $x_0<2$, the asymptotic expansions like (\ref{asEll52d}) and the equivalents of (\ref{asHyp54a})--(\ref{asHyp54d}) must be computed up to $\xi^2$, since now $\sqrt{2m_0-k_0}=0$. The result (valid for the expanding and collapsing phases) is
\begin{equation} \tilde L(\xi) \approx \tilde L(0)+\frac{1}{2}\tilde L_{,\xi\xi}(0)\,\xi^2=1-\frac{m_0}{2}\,\xi^2,\label{Laprx02}\end{equation}
where $\tilde L_{,\xi\xi}(0)=[{\rm{d}}^2\tilde L/{\rm{d}}\xi^2]_{\xi=0}$ was computed by implicit derivation of (\ref{asepat3}) with respect to $\xi$. The asymptotic form for ${\cal{H}}_q$ follows readily from (\ref{asEll53b}) with $k_0=2m_0$ and (\ref{Laprx02})
\begin{equation}\fl {\cal{H}}_q \sim \frac{\sqrt{2m_0}\,(1-\tilde L)^{1/2}}{r^{\gamma/2}\,\tilde L^{3/2}}\approx \frac{m_0\,\xi}{1-3m_0\xi^2/4}\approx m_0\,\xi\,\left[1+\frac{3m_0}{4}\,\xi+O(\xi^2)\right]\to 0.\label{Hqaprx02}\end{equation}
The asymptotic form for $\Gamma$ cannot be computed from (\ref{Ghe}) because $\xi=0$ at $t=t_i$, hence ${\cal{H}}_{qi}\approx 0$. Instead, we use the definition of $\Gamma$ in (\ref{LGdef}) applied to $\tilde L$ in (\ref{Laprx02})
\begin{equation} \Gamma \sim 1+\frac{r\tilde L'}{\tilde L} \approx \frac{1+\gamma\,m_0\,\xi^2/2}{1+m_0\,\xi^2/2}\approx 1+\frac{(\gamma-1) m_0}{2}\,\xi^2\to 1. \label{Gaprx02}\end{equation}
Bearing in mind that $\tilde\delta_i^{(m)} =\tilde\delta_i^{(k)} =-\gamma/3$, the following asymptotic forms follow readily
\numparts\begin{eqnarray} \delta^{(m)} \approx -\frac{\gamma}{3}\,\left[1+\frac{(3-\gamma)\,m_0}{2}\,\xi^2+O(\xi^3)\right]\to -\frac{\gamma}{3},\label{Dmaprx02}\\
\delta^{(k)} \approx -\frac{\gamma}{3}\,\left[1+\frac{(2-\gamma)\,m_0}{2}\,\xi^2+O(\xi^3)\right]\to -\frac{\gamma}{3},\label{Dkaprx02}\\
2\delta^{(\HH)} \approx -\frac{2\gamma}{3}\,\left[1+\frac{(2-\gamma)\,m_0}{4}\,\xi^2+O(\xi^3)\right]\to -\frac{2\gamma}{3}.\label{Dhaprx02}\end{eqnarray}\endnumparts
Hence, the case $x_0=2$ yields the same asymptotic states as $x_0<2$, but with a faster decay $O(\xi^2)$ to the Minkowski limit. The models are also asymptotic to Minkowski in generalized Milne coordinates when $2<\gamma<3$, with $\gamma=2,3$ corresponding to models asymptotic to self--similar and Schwarzschild--Kruskal spacetimes.
\section{Simultaneous big bang or maximal expansion.}
So far we have assumed that the initial curvature singularity ($L=0$) is not simultaneous, but given by the curve $[ct_{\textrm{\tiny{bb}}}(r),r]$ in the $(ct,r)$ plane. While a constant $t_{\textrm{\tiny{bb}}}$ is incomplatible with parabolic models, non--trivial and perfectly regular hyperbolic and elliptic models exist for which $t_{\textrm{\tiny{bb}}}'=0$ (see \cite{KH4,ltbstuff,suss02}). As shown in \cite{suss10}, the Hellaby--Lake conditions to avoid shell crossings are simply (\ref{noshxGh}) and (\ref{noshxGe}) with $t_{\textrm{\tiny{bb}}}'=0$. Also, the locus $[ct_{\textrm{\tiny{max}}}(r),r]$ of the maximal expansion time (${\cal{H}}_q = 0$) in elliptic models or regions is, in general, not simultaneous, though regular models exist with $ct_{\textrm{\tiny{max}}}'=0$ \cite{KH4,suss10}. We examine in this section the radial asymptotic behavior of regular models with these characteristics.
\subsection{Simultaneous big bang.}
Let $t_{\textrm{\tiny{bb}}}=t_{(0)}$ denote the constant time value associated with $L=0$, then from (\ref{tbbh}) and (\ref{tmc}), the initial value functions $m_{qi}$ and $k_{qi}$ are necessarily linked by the constraints
\numparts\begin{eqnarray} \fl c(t_i-t_{(0)})=F_h(m_{qi},|k_{qi}|)\equiv m_{qi}\frac{Z_h(|k_{qi}|/m_{qi})}{|k_{qi}|^{3/2}},\qquad\hbox{hyperbolic},\label{tbbch}\\
\fl c(t_i-t_{(0)})=F_h(m_{qi},k_{qi})\equiv m_{qi}\frac{Z_e(k_{qi}/m_{qi})}{k_{qi}^{3/2}},\qquad\hbox{elliptic},\label{tbbce}\end{eqnarray}\endnumparts
where the functions $Z_h$ and $Z_e$ are given by (\ref{hypZ1a}) and (\ref{ellZ1a}), and we have used (\ref{phih}) and (\ref{phie}) to express $x_i$ and $y_i$ in terms of $m_{qi}$ and $k_{qi}$. We can prescribe one of the pair $m_{qi}, k_{qi}$ and find the other by solving the constraints (\ref{tbbch}) and (\ref{tbbce}). While these constraints cannot be solved analytically, we examine below the radial asymptotics of models with $t_{\textrm{\tiny{bb}}}=t_{(0)}$ by looking at them qualitatively.
Since the functions $m_{qi},k_{qi}$ that solve (\ref{tbbch}) and (\ref{tbbce}) for given values $c(t_i-t_{(0)})$ define the level curves of $F_h(m_{qi},|k_{qi}|)$, we need to verify first if the admissible assumptions of asymptotic convergence of $m_{qi},k_{qi}$ are compatible with these functions lying in a level curve, and then if their limit as $r\to\infty$ (in the $(m_{qi},k_{qi})$ plane) fixes the value of $c(t_i-t_{(0)})$. Considering these assumptions, the following mutually exclusive cases emerge:
\begin{itemize}
\item $m_{qi}\sim m_0,\,k_{qi}\sim k_0$. These assumptions are only possible in hyperbolic models, hence we have $F_h(m_{qi},|k_{qi}|)\sim F_h(k_0,m_0) =$ constant, which fixes the level curve $c(t_i-t_{(0)})$. Hence, all hyperbolic models with these initial value functions and $t_{\textrm{\tiny{bb}}}=t_{(0)}$ are asymptotic to negatively curved FLRW dust cosmology, as the latter follows from these asymptotic convergence forms.
\item $k_{qi}\to 0$ and $m_{qi}\sim m_0$. For models with a simultaneous big bang having these convergence forms we need to find the limit of $F_h$ and $F_e$ as $k_{qi}\to 0$ with $m_{qi}= m_0$ fixed:
\begin{equation}
\mathop {\lim }\limits_{|k_{qi} | \to 0} F_h (|k_{qi} |,m_0 ) = \mathop {\lim }\limits_{k_{qi} \to 0} F_e (k_{qi} ,m_0 ) = \frac{2}
{{3\sqrt {2m_0 } }}
\end{equation}
which fixes the constant bang time in terms of the level curve $c(t_i-t_{(0)}) = 2/(3\sqrt{2m_0})$. As a consequence, hyperbolic and elliptic models with $t_{\textrm{\tiny{bb}}}=t_{(0)}$ and these initial value functions are asymptotic to a spatially flat FLRW dust cosmology, as the latter corresponds to these asymptotic convergence forms of $m_{qi}$ and $k_{qi}$.
\item $m_{qi}\to 0$ and $k_{qi}\sim k_0$. This asymptotic convergence only occurs in hyperbolic models ($k_{qi}<0$), since in elliptic models $0<k_{qi}\leq 2m_{qi}$ must hold. To examine this case in models with a simultaneous big bang, we evaluate the limit of $F_h$ as $m_{qi}\to 0$ with $k_{qi}= k_0$ fixed:
\begin{equation}
\mathop {\lim }\limits_{m_{qi} \to 0} F_e (k_0 ,m_{qi} ) = \frac{1}
{{\sqrt {k_0 } }}
\end{equation}
which fixes the constant bang time in terms of the level curve $c(t_i-t_{(0)}) =1/\sqrt{k_0}$. Thus, hyperbolic models with $t_{\textrm{\tiny{bb}}}=t_{(0)}$ and these initial value functions are asymptotic to the Milne Universe, which corresponds to these asymptotic convergence forms of $m_{qi}$ and $k_{qi}$.
\item $k_{qi}\to 0$ and $m_{qi}\to 0$. In order to examine this asymptotic convergence for models with a simultaneous big bang, we explore the behavior of $F_h$ and $F_e$ in the direction in the $(m_{qi},k_{qi})$ plane given by the ray $k_{qi}=\alpha_0 m_{qi}$, where $\alpha_0>0$ is a constant ($\alpha_0\leq 2$ for elliptic models). Hence, $F_h$ and $F_e$ along this direction are given by
\begin{equation}F_h(|k_{qi}|)=\frac{\alpha_0\,Z_h(1/\alpha_0)}{|k_{qi}|^{1/2}},\qquad F_e(k_{qi})=\frac{\alpha_0\,Z_e(1/\alpha_0)}{k_{qi}^{1/2}},\end{equation}
Thus, in the limit as $m_{qi}\to 0$ and $k_{qi}\to 0$ (which reduces in this direction to the limit as $k_{qi}$ tends to zero) we have $F_h\to\infty$ and $F_e\to\infty$. In fact, both $F_h$ and $F_e$ only diverge as $(m_{qi},k_{qi})\to (0,0)$. All this means that $c(t_i-t_{(0)})\to\infty$ for any choice of $m_{qi},k_{qi}$ in which both tend to zero asymptotically. As a consequence, $m_{qi}$ and $k_{qi}$ cannot be fit to any level curve of these functions, and so a simultaneous big bang is incompatible with both $k_{qi}\to 0$ and $m_{qi}\to 0$ as $r\to\infty$. Since this radial asymptotic behavior of initial value functions comprises models asymptotic to Minkowski in generalized Milne coordinates, self--similar dust solutions and Schwarzschild--Kruskal, models with a simultaneous big bang cannot be asymptotic to any of these spacetimes. This is not surprising since $ct_{\textrm{\tiny{bb}}}\to -\infty$ holds in all these cases.
\end{itemize}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=4in]{config_HP.pdf}
\caption{{\bf Regularity domain of parabolic and hyperbolic models.} The figure displays the $(ct,r)$ plane with the locus of the initial curvature singularity (big bang) $L=0$, marked by the thick curve $[ct_{\textrm{\tiny{bb}}}(r),r]$, together with the initial slice ${}^3{\cal{T}}[t_i]$ (dotted line). For all models asymptotic to a FLRW or Milne spacetimes (panel (a)), $ct_{\textrm{\tiny{bb}}}$ tends asymptotically to a constant, while models asymptotic to Minkowski in generalized Milne coordinates, self--similar solutions or Schwarzschild--Kruskal, we have $ct_{\textrm{\tiny{bb}}}\to-\infty$ in this limit. Notice that space slices ${}^3{\cal{T}}[t]$ are fully regular only for $ct>B$.}
\label{fig1}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=4in]{config_EL.pdf}
\caption{{\bf Regularity domain of open elliptic models.} The figure displays the $(ct,r)$ plane with the locus of the initial and collapse curvature singularities corresponding to $L=0$, marked by the thick curves $[ct_{\textrm{\tiny{bb}}}(r),r]$ and $[ct_{\textrm{\tiny{coll}}}(r),r]$, together with the initial slice ${}^3{\cal{T}}[t_i]$ (dotted line) and the maximal expansion time (${\cal{H}}_q=0$) marked by the thick dotted curve $[ct_{\textrm{\tiny{max}}}(r),r]$. Panel (a) corresponds to models asymptotic to the spatially flat FLRW cosmology, so that $ct_{\textrm{\tiny{bb}}}$ tends asymptotically to a constant (while it tends to $-\infty$ in all other cases). Panel (c) corresponds to the case $k_{qi}/m_{qi}\to 2$ examined in section 11.5, for which $ct_{\textrm{\tiny{max}}}$ tends asymptotically to a constant. Panel (b) corresponds to all other cases. For the collapse time we have $ct_{\textrm{\tiny{coll}}}\to\infty$ in all cases. Notice that in cases (a) and (b) the full asymptotic radial range is contained in the expanding phase ($ct<ct_{\textrm{\tiny{max}}}$). Also, the space slices ${}^3{\cal{T}}[t]$ are fully regular only for $B<ct<C$.}
\label{fig2}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=3in]{LevCur_tmaxPV.pdf}
\caption{{\bf Level curves of $F$ in (\ref{tmce}).} The figure displays the level curves in the $(k_{qi},m_{qi})$ plane. We have $F=0$ in the line $k_{qi} =2m_{qi}$, while $F\to\infty$ as $k_{qi}\to 0$ with $m_{qi}> 0$, while all curves converge at $(k_{qi},m_{qi})\to (0,0)$. As a consequence, only the asymptotic convergence $k_{qi}\to 0$ and $m_{qi}\to 0$ is possible, and so elliptic models with a simultaneous maximal expansion cannot be asymptotic to a FLRW cosmology.}
\label{fig3}
\end{center}
\end{figure}
\subsection{Simultaneous maximal expansion.}
Let $t_{\textrm{\tiny{max}}}=t_{(m)}$ denote the constant time value associated with ${\cal{H}}_q=0$, then (\ref{tmc}) implies that $m_{qi}$ and $k_{qi}$ are necessarily linked by the constraint
\begin{equation}c(t_{(m)}-t_i) = F(k_{qi},m_{qi})\equiv \frac{m_{qi}\,[\pi-Z_e(k_{qi}/m_{qi})]}{k_{qi}^{3/2}},\label{tmce}\end{equation}
where $Z_e$ follows from (\ref{ellZ1a}) and we have expressed $x_i,\,y_i$ in terms of $m_{qi}$ and $k_{qi}$ by means of (\ref{phie}). This constraint is very similar to (\ref{tbbce}), hence the compatibility with assumptions on radial convergence of $m_{qi}$ and $k_{qi}$ follows from a qualitative study of the level curves of the function $F$ in (\ref{tmce}).
As shown in figure 3, the level curves of this function do not allow for $k_{qi}\to 0$ to occur with $m_{qi}\to m_0>0$, since $k_{qi}\to 0$ necessarily implies $m_{qi}\to 0$. Therefore, models with in which $ct_{\textrm{\tiny{max}}}$ is constant must exhibit the same asymptotic states as the models examined in section 11.5 ($x_i=k_{qi}/m_{qi}\to 2$), in which $ct_{\textrm{\tiny{max}}}$ tends to a constant as $r\to \infty$. In fact, the models of section 11.5 asymptotically converge to models with a simultaneous $t_{\textrm{\tiny{max}}}$. Depending on the power law decay of $m_{qi}$ and $k_{qi}$, these models are all asymptotic to Minkowski in generalized Milne coordinates, or to self--similar or Schwarzschild--Kruskal spacetimes.
\section{Summary, conclusion and final discussion.}
We have examined the asymptotic regime along the radial direction for regular parabolic, hyperbolic and open LTB models admitting a symmetry center and a covariant time slicing that defines a one--parameter family of 3--dimensional hypersurfaces, ${}^3{\cal{T}}[t]$, orthogonal to the 4--velocity and marked by constant $t$. With the help of suitably defined covariant quasi--local scalars, $m_q,\,k_q,\,{\cal{H}}_q$ and their relative fluctuations, this time slicing allows us to study LTB models under a well posed initial value framework in terms of a fiducial initial hypersurface ${}^3{\cal{T}}[t_i]$ marked by $t=t_i$. Since radial rays in the ${}^3{\cal{T}}[t]$ are spacelike geodesics of the LTB metric whose affine parameter is proper length, $\ell$, asymptotic conditions naturally correspond to the limit $\ell\to\infty$, which is related to the limit $R\to\infty$ along the radial rays (see section 5). However, $\ell=\ell(t,r)$ and $R=R(t,r)$, and thus it is very difficult to use these invariant quantities to probe the behavior of covariant scalars in this asymptotic limit. Hence, we have provided in section 4 the appropriate conditions in which $\ell\to\infty$ corresponds to the limit $r\to\infty$, so that a regular asymptotic regime follows by demanding that covariant scalars (local and quasi--local) remain bounded in this limit. The asymptotic limit along radial rays then follows from looking at the uniform convergence of initial value functions $m_{qi}(r),\,k_{qi}(r)$ to analytic trial functions $\tilde m_{qi}(r),\,\tilde k_{qi}(r)$ as $r\to\infty$, under the assumption of the radial coordinate gauge (\ref{Rgauge}), which fixes the radial coordinate as proportional to $R_i=R(t_i,r)$ (sections 6 and 7).
\subsection{Classification in terms of their asymptotic limits and states.}
By assuming asymptotic convergence forms $\tilde m_{qi},\,\tilde k_{qi}$ for $m_{qi},\,k_{qi}$, we examined separately in sections 9--12 the radial asymptotic behavior of the metric functions and the covariant scalars, for parabolic, hyperbolic, open elliptic models and models with a simultaneous big bang or maximal expansion, making the distinction between the ``asymptotic limit'' as the strict limit as $r\to\infty$, and the ``asymptotic state'' as the set of asymptotic expansions of all scalars around the asymptotic limit up to leading terms on $\tilde m_{qi}(r),\,\tilde k_{qi}(r)$ (see section 8). The summary of the results is as follows:\\
\noindent
{\bf{Asymptotic Limit.}} All open LTB models can be classified in terms of two broad classes: asymptotic limit to FLRW or Minkowski. Considering $A=m,\,{\cal{H}},\,k$ and $A_q=m_q,\,{\cal{H}}_q,\,k_q$, the ``conventional'' variables $M,\,E$ and $ct_{\textrm{\tiny{bb}}}$ (bang time), as well as $ct_{\textrm{\tiny{coll}}},\,ct_{\textrm{\tiny{max}}}$ (collapse and maximal expansion times, see (\ref{tmc})), together with the metric functions $L$ and $\Gamma$ in (\ref{LTB2}), and using the radial coordinate gauge (\ref{Rgauge}), we have:
\begin{itemize}
\item Models whose asymptotic limit is a FLRW cosmology, including Milne Universe as a particular vacuum case (see figures 1a and 2a):
\numparts\begin{eqnarray} \fl L\to \tilde L(t),\qquad R\to R_0\,r\,\tilde L(t)\to\infty,\qquad \Gamma\to 1,\\
\fl A_q \to \tilde A_q(t),\qquad A\to \tilde A_q(t),\qquad\qquad \delta^{(A)}\to 0,\\
\fl M\sim m_0 R_0^3 r^3,\qquad E\sim \pm k_0 R_0^2 r^2, \qquad ct_{\textrm{\tiny{bb}}}\to \hbox{const.},\\
\fl ct_{\textrm{\tiny{coll}}}\to\infty,\qquad ct_{\textrm{\tiny{max}}}\to\infty\qquad (\hbox{only elliptic models}).\end{eqnarray}\endnumparts
\item Models whose asymptotic limit is Minkowski, containing self--similar, Schwarzschild--Kruskal and generalized Milne asymptotic states (see figures 1b and 2b--2c):
\numparts\begin{eqnarray} \fl L\to 1,\qquad R\to R_0\,r\to\infty,\qquad \Gamma\to 1,\\
\fl A_q \to 0,\qquad A\sim A_q(1+\delta^{(A)})\to 0,\qquad\delta^{(A)}\to \tilde\delta^{(A)}\ne 0\;\;(\hbox{in general}),\\
\fl M\sim \tilde m_{qi}r^3R_0^3,\qquad E\sim \tilde k_{qi} r^2 R_0^2, \qquad ct_{\textrm{\tiny{bb}}}\to -\infty,\\
\fl ct_{\textrm{\tiny{coll}}}\to\infty,\qquad ct_{\textrm{\tiny{max}}}\to\infty\quad (\hbox{or}\;\to \hbox{const})\qquad (\hbox{only elliptic models}).\label{ctmaxcons}\end{eqnarray}\endnumparts
\end{itemize}
where the special case $ct_{\textrm{\tiny{max}}} \to$ const. in (\ref{ctmaxcons}) has been discussed in sections 11.5 and 12.2.
\bigskip
\noindent {\bf{Asymptotic States.}} It is useful to list these states for each kinematic class: \\
\noindent {\bf{Parabolic models}} (section 9) are radially asymptotic to
\begin{itemize}
\item FLRW dust with zero spatial curvature (section 9.1)
\item Minkowski in generalized Milne coordinates (section 9.2)
\item Self--similar dust solution with zero spatial curvature (section 9.2)
\item Schwarzschild--Kuskal in Lema\^\i tre coordinates, built by radial geodesics with zero binding energy (section 9.2)
\end{itemize}
\noindent {\bf{Hyperbolic models}} (section 10) are radially asymptotic to
\begin{itemize}
\item A subclass of parabolic models (section 10.3), which are radially asymptotic to
\begin{itemize}
\item FLRW dust with zero spatial curvature
\item Minkowski in generalized Milne coordinates
\end{itemize}
\item Milne Universe (section 10.4)
\item FLRW dust with negative spatial curvature (section 10.5)
\item Minkowski in generalized Milne coordinates (sections 10.4 and 10.5)
\item Self--similar dust solution with negative spatial curvature (section 10.5)
\item Schwarzschild--Kuskal in coordinates given by radial geodesics with positive binding energy (section 10.4)
\end{itemize}
\noindent {\bf{Open elliptic models}} (section 11) are radially asymptotic to
\begin{itemize}
\item A subclass of parabolic models (section 11.3), which are radially asymptotic to
\begin{itemize}
\item FLRW dust with zero spatial curvature
\item Minkowski in generalized Milne coordinates
\end{itemize}
\item Minkowski in generalized Milne coordinates (sections 11.4 and 11.5)
\item Self--similar dust solution with positive spatial curvature (sections 11.4 and 11.5)
\item Schwarzschild--Kuskal in Novikov coordinates, built by radial geodesics with negative binding energy (sections 11.4 and 11.5)
\end{itemize}
\noindent {\bf{Models with a simultaneous big bang}} (hyperbolic and open elliptic) (section 12.1) are radially asymptotic to
\begin{itemize}
\item FLRW dust with zero spatial curvature (as in sections 10.3, 10.5 and 11.3)
\item FLRW dust with negative spatial curvature (as in section 10.5)
\item Milne Universe (as in section 10.4)
\end{itemize}
\noindent {\bf {Models with a simultaneous maximal expansion}} (elliptic) (section 12.2) are radially asymptotic to the same spacetimes as those of section 11.5.
\subsection{Cosmological considerations vs radial asymptotics.}
It is usual to consider any given LTB configuration as a model of a spherical inhomogeneity somehow immersed in a cosmic ``background''. In order to discuss this issue in terms of the results obtained in this article, it is illustrative to relate the asymptotic forms of LTB scalars with observational parameters of a FLRW cosmology.
If we denote by $\Omega(t)$ the FLRW Omega parameter, a possible generalization for LTB models can be given in terms of quasi--local scalars:
\begin{equation} \hat\Omega \equiv \frac{\kappa\,\rho_q}{3{\cal{H}}_q^2}=\frac{2m_q}{{\cal{H}}_q^2}=\frac{2m_q}{2m_q-k_q}=\frac{2m_{qi}}{2m_{qi}-Lk_{qi}},\end{equation}
so that each kinematic class (parabolic, hyperbolic or elliptic) follows from the sign of $\hat\Omega-1$
\begin{equation} \hat\Omega-1=\frac{k_q}{{\cal{H}}_q^2}=\frac{k_{qi} L}{2m_{qi}-Lk_{qi}}. \end{equation}
While $\hat\Omega$ is a covariant quantity (because $m_q$ and ${\cal{H}}_q$ are covariant), it is not a quasi--local scalar: notice that we do not use the symbol $\Omega_q$ because $\hat\Omega$ cannot be obtained from applying (\ref{aveq_def}) to the ratio $2m/{\cal{H}}^2$. In fact, other expressions for the Omega and Hubble factors for LTB models have been suggested in the literature \cite{moftar,HMM} (see \cite{suss10} for further discussion on this issue). However, it is evident that for all LTB models whose radial asymptotic limits and states are FLRW cosmologies we have as $r\to\infty$
\begin{equation} \hat\Omega \sim \Omega(t), \end{equation}
which clearly suggests that LTB models with this radial asymptotic behavior are suitable to model cosmological inhomogeneities asymptotically converging to a FLRW ``background''. This means that, irrespective of how we choose to generalize the observational parameters, they will tend asymptotically in the radial direction to the FLRW parameters. Notice that we have reached this conclusion by looking at the asymptotic states of LTB models defined all along the radial rays, without considering the rather artificial situation in which a FLRW radial asymptotic state of an LTB model is forced by matching to it to a FLRW cosmology at a finite fixed comoving boundary $r=r_b$. Though such configurations can also be constructed with LTB models, leading to smooth and fully relativistic generalizations of the Newtonian ``spherical collapse'' or ``top hat'' models used in qualitative studies of structure formation (see \cite{suss09} for discussion and references on this type of models and their relativistic generalizations).
The cosmological interpretation of the radial asymptotics of LTB models with asymptotic limit to Minkowski is more complicated. To discuss these cases it is convenient to introduce a ``local' equivalent of $\hat\Omega$ defined with the local scalars $m=\kappa\rho/3$ and ${\cal{H}}=\Theta/3$ as
\begin{equation}\fl\hat\Omega_{\rm{loc}}\equiv \frac{\kappa\rho}{3{\cal{H}}^2}=\frac{2m}{2m-k+({\cal{H}}_q\delta^{(\HH)})^2}=1+\frac{(\hat\Omega-1)\,(1+\delta^{(k)})-(\delta^{(\HH)})^2}{(1+\delta^{(\HH)})^2},\end{equation}
where we eliminated ${\cal{H}}$ and $k$ from (\ref{qltransf}) and used the constraint
\begin{equation} {\cal{H}}^2 = 2m-k+\Sigma^2=2m-k+({\cal{H}}_q\delta^{(\HH)})^2, \end{equation}
which follows from the definitions of $\Theta$ and $\Sigma$ in (\ref{ThetaRR}) and (\ref{SigEE1}). For models with an asymptotic FLRW limit, $\delta^{(\HH)}\to 0$ and $\delta^{(k)}\to 0$ hold, so $\hat\Omega_{\rm{loc}}\to \hat\Omega$ and the FLRW $\Omega(t)$ results. However, in models with a Minkowski limit $\delta^{(\HH)},\,\delta^{(k)}$ do not tend to zero, resulting in a different form for $\hat\Omega_{\rm{loc}}$.
In parabolic models with asymptotic limit to Minkwoski we have $\hat\Omega =1$ everywhere, while $\delta^{(\HH)}\to\delta_i^{(\HH)}=\delta_i^{(m)}/2$ (from (\ref{asPar3d})), hence
\begin{equation} \hat\Omega_{\rm{loc}}\approx 1-\left(\frac{\delta_i^{(\HH)}}{1-\delta_i^{(\HH)}}\right)^2 \leq 1,\label{Omlocp}\end{equation}
Since we have $\hat\Omega\to 1$ for hyperbolic with a matter dominated Minkowski limit (section 10.4) and for all open elliptic models with a Minkowski limit (which is also matter dominated), we obtain for these cases the same limit as in (\ref{Omlocp}). For all hyperbolic models with a vacuum dominated Minkowski limit we have $\hat\Omega\to 0$, hence
\begin{equation} \hat\Omega_{\rm{loc}}\approx \frac{2\delta^{(\HH)}-\delta^{(k)}}{(1+\delta^{(\HH)})^2}.\end{equation}
Therefore, considering (\ref{asHyp23d})--(\ref{asHyp23e}), (\ref{asHyp33c})--(\ref{asHyp33d}) and (\ref{asHyp54c})--(\ref{asHyp54d}), we have $\delta^{(k)}\to 2\delta^{(\HH)}$, and so $\hat\Omega_{\rm{loc}}\to 0$ holds for models asymptotic to Milne, generalized Milne and self--similar solutions, but in models asymptotic to Schwarzschild--Kruskal we have $\delta^{(k)}\to \delta^{(\HH)}$ (see (\ref{asHyp43d})--(\ref{asHyp43e})), so for these models $\hat\Omega_{\rm{loc}}$ has an asymptotic form as in (\ref{Omlocp}).
Since the asymptotic forms for either $\hat\Omega$ or $\hat\Omega_{\rm{loc}}$ are independent of $t$ (being either zero or a constant $\leq 1$), a cosmological interpretation for the radial asymptotics of models with a Minkowski limit cannot be given in terms of a smooth convergence or transition to a FLRW model in the radial direction. Instead of a cosmic FLRW background, the external realm of these LTB models could be a large void region. A similar and appealing interpretation of radial asymptotics to Minkowski follows naturally from the notion of ``finite infinite'' (fi), introduced by Ellis \cite{fiEllis} (see also comments in \cite{fiWilt}) to represent a near Minkowskian timelike envelope around bound structures, which defines a scale in which these structures can be studied as almost asymptotically flat systems without considering the effect of cosmic expansion. Such a scale would be intermediate between characteristic lengths of bound structures and a cosmic scale where expansion cannot be ignored. Evidently, LTB models with asymptotic limit to Minkowski approximate spherically symmetric realizations of this cosmic structure in such a near Minkowski envelope.
The results of this article are useful for probing supernovae and CMB observations by means of LTB models. They are essential for the study of radial profiles of covariant scalars, namely: the ``clump'' or ``void'' profiles and the possibility of profile inversions as the models evolve in time. This study has been undertaken in a separate article \cite{rprofs}. The radial asymptotic properties of the models have also theoretical and practical consequences in the application of Buchert's scalar averaging formalism \cite{ave_review} to LTB models, as vacuum dominated LTB models are the LTB configurations most likely to yield an ``effective'' acceleration that mimics dark energy emerging from back--reaction terms in the context this formalism. These connections have been already remarked \cite{LTBave2,LTBave3} and are currently under further investigation.
\begin{appendix}
\section{Analytic solutions in the conventional variables.}
The solutions of the Friedman--like equation (\ref{fieldeq1}) in terms of the conventinal free functions $M,\,E,\,ct_{\textrm{\tiny{bb}}}$ are:
\begin{itemize}
\item{Parabolic models or regions: \,\, $E=0$.}
\begin{equation}c(t-t_{\textrm{\tiny{bb}}}) = \frac{2}{3}\,\eta^3,\qquad R=(2M)^{1/3}\,\eta^2,\label{par1}\end{equation}
\item{Hyperbolic models or regions: \,\, $E\geq 0$.}
\begin{eqnarray}
\fl R =\frac{M}{E}\,\left(\cosh\,\eta-1\right),\qquad c(t-t_{\textrm{\tiny{bb}}})=\frac{M}{E^{3/2}}\,\left(\sinh\,\eta-\eta\right)\label{hypRt}\\
\fl \frac{E^{3/2}}{M}\,c(t_i-t_{\textrm{\tiny{bb}}})=Z_h(U),\qquad U\equiv \frac{ER}{M},\label{hyptR}
\end{eqnarray}
\item{Elliptic models or regions: \,\, $E\leq 0$.}
\begin{eqnarray}
\fl R =\frac{M}{|E|}\,\left(1-\cos\,\eta\right),\qquad c(t-t_{\textrm{\tiny{bb}}})=\frac{M}{|E|^{3/2}}\,\left(\eta-\sin\,\eta\right),\label{ellRt}\\
\fl \frac{|E|^{3/2}}{M}\,c(t_i-t_{\textrm{\tiny{bb}}})=\left\{ \begin{array}{l}
Z_e(U)\quad \hbox{expanding phase}\quad \dot R> 0 \\
2\pi-Z_e(U) \quad \hbox{collapsing phase}\quad \dot R< 0 \\
\end{array} \right.,\quad U\equiv \frac{|E|R}{M},\nonumber\\\label{elltR}
\end{eqnarray}
\end{itemize}
where $Z_h$ and $Z_e$ are respectively given by (\ref{hypZ1a}) and (\ref{ellZ1a}). The solutions presented above are given in terms of the new variables by (\ref{par2}), (\ref{phi_L_h}) and (\ref{ellZ2}).
Another particular solution of (\ref{fieldeq1}) follows by assuming $M=0$ and $E>0$ (but otherwise arbitrary), leading to the case ``[s2]'' in \cite{ltbstuff}
\begin{equation} R = E\,c(t-t_{\textrm{\tiny{bb}}}),\label{s2case}\end{equation}
which are locally Minkowski solutions (sections of Minkowski spacetime in coordinates that generalize Milne's Universe).
\section{Prescribing local initial value functions.}
In section 7 we assumed an asymptotic convergence for for $A_{qi}$ and obtained the forms for $A_i$ and $\delta^{(A)}_i$. We examine here the opposite situation. If we assume that $A_i\sim \tilde A_i$, then for $r>y$ and with the help from (\ref{aveq_def}) and (\ref{Rgauge}) we have
\begin{equation} A_{qi}=\frac{3}{r^3}\int_0^r{A_i x^2 {\rm{d}} x}\sim \frac{3}{r^3}\left[\int_0^y{A_i x^2 {\rm{d}} x}+\int_y^r{\tilde A_i x^2 {\rm{d}} x}\right],\label{Aqsim}\end{equation}
while $\delta^{(A)}_i$ follows directly from (\ref{Dadef}):
\begin{equation} \delta^{(A)}_i \sim \frac{\tilde A_i(r) r^3-3\int_0^y{A_i x^2 {\rm{d}} x}-3\int_y^r{\tilde A_i x^2 {\rm{d}} x}}{3\int_0^y{A_i x^2 {\rm{d}} x}+3\int_y^r{\tilde A_i x^2 {\rm{d}} x}}.\label{Daas}\end{equation}
A comparison of (\ref{Asim})--(\ref{Daasq}) and (\ref{Aqsim})--(\ref{Daas})shows that it is much easier to obtain the asymptotic limit of $\delta^{(A)}_i$ from $A_{qi}$ than from $A_i$. Prescribing $A_{qi}$ is also more useful in practice, since the initial value functions in the analytic solutions given in sections 9--11 are $m_{qi},\,k_{qi}$.
We now examine the asymptotic forms for $m_{qi}$ and $\delta_i^{(m)}$ that follow given a prescribed asymptotic form for $m_i$. Consider a power law decay $m_i\sim m_0 r^{-\alpha}$, then (\ref{Aqsim}) and (\ref{Daas}) yield
\begin{eqnarray}\fl m_{qi}\sim m_0\left[\frac{I(y)}{r^3}+\frac{3}{(3-\alpha)\,r^\alpha}\right],\qquad \delta_i^{(m)}\sim \frac{-\alpha\,r^{-\alpha}-(3-\alpha)I(y)r^{-3}}{(3-\alpha)I(y)r^{-3}+3\,r^{-\alpha}},\label{AqDa}\\
\hbox{where:}\qquad I(y)=\frac{3}{m_0}\int_0^y{m_i x^2 {\rm{d}} x}-\frac{3}{3-\alpha}y^{3-\alpha},\label{Idef}\nonumber\end{eqnarray}
Notice that if $\alpha<3$ then $r^{-\alpha}$ is the dominant term, whereas if $\alpha>3$ then $r^{-3}$ is dominant. Hence, we obtain $m_{qi}\propto r^{-\alpha}$ for $\alpha<3$ and for a logarithmic decay, while $m_{qi}\propto r^{-3}$ results if we assume that $\alpha\geq 3$ or an exponential decay. However, from Lemma 3, both $m_i$ and $m_{qi}$ tend to zero. By assuming $r\gg y$ we obtain the following asymptotic limit for $\delta_i^{(m)}$
\begin{equation} \delta_i^{(m)} \sim \tilde\delta_i^{(m)} =\left\{ \begin{array}{l}
- \alpha /3,\quad \hbox{if}\quad \alpha < 3 \\
- 1\quad \quad \;\;\hbox{if}\quad \alpha \ge 3 \\
\end{array} \right.,\label{Da00}\end{equation}
It is straightforward to show that if $m_i$ decays to zero exponentially, then $\tilde\delta_i^{(m)}\to-1$, while a slow logarithmic decay yields $\tilde\delta_i^{(m)}\to 0$. As opposed to prescribing an asymptotic decay for $m_{qi}\to 0$ as in (\ref{decmq2})--(\ref{decmq3}), all assumptions on the decay of $m_i\to 0$ necessarily yield $m_i$ and $m_{qi}$ positive in the full range $r>y$. This is so because $m_{qi}\geq m_i\geq 0$ holds in this range (as $m'_i$ and $m'_{qi}$ are negative, see (\ref{propq2})).
\section{Particular cases of open LTB models.}
Given the radial coordinate gauge (\ref{Rgauge}), all open LTB models can be univocally characterized by the initial value functions $m_{qi},\,k_{qi}$. Hence, spacetimes that emerge as particular cases of these models emerge by specific specialization of these functions. The line elements of these spacetimes follow directly from (\ref{LTB2}) by recalling the particular cases of these functions accordingly, and define particular case spacetimes which have served as references to the possible asymptotic states of all open LTB models in the limit $r\to\infty$.
\subsection{FLRW dust spacetimes}
Homogeneous and isotropic dust FLRW cosmologies follow as particular cases of LTB models by setting
\begin{equation} m_{qi} = m_0>0\quad\hbox{and}
\quad k_{qi}=k_0,\label{FLRW}\end{equation}
where $m_0,\,k_0$ are constants. From (\ref{ME}), (\ref{tbbpar}), (\ref{tbbh}) and (\ref{tmc}), we have
\begin{equation} \fl M=m_0\,r^3,\quad E=-k_0\,r^2,\quad t_{\textrm{\tiny{bb}}},\,t_{\textrm{\tiny{max}}},\,t_{\textrm{\tiny{coll}}}\quad\hbox{constants}.\label{MEtbb1}\end{equation}
Evidently, the spatially flat case corresponds to $k_0=0$, while positive/negative $k_0$ are FLRW models with positive/negative spatial curvature. It is evident from (\ref{par2}) that (\ref{FLRW}) yields for $k_0=0$
\begin{equation} L=L(t) = \left[1+\frac{3}{2}\sqrt{2m_0}\,c(t-t_i)\right]^{2/3},\qquad k_0=0,\label{FLRW_L0}\end{equation}
while for $k_0\ne 0$ the function $L=L(t)$ follows from the implicit solutions (\ref{phi_L_h}) and (\ref{ellZ2}) under the specialization (\ref{FLRW}):
\begin{equation} \fl \phi(t)=y_0\,c(t-t_i)+Z(x_0) = Z(x_0\,L), \qquad x_0=\frac{|k_0|}{m_0},\quad y_0=\frac{|k_0|^{3/2}}{m_0},\label{FLRW_L}\end{equation}
with $Z=Z_h$ or $Z=Z_e$ given by (\ref{hypZ1a}) and (\ref{ellZ1a}) for hyperbolic (negative curvature) and elliptic (positive curvature) FLRW models. For all particular cases (\ref{FLRW}), equations (\ref{Dadef}) and (\ref{qltransf}) imply $\delta_i^{(m)}=\delta_i^{(k)}=\delta_i^{(\HH)}=0$, so that $m_i=m_{qi},\,k_i=k_{qi}$ and ${\cal{H}}_i={\cal{H}}_{qi}$ hold. Also, (\ref{Hq}), (\ref{Gp}) and (\ref{Ghe}) imply $\Gamma=1$, and thus (\ref{slawDm}) and (\ref{slawDk}) yield $\delta^{(m)}=\delta^{(k)}=0$, and so, $\delta^{(\HH)}=0,\,m=m_q,\,k=k_q$ and ${\cal{H}}={\cal{H}}_q$ follow from (\ref{slaw1}), (\ref{slaw2}) and (\ref{slawDh}). As a consequence, (\ref{LTB2}) becomes a FLRW line element and all scalars are either constants or functions of $t$ only.
\subsection{Minkowskian particular cases.}
Considering the Riemann tensor (see Appendix A2 of \cite{suss10}), any particular LTB model in which $m_q=m=0$ is locally equivalent to Minkowski spacetime given in curvilinear coordinates that generalize the well known Milne Universe. These are the locally Minkowskian solutions (\ref{s2case}), corresponding to the case ``[s2]'' in the classification of \cite{ltbstuff}, and are given by the specialization:
\begin{equation} m_{qi}=0 \quad\hbox{and}\quad k_{qi}<0\quad\hbox{but otherwise arbitrary.}\label{Mink}\end{equation}
so that $m_q=m=\delta^{(m)}=0$, but $\delta^{(k)}\ne 0,\,\delta^{(\HH)}\ne 0$. The metric functions $L$ and $\Gamma$ follow from substituting (\ref{Mink}) into the equation for $\dot L$ in (\ref{Hq}) and substituting in the definition of $\Gamma$ in (\ref{LGdef}):
\begin{equation}\fl L = 1+|k_{qi}|^{1/2}\,c(t-t_i),\qquad \Gamma= \frac{1+|k_{qi}|^{1/2}\left(1+\frac{3}{2}\delta_i^{(k)}\right)\,c(t-t_i)}{1+|k_{qi}|^{1/2}\,c(t-t_i)}\label{Mink_L}\end{equation}
where we used (\ref{Dadef}) and the coordinate gauge (\ref{Rgauge}). Evidently, we have $M=0$, while $E$ is arbitrary as in (\ref{s2case}). The scalars associated with spatial curvature and expansion correspond to hypersurfaces ${}^3{\cal{T}}$ and a 4--velocity associated with test observers. In general, these coordinates do not cover the full Minkowski spacetime, and thus $L=0$, marked by $ct=ct_{\textrm{\tiny{bb}}}(r)=ct_i-1/|k_{qi}|^{1/2}$, corresponds to a caustic surface for this congruence, not to a curvature singularity. Milne's Universe, which can also be considered as a vacuum FLRW cosmology, is the particular case of (\ref{Mink}) and (\ref{Mink_L}) given by
\begin{eqnarray} m_{qi}=0 \quad\hbox{and}\quad |k_{qi}|=k_0=\hbox{constant},\label{Milne}\\
L = L(t)= 1+\sqrt{k_0}\,c(t-t_i),\label{Milne_L}\\
M=0,\quad E={\cal{F}}^2-1=|k_0|\,r^2,\quad t_{\textrm{\tiny{bb}}}=\hbox{constant}.\label{MEtbb2}\end{eqnarray}
Since $\delta_i^{(k)}=0$ in this case vanish, then $\Gamma=1$ and all the $\delta^{(A)}$ also vanish. Also, we have $m_q=m=0$ for all $t$, though $k_q=k<0$ and ${\cal{H}}_q={\cal{H}}$ are functions of time. In the general case (\ref{Mink}) $k_q=k<0$ and ${\cal{H}}_q={\cal{H}}$ are also functions of $r$.
\subsection{Schwarzschild--Kruskal spacetime.}
The Schwarzschild--Kruskal spacetime in comoving coordinates constructed by radial geodesics (Lema\^\i tre and Novikov coordinates, see page 332 of \cite{kras2}) follows by setting
\begin{equation} m_{qi}=m_0\,r^{-3}\quad\hbox{and}\quad k_{qi}\:\:\hbox{arbitrary},\label{Schw}\end{equation}
Hence, (\ref{ME}) and (\ref{Rgauge}) imply
\begin{equation} M=M_0=m_0\,R_0^3,\qquad E={\cal{F}}^2-1\:\:\hbox{arbitrary},\label{MEtbb3}\end{equation}
while (\ref{Dadef}) and (\ref{slawDm}) yield $\delta_i^{(m)}=\delta^{(m)}=-1$, so that $2m_q=2M_0/R^3$ and $m_i=m=0$ follows from (\ref{qltransf}) and (\ref{slaw1}). It is straightforward to prove that the Riemann tensor (see Appendix A2 of \cite{suss10}) with $m=0$ and $m_q\ne 0$ leads to a vanishing Ricci tensor, and thus we can identify this case as a vacuum Schwarzschild--Kruskal solution where $2M_0$ is the Schwarzschild radius. The variables associated with spatial curvature ($k,\,k_q$ and $\delta^{(k)}$) and the expansion (${\cal{H}},\,{\cal{H}}_q$ and $\delta^{(\HH)}$) remain arbitrary because a spatial curvature and an expansion can always be associated to a congruence of radial geodesic test observers with arbitrary binding energy. In fact, by setting $M=M_0$ in (\ref{fieldeq1}) we obtain the equation for radial timelike geodesics in a Schwarzschild spacetime if we identify ${\cal{F}}^2-1$ with the binding energy of the comoving test observers.
The form of $L$ depends on the choice of $k_{qi}$. If we choose geodesic observers with zero binding energy $k_{qi}=0$, then $L$ and $t_{\textrm{\tiny{bb}}}$ follow from (\ref{par2}) and (\ref{tbbpar}) with $m_{qi}$ given by (\ref{Schw}):
\begin{equation}\fl L(t,r)=\left[1+\frac{3}{2}\sqrt{2m_0}\,r^{-3/2}\,c(t-t_i)\right]^{2/3},\qquad ct_{\textrm{\tiny{bb}}}= ct_i- \frac{2\,r^{3/2}}{3\sqrt{2m_0}}.\label{Schw_L0}\end{equation}
If we choose $k_{qi}\ne 0$, then $L$ follows from the implicit solutions (\ref{phi_L_h}) and (\ref{ellZ2}) under the specialization (\ref{Schw}):
\begin{equation} \fl \phi(t,r)=y_i\,c(t-t_i)+Z(x_i) = Z(x_i\,L), \qquad x_i=\frac{k_{qi}\,r^3}{m_0},\quad y_i=\frac{k_{qi}^{3/2}r^3}{m_0}.\label{Schw_L}\end{equation}
where $Z$ is either one of the functions (\ref{hypZ1a}) or (\ref{ellZ1a}). Notice that $t_{\textrm{\tiny{bb}}}\to-\infty$ as $r\to\infty$ in (\ref{Schw_L0}), whereas the form of $t_{\textrm{\tiny{bb}}}$ (and $t_{\textrm{\tiny{max}}},\,t_{\textrm{\tiny{coll}}}$ if $k_{qi}> 0$) depend on the choice of $k_{qi}$. As shown in sections 10 and 11, if $k_{qi}^{3/2}r^3\to 0$, then $L$ tends asymptotically to (\ref{Schw_L0}). We also remark that the Schwarzschild--Kruskal spacetimer lacks a regular symmetry center, hence $r=0$ is not a special worldline. Also, the comoving coordinates in the cases $k_{qi}\leq 0$ do not cover the full Schwarzschild-Kruskal manifold.
\subsection{Self--similar solutions.}
The self--similar LTB solutions have been popular models to study gravitational collapse \cite{sscoll}. They have been classified in \cite{selfsim}. In terms of the parameters used in this article these solutions follow as the particular case
\begin{equation} m_{qi}=m_0\,r^{-2}\quad\hbox{and}\quad k_{qi}=k_0\,r^{-2},\label{SS}\end{equation}
where $m_0>0$ and $k_0$ are constants. The form of $L$ follows from (\ref{par2}) and from the implicit solutions (\ref{phi_L_h}) and (\ref{ellZ2}) under the specialization (\ref{SS}):
\begin{eqnarray}\fl L=\left[1+\frac{3}{2}\sqrt{2m_0}\,\zeta\right]^{2/3},\qquad k_0=0\label{SS_L0}\\
\fl \phi(t,r)=y_0\,\zeta+Z(x_0) = Z(x_0\,L), \qquad x_0=\frac{k_0}{m_0},\quad y_0=\frac{k_0^{3/2}}{m_0},\quad k_0\ne 0\label{SS_L}\end{eqnarray}
where $Z=Z_h$ or $Z=Z_e$ given by (\ref{hypZ1a}) and (\ref{ellZ1a}) for hyperbolic and elliptic models, and we can identify the self similar variable
\begin{equation} \zeta \equiv \frac{c(t-t_i)}{r}.\label{SSvar}\end{equation}
The functions $M,\,E$ and $t_{\textrm{\tiny{bb}}}$ follow from applying (\ref{SS}) to (\ref{ME}), (\ref{tbbpar}) and (\ref{tbbh})
\begin{eqnarray} \fl M=m_0 R_0^3\,r,\qquad E={\cal{F}}^2-1 =-k_0\,R_0^2,\label{MEtbb41}\\
\fl ct_{\textrm{\tiny{bb}}}=ct_i-\frac{2\,r}{3\,\sqrt{2m_0}},\quad k_0=0,\qquad ct_{\textrm{\tiny{bb}}}=ct_i-\frac{r}{y_0}\,Z(x_0),\quad k_0\ne 0.\label{MEtbb42} \end{eqnarray}
As in the Schwarzschild--Kruskal case, we have $t_{\textrm{\tiny{bb}}}\to-\infty$ as $r\to\infty$ (see section 13). As a consequence of (\ref{SS}) we have (in general) $\delta_i^{(m)}=\delta_i^{(k)}=-2/3$. The forms for $m_q,\,{\cal{H}}_q,\,k_q,\,\Gamma,\,\delta^{(m)},\,\delta^{(k)},\,\delta^{(\HH)}$ and the remaining scalars follow by applying (\ref{SS}) to the appropriate equations. It is straightforward to verify that the metric (\ref{LTB2}) takes the appropriate self--similar form, admitting the homothetic vector $z^a=c(t-t_i)\delta^a_0+r\delta^a_r$. Also, all dimensionless variables ({\it{i.e.}} of the form $A/A_i$, as well as $L,\,\Gamma$ and ${\cal{H}}_q\,c(t-t_i)$) depend only on the self--similar variable.
\end{appendix}
\section*{References}
|
\section{Introduction}
Understanding how bulges form is of great importance to develop a
complete picture of galaxy formation. The processes by which bulges
form are still debated. On the one hand, the merger of dwarf-sized
galactic subunits has been suggested as the main path for bulge
formation \citep{kau_etal_93}, which is supported by the relatively
homogeneous bulge stellar populations of the Milky Way and M31
\citep{zoc_etal_03}. Bulges formed in such mergers are termed
'classical' bulges. Alternatively, bulges may form via internal
`secular' processes such as bar-driven gas inflows, bending
instabilities, clump instabilities, etc. \citep{com_san_81,
deb_etal_06}. Evidence for secular bulge formation includes the
near-exponential bulge light profiles \citep{and_san_94}, a
correlation between bulge and disk scale lengths \citep{mac_etal_03,
men_etal_08}, the similar colors of bulges and inner disks
\citep{pel_bal_96}, substantial bulge rotation \citep{kor_ken_04}, and
the presence of box/peanut (B/P)-shaped bulges in $\sim45$\% of
edge-on disk galaxies \citep{lut_etal_00}. A review of secular
`pseudo-bulge' formation and evidence for it can be found in
\citet{kor_ken_04}. Standard cold dark matter cosmology predicts that
galaxies without classical bulges should be rare \citep{don_bur_04}.
Not only are they not rare in nature, but the formation of
pseudo-bulges means that some fraction of bulged galaxies lack a
classical bulge, exacerbating the disagreement between theory and
observations \citep{deb_etal_06}. It is therefore important to
determine which bulges are of the classical versus pseudo variety, and
which are a mix of both.
Of primary concern here, several pieces of evidence point to the
identification of most B/P bulges in edge-on spiral galaxies with the
bars of barred spirals. N-body simulations show that barred
galaxies have a tendency to develop B/P bulges. The observed
incidence of B/P bulges is consistent with that expected if they are
associated with relatively strong bars. A recent work by
\citet{lut_etal_00} demonstrated that $45$\% of
all bulges are B/P, while amongst those the shape of the bulge
depends mainly on the viewing angle to the bar. As shown by the
numerical simulations, true peanuts are bars seen side-on, that is,
with the major-axis of the bar perpendicular to the line of sight.
For less-favourable viewing angles, the bulge/bar looks boxy, and if
the bar is seen end-on it looks almost spherical.
The presence of bars in edge-on galaxies with B/P bulges has been
observationally established by the kinematics of gas and stars
\citep[see][and references therein]{bur_ath_05}. The kinematics of
discs harbouring a B/P bulge, as measured from both ionized-gas
emission lines and stellar absorption lines, show the behaviour
expected of barred spirals viewed edge-on. However, the degeneracy
inherent in deprojecting edge-on galaxies makes it difficult to study
other properties of the host galaxy. For example, simulations show
that a bar can produce a B/P shape even if a massive classical bulge
formed before the disk \citep{deb_etal_05}.
Understanding the relative importance of classical and pseudo-bulges
requires an attempt at a cleaner separation of bulges, bars and
peanuts, which is easiest to accomplish in less inclined systems.
Recently, \citet{deb_etal_05} proposed a kinematic diagnostic of B/P
bulges in face-on ($i<30^{\circ}$) galaxies, namely a double minimum
in the fourth-order Gauss-Hermite moment, $h_4$, along the major-axis
of the bar. These minima occur because at the location of the B/P
shape, the vertical density distribution of stars becomes broader,
which leads to a double minimum in $z_4$, the fourth-order
Gauss-Hermite moment of the vertical density distribution. The
kinematic moment $h_4$ is then found to be an excellent proxy for the
unobservable $z_4$. In contrast, the increase in the vertical
scale-height does not produce any distinct signature of a B/P bulge
and the vertical velocity dispersion, $\sigma_z$, is too strongly
dependent on the radial density variation to provide a useful B/P
bulge diagnostic. \citet{deb_etal_06} showed that the diagnostic
continues to hold even when gas is present since this sinks to a
radius smaller than that of the B/P bulge.
This new kinematic diagnostic has been confirmed observationally by
\citet{mendezabreu_08}. Their main results are given here.
\begin{figure*}[!t]
\begin{center}$
\begin{array}{ccc}
\includegraphics[width=4cm, bb=110 360 410 1105]{mendezabreu_f01a.ps} &
\includegraphics[width=4cm, bb=110 360 410 1105]{mendezabreu_f01b.ps} &
\includegraphics[width=4cm, bb=110 360 410 1105]{mendezabreu_f01c.ps}
\end{array}$
\end{center}
\caption{\footnotesize
Morphology and stellar kinematics of NGC~98 (left panels), NGC~600
(central panels), and NGC~1703 (right panels).
For each galaxy the top panel shows the VLT/FORS2 $R-$band image.
The slit position and image orientation are indicated. The inset
shows the portion of the galaxy image marked with a white box. The
gray scale and isophotes were chosen to enhance the features
observed in the central regions.
The remaining panels show from top to bottom the profiles of surface
density, ${\cal B}(0.7,0.9)$, $h_4$, $h_3$, $\sigma$, and mean
velocity after the subtraction of the systemic velocity, $v$. The
two vertical lines indicate the location of the $h_4$ minima in
NGC~98.}
\label{fig:kine}
\end{figure*}
\section{Observations, data-reduction and analysis}
The barred galaxies NGC~98 and NGC~600 were selected as bright and
undisturbed galaxies with a low inclination, a large bar, and modest
extinction in the bar. The nearly face-on unbarred galaxy NGC~1703 was
added as a control object.
The spectroscopic observations were carried out in service mode at the
Very Large Telescope (VLT) at the European Southern Observatory (ESO)
using The Focal Reducer Low Dispersion Spectrograph 2 (FORS2).
The spectra were taken along the bar major axis of NGC~98 and NGC~600
and along the disk major axis of NGC~1703. The total integration time
for each galaxy was 3 hours, in four exposures of 45 minutes each.
Using standard IRAF routines, all the spectra were bias subtracted,
flat-field corrected, cleaned of cosmic rays, corrected for bad
pixels, and wavelength calibrated. The instrumental resolution was
$1.84\pm0.01$ \AA\ (FWHM) corresponding to $\sigma_{\it inst} = 27$ km
s$^{-1}$ at 8552 \AA .
\subsection{Photometry}
We analyzed the uncalibrated acquisition images from the VLT to derive
the photometric properties of the sample galaxies.
Isophote-fitting with ellipses, after masking foreground stars and bad
pixels, was carried out using the IRAF task ELLIPSE. Under the
assumption that the outer disks are circular, their inclination was
determined by averaging the outer isophotes. All 3 galaxies have $i <
30^{\circ}$.
The semi-major axis length, $a_{\rm B}$ of the two bars was measured
from a Fourier decomposition as in \citet{agu_etal_00}.
\subsection{Kinematics}
The stellar kinematics of the three galaxies were measured from the
absorption features present in the wavelength range centered on the
CaII triplet ($\lambda\lambda\,8498,\,8542,\,8662\,$\AA ) using the
Penalized Pixel Fitting method \citep[pPXF,][]{cap_ems_04}. The
spectra were rebinned along the dispersion direction to a logarithmic
scale, and along the spatial direction to obtain a signal-to-noise
ratio $S/N \ga 40$ \AA$^{-1}$.
A linear combination of the template stellar spectra, convolved with
the line-of-sight velocity distribution (LOSVD) described by a
Gauss-Hermite expansion \citep{gerhar_93, vdm_fra_93} was fitted to
each galaxy spectrum by $\chi^2$ minimization in pixel space. This
allowed us to derive profiles of the mean velocity ($v$), velocity
dispersion ($\sigma$), third ($h_3$), and fourth-order ($h_4$)
Gauss-Hermite moments. The uncertainties on the kinematic parameters
were estimated by Monte Carlo simulations with photon, read out and
sky noise.
Extensive testing on simulated galaxy spectra was performed to provide
an estimate of the biases of the pPXF method with the adopted
instrumental setup and spectral sampling. The simulated spectra were
obtained by convolving the template spectra with a LOSVD parametrized
as a Gauss-Hermite series and measured as if they were real. No bias
was found in the ranges of $S/N$ and $\sigma$ which characterize the
spectra of the sample galaxies. The values of $h_3$ and $h_4$ measured
for the simulated spectra differ from the intrinsic ones only within
the measured errors \citep[see also][]{ems_etal_04}.
\section{Conclusions}
We have identified for the first time a B/P-shaped bulge in a
face-on galaxy. In the unbarred galaxy NGC~1703 we had not expected
to find a B/P bulge and we included it in our sample as a control.
The failure to find a double-minimum in $h_4$ is therefore fully
consistent with previous results \citep{chu_bur_04}. Of the two
barred galaxies, NGC~98 has clear evidence of a B/P bulge while
NGC~600 does not. The absence of a B/P shape in this galaxy is not
surprising since it appears to not have a bulge.
If we identify the radius of the B/P bulge, $r_{\rm bp}$, with the
location of the minimum in $h_4$, as found in simulations
\citep{deb_etal_05}, then we find $r_{\rm bp} = 0.35$ \mbox{$a_{\rm
B}$}, where \mbox{$a_{\rm B}$} is the bar semi-major axis.
Similarly, \citet{kor_ken_04} noted that the maximum radius of the
boxy bulge is about one-third of the bar radius. Simulations also
produce B/P-bulges which are generally smaller than the bar . Since
B/P bulges are supported by resonant orbits, it would be very
instructive to measure this ratio for a sample of barred galaxies.
\bibliographystyle{aa}
|
\section{Introduction}
The purpose of this article is to investigate some applications of
commutator subgroups to homotopy groups and geometric groups.
Recall ~\cite[p. 288-289]{MKS} that \textit{a bracket arrangement of weight $n$} in a group $G$ is a map
$\beta^n\colon G^n\to G$ which is defined inductively as follows:
$$\beta^1=\mathrm{id}_G,\,\, \beta^2(a_1,a_2)=[a_1,a_2]$$
for any $a_1,a_2\in G$, where $[a_1,a_2]=a_1^{-1}a_2^{-1}a_1a_2$. Suppose that the bracket arrangements of weight $k$ are
defined for $1\leq k<n$ with $n\geq 3$. A map $\beta^n\colon G^n\to G$ is called
a bracket arrangement of weight $n$ if $\beta^n$ is the composite
\begin{diagram}
G^n=G^k\times G^{n-k}&\rTo^{\beta^k\times\beta^{n-k}}&G\times G&\rTo^{\beta^2}&G
\end{diagram}
for some bracket arrangements $\beta^k$ and $\beta^{n-k}$ of weight $k$ and
$n-k$, respectively, with $1\leq k<n$. For instance, if $n=3$, there are two bracket arrangements given by
$[[a_1,a_2],a_3]$ and $[a_1,[a_2,a_3]]$.
Let $R_{j}$ be a sequence of subgroups of $G$ for
$1\leq j\leq n$. The \textit{fat commutator subgroup} $[[R_{1}, R_2, \dots,R_{n}]]$ is defined to be the subgroup of $G$ generated by all of
the commutators
\begin{equation}\label{equation1.1}
\beta^t(g_{i_1},\ldots,g_{i_t}),
\end{equation}
where
\begin{enumerate}
\item[1)] $1\leq i_s\leq n$;
\item[2)] $\{i_1,\ldots,i_t\}=\{1,\ldots,n\}$, that is each integer in $\{1,2,\cdots,n\}$ appears as at least one of the
integers $i_s$;
\item[3)] $g_j\in R_j$;
\item[4)] $\beta^t$ runs over all of the bracket arrangements of weight $t$ (with $t\geq n$).
\end{enumerate}
For convenience, let $[[R_1]]=R_1$. The fat commutator subgroups have important applications in homotopy theory~\cite{BCWW,BL,EM,Wu1}. Let $G=F_n$ be the free group of rank $n$ with a basis $x_1,\ldots,x_n$. Let $R_i=\la\la x_i\ra\ra$ be the normal closure of $x_i$ in $F_n$ for $1\leq i\leq n$ and let $R_{n+1}=\la\la x_1\cdots x_n\ra\ra$ be the normal closure of the product element $x_1x_2\cdots x_n$ in $F_n$. According to~\cite[Theorem 1.7]{Wu1}, the homotopy group $\pi_{n+1}(S^2)$ is isomorphic to the quotient group $$ (R_1\cap R_2\cap\cdots\cap R_{n+1})/[[R_1,R_2,\ldots,R_{n+1}]]$$ for each $n$.
To understand the fat commutator subgroup, we consider the
\textit{symmetric commutator subgroup}\footnote{Symmetric commutator
subgroup was named by Roman Mikhailov during the communications.}
$[[R_1, R_2],\ldots,R_n]_S$ defined by
$$
[[R_1,R_2],\ldots,R_n]_S=\prod_{\sigma\in \Sigma_n}[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n)}],
$$
where $[[R_{\sigma(1)},R_{\sigma(2)}],\ldots, R_{\sigma(n)}]$ is the subgroup generated by the left iterated commutators
$$
[[[g_1,g_2],g_3],\ldots,g_n]
$$
with $g_i\in R_{\sigma(i)}$. For convenience, let $[R_1]_S=R_1$. From the definition, the symmetric commutator subgroup is a subgroup of the fat commutator subgroup. Our first result states that the fat commutator subgroup is in fact the same as the symmetric commutator subgroup.
\begin{thm}\label{theorem1.1}
Let $R_j$ be any normal subgroup of a group $G$ with $1\leq j\leq n$. Then
$$
[[R_1,R_2,\ldots,R_n]]=[[R_1,R_2],\ldots,R_n]_S.
$$
\end{thm}
The symmetric commutator subgroup can be simplified as follows:
\begin{thm}\label{theorem1.2}
Let $R_j$ be any normal subgroup of a group $G$ with $1\leq j\leq n$. Then
$$
[[R_1,R_2],\ldots,R_n]_S=\prod_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}],
$$
where $\Sigma_{n-1}$ acts on $\{2,3,\ldots,n\}$.
\end{thm}
Our next step is to give a generalization of~\cite[Theorem 1.7]{Wu1}. This will give more connections between homotopy groups and symmetric commutator subgroups.
Let $(X,A)$ be a pair of spaces. An \textit{$n$-partition} of $X$ relative to $A$ means a sequence of subspaces $(A_1,\ldots,A_n)$ of $X$ such that
\begin{enumerate}
\item $A=A_i\cap A_j$ for each $1\leq i<j\leq n$ and
\item $X=\bigcup\limits_{i=1}^n A_i$.
\end{enumerate}
An $n$-partition $(A_1,\ldots,A_n)$ of $X$ relative to $A$ is called \textit{cofibrant} if the inclusion
$$
\bigcup_{i\in I}A_i \hookrightarrow \bigcup_{j\in J}A_j
$$
is a cofibration for any $I\subseteq J\subseteq\{1,2,\ldots,n\}$. Note that for a cofibrant partition, each union $A_I=\bigcup_{i\in I}A_i$ is the homotopy colimit of the diagram given by the inclusions $A_{I'}\hookrightarrow A_{I''}$ for $\emptyset\subseteq I'\subseteq I''\subsetneqq I$.
\begin{thm}\label{theorem1.3}
Let $(X,A)$ be a pair of spaces and let $(A_1,\ldots,A_n)$ be a cofibrant $n$-partition of $X$ relative to $A$ with $n\geq 2$. Suppose that
\begin{enumerate}
\item[i)] For any proper subset $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$, the union
$
\bigcup_{i\in I}A_i
$
is a path-connected $K(\pi,1)$-space.
\item[ii)] The inclusion $A \to A_i$ induces an epimorphism of the corresponding fundamental groups for each $1\leq i\leq n$.
\end{enumerate}
Let $R_i$ be the kernel of $\pi_1(A)\to \pi_1(A_i)$ for $1\leq i\leq n$. Then
\begin{enumerate}
\item For any proper subset $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$,
$$
R_{i_1}\cap\cdots\cap R_{i_k}=[[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S.
$$
\item For any $1<k\leq n$ and any subset $I=\{i_1,\ldots,i_k\}\subseteq \{1,2,\ldots,n\}$, there is an isomorphism of groups
$$
\pi_k(X)\cong \left.\left(\bigcap_{s=1}^k\left(R_{i_s}\cdot \prod_{j\in J}R_j\right)\right)\right/\left([[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S\cdot \prod_{j\in J}R_j \right),
$$
where $J= \{1,2,\ldots,n\}-I$.
In particular, $$\pi_n(X)\cong (R_1\cap R_2\cap \cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S.$$
\end{enumerate}
\end{thm}
For the special case where $n=2$, Theorem~\ref{theorem1.3} is the classical Brown-Loday Theorem~\cite[Corollary 3.4]{BL} with a generalization recently given in~\cite{EM}. But the connectivity hypothesis on the subgroups in~\cite{EM} seems difficult to check. Theorem~\ref{theorem1.3} emphasizes the point that, for any space $X$ that admits a cofibrant $K(\pi,1)$-partition, the higher homotopy groups of $X$ measure the difference between the intersection subgroups and the symmetric commutator subgroups for certain subgroups in the fundamental groups of the partition spaces. Moreover Theorem~\ref{theorem1.3} admits many applications. A direct consequence is to give an interesting connection between link groups and higher homotopy groups.
\begin{cor}\label{corollary1.4}
Let $M$ be a path-connected $3$-manifold and $L$ be a proper $m$-link in $M$ with $m\geq 2$. Suppose that for any nonempty sub-link $L'$ of $L$, the link complement $M\smallsetminus |L'|$ is a $K(\pi,1)$-space. Let $\Lambda_1,\Lambda_2,\ldots,\Lambda_n$ be any subsets of $\{1,2,\ldots,m\}$, with $n\geq 2$, such that
\begin{enumerate}
\item[(i)] $\Lambda_i\not=\emptyset$ for each $1\leq i\leq n$.
\item[(ii)] $\Lambda_i\cap \Lambda_j=\emptyset$ for $i\not=j$.
\item[(iii)] $\bigcup_{i=1}^n \Lambda_i=\{1,2,\ldots,m\}$.
\end{enumerate}
Let $\alpha_j$ be the $j\,$th meridian of the link $L$ and let
$$
R_i=\la\la \alpha_j \ | \ j\in \Lambda_i\ra\ra
$$
be the normal closure of $\alpha_j$ with $j\in \Lambda_i$ in $\pi_1(M\smallsetminus |L|)$. Then
\begin{enumerate}
\item For any proper subset $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$,
$$
R_{i_1}\cap\cdots\cap R_{i_k}=[[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S.
$$
\item There is an isomorphism of groups
$$
(R_1\cap R_2\cap\cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S\cong \pi_n(M).
$$\hfill $\Box$
\end{enumerate}
\end{cor}
There are examples of links whose complements are $K(\pi,1)$-space
s. For instance, let $M=S^3$ and let $\pi\colon S^3\to S^2$ be the
Hopf fibration. Let $q_1,\ldots,q_m$ be $m$ distinct points in
$S^2$. Let $L=\pi^{-1}(\{q_1,q_2,\ldots,q_m\})$. Then $L$ is an
$m$-link in $S^3$ with the property that $S^3\smallsetminus |L'|$ is
a $K(\pi,1)$-space for any non-empty sub-link $L'$ of $L$.
Consider the special case where $n=m$ with $\Lambda_i=\{i\}$ in Corollary~\ref{corollary1.4}. Then $R_i$ is the normal closure generated by the $i\,$th meridian in $\pi_1(M\smallsetminus |L|)$. Any element in the intersection subgroup $R_1\cap R_2\cap\cdots\cap R_n$ can be represented by a $1$-link $l$. The union $L\cup \{l\}$ gives an $(n+1)$-link in $M$ related to Brunnian links. Thus Corollary~\ref{corollary1.4} gives a connection between higher homotopy groups and Brunnian links.
For more applications of Theorem~\ref{theorem1.3}, we consider certain subgroups of surface groups and braid groups whose intersection subgroups modulo symmetric commutator subgroups are given by the homotopy groups. Our results on the braid groups are described as follows.
Let $M$ be a manifold. Recall that the $m\,$th ordered configuration space $F(M,m)$ is defined by
$$
F(M,m)=\{(z_1,\ldots,z_m)\in M^m\ | \ z_i\not=z_j\textrm{ for } i\not=j\}
$$
with subspace topology. Recall that the Artin braid group $B_n$ is generated by $\sigma_1,\ldots,\sigma_{n-1}$ with defining relations given by $\sigma_i\sigma_j=\sigma_j\sigma_i$ for $|i-j|\geq 2$ and $\sigma_i\sigma_{i+1}\sigma_i=\sigma_{i+1}\sigma_i\sigma_{i+1}$ for each $i$. The pure braid group $P_n$ is defined to be the kernel of the canonical quotient homomorphism from $B_n$ to the symmetric group $\Sigma_n$, with a set of generators given by
$$
A_{i,j}=\sigma_{j-1}\sigma_{j-2}\cdots \sigma_{i+1}\sigma_i^2\sigma_{i+1}^{-1}\cdots\sigma_{j-2}^{-1}\sigma_{j-1}^{-1}
$$
for $1\leq i<j\leq n$. Let
\begin{align*}
A_{0,j} & =(A_{j,j+1}A_{j,j+2}\cdots A_{j,n})^{-1}(A_{1,j}\cdots
A_{j-1,j})^{-1}\\
& =(\sigma_{j}\sigma_{j+1}\cdots\sigma_{n-2}\sigma_{n-1}^{2}\sigma
_{n-2}\cdots\sigma_{j})^{-1}\cdot(\sigma_{j-1}\cdots\sigma_{2}\sigma_{1}
^{2}\sigma_{1}\cdots\sigma_{j-1})^{-1}.
\end{align*}
Then we have the following:
\begin{thm}\label{theorem1.5}
Let $\Lambda_1,\Lambda_2,\ldots,\Lambda_n$ be any subsets of $\{1,2,\ldots,m\}$, with $n\geq 2$, such that
\begin{enumerate}
\item[(i)] $\Lambda_i\not=\emptyset$ for each $1\leq i\leq n$.
\item[(ii)] $\Lambda_i\cap \Lambda_j=\emptyset$ for $i\not=j$.
\item[(iii)] $\bigcup_{i=1}^n \Lambda_i=\{1,2,\ldots,m\}$.
\end{enumerate}
Let $R_i=\la\la A_{0,j} \ | \ j\in \Lambda_i\ra\ra$ be the normal closure in the pure braid group $P_m$. Then
\begin{enumerate}
\item For any proper subset $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$,
$$
R_{i_1}\cap\cdots\cap R_{i_k}=[[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S.
$$
\item There is an isomorphism of groups
$$
(R_1\cap R_2\cap\cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S\cong \pi_n(F(S^2,m)).
$$
\end{enumerate}
\end{thm}
For instance, for $n=2$ with $m\geq 3$, $(R_1\cap R_2)/[R_1,R_2]=\pi_2(F(S^2,m))=0$, where $\pi_2(F(S^2,m))=0$ for $m\geq 3$ is given in~\cite[p. 244]{FB}. For $m\geq n\geq 3$, from~\cite[Theorem 1]{Fadell}, we have $\pi_n(F(S^2,m))\cong \pi_n(S^2)$ and so
$$
(R_1\cap R_2\cap\cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S\cong \pi_n(F(S^2,m))\cong \pi_n(S^2).
$$
The braided descriptions of the homotopy groups $\pi_*(S^2)$ have
been studied in~\cite{BCWW,BMVW, CW1,CW2,LW, Wu2}.
Theorem~\ref{theorem1.5} is a new braided description of the
homotopy groups.
The article is organized as follows. The proofs of Theorems~\ref{theorem1.1} and~\ref{theorem1.2} are given in section~\ref{section2}. In section~\ref{section3}, we provide the proof of Theorem~\ref{theorem1.3}. In section~\ref{section4}, we discuss some applications to the free groups and surface groups. The applications to the braid groups and the proof of Theorem~\ref{theorem1.5} are given in section~\ref{section5}.
\section{The Proofs of Theorems~\ref{theorem1.1} and~\ref{theorem1.2}}\label{section2}
\subsection{Some Lemmas}
In this subsection, we give some useful lemmas on commutator subgroups. The following lemma is elementary.
\begin{lem}\label{lemma2.1}
Let $G$ be a group and let $A,B,C$ be normal subgroups of $G$. Then $[AB,C]=[A,C][B,C]$.\hfill $\Box$
\end{lem}
The following classical theorem can be found in~\cite[Theorem 5.2, p.290]{MKS}.
\begin{thm}[Hall's Theorem]
Let $G$ be a group and let $A,B,C$ be normal subgroups of $G$. Then any one of the subgroups $[A,[B,C]]$, $[[A,B],C]$ and $[[A,C],B]$ is a subgroup of the product of the other two.
\hfill $\Box$
\end{thm}
Now let $R_1,\ldots,R_n$ be subgroups of $G$. Recall that the symmetric commutator subgroup $[[R_1,R_2],\ldots,R_n]_S$ is defined by
$$
[[R_1,R_2],\ldots,R_n]_S=\prod_{\sigma\in \Sigma_n}[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n)}].
$$
If each $R_i$ is a normal subgroup of $G$, then $[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n)}]$ is normal in $G$ and so is $[[R_1,R_2],\ldots,R_n]_S$.
\begin{lem}\label{lemma2.3}
Let $R_1,\ldots,R_n$ be normal subgroups of $G$. Let $g_j\in R_j$ for $1\leq j\leq n$. Then
$$
\beta^n(g_{\sigma(1)},\ldots,g_{\sigma(n)})\in [[R_1,R_2],\ldots,R_n]_S
$$
for any $\sigma\in \Sigma_n$ and any bracket arrangement $\beta^n$ of weight $n$.
\end{lem}
\begin{proof}
The proof is given by double induction. The first induction is on $n$. Clearly the assertion holds for $n=1$. Suppose that the assertion holds for $m$ with $m<n$. Given an element $\beta^n(g_{\sigma(1)},\ldots,g_{\sigma(n)})$ as in the lemma. Then
$$
\beta^n(g_{\sigma(1)},\ldots,g_{\sigma(n)})=[\beta^p(g_{\sigma(1)},\ldots,g_{\sigma(p)}),\beta^{n-p}(g_{\sigma(p+1)},\ldots,g_{\sigma(n)})]
$$
for some bracket arrangements $\beta^p$ and $\beta^{n-p}$ with $1\leq p\leq n-1$. The second induction is on $q=n-p$. If $q=1$, we have
$$
\beta^{n-1}(g_{\sigma(1)},\ldots,g_{\sigma(n-1)})\in [[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n-1)}]_S
$$
by the first induction and so
$$
\begin{array}{rcl}
\beta^n(g_{\sigma(1)},\ldots,g_{\sigma(n)})&=&[\beta^{n-1}(g_{\sigma(1)},\ldots,g_{\sigma(n-1)}),g_{\sigma(n)}]\\
&\in& [[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n-1)}]_S, R_{\sigma(n)}]\\
\end{array}
$$
with
$$
\begin{array}{rl}
& [[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n-1)}]_S, R_{\sigma(n)}]\\
=&\left[\prod_{\tau\in\Sigma_{n-1}}[[R_{\sigma(\tau(1))},R_{\sigma(\tau(2))}],\ldots,R_{\sigma(\tau(n-1))}],R_{\sigma(n)}\right]\\
\rEq^{\textrm{by Lemma~\ref{lemma2.1}}}&\prod_{\tau\in\Sigma_{n-1}}[[[R_{\sigma(\tau(1))},R_{\sigma(\tau(2))}],\ldots,R_{\sigma(\tau(n-1))}], R_{\sigma(n)}]\\
\leq& [[R_1,R_2],\ldots,R_n]_S.\\
\end{array}
$$
Now suppose that the assertion holds for $q'=n-p<q$. By the first induction, we have
$$
\beta^p(g_{\sigma(1)},\ldots,g_{\sigma(p)})\in [[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(p)}]_S
$$
and
$$
\beta^{n-p}(g_{\sigma(p+1)},\ldots,g_{\sigma(n)})\in [[R_{\sigma(p+1)},R_{\sigma(p+2)}],\ldots,R_{\sigma(n)}]_S.
$$
Thus
$$
\beta^{n}(g_{\sigma(1)},\ldots,g_{\sigma(n)})\in \left[[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(p)}]_S, [[R_{\sigma(p+1)},R_{\sigma(p+2)}],\ldots,R_{\sigma(n)}]_S\right].
$$
By Lemma~\ref{lemma2.1}, $\beta^{n}(g_{\sigma(1)},\ldots,g_{\sigma(n)})$ lies in the product subgroup
$$
T=\prod\limits_{
\begin{array}{c}
\tau\in\Sigma_p\\
\rho\in\Sigma_{n-p}\\
\end{array}}
\left[[[R_{\sigma(\tau(1))},\ldots,R_{\sigma(\tau(p))}], [[R_{\sigma(\rho(p+1))},\ldots,R_{\sigma(\rho(n))}]\right],
$$
where $\Sigma_{n-p}$ acts on $\{p+1,\ldots,n\}$. By applying Hall's Theorem, we have
$$
\begin{array}{rl}
&\left[[[R_{\sigma(\tau(1))},\ldots,R_{\sigma(\tau(p))}], [[R_{\sigma(\rho(p+1))},\ldots,R_{\sigma(\rho(n))}]\right]\\
=&\left[[[R_{\sigma(\tau(1))},\ldots,R_{\sigma(\tau(p))}], \left[[[R_{\sigma(\rho(p+1))},\ldots, R_{\sigma(\rho(n-1))}],R_{\sigma(\rho(n))}\right]\right]\\
\leq &\left[\left[[[R_{\sigma(\tau(1))},\ldots,R_{\sigma(\tau(p))}], [[R_{\sigma(\rho(p+1))},\ldots, R_{\sigma(\rho(n-1))}]\right],R_{\sigma(\rho(n))}\right]\\
&\cdot\left[ \left[[[R_{\sigma(\tau(1))},\ldots,R_{\sigma(\tau(p))}],R_{\sigma(\rho(n))}\right], [[R_{\sigma(\rho(p+1))},\ldots, R_{\sigma(\rho(n-1))}]\right].\\
\end{array}
$$
Note that $A=\left[\left[[[R_{\sigma(\tau(1))},\ldots,R_{\sigma(\tau(p))}], [[R_{\sigma(\rho(p+1))},\ldots, R_{\sigma(\rho(n-1))}]\right],R_{\sigma(\rho(n))}\right]$ is generated by the elements of the form
$$
\left[\left[[[g'_{\sigma(\tau(1))},\ldots,g'_{\sigma(\tau(p))}], [[g'_{\sigma(\rho(p+1))},\ldots, g'_{\sigma(\rho(n-1))}]\right],g'_{\sigma(\rho(n))}\right]
$$
with $g'_j\in R_j$. By the second induction on case that $q=1$, the above elements lie in
$[[R_1,R_2],\ldots,R_n]_S$ and so
$$
A\leq [[R_1,R_2],\ldots,R_n]_S.
$$
Similarly, by the second induction hypothesis, $$\left[ \left[[[R_{\sigma(\tau(1))},\ldots,R_{\sigma(\tau(p))}],R_{\sigma(\rho(n))}\right], [[R_{\sigma(\rho(p+1))},\ldots, R_{\sigma(\rho(n-1))}]\right]$$ is a subgroup of $[[R_1,R_2],\ldots,R_n]_S$. It follows that
$$
T\leq [[R_1,R_2],\ldots,R_n]_S
$$
and so
$$
\beta^{n}(g_{\sigma(1)},\ldots,g_{\sigma(n)})\in [[R_1,R_2],\ldots,R_n]_S.
$$
Both the first and second inductions are finished, hence the result.
\end{proof}
\begin{lem}\label{lemma2.4}
Let $G$ be a group and let $R_1,\ldots,R_n$ be normal subgroups of $G$. Let $(i_1,i_2,\ldots,i_p)$ be a sequence of integers with $1\leq i_s\leq n$. Suppose that $$\{i_1,i_2,\ldots,i_p\}=\{1,2,\ldots,n\}.$$ Then
$$
[[R_{i_1},R_{i_2}],\ldots,R_{i_p}]\leq [[R_1,R_2],\ldots,R_n]_S.
$$
\end{lem}
\begin{proof}The proof is given by double induction. The first induction is on $n$.
The assertion clearly holds for $n=1$. Suppose that the assertion holds for $n-1$ with $n>1$.
From the hypothesis that $\{i_1,i_2,\ldots,i_p\}=\{1,2,\ldots,n\}$, we have $p\geq n$. When $p=n$, $(i_1,\ldots,i_n)$ is a permutation of $(1,\ldots,n)$ and so $$[[R_{i_1},R_{i_2}],\ldots,R_{i_n}]\leq [[R_1,R_2],\ldots,R_n]_S.$$ Suppose that
$$
[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]\leq [[R_1,R_2],\ldots,R_n]_S
$$
for any sequence $(j_1,\ldots,j_q)$ with $q<p$ and $\{j_1,\ldots,j_q\}=\{1,\ldots,n\}$. Let $(i_1,\ldots,i_p)$ be a sequence with $\{i_1,\ldots,i_p\}=\{1,\ldots,n\}$.
If $i_p\in \{i_1,\ldots,i_{p-1}\}$, then $\{i_1,\ldots,i_{p-1}\}=\{1,\ldots,n\}$ and so
$$
[[R_{i_1},R_{i_2}],\ldots,R_{i_{p-1}}]\leq [[R_1,R_2],\ldots,R_n]_S
$$
by the second induction hypothesis. It follows that
$$
[[R_{i_1},R_{i_2}],\ldots,R_{i_p}]\leq [[R_1,R_2],\ldots,R_n]_S.
$$
If $i_p\not\in \{i_1,\ldots,i_{p-1}\}$ , we may assume that $i_p=n$. Then $$\{i_1,\ldots,i_{p-1}\}=\{1,\ldots,n-1\}$$ and so
$$
[[R_{i_1},R_{i_2}],\ldots,R_{i_{p-1}}]\leq [[R_1,R_2],\ldots,R_{n-1}]_S
$$
by the first induction hypothesis. From Lemma~\ref{lemma2.3}, we have
$$
[[R_{i_1},R_{i_2}],\ldots,R_{i_p}]\leq [[R_1,R_2],\ldots,R_n]_S.
$$
The inductions are finished, hence the result holds.
\end{proof}
\begin{lem}\label{lemma2.5}
Let $G$ be a group and let $R_1,\ldots,R_n$ be normal subgroups of $G$ with $n\geq 2$. Let $(i_1,\ldots,i_p)$ and $(j_1,\ldots,j_q)$ be sequences of integers such that $\{i_1,\ldots,i_p\}\cup\{j_1,\ldots,j_q\}=\{1,2,\ldots,n\}$. Then
$$
[[[R_{i_1},R_{i_2}],\ldots,R_{i_p}],[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]]\leq [[R_1,R_2],\ldots,R_n]_S.
$$
\end{lem}
\begin{proof}
The proof is given by double induction on $n$ and $q$ with $n\geq 2$ and $q\geq 1$. First we prove the assertion holds for $n=2$. If $\{i_1,\ldots,i_p\}=\{1,2\}$
or $\{j_1,\ldots,j_q\}=\{1,2\}$, we have
$$
[[R_{i_1},R_{i_2}],\ldots,R_{i_p}]\leq [[R_1,R_2]_S \textrm{ or }[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]\leq [[R_1,R_2]_S$$
by Lemma~\ref{lemma2.4} and so $$
[[[R_{i_1},R_{i_2}],\ldots,R_{i_p}],[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]]\leq [[R_1,R_2]_S.
$$
Otherwise, $i_1=\cdots=i_p$ and $j_1=\cdots=j_q $, since $\{i_1,\ldots,i_p\}\cup\{j_1,\ldots,j_q\}=\{1,2\}$, we may assume that $i_1=\cdots=i_p=1, j_1=\cdots=j_q =2$, then $$[[R_{i_1},R_{i_2}],\ldots,R_{i_p}]\leq R_1 \textrm{ and }[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]\leq R_2$$ and so
$$
[[[R_{i_1},R_{i_2}],\ldots,R_{i_p}],[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]]\leq [[R_1,R_2]_S.
$$
Suppose the assertion holds for $n-1$, that is $$
[[[R_{i_1},R_{i_2}],\ldots,R_{i_p}],[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]]\leq [[R_1,R_2],\ldots,R_{n-1}]_S.
$$ when $\{i_1,\ldots,i_p\}\cup\{j_1,\ldots,j_q\}=\{1,2,\ldots,n-1\}$. We will use the second induction on $q$ to prove the assertion holds for $n$.
When $q=1$, the assertion follows by Lemma~\ref{lemma2.4}. Suppose that the assertion holds for $q-1$. By the Hall Theorem, $[[[R_{i_1},R_{i_2}],\ldots,R_{i_p}],[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]]$ is a subgroup of the product
$$
[[[[R_{i_1},\ldots,R_{i_p}],[[R_{j_1},\ldots,R_{j_{q-1}}]],R_{j_q}]\cdot [[[[R_{i_1},\ldots,R_{i_p}],R_{j_q}],[[R_{j_1},\ldots,R_{j_{q-1}}]].
$$
By the second induction we have $$ [[[[R_{i_1},\ldots,R_{i_p}],R_{j_q}],[[R_{j_1},\ldots,R_{j_{q-1}}]]\leq [[R_1,R_2],\ldots,R_n]_S.$$
If $\{i_1,\ldots,i_p\}\cup\{j_1,\ldots,j_{q-1}\}=\{1,2,\ldots,n\}$, by the second induction $$[[[R_{i_1},\ldots,R_{i_p}],[[R_{j_1},\ldots,R_{j_{q-1}}]]\leq [[R_1,R_2],\ldots,R_n]_S$$ and hence $$[[[[R_{i_1},\ldots,R_{i_p}],[[R_{j_1},\ldots,R_{j_{q-1}}]],R_{j_q}]\leq [[R_1,R_2],\ldots,R_n]_S. $$
If $\{i_1,\ldots,i_p\}\cup\{j_1,\ldots,j_{q-1}\}\neq \{1,2,\ldots,n\}$, we may assume that $$\{i_1,\ldots,i_p\}\cup\{j_1,\ldots,j_{q-1}\}=\{1,2,\ldots,n-1\}$$ and $j_q=n$. By the first induction, $$[[[R_{i_1},\ldots,R_{i_p}],[[R_{j_1},\ldots,R_{j_{q-1}}]]\leq [[R_1,R_2],\ldots,R_{n-1}]_S.$$
Then $$[[[[R_{i_1},\ldots,R_{i_p}],[[R_{j_1},\ldots,R_{j_{q-1}}]],R_{j_q}]\leq [[R_1,R_2],\ldots,R_n]_S.$$
It follows that $[[[R_{i_1},R_{i_2}],\ldots,R_{i_p}],[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]]\leq [[R_1,R_2],\ldots,R_n]_S$. The double induction is finished, hence the result.
\end{proof}
\subsection{Proof of Theorem~\ref{theorem1.1}}
Clearly $[[R_1,R_2],\ldots,R_n]_S\leq [[R_1,R_2,\ldots, R_n]]$. We prove by induction on $n$ that $$[[R_1,R_2,\ldots,R_n]]\leq [[R_1,R_2],\ldots,R_n]_S.$$ The assertion holds for $n=1$.
\begin{hypothesis}\label{induction1}
Suppose that
$$
[[R_1,R_2,\ldots, R_s]]\leq [[R_1,R_2],\ldots, R_s]_S
$$
for any normal subgroups $R_1,\ldots,R_s$ of $G$ with $1\leq s<n$.
\end{hypothesis}
Let $R_1,\ldots, R_n$ be any normal subgroups of $G$. By definition, $[[R_1,R_2,\ldots,R_n]]$ is generated by all commutators $$\beta^t(g_{i_1},\ldots,g_{i_t})$$ of weight $t$ such that
$\{i_1,i_2,\ldots,i_t\}=\{1,2,\ldots,n\}$ with $g_j\in R_j$. To prove that each generator $\beta^t(g_{i_1},\ldots,g_{i_t})\in [[R_1,R_2],\ldots,R_n]_S$, we start the second induction on the weight $t$ of $\beta^t$ with $t\geq n$. If $t=n$, then $(i_1,\ldots,i_n)$ is a permutation of $(1,\ldots,n)$ and so the assertion holds by Lemma~\ref{lemma2.3}. Now assume that the following hypothesis holds:
\begin{hypothesis}\label{induction2}
Let $n\leq k<t$ and let
$$
\beta^k(g'_{i_1},\ldots,g'_{i_k})
$$
be any bracket arrangement of weight $k$ such that
where
\begin{enumerate}
\item[1)] $1\leq i_s\leq n$;
\item[2)] $\{i_1,\ldots,i_k\}=\{1,\ldots,n\}$;
\item[3)] $g'_j\in R_j$;
\end{enumerate}
Then $\beta^k(g'_{i_1},\ldots,g'_{i_k})\in [[R_1,R_2],\ldots,R_n]_S$.
\end{hypothesis}
Let $\beta^t(g_{i_1},\ldots,g_{i_t})$ be any bracket arrangement of weight $t$ with $\{i_1,\ldots,i_t\}=\{1,\ldots,n\}$ and $g_j\in R_j$ for $1\leq j\leq n$. From the definition of bracket arrangement, we have
$$
\beta^t(g_{i_1},\ldots,g_{i_t})=[\beta^p(g_{i_1},\ldots,g_{i_p}), \beta^{t-p}(g_{i_{p+1}},\ldots,g_{i_t})]
$$
for some bracket arrangements $\beta^p$ and $\beta^{t-p}$ of weight $p$ and $t-p$, respectively, with $1\leq p\leq n-1$. Let
$$
A=\{i_1,\ldots,i_p\}\textrm{ and } B=\{i_{p+1},\ldots,i_t\}.
$$
Then both $A$ and $B$ are the subsets of $\{1,\ldots,n\}$ with $A\cup B=\{1,\ldots,n\}$.
Suppose that the cardinality $|A|=n$ or $|B|=n$. We may assume that $|A|=n$. By Hypothesis~\ref{induction2},
$$
\beta^p(g_{i_1},\ldots,g_{i_p})\in [[R_1,R_2],\ldots,R_n]_S.
$$
Since $[[R_1,R_2],\ldots,R_n]_S$ is a normal subgroup of $G$, we have
$$
\beta^t(g_{i_1},\ldots,g_{i_t})=[\beta^p(g_{i_1},\ldots,g_{i_p}), \beta^{t-p}(g_{i_{p+1}},\ldots,g_{i_t})]\in [[R_1,R_2],\ldots,R_n]_S.
$$
This proves the result in this case.
Suppose that $|A|<n$ and $|B|<n$. Let $A=\{l_1,\ldots,l_a\}$ with $1\leq l_1<l_2<\cdots<l_a\leq n$ and $1\leq a<n$, and let $B=\{k_1,\ldots,k_b\}$ with $1\leq k_1<k_2<\cdots< k_b$ and $1\leq b<n$. Observe that
$$
\beta^p(g_{i_1},\ldots,g_{i_p})\in [[R_{l_1},R_{l_2},\ldots,R_{l_a}]].
$$
By Hypothesis~\ref{induction1},
$$
[[R_{l_1},R_{l_2},\ldots,R_{l_a}]]=[[R_{l_1},R_{l_2}],\ldots,R_{l_a}]_S.
$$
Thus
$$
\beta^p(g_{i_1},\ldots,g_{i_p})\in [[R_{l_1},R_{l_2}],\ldots,R_{l_a}]_S.
$$
Similarly
$$
\beta^{t-p}(g_{i_{p+1}},\ldots,g_{i_t})\in [[R_{k_1},R_{k_2}],\ldots,R_{k_b}]_S.
$$
It follows that the element
$\beta^t(g_{i_1},\ldots,g_{i_t})$ lies in the commutator subgroup
$$
\left[[[R_{l_1},R_{l_2}],\ldots,R_{l_a}]_S, [[R_{k_1},R_{k_2}],\ldots,R_{k_b}]_S\right].
$$
From Lemma~\ref{lemma2.5}, we have
$$
\left[[[R_{l_{\sigma(1)}},R_{l_{\sigma(2)}}],\ldots,R_{l_{\sigma(a)}}], [[R_{k_{\tau(1)}},R_{k_{\tau(2)}}],\ldots,R_{k_{\tau(b)}}]\right]\leq [[R_1,R_2],\ldots,R_n]_S
$$
for all $\sigma\in\Sigma_a$ and $\tau\in\Sigma_b$ because $\{l_1,\ldots,l_a\}\cup \{k_1,\ldots,k_b\}=A\cup B=\{1,2,\ldots,n\}$. It follows from Lemma~\ref{lemma2.1} that
$$
\left[[[R_{l_1},R_{l_2}],\ldots,R_{l_a}]_S, [[R_{k_1},R_{k_2}],\ldots,R_{k_b}]_S\right]\leq [[R_1,R_2],\ldots,R_n]_S.
$$
Thus
$$
\beta^t(g_{i_1},\ldots,g_{i_t})\in [[R_1,R_2],\ldots,R_n]_S.
$$
The inductions are finished, hence Theorem~\ref{theorem1.1}.
\subsection{Proof of Theorem~\ref{theorem1.2}}
Clearly $$\prod_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}]\leq [[R_1,R_2],\ldots,R_n]_S.$$
We prove by induction on $n$ that $$[[R_1,R_2],\ldots,R_n]_S\leq \prod_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}].$$ The assertion holds for $n=1$.
Suppose that
$$
[[R_1,R_2],\ldots, R_s]_S\leq \prod_{\sigma\in \Sigma_{s-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(s)}]
$$
for any normal subgroups $R_1,\ldots,R_s$ of $G$ with $1\leq s<n$.
Let $R_1,\ldots, R_n$ be any normal subgroups of $G$. By definition, $$[[R_1,R_2],\ldots,R_n]_S=\prod\limits_{\tau\in \Sigma_{n}}[[R_{\tau(1)},R_{\tau(2)}],\ldots,R_{\tau(n)}].$$ It suffices to prove that for any $\tau\in \Sigma_n$, $$[[R_{\tau(1)},R_{\tau(2)}],\ldots,R_{\tau(n)}]\leq \prod_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}].$$
The assertion holds for $\tau(1)=1$.
Suppose that the assertion holds for $\tau(k-1)=1$ with $1\leq k-1<n$.
When $\tau(k)=1$, consider the subgroup $$[[[[[R_{\tau(1)},\ldots,R_{\tau(k-2)}],R_{\tau(k-1)}],R_1],R_{\tau(k+1)},\ldots,R_{\tau(n)}].$$
Following from the Hall Theorem,
$$
\begin{array}{rl}
&[[[[[R_{\tau(1)},\ldots, R_{\tau(k-2)}],R_{\tau(k-1)}],R_1],R_{\tau(k+1)},\ldots,R_{\tau(n)}]\\
&\\
\leq &[[[[[R_{\tau(1)},\ldots,R_{\tau(k-2)}],R_1],R_{\tau(k-1)}], R_{\tau(k+1)},\ldots,R_{\tau(n)}]]\\
&\cdot[ [[R_{\tau{(k-1)}},R_1], [[R_{\tau(1)},\ldots, R_{\tau(k-2)}]], R_{\tau(k+1)},\ldots,R_{\tau(n)} ].\\
\end{array}
$$
By the second induction $$
\begin{array}{rl}
&[[[[[R_{\tau(1)},\ldots,R_{\tau(k-2)}],R_1],R_{\tau(k-1)}], R_{\tau(k+1)},\ldots,R_{\tau(n)}]]\\
\leq &\prod\limits_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}].\\
\end{array}
$$
From Lemma~\ref{lemma2.5} and the first induction, we have
$$
\begin{array}{rcl}
[[R_{\tau{(k-1)}},R_1], [[R_{\tau(1)},\ldots, R_{\tau(k-2)}]]&\leq& [[R_1,R_{l_1}],R_{l_2},\ldots,R_{l_{k-1}}]_S\\
&\leq& \prod\limits_{\sigma\in \Sigma_{k-1}}[[R_{1},R_{\sigma(l_1)}],\ldots,R_{\sigma(l_{k-1})}],\\
\end{array} $$
where
$\{1,\tau(1),\ldots,\tau{(k-1)}\}=\{1,l_1,l_2,\ldots,l_{k-1}\}$ with $1<l_1<l_2<\cdots<l_{k-1}$.
By Lemma~\ref{lemma2.1}
$$
\begin{array}{rl}
&[ [[R_{\tau{(k-1)}},R_1], [[R_{\tau(1)},\ldots, R_{\tau(k-2)}]], R_{\tau(k+1)},\ldots,R_{\tau(n)} ]\\
&\\
\leq &\prod\limits_{\sigma\in \Sigma_{k-1}}[ [[R_{1},R_{\sigma(l_1)}],\ldots,R_{\sigma(l_{k-1})}],R_{\tau(k+1)},\ldots,R_{\tau(n)} ]\\
&\\
\leq &\prod\limits_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}]\\
\end{array}
$$
because for each $\sigma\in \Sigma_{k-1}$, the sequence $(\sigma(l_1),\ldots,\sigma(l_{k-1}),\tau(k+1),\ldots,\tau(n))$ is a permutation of $(\tau(1),\ldots,\tau(k-1),\tau(k+1),\ldots,\tau(n))$ which is a permutation of $(2,\ldots,n)$.
It follows that
$$
\begin{array}{rl}
&[[[[[R_{\tau(1)},\ldots,R_{\tau(k-2)}],R_{\tau(k-1)}],R_1],R_{\tau(k+1)},\ldots,R_{\tau(n)}]\\
&\\
\leq &\prod\limits_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}]\\
\end{array}
$$
and so $$[[R_1,R_2],\ldots,R_n]_S\leq \prod_{\sigma\in \Sigma_{n-1}}[[R_{1},R_{\sigma(2)}],\ldots,R_{\sigma(n)}].$$ This finishes the proof.
\section{Proof of Theorem~\ref{theorem1.3}}\label{section3}
\subsection{Ellis-Mikhailov Theorem.}
In this subsection, we review some terminology and the main result in~\cite{EM}. Let $G$ be a group. An $m$-tuple of normal subgroups $(R_1,\ldots,R_m)$ of $G$ is called \textit{connected} if either
\begin{enumerate}
\item $m\leq 2$ or
\item $m\geq 3$ with the property that: for all subsets $I, J \subseteq \{1,\cdots,m\}$ with $ |I|\geq 2,
|J|\geq 1$
\begin{equation}\label{connectivity}
\left(\bigcap_{i\in I} R_i \right)\cdot \prod_{j\in J}R_j =
\bigcap_{i\in I} \left( R_i\cdot \prod_{j\in J}R_j \right).
\end{equation}
\end{enumerate}
Let $G$ be a group with normal subgroups $R_1,\ldots,R_n$. Let $X(G;R_1,\ldots,R_n)$ be the homotopy colimit of the cubical diagram obtained from classifying spaces $B(G/\prod_{i\in I} R_i)$ with the maps
$$
B(G/\prod_{i\in I} R_i)\to B(G/\prod_{i'\in I'}R_{i'})
$$
induced by the canonical quotient homomorphism $G/\prod_{i\in I} R_i \twoheadrightarrow G/\prod_{i'\in I'}R_{i'}$ for $I\subseteq I'$, where $I$ ranges over all proper subsets $I \subsetneq \{1,\ldots,n\}$.
In the following theorem and the rest of the article, the notation $(\ \cdots \ \hat a \ \cdots)$ means to remove letter $a$.
\begin{thm}~\cite[Theorem 1]{EM}\label{theorem3.1}
Let $G$ be a group with normal subgroups $R_1,\ldots,R_n$ with $n\geq2$. Let $X=X(G;R_1,\ldots,R_n)$.
Suppose that the $(n-1)$-tuple $$(R_1,\ldots,\hat R_i,\ldots,R_n)$$ is connected for each $1\leq i\leq n$. Then
$$
\pi_n(X) \cong \frac{R_1\cap \cdots \cap R_n}{\prod_{I\cup J=\{1,\ldots,n\},
{I\cap J=\emptyset}}
[\cap_{i\in I}R_i,\cap_{j\in J}R_j]}.
$$\hfill $\Box$
\end{thm}
\subsection{Proof of Theorem~\ref{theorem1.3}}
Theorem~\ref{theorem1.3} is part of the following statement.
\begin{thm}
Let $(X,A)$ be a pair of spaces and let $(A_1,\ldots,A_n)$ be a cofibrant $n$-partition of $X$ relative to $A$ with $n\geq 2$. Suppose that
\begin{enumerate}
\item[i)] For any proper subset $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$, the union
$
\bigcup_{i\in I}A_i
$
is a path-connected $K(\pi,1)$-space.
\item[ii)] The inclusion $A \to A_i$ induces an epimorphism of the corresponding fundamental groups for each $1\leq i\leq n$.
\end{enumerate}
Let $R_i$ be the kernel of $\pi_1(A)\to \pi_1(A_i)$ for $1\leq i\leq n$. Then
\begin{enumerate}
\item For any proper subset $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$,
$$
R_{i_1}\cap\cdots\cap R_{i_k}=[[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S.
$$
\item The $(n-1)$-tuple $(R_1,\ldots,\hat R_i,\ldots,R_n)$ is connected for each $1\leq i\leq n$.
\item There is an isomorphism of groups
$$
\pi_n(X)\cong (R_1\cap R_2\cap \cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S.
$$
\item For $1< k< n$, $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$ and $J= \{1,2,\ldots,n\}\smallsetminus I$, there is an isomorphism
$$
\pi_k(X)\cong \left.\left(\bigcap_{s=1}^k(R_{i_s}\cdot \prod_{j\in J}R_j)\right)\right/\left(([[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S)\cdot \prod_{j\in J}R_j \right).
$$
\end{enumerate}
\end{thm}
\begin{proof}
We prove assertions (1)-(3) by induction on $n$. If $n=2$, assertions (1) and (2) hold obviously and assertion (3) follows from the classical Brown-Loday Theorem~\cite{BL}.
Suppose that assertions (1)-(3) hold for all cofibrant $m$-partitions $(B_1,\ldots,B_m)$ of any space $Y$ relative to $B$ satisfying conditions (i) and (ii) with $m<n$.
Let $(X,A)$ be a pair of spaces and let $(A_1,\ldots,A_n)$ be a cofibrant $n$-partition of $X$ relative to $A$ with $n>2$.
(1). Let $I=\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\}$ be a proper subset. If $k=1$, then $R_{i_1}=[R_{i_1}]_S$ by definition. We may assume that $2\leq k<n$. Let $B_s=A_{i_s}$ for $s=1,\ldots, k$ and let $Y=B_1\cup\cdots\cup B_{k}$. Then $(B_1,\ldots,B_k)$ is a cofibrant $k$-partition of $Y$ relative to $A$ satisfying conditions (i) and (ii). Notice that
$$
\mathrm{K er}(\pi_1(A)\to \pi_1(B_s))=\mathrm{K er}(\pi_1(A)\to \pi_1(A_{i_s}))=R_{i_s}.
$$
By induction hypothesis, there is an isomorphism
$$
\pi_k(Y)\cong (R_{i_1}\cap R_{i_2}\cap \cdots\cap R_{i_k})/[[R_{i_1},R_{i_2}],\ldots,R_{i_k}]_S.
$$
From condition (i), $Y=A_{i_1}\cup\cdots\cup A_{i_k}$ is a $K(\pi,1)$-space (as $k<n$) and so $\pi_k(Y)=0$ (as $k\geq2$). Thus
$$
R_{i_1}\cap R_{i_2}\cap \cdots\cap R_{i_k}=[[R_{i_1},R_{i_2}],\ldots,R_{i_k}]_S,
$$
which is assertion~(1).
(2). We show that the $(n-1)$-tuple $(R_1,\ldots,R_{n-1})$ is connected. Let $m=n-1$. If $m=2$, then $(R_1,R_2)$ is connected by definition. Assume that $m\geq 3$. Let $I, J \subseteq \{1,\cdots,m\}$ with $ |I|\geq 2$ and $|J|\geq 1$. We have to show the connectivity condition that
$$
\left(\bigcap_{i\in I} R_i \right)\cdot \prod_{j\in J}R_j =
\bigcap_{i\in I} \left( R_i\cdot \prod_{j\in J}R_j \right).
$$
Clearly
$$
\left(\bigcap_{i\in I} R_i \right)\cdot \prod_{j\in J}R_j\leq
\bigcap_{i\in I} \left( R_i\cdot \prod_{j\in J}R_j \right).
$$
Now we show the other direction. Let $I=\{i_1,\ldots,i_k\}$ with $2\leq k<n$. Let $B=\bigcup_{j\in J}A_j$ and let $B_s=B\cup A_{i_s}$ for $s=1,2,\ldots,k$. Let
$$
Y=\bigcup_{s=1}^kB_s=\bigcup_{i\in I, j\in J}A_i\cup A_j.
$$
Then $(B_1,\ldots,B_k)$ is a cofibrant $k$-partition of $Y$ relative to $B$. Clearly condition (i) holds. Let $G=\pi_1(A)$. Then $\pi_1(B)=G/\prod_{j\in J}R_j$ and
$$
\pi_1(B_s)=G/(R_{i_s}\cdot\prod_{j\in J}R_j).
$$
Thus condition (ii) holds. Let
$$
\begin{array}{rcl}
N_s&=&\mathrm{K er}(\pi_1(B)\to \pi_1(B_s))\\
&=&\mathrm{K er}(G/(\prod_{j\in J}R_j)\to G/(R_{i_s}\cdot\prod_{j\in J}R_j))\\
&=&(R_{i_s}\cdot\prod_{j\in J}R_j)/\prod_{j\in J}R_j\\
&=&R_{i_s}/(R_{i_s}\cap \prod_{j\in J}R_j).\\
\end{array}
$$
From induction hypothesis, we have
$$
\pi_k(Y)=(N_1\cap N_2\cap\cdots\cap N_k)/[[N_1,N_2],\ldots,N_k]_S.
$$
Since $
Y=\bigcup_{i\in I, j\in J}A_i\cup A_j$ with $I,J\subseteq\{1,2,\ldots,n-1\}$, the space $Y$ is a $K(\pi,1)$-space from condition (i). Thus $\pi_k(Y)=0$ and so
\begin{equation}\label{equation3.2}
N_1\cap N_2\cap\cdots\cap N_k=[[N_1,N_2],\ldots,N_k]_S
\end{equation}
with $N_s=R_{i_s}/(R_{i_s}\cap \prod_{j\in J}R_j)$ in $G/\prod_{j\in J}R_j$. Consider the quotient homomorphism
$$
\phi\colon G\longrightarrow G/\prod_{j\in J}R_j.
$$
Then
$$
N_s=R_{i_s}/(R_{i_s}\cap \prod_{j\in J}R_j)=\phi(R_{i_s})
$$
and so
$$
\begin{array}{rcl}
\phi\left([[R_{i_1},R_{i_2}],\ldots,R_{i_k}]_S\right)&=&[[\phi(R_{i_1}),\phi(R_{i_2})],\ldots,\phi(R_{i_k})]_S\\
&=&[[N_1,N_2],\ldots,N_k]_S\\
&=&N_1\cap N_2\cap\cdots\cap N_k \textrm{ by equation~(\ref{equation3.2})}.\\
\end{array}
$$
From the fact that
$$
\phi\left(\bigcap_{i\in I} \left( R_i\cdot \prod_{j\in J}R_j \right)\right)\leq \bigcap_{s=1}^k N_s
$$
together with equation~(\ref{equation3.2}),
we have
$$
\begin{array}{rcl}
\bigcap_{i\in I} \left( R_i\cdot \prod_{j\in J}R_j \right)&\leq&\phi^{-1}\left(\phi\left([[R_{i_1},R_{i_2}],\ldots,R_{i_k}]_S\right)\right)\\ &=&[[R_{i_1},R_{i_2}],\ldots,R_{i_k}]_S\cdot \prod_{j\in J}R_j\\
&\leq& \left(\bigcap_{s=1}^k R_{i_s}\right)\cdot\prod_{j\in J} R_j\\
&=&\left(\bigcap_{i\in I}R_i\right)\cdot \prod_{j\in J} R_j.\\
\end{array}
$$
This proves that $(R_1,\ldots,R_{n-1})$ is connected. Similarly, each $(R_1,\ldots, \hat{R}_i,\ldots, R_n)$ is connected for $1\leq i< n$ and
hence assertion (2).
(3). From assertion (2), each $(R_1,\ldots, \hat{R}_i,\ldots, R_n)$ is connected for $1\leq i\leq n$. Since $(A_1,\ldots,A_n)$ is a cofibrant partition of $X$ relative to $A$, $X$ is the homotopy colimit of the diagram given by the inclusions
$$
A\subseteq A_{j_1}\cup A_{j_2}\cup\cdots\cup A_{j_q}\subseteq A_{i_1}\cup A_{i_2}\cup\cdots\cup A_{i_p}
$$
with $\{j_1,\ldots,j_q\}\subseteq \{i_1,\ldots,i_p\}\subsetneq \{1,2,\ldots,n\}$. From Van Kampen's theorem
$$
\pi_1(A_{i_1}\cup \cdots\cup A_{i_k})=\pi_1(A)/\prod_{s=1}^k R_{i_s}.
$$
Thus $X=X(\pi_1(A);R_1,R_2,\ldots,R_n)$. From Theorem~\ref{theorem3.1}, we have
$$
\pi_n(X)\cong \frac{R_1\cap \cdots \cap R_n}{\prod_{I\cup J=\{1,\ldots,n\},{I\cap J=\emptyset}}
[\cap_{i\in I}R_i,\cap_{j\in J}R_j]}.
$$
It suffices to show that
$$
\prod_{I\cup J=\{1,\ldots,n\}, {I\cap J=\emptyset}}
\left[\bigcap_{i\in I}R_i,\bigcap_{j\in J}R_j\right]=[[R_1,R_2],\ldots,R_n]_S.
$$
Recall that $$[[R_1,R_2],\ldots,R_n]_S=\prod_{\sigma\in \Sigma_n}[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n)}].$$
For each $\sigma\in\Sigma_n$, let $I=\{\sigma(1),\ldots,\sigma(n-1)\}, J=\{\sigma(n)\}$, then
$$
\begin{array}{rcl}
[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n)}]&=&[[[R_{\sigma(1)},R_{\sigma(2)}],\ldots,R_{\sigma(n-1)}],R_{\sigma(n)}]\\
&&\\
&\leq &\left[\bigcap_{i\in I}R_i,\bigcap_{j\in J}R_j\right]\\
\end{array}
$$
and so
$$
[[R_1,R_2],\ldots,R_n]_S\leq \prod_{I\cup J=\{1,\ldots,n\}, {I\cap J=\emptyset}}
\left[\bigcap_{i\in I}R_i,\bigcap_{j\in J}R_j\right].
$$
Conversely let $I=\{i_1,\ldots,i_p\}$ and $J=\{j_1,\ldots,j_q\}$ with $1\leq p,q\leq n-1$, $I\cup J=\{1,\ldots,n\}$ and $I\cap J=\emptyset$. By assertion (1),
$$
\bigcap_{i\in I}R_i=[[R_{i_1},R_{i_2}],\ldots,R_{i_p}]_S
\textrm{
and }
\bigcap_{j\in J}R_j=[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]_S.
$$
Thus
$$
\begin{array}{rcl}
\left[\bigcap_{i\in I}R_i,\bigcap_{j\in J}R_j\right]&=&\left[ [[R_{i_1},R_{i_2}],\ldots,R_{i_p}]_S,[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]_S\right]\\
&\leq&[[R_1,R_2],\ldots,R_n]_S\\
\end{array}
$$
by Theorem~\ref{theorem1.1} because
$$
\left[ [[R_{i_1},R_{i_2}],\ldots,R_{i_p}]_S,[[R_{j_1},R_{j_2}],\ldots,R_{j_q}]_S\right]\leq [[R_1,R_2,\ldots,R_n]].
$$
This finishes the proof of assertion (3).
Assertion (4) follows from assertion (3) by constructing a new partition as follows: let $B=\bigcup_{j\in J}A_j$ and let $B_s=B\cup A_{i_s}$ for $s=1,2,\ldots,k$. Then
$$
X=\bigcup_{s=1}^kB_s=\bigcup_{i\in I, j\in J}A_i\cup A_j.
$$
and $(B_1,\ldots,B_k)$ is a cofibrant $k$-partition of $X$ relative to $B$. Condition (i) and (ii) hold similar as in the proof of assertion (2). Let
$$
\begin{array}{rcl}
N_s&=&\mathrm{K er}(\pi_1(B)\to \pi_1(B_s))\\
&=&R_{i_s}/(R_{i_s}\cap \prod_{j\in J}R_j).\\
\end{array}
$$
From assertion (3), we have $$
\pi_k(X)\cong \left(\bigcap_{s=1}^kN_s\right)/[[N_{1}, N_2],\ldots,N_k]_S.
$$
To finish the proof, it suffices to show that
\begin{equation}\label{equation3.3}
\begin{array}{rl}
&\left(\bigcap_{s=1}^kN_s\right)/[[N_{1}, N_2],\ldots,N_k]_S\\
\cong &(\bigcap_{s=1}^k(R_{i_s}\cdot \prod_{j\in J}R_j))/([[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S\cdot \prod_{j\in J}R_j ).\\
\end{array}
\end{equation}
Let $\phi\colon G\longrightarrow G/\prod_{j\in J}R_j$ be the quotient homomorphism. Then it is straightforward to check that
$$
\begin{array}{rcl}
\phi^{-1}(N_s)&=&R_{i_s}\cdot \prod_{j\in J}R_j,\\
&&\\
\phi^{-1}\left(\bigcap_{s=1}^kN_s\right)&=& \bigcap_{s=1}^k(R_{i_s}\cdot \prod_{j\in J}R_j),\\
&&\\
\phi^{-1}([[N_{1}, N_2],\ldots,N_k]_S)&=&[[R_{i_1}, R_{i_2}],\ldots,R_{i_k}]_S\cdot \prod_{j\in J}R_j.\\
\end{array}
$$
Thus equation~(\ref{equation3.3}) holds, hence assertion (4).
The proof is finished.
\end{proof}
\section{Applications to the Free Groups and Surface Groups}\label{section4}
\subsection{Subgroups of the Surface Groups}
Let $X=S$ be a path-connected compact $2$-manifold with or without boundary. Let $Q_i$ be a set of finite points in $S\smallsetminus \partial S$, $1\leq i\leq n$, such that
\begin{enumerate}
\item $Q_i\not=\emptyset$ for each $1\leq i\leq n$ and
\item $Q_i\cap Q_j=\emptyset$ for $i\not=j$.
\end{enumerate}
Let $
A=S\smallsetminus\left(\bigcup_{i=1}^n Q_i\right)
$
be the punctured surface and let $A_i=A\cup Q_i$. Then $(A_1,\ldots,A_n)$ is a cofibrant $n$-partition\footnote{One needs to do some modifications such that the cofibrant hypothesis holds: By replacing each punctured point by a small open disk in $S$, the resulting punctured surfaces become compact $2$-manifolds and so the cofibrant hypothesis for the partition $(A_1,\ldots,A_n)$ holds.}\label{footnote1} of $S$ relative to $A$. For any subset $$\{i_1,\ldots,i_k\}\subsetneq \{1,2,\ldots,n\},$$ the space
$$
\bigcup_{s=1}^k A_{i_s}
$$
is a $K(\pi,1)$-space because it is a surface punctured by at least
one point. Observe that each homomorphism
$$
\pi_1(A)\longrightarrow \pi_1(A_i)
$$
is an epimorphism. Let $R_i$ be the kernel of the epimorphism $\pi_1(A)\to \pi_1(A_i)$. By Theorem~\ref{theorem1.3}, we have
\begin{equation}\label{equation4.1}
\pi_n(S)\cong (R_1\cap R_2\cap\cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S.
\end{equation}
We give a group theoretic interpretation of this isomorphism.
\bigskip
\noindent\textbf{Case 1.} $X=S^2$. Let $Q=\bigcup_{i=1}^nQ_i$ with
$$
Q=\{q_1,q_2,\ldots,q_m\}
$$
with a choice of order that $q_i<q_j$ for $i<j$. For each $q\in Q$, let $c_q$ be a generator in $\pi_1(A)$ represented by a small circle around the point $q$ with a choice of orientation such that $\pi_1(A)$ admits the presentation
$$
\pi_1(A)=\left\la c_q \ \left| \ q\in Q\ \prod_{j=1}^mc_{q_j}=1\right. \right\ra.
$$
Then $R_i=\la\la c_q \ | \ q\in Q_i\ra\ra$. From equation~(\ref{equation4.1}), we have the following result.
\begin{thm}\label{theorem4.1}
Let $G=\la x_1,x_2,\ldots,x_m \ | \ x_1x_2\cdots x_m=1\ra$ be the free group of rank $m-1$ with $m\geq2$. Let $n\geq 2$ and let $P_i$ be any subset of $\{x_1,x_2,\ldots,x_m\}$, $1\leq i\leq n$, such that
\begin{enumerate}
\item[(i)] $P_i\not=\emptyset$ for each $1\leq i\leq n$;
\item[(ii)] $P_i\cap P_j=\emptyset$ for $i\not=j$ and
\item[(iii)] $\bigcup_{i=1}^n P_i=\{x_1,x_2,\ldots,x_m\}$.
\end{enumerate}
Let $R_i=\la\la P_i\ra\ra$ be the normal closure of $P_i$ in $G$. Then there is an isomorphism of groups
$$
(R_1\cap R_2\cap \cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S\cong\pi_n(S^2).
$$\hfill $\Box$
\end{thm}
If $n=m$ with $P_i=\{x_i\}$, then there is an isomorphism of groups
$$
(\la\la x_1\ra\ra\cap \la\la x_2\ra\ra\cap\cdots\cap \la\la x_n)/[[\la\la x_1\ra\ra,\la\la x_2\ra\ra], \ldots,\la\la x_n\ra\ra]_S\cong \pi_n(S^2),
$$
which is ~\cite[Theorem 1.7]{Wu1}. One interesting point of Theorem~\ref{theorem4.1} is that the factor group
$$
(R_1\cap R_2\cap\cdots \cap R_n)/[[R_1,R_2],\ldots,R_n]_S
$$
only depends on the length of the partition $(P_1,\ldots,P_n)$ of
$\{x_1,\ldots,x_m\}$, which does not seem obvious from the group
theoretic point of view.
\bigskip
\noindent\textbf{Case 2.} $X=\mathbb{R}\mathrm{P}^2$. Let $Q=\bigcup_{i=1}^nQ_i$ with
$$
Q=\{q_1,q_2,\ldots,q_m\}.
$$
Then $\pi_1(A)$ admits a presentation
$$
\pi_1(A)=\left\la a_1, c_q \ \left| \ q\in Q,\ a_1^2=\prod_{i=1}^mc_{q_i}\right.\right\ra
$$
with $R_i=\la\la c_q \ | \ q\in Q_i\ra\ra$. Note that $\pi_n(\mathbb{R}\mathrm{P}^2)\cong \pi_n(S^2)$ for $n\geq 2$. From equation~(\ref{equation4.1}), we have the following result.
\begin{thm}\label{theorem4.2}
Let $G=\la a_1, x_1,x_2,\ldots,x_m \ | \ a_1^2=x_1x_2\cdots x_m\ra$ with $m\geq 2$. Let $n\geq 2$ and let $P_i$ be any subset of $\{x_1,x_2,\ldots,x_m\}$, $1\leq i\leq n$, such that
\begin{enumerate}
\item[(i)] $P_i\not=\emptyset$ for each $1\leq i\leq n$;
\item[(ii)] $P_i\cap P_j=\emptyset$ for $i\not=j$ and
\item[(iii)] $\bigcup_{i=1}^n P_i=\{x_1,x_2,\ldots,x_m\}$.
\end{enumerate}
Let $R_i=\la\la P_i\ra\ra$ be the normal closure of $P_i$ in $G$. Then there is an isomorphism of groups
$$
(R_1\cap R_2\cap \cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S\cong\pi_n(S^2).
$$\hfill $\Box$
\end{thm}
\noindent\textbf{Remark.} In the above theorem, $G\cong \pi_1(\mathbb{R}\mathrm{P}^2\smallsetminus Q_m)$ is a free group of rank $m$ with the presentation given in the form as in the statement. Thus the subgroups $R_i$ are different from those given in Theorem~\ref{theorem4.1}.
\bigskip
\noindent\textbf{Case 3.} $X\not=S^2$ or $\mathbb{R}\mathrm{P}^2$. Let $Q=\bigcup_{i=1}^nQ_i$ with
$$
Q=\{q_1,q_2,\ldots,q_m\}.
$$
If $X$ is an oriented surface of genus $g$ with $t$ boundary components, then $\pi_1(A)$ admits a presentation
$$
\pi_1(A)=\left\la a_1,b_1,\ldots,a_g,b_g, d_1,\ldots,d_t, c_q \ \left| \ q\in Q \ \prod_{i=1}^g[a_i,b_i]=\prod_{j=1}^td_j\cdot \prod_{i=1}^m c_{q_i}\right. \right\ra
$$
with $R_i=\la\la c_q \ | \ q\in Q_i\ra\ra$. If $X$ is a non-oriented surface of genus $h$ with $t$ boundary components, then $\pi_1(A)$ admits a presentation
$$
\pi_1(A)=\left\la a_1,\ldots,a_h, d_1,\ldots,d_t,c_q \ \left| \ q\in Q \ \prod_{i=1}^h a_i^2=\prod_{j=1}^td_j\cdot \prod_{i=1}^m c_{q_i}\right. \right\ra
$$
with $R_i=\la\la c_q \ | \ q\in Q_i\ra\ra$. Note that $\pi_n(X)=0$ for $n\geq 2$. From equation~(\ref{equation4.1}), we have the following result.
\begin{thm}\label{theorem4.3}
Let $G$ be one of the following surface groups:
\begin{enumerate}
\item[(a)] $\la a_1,b_1,\ldots,a_g,b_g, y_1,\ldots,y_t, x_1,\ldots,x_m \ | \ \prod_{i=1}^g[a_i,b_i]=\prod_{j=1}^ty_j\cdot \prod_{i=1}^m x_i\ra$
with $g>0$ or $t>0$.
\item[(b)] $\la a_1,\ldots,a_h, y_1,\ldots,y_t, x_1,\ldots,x_m \ | \ \prod_{i=1}^h a_i^2=\prod_{j=1}^ty_j\cdot \prod_{i=1}^m x_i\ra$
with $h>1$ or $t>0$.
\end{enumerate}
Let $P_i$ be any subset of $\{x_1,x_2,\ldots,x_m\}$, $1\leq i\leq n$, such that
\begin{enumerate}
\item[(i)] $P_i\not=\emptyset$ for each $1\leq i\leq n$;
\item[(ii)] $P_i\cap P_j=\emptyset$ for $i\not=j$ and
\item[(iii)] $\bigcup_{i=1}^n P_i=\{x_1,x_2,\ldots,x_m\}$.
\end{enumerate}
Let $R_i=\la\la P_i\ra\ra$ be the normal closure of $P_i$ in $G$. Then
$$
R_1\cap R_2\cap\cdots\cap R_n=[[R_1,R_2],\ldots,R_n]_S.
$$\hfill $\Box$
\end{thm}
\subsection{Homotopy Groups of Higher Dimensional Spheres and Free Products of Surface Groups}
It is a natural question whether one can get similar group
theoretical descriptions of the (general) homotopy groups of
higher-dimensional spheres. We give some remarks that the homotopy
groups of certain $2$-dimensional complexes contains the homotopy
groups of all of higher-dimensional spheres as summands. From this,
we can answer the above question in some sense.
Let $X$ and $Y$ be path-connected spaces. According to~\cite{Gray}, there is a homotopy decomposition
\begin{equation}\label{equation4.2}
\Omega(X\vee Y)\simeq \Omega X\times \Omega Y\times \Omega\Sigma(\Omega X\wedge \Omega Y).
\end{equation}
Now let $X=S^2$ and let $Y$ be any surface. Equation~(\ref{equation4.2}) implies the following homotopy decomposition
\begin{equation}\label{equation4.3}
\Omega(S^2\vee Y)\simeq \Omega S^2\times \Omega Y\times\Omega(\Sigma \Omega S^2\wedge \Omega Y).
\end{equation}
From the classical James Theorem~\cite{James1}, there is homotopy decomposition
$$
\Sigma\Omega\Sigma Z\simeq \bigvee_{k=1}^\infty \Sigma Z^{\wedge k},
$$
where $Z^{\wedge k}$ is the $k$-fold self smash product of $Z$. Note that the $k$-fold self smash product of $S^1$ is $S^k$. There is a homotopy decomposition
$$
\Sigma\Omega S^2=\Sigma\Omega\Sigma S^1\simeq \bigvee_{k=1}^\infty \Sigma S^k=\bigvee_{k=2}^\infty S^k.
$$
By substituting this decomposition into formula~(\ref{equation4.3}), there is a homotopy decomposition
$$
\begin{array}{rcl}
\Omega(S^2\vee Y)&\simeq & \Omega S^2\times \Omega Y\times\Omega \left(\bigvee_{k=2}^\infty S^k\wedge\Omega Y\right)\\
&\simeq& \Omega S^2\times \Omega Y \times \left(\prod_{k=2}^\infty \Omega (S^k\wedge \Omega Y) \right) \times\textrm{ other factors}.\\
\end{array}
$$
If $Y$ is a surface with $Y\not=S^2$, $\mathbb{R}\mathrm{P}^2$ and $D^2$, then $\Omega Y$ is homotopy equivalent to a discrete space of a countably infinite set. Then $S^k\wedge Y$ is a wedge of countably infinite copies of $S^k$. Thus given any $k\geq2$,
$\pi_q(S^k)$ is a summand of $\pi_q(S^2\vee Y)$ for each $q$ with infinite multiplicity. If $Y=S^2$ or $\mathbb{R}\mathrm{P}^2$, then we can repeat the above decomposition formula and obtain the fact that, given any $k\geq 2$, $\pi_q(S^k)$ is a summand of $\pi_q(S^2\vee Y)$ for each $q$ (with infinite multiplicity).
Hence it suffices to consider $\pi_*(S^2\vee Y)$ for any surface $Y\not=D^2$. By the same arguments in the previous subsection, we can get partitions of the space $S^2\vee Y$ by removing points from $S^2$ and $Y$. This gives the following result.
\begin{thm}
Let $G_1=\la x_{11},x_{12},\ldots,x_{1m} \ | \ x_{11}x_{12}\cdots x_{1m}=1\ra$ and let $G_2$ be one of the following surface groups:
\begin{enumerate}
\item[(a)] $\la a_1,b_1,\ldots,a_g,b_g, y_1,\ldots,y_t, x_{21},\ldots,x_{2m} \ | \ \prod\limits_{i=1}^g[a_i,b_i]=\prod\limits_{j=1}^ty_j\cdot \prod\limits_{i=1}^m x_{2i}\ra$.
\item[(b)] $\la a_1,\ldots,a_h, y_1,\ldots,y_t, x_{21},\ldots,x_{2m} \ | \ \prod\limits_{i=1}^h a_i^2=\prod\limits_{j=1}^ty_j\cdot \prod\limits_{i=1}^m x_{2i}\ra$.
\end{enumerate}
Let $P_{ji}$ be any subset of $\{x_{j1},x_{j2},\ldots,x_{jm}\}$, $j=1,2$ and $1\leq i\leq n$, such that
\begin{enumerate}
\item[(i)] $P_{ji}\not=\emptyset$ for each $1\leq i\leq n$;
\item[(ii)] $P_{ji}\cap P_{jk}=\emptyset$ for $i\not=k$ and
\item[(iii)] $\bigcup_{i=1}^n P_{ji}=\{x_{j1},x_{j2},\ldots,x_{jm}\}$.
\end{enumerate}
Let $R_i=\la\la P_{1i},P_{2i}\ra \ra$ be the normal closure of $P_{1i}\cup P_{2i}$ in the free product $G=G_1\ast G_2$. Then there is an isomorphism of groups
$$
\pi_n(S^2\vee Y)\cong (R_1\cap R_2\cap\cdots\cap R_n)/[[R_1,R_2],\ldots,R_n]_S,
$$
where $Y$ is a surface such that the fundamental group of $Y$ punctured by $m$ points is the group $G_2$.\hfill $\Box$
\end{thm}
\section{Applications to Braid Groups and the Proof of Theorem~\ref{theorem1.5}}\label{section5}
\subsection{Applications to the Braid Groups}
Recall that a braid of $n$ strands is called \textit{Brunnian} if deleting any one of the strands produces a trivial braid of $(n-1)$-strands. Let $\mathrm{Brun}_n$ denote the Brunnian subgroup of the $n\,$th pure Artin braid group $P_n$. (\textbf{Note.} An $n$-strand Brunnian braid is always a pure braid for $n\geq3$ according to ~\cite[Proposition 4.2.2]{BCWW}.) The group $\mathrm{Brun}_n$ has been characterized by Levinson ~\cite{Levinson1,Levinson2} for $n\leq4$. A classical question proposed by Makanin~\cite{Makanin} in 1980 is to determine a set of generators for $\mathrm{Brun}_n$. This question has been answered by Johnson~\cite{Johnson} and Stanford~\cite{Stanford}. An answer using fat commutators was given in~\cite[Theorem 8.6.1]{BCWW}. As an application, we gave a generalization of Levinson's result~\cite[Theorem 2]{Levinson2}.
Recall that the braid group $B_n$ is generated by $\sigma_1,\ldots,\sigma_{n-1}$ with defining relations given by
\begin{enumerate}
\item $\sigma_i\sigma_j=\sigma_j\sigma_i$ for $|i-j|\geq 2$;
\item $\sigma_i\sigma_{i+1}\sigma_i=\sigma_{i+1}\sigma_i\sigma_{i+1}$.
\end{enumerate}
Following Levinson's notation in~\cite{Levinson2}, let
$$
t_i=\sigma_i\sigma_{i+1}\cdots\sigma_{n-2}\sigma_{n-1}^2\sigma_{n-2}^{-1}\cdots\sigma_i^{-1}
$$
for $1\leq i\leq n-1$. Intuitively $t_i$ is the braid that links strand $i$ and strand $n$ in front of all other strands. Let $R_i=\la\la t_i\ra\ra$ be the normal closure of $t_i$ in $P_n$. (\textbf{Note.} Levinson uses the notation $\theta_i$.)
\begin{thm}\label{theorem5.1}
For each $n\geq2$, $\mathrm{Brun}_n=[[R_1,R_2],\ldots,R_{n-1}]_S$.
\end{thm}
When $n=4$, we have $\mathrm{Brun}_4=[[R_1,R_2],R_3]_S=[[R_1,R_2],R_3]\cdot [[R_1,R_3],R_2]$, which is exactly Levinson's theorem~\cite[Theorem 2]{Levinson2}. Let $M$ be any manifold. Recall that the coordinate projection
$$
F(M,n)\longrightarrow F(M,k)\quad (z_1,\ldots,z_n)\mapsto (z_{i_1},\ldots,z_{i_k}) \textrm{ for given } i_1<\cdots<i_k
$$
is a fibration by Fadell-Neuwirth's Theorem~\cite{FN}. The coordinate projection
$$
d_j\colon F(\ensuremath{\mathbb{R}}^2,n)\longrightarrow F(\ensuremath{\mathbb{R}}^2,n-1),\quad (z_1,\ldots,z_n)\mapsto (z_1,\ldots,\hat z_j,\ldots, z_n)$$ induces a group homomorphism
$$
d_{j*}\colon P_n=\pi_1(F(\ensuremath{\mathbb{R}}^2,n))\longrightarrow P_{n-1}=\pi_1(F(\ensuremath{\mathbb{R}}^2,n-1))
$$
which is given by removing the $j\,$th strand.
\begin{proof}[Proof of Theorem~\ref{theorem5.1}]
The basepoint $(q_1, q_2, \ldots, q_n)$ of $F(\ensuremath{\mathbb{R}}^2,n)$ is chosen in the way that the points $q_i\in\ensuremath{\mathbb{R}}^1\subseteq \ensuremath{\mathbb{R}}^2$ with the order $q_1<q_2<\cdots<q_n$. Let $Q_n=\{q_1,\ldots,q_n\}$ and let $Q_{n,i}=\{q_1,\ldots,\hat q_i,\ldots,q_n\}$. For each $1\leq i\leq n-1$, there is a commutative diagram of fibrations
\begin{diagram}
\ensuremath{\mathbb{R}}^2\smallsetminus Q_{n-1}&\rInto^{f}& F(\ensuremath{\mathbb{R}}^2,n)&\rOnto^{d_n}& F(\ensuremath{\mathbb{R}}^2,n-1)\\
\dInto>{g_i}&&\dTo>{d_i}&&\dTo>{d_i}\\
\ensuremath{\mathbb{R}}^2\smallsetminus Q_{n-1,i}&\rInto^{h_i}& F(\ensuremath{\mathbb{R}}^2,n-1)&\rOnto^{d_{n-1}}& F(\ensuremath{\mathbb{R}}^2,n-2).\\
\end{diagram}
By taking fundamental groups and using the fact that $F(\ensuremath{\mathbb{R}}^2,m)$ is a $K(\pi,1)$-space, there is a commutative diagram of short exact sequences of groups
\begin{diagram}
\pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{n-1})&\rInto^{f_{*}}& P_n=\pi_1(F(\ensuremath{\mathbb{R}}^2,n))&\rOnto^{d_{n*}}& P_{n-1}=\pi_1(F(\ensuremath{\mathbb{R}}^2,n-1))\\
\dInto>{g_{i*}}&&\dTo>{d_{i*}}&&\dTo>{d_{i*}}\\
\pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{n-1,i})&\rInto^{h_{i*}}& P_{n-1}=\pi_1(F(\ensuremath{\mathbb{R}}^2,n-1))&\rOnto^{(d_{n-1})_*}& P_{n-2}=\pi_1(F(\ensuremath{\mathbb{R}}^2,n-2)).\\
\end{diagram}
Let
$$
R'_i=\mathrm{K er}(g_{i*}\colon \pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{n-1})\to \pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{n-1,i})).
$$
Since $h_{i*}$ is a monomorphism,
$$
f_{*}|\colon R'_i\longrightarrow \mathrm{K er}(d_{n*}\colon P_n\to P_{n-1})\cap\mathrm{K er}(d_{i*}\colon P_n\to P_{n-1})
$$
is an isomorphism. Observe that $R'_i$ is the normal closure of the elements in $\pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{n-1})$ represented by the small circles around the point $q_i$. By using the terminology of braids, the small circles around the point $q_i$ is represented by ~$t_i^{\pm}$. Thus
$$
\mathrm{K er}(d_{n*}\colon P_n\to P_{n-1})\cap\mathrm{K er}(d_{i*}\colon P_n\to P_{n-1})\leq R_i.
$$
On the other hand, since $d_{n*}(t_i)=d_{i*}(t_i)=1$, we have
$$
R_i\leq \mathrm{K er}(d_{n*}\colon P_n\to P_{n-1})\cap\mathrm{K er}(d_{i*}\colon P_n\to P_{n-1})
$$
and so
\begin{equation}\label{R_i}
\mathrm{K er}(d_{n*}\colon P_n\to P_{n-1})\cap\mathrm{K er}(d_{i*}\colon P_n\to P_{n-1})=R_i
\end{equation}
with the property that
$$
f_*\colon R'_i\longrightarrow R_i
$$
is an isomorphism. If $w\in \bigcap\limits_{i=1}^{n-1}R_i$, then $w\in \mathrm{Im}(f_*)$ and $f_*^{-1}(w)\in R'_i$ for each $1\leq i\leq n-1$. Thus the monomorphism $f_*$ induces an isomorphism
$$
f_*\colon \bigcap_{i=1}^{n-1}R'_i\longrightarrow \bigcap_{i=1}^{n-1} R_i.
$$
By applying Theorem~\ref{theorem4.3} to the punctured disks, we have
$$
R'_1\cap R'_2\cap\cdots\cap R'_{n-1}=[[R'_1,R'_2],\ldots,R'_{n-1}]_S.
$$
It follows that
$$
R_1\cap R_2\cap\cdots\cap R_{n-1}=[[R_1,R_2],\ldots,R_{n-1}]_S.
$$
From equality~(\ref{R_i}), we have
$$
\bigcap_{i=1}^{n-1} R_i=\bigcap_{i=1}^{n-1}\mathrm{K er}(d_{n*})\cap\mathrm{K er}(d_{i*})=\bigcap_{j=1}^{n}\mathrm{K er}(d_{j*}\colon P_n\to P_{n-1})=\mathrm{Brun}_n,
$$
hence the result.
\end{proof}
\subsection{Proof of Theorem~\ref{theorem1.5}}
The key point to proving Theorem~\ref{theorem1.5} is to construct a $K(\pi,1)$-partition for the configuration space $F(S^2,n)$.
Before giving the proof, we need some lemmas. Let $q_1,q_2,\ldots$ be strictly increasing real numbers that lie in the open interval $(-1,1)\subseteq \ensuremath{\mathbb{R}}^1\subseteq \ensuremath{\mathbb{R}}^2\subseteq S^2$. The base point of the configuration spaces $F(\ensuremath{\mathbb{R}}^2,m)\subseteq F(S^2,m)$ is chosen to be $(q_1,\ldots,q_m)$ for each $m\geq 1$.
\begin{lem}\label{lemma5.2}
Let $1\leq i\leq m$ and let
$$
B_i=\{(z_1,\ldots,z_m)\in F(S^2,m)\ | \ z_p\not=\infty\textrm{ for } p\not=i\}.
$$
Let $f\colon F(\ensuremath{\mathbb{R}}^2,m)\to B_i$ be the inclusion. Then the homomorphism
$$
f_*\colon P_m=\pi_1(F(\ensuremath{\mathbb{R}}^2,m))\longrightarrow \pi_1(B_i)
$$
is an epimorphism with $\mathrm{K er}(f_*)=\la\la A_{0,i}\ra\ra^{P_m}$, the normal closure of $A_{0,i}$ in $P_m$.
\end{lem}
\begin{proof}
The assertion holds trivially for $m=1$. We may assume that $m\geq 2$. Consider the Fadell-Neuwirth fibration
$$
d_i\colon F(S^2,m)\longrightarrow F(S^2,m-1)\quad (z_1,\ldots,z_n)\mapsto (z_1,\ldots,\hat z_i,\ldots,z_m).
$$
Then $B_i=d_i^{-1}(F(\ensuremath{\mathbb{R}}^2,m-1))$ by the construction of $B_i$. Consider the commutative diagram of the fibrations
\begin{diagram}
\ensuremath{\mathbb{R}}^2\smallsetminus Q_{m,i}&\rInto& F(\ensuremath{\mathbb{R}}^2,m)&\rTo& F(\ensuremath{\mathbb{R}}^2,m-1)\\
\dInto>{f|}&&\dInto>{f}&&\dEq\\
S^2\smallsetminus Q_{m,i}&\rInto& B_i&\rTo&F(\ensuremath{\mathbb{R}}^2,m-1),\\
\end{diagram}
where $Q_{m,i}=\{q_1,\ldots,q_{i-1},q_{i+1},\ldots,q_m\}$. Thus there is a commutative diagram of short exact sequence of groups
\begin{diagram}
\pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{m,i})&\rInto& \pi_1(F(\ensuremath{\mathbb{R}}^2,m))&\rOnto& \pi_1(F(\ensuremath{\mathbb{R}}^2,m-1))\\
\dTo>{f|_*}&&\dTo>{f_*}&&\dEq\\
\pi_1(S^2\smallsetminus Q_{m,i})&\rInto& \pi_1(B_i)&\rOnto&\pi_1(F(\ensuremath{\mathbb{R}}^2,m-1)).\\
\end{diagram}
Since $f|_*\colon \pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{m,i})\to \pi_1(S^2\smallsetminus Q_{m,i})$ is an epimorphism, the middle map
$$
f_*\colon \pi_1(F(\ensuremath{\mathbb{R}}^2,m))\longrightarrow \pi_1(B_i)
$$
is an epimorphism with $\mathrm{K er}(f_*)=\mathrm{K er}(f|_*)$. Note that $\mathrm{K er}(f|_*)$ is the normal closure of the element $[\omega_i]$ in $\pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{m,i})$, where the loop $\omega_i$ is given by the following left picture:
\begin{center}
\epsfig{file=fig, bbllx=0, bblly=0, bburx=488, bbury=72, width=4in,
height=4in, clip=}
\end{center}
\vspace*{-5cm}
In the above picture, the braid $A_{0,i}$ is obtained
from the loop $\omega_i$ by drawing it as a braid. Thus
$[\omega_i]=A_{0,i}$ and so
$$
\mathrm{K er}(f_*)=\mathrm{K er}(f|_*)=\la\la A_{0,i}\ra\ra^{P_m}\cap \pi_1(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{m,i})\leq \la\la A_{0,i}\ra\ra^{P_m}.
$$
On the other hand, since $\mathrm{K er}(f_*)$ is normal in $P_m$ with $A_{0,i}\in \mathrm{K er}(f_*)$, we have $\la\la A_{0,i}\ra\ra^{P_m}\leq \mathrm{K er}(f_*)$. It follows that
$$
\mathrm{K er}(f_*)=\la\la A_{0,i}\ra\ra^{P_m},
$$
hence the result holds.
\end{proof}
\begin{lem}\label{lemma5.3}
Let $I=\{i_1,i_2,\ldots,i_k\}$ be a subset of $\{1,2,\ldots,m\}$. Let
$$
B_I=\{(z_1,\ldots,z_m)\in F(S^2,m) \ | \ z_p\not=\infty\textrm{ for } p\not\in I\}
$$
and let $f\colon F(\ensuremath{\mathbb{R}}^2,m)\to B_I$ be the inclusion. Then the homomorphism
$$
f_*\colon \pi_1(F(\ensuremath{\mathbb{R}}^2,m))\to \pi_1(B_I)
$$
is an epimorphism whose kernel
$
\mathrm{K er}(f_*)=\la\la A_{0,i}\ | \ i\in I\ra\ra^{P_m}.
$
\end{lem}
\begin{proof}
Note that
$$
B_I=\bigcup_{s=1}^k B_{i_s}.
$$
By Lemma~\ref{lemma5.2}, $P_m=\pi_1(F(\ensuremath{\mathbb{R}}^2,m))\to \pi_1(B_{i_s})$ is an epimorphism whose kernel is given by $\la\la A_{0,i_s}\ra\ra^{P_m}$ for each $1\leq s\leq k$. From Van Kampen's Theorem,
$$
f_*\colon P_m=\pi_1(F(\ensuremath{\mathbb{R}}^2,m))\longrightarrow \pi_1(B_I)
$$
is an epimorphism with
$$
\mathrm{K er}(f_*)=\prod_{s=1}^k \la\la A_{0,i_s}\ra\ra=\la\la A_{0,i_s}\ | \ 1\leq s\leq k\ra\ra^{P_m},
$$
hence the result.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{theorem1.5}]
Let $A=F(\ensuremath{\mathbb{R}}^2,m)$ be the subspace of $F(S^2,m)$. Let
$$
A_i=\{(z_1,z_2,\ldots,z_m)\in F(S^2,m)\ | \ z_p\not=\infty \textrm{ for } p\not\in \Lambda_i\},
$$
where the $\Lambda_i$ are as defined in the statement of the Theorem.
Since $\Lambda_i\cap \Lambda_j=\emptyset$ for $i\not=j$, we have
$$
A_i\cap A_j=F(\ensuremath{\mathbb{R}}^2,m)=A
$$
for $i\not=j$. From the assumption that $\bigcup_{i=1}^n \Lambda_i=\{1,2,\ldots,m\}$, we have
$$
\bigcup_{i=1}^n A_i=F(S^2,m).
$$
Thus $(A_1,A_2,\ldots,A_n)$ is a cofibrant $n$-partition \footnote{Similar to footnote$^2$, one needs to do some modifications such that the cofibrant hypothesis holds: Take a triangulation on $F(S^2,m)$ and deform linearly $A$ and each $A_i$ into certain subcomplexes with necessary subdivisions for having a cofibrant $n$-partition of $F(S^2,m)$.} of the space $F(S^2,m)$ relative to $F(\ensuremath{\mathbb{R}}^2,m)$. For a subset $I=\{i_1,i_2,\ldots,i_k\}\subsetneq\{1,2,\ldots,n\}$, let
$$
A_I=\bigcup_{s=1}^k A_{i_s}
$$
with $A_{\emptyset}=A$. We now check that $A_I$ is a $K(\pi,1)$-space for $I\subsetneq \{1,2,\ldots,n\}$. Let
$$
J=\{j_1,j_2,\ldots,j_{t}\}=\{1,2,\ldots,m\}\smallsetminus \left(\bigcup_{s=1}^k\Lambda_{i_s}\right)
$$
with $j_1<j_2<\ldots<j_t$. Since $I\not=\{1,2,\ldots,n\}$, $J$ is not empty. Let $(z_1,\ldots,z_m)\in F(S^2,m)$. Observe that $(z_1,\ldots,z_m)\in A_I$ if and only if $z_{j_s}\not=\infty$ for $1\leq s\leq t$. Consider the Fadell-Neuwirth fibration~\cite{FN}
$$
d_I\colon F(S^2,m)\longrightarrow F(S^2,t).
$$
There is a pull-back diagram
\begin{diagram}
A_I&\rInto& F(S^2,m)\\
d_I|_{A_I}\dTo&\mathrm{pull}&\dTo>{d_I}\\
F(\ensuremath{\mathbb{R}}^2,t)&\rInto&F(S^2,t)\\
\end{diagram}
with a fibration
$$
F(S^2\smallsetminus Q_t, m-t)\rTo A_I\rTo F(\ensuremath{\mathbb{R}}^2,t),
$$
where $Q_t$ is a set of $t$ distinct points in $S^2$. Since $t\geq 1$,
$$
F(S^2\smallsetminus Q_t,m-t)\cong F(\ensuremath{\mathbb{R}}^2\smallsetminus Q_{t-1}, m-t)
$$
is a $K(\pi,1)$-space. Together with the fact that the base space $F(\ensuremath{\mathbb{R}}^2,t)$ is a $K(\pi,1)$-space, the total space $A_I$ is a $K(\pi,1)$-space for any $I\subsetneq \{1,2,\ldots,n\}$.
By Lemma~\ref{lemma5.3}, the inclusion $f\colon A=F(\ensuremath{\mathbb{R}}^2,m)\hookrightarrow A_i$ induces an epimorphism
$$
f_*\colon P_m=\pi_1(A)\rOnto \pi_1(A_i)
$$
whose kernel
$$
R_i=\mathrm{K er}(f_*)=\la\la A_{0,j}\ | \ j\in \Lambda_i\ra\ra^{P_m}.
$$
The assertion follows from Theorem~\ref{theorem1.3}.
\end{proof}
\noindent\small{\textbf{Acknowledgements.}} The authors would like to thank the referee for pointing out some errors in the original manuscript. The authors also would like to thank Fred Cohen for his help on the writing of this article.
|
\section{Introduction}
\label{sec:introduction}
Weak gravitational lensing refers to the mild distortion of the light from distant sources by the large-scale matter inhomogeneity between the source and the observer. One observable effect of weak gravitational lensing is the coherent shape distortion of the light sources, known as cosmic shear. Cosmic shear is sensitive to all cosmological parameters which have influence on the density perturbations and/or the geometry of the universe, including those concerning properties of dark energy, which have been a key concern after the discoveries made by observations of supernovae, the cosmic microwave background, and the large-scale structure \citep[for a review see e.g.][]{munshi08}.
Since its first detection in 2000 \citep{bacon00, kaiser00, vwaer00, wittman00}, cosmic shear has been developed into a competitive cosmological probe. Its constraining power on cosmological parameters is now comparable to other probes \citep[e.g.][]{spergel07,fu08}.
With forthcoming large field multicolor imaging surveys (e.g. DES\footnote{\texttt{http://www.darkenergysurvey.org/}}, KIDS\footnote{\texttt{http://www.astro-wise.org/projects/KIDS/}}, EUCLID\footnote{\texttt{http://sci.esa.int/science-e/www/area/\\index.cfm?fareaid=102}}, etc), photometric redshift and shape information of a huge number of galaxies will be available, rendering cosmic shear even greater statistical power. In particular, cosmic shear is considered to be one of the most promising dark energy probes \citep{TaskForce06,ESO-ESA06} when the results of these surveys become available.
Such constraining power will be further enhanced by the use of higher-order statistics. Second-order statistics measure only the Gaussian signature of a random field. Even if the primordial cosmic density field is Gaussian, non-Gaussianity will be generated due to the nonlinear nature of gravitational clustering. Such non-Gaussianity in the cosmic density field will then show up via its lensing effect, leading to non-Gaussian signals in the cosmic shear field. A common way of measuring non-Gaussianity is to use higher-order statistics. In cosmic shear studies, several authors have shown that the lowest order of them, i.e. the third-order statistics, already provide a valuable probe for cosmological parameter estimates; in particular it can break the near degeneracy between the density parameter $\Omega_{\rm m}$ and the power spectrum normalization $\sigma_{\rm 8}$ \citep{ber97,jain97,vwaer99,hui99}. A more recent study by \citet{TJ04} (TJ04 afterwards) showed that including third-order statistics can improve parameter constraints significantly, typically by a factor of three.
But the ultimate performance of these future surveys still largely depends on the how well the systematic errors can be controlled \citep[e.g.][]{hut06}. In this paper we focus on a particularly worrisome systematic error in cosmic shear studies: the intrinsic-shear alignment, and demonstrate a way to control it for shear three-point statistics.
In the weak lensing limit the observed ellipticity of a galaxy $\epsilon_{\rm obs}$ can be written as the sum of the intrinsic ellipticity $\epsilon_{\rm I}$ of the galaxy, and the shear $\gamma$ which is caused by gravitational lensing of the foreground matter distribution. Here $\epsilon_{\rm obs}$, $\epsilon_{\rm I}$ and $\gamma$ are complex quantities. Intrinsic-shear alignment is defined in two-point cosmic shear statistics as the correlation between the intrinsic ellipticity of one galaxy and the shear of another galaxy (the GI term, \citealp{hirata04}). Three-point statistics $\ba{\epsilon_{\rm obs}^i\epsilon_{\rm obs}^j\epsilon_{\rm obs}^k}$, a correlator of ellipticities of three galaxy images $i$, $j$ and $k$, can also be expanded into lensing (GGG), intrinsic-shear (GGI and GII), and intrinsic (III) terms:
\eq{
\label{eq:split}
\ba{\epsilon_{\rm obs}^i \epsilon_{\rm obs}^j \epsilon_{\rm obs}^k} = \rm{GGG} + \rm{GGI} + \rm{GII} + \rm{III} \;,\mbox{with}
}
\eq{
\textrm{GGG} = \ba{\gamma^i \gamma^j \gamma^k} \;,
}
\eq{
\textrm{GGI} = \ba{\epsilon^i_{\textrm I} \gamma^j \gamma^k} + \ba{\epsilon^j_{\textrm I} \gamma^k \gamma^i} +\ba{\epsilon^k_{\textrm I} \gamma^i \gamma^j} \;,
}
\eq{
\textrm{GII} = \ba{\epsilon^i_{\textrm I} \epsilon^j_{\textrm I} \gamma^k} + \ba{\epsilon^j_{\textrm I} \epsilon^k_{\textrm I} \gamma^i} +\ba{\epsilon^k_{\textrm I} \epsilon^i_{\textrm I} \gamma^j} \;,
}
\eq{
\textrm{III} = \ba{\epsilon^i_{\textrm I} \epsilon^j_{\textrm I} \epsilon^k_{\textrm I}} \;.
}
Physically, if one assumes that galaxies are randomly oriented on the sky, only the desired GGG term remains on the right-hand side of (\ref{eq:split}). However, when these galaxies are subject to the tidal gravitational force of the same matter structure (e.g. they formed under the influence of the same massive dark matter halo), their shapes can intrinsically align and become correlated, giving rise to a nonvanishing III term. Furthermore, GGI and GII terms can be generated when a matter structure tidally influences close-by galaxies and at the same time contributes to the shear signal of background objects, leading to correlations among them.
In two-point statistics, the corresponding intrinsic (II) and intrinsic-shear (GI) terms have been subject to detailed studies both theoretically (e.g. \citet{catelan01, croft00, heavens00, hui02, mackey02, jing02, hirata04, heymans06, bridle07, schneider09})
and observationally \citep{brown02, heymans04, mandel06, mandel09, hirata07, fu08, brainerd09, okumura09a, okumura09b}. Although the results of these studies show large variations, most of them are consistent with a 10$\,\%$ contamination by both II and GI correlations for future surveys with photometric redshift information. Especially, neglecting these correlations can bias the dark energy equation of state parameter $w_0$ by as much as $50\,\%$ \citep{bridle07b} for a ``shallow'' survey described in \citet{amara07}. For three-point shear statistics, there have been few measurements up to now \citep{ber02,pen03,jarvis04}. However the potential systematics level in these studies is found to be high. A recent numerical study by \citet{sembo08} showed that intrinsic alignments affect three-point weak lensing statistics more strongly than at the two-point level for a given survey depth. In particular, neglecting GGI and GII systematics would lead to an underestimation of the GGG signal by ~$5-10\,\%$ for a moderately deep survey like the CFHTLS Wide. Therefore, to match the statistical power expected for cosmic shear in the future surveys, it is essential to control these systematics.
The intrinsic alignment, II\,(III) in the two-\,(three-) point case, is relatively straightforward to eliminate, since it requires that the galaxies in consideration are physically close to each other, i.e. have very similar redshifts and angular positions \citep{king02,king03,HH03,TW04}. The control of intrinsic-shear systematics, GI for the two-point case and GGI in the three-point case (GII also requires that two of the three galaxies are physically close and thus can be eliminated in the same way as II and III), turns out to be a much greater challenge. However, as already pointed out by HS04, the characteristic dependence on galaxy redshifts is a valuable piece of information that helps to control the intrinsic-shear alignments.
Several methods for this have already been constructed in the context of two-point statistics. They can be roughly classified into three categories: modeling \citep{king05,bridle07b}, nulling (\citealp{JS08b}, JS08 hereafter; \citealp{joachimi09a}) and self-calibration \citep{zhang08, joachimi09c}. Modeling separates cosmic shear from the intrinsic-shear alignment effect by constructing template functions for the latter. It suffers from uncertainties of the model due to the lack of knowledge of the angular scale and redshift dependence of the intrinsic-shear signal. The nulling technique employs the characteristic redshift dependence of the intrinsic-shear signal to ``null it out''. It is a purely geometrical method and is model-independent, but suffers from a significant information loss. Self-calibration intends to solve the problem of information loss by using additional information from the galaxy distribution to ``calibrate'' the signal. The original form of self-calibration, proposed by \citet{zhang08}, is model-independent but strong assumptions have been made. \citet{joachimi09c} then develop it into a modeling method, by treating intrinsic alignments and galaxy biasing as free functions of scale and redshift.
All these methods have the potential of being generalized to three-point statistics. In this paper we focus on the nulling technique, and establish it as a method to reduce the three-point intrinsic-shear alignments GGI and GII. Since GII can be removed by discarding close pairs of galaxies as in the case of II controlling \citep[e.g.][]{HH03}, we focus on the control of GGI systematics.
As known from the case of two-point statistics, the nulling technique introduces significant information loss while (in principle) completely removing the intrinsic-shear alignment from the signal. In this work we compare the nulling technique to an unconditioned linear compression of the data, distinguish different sources of such information loss, and discuss the possible ways of reducing it. We also study the combined constraints on cosmological parameters with both two- and three-point cosmic shear statistics.
In Sect.$\,$2 we demonstrate why and how the nulling technique can be applied to three-point lensing statistics. We then apply the nulling technique to the modeled lensing bispectrum which we contaminate by intrinsic-shear alignment. The modeling details are described in Sect.$\,$3. The method of nulling weights construction and the corresponding results are shown in Sect.$\,$4, while the results concerning the constraints on cosmological parameters are presented in Sect.$\,$5. We conclude in Sect.$\,$6.
We will work in the context of a spatially flat CDM cosmology with a variable dark energy whose equation of state $w$ is parameterized as $w = w_0 + w_a (1-a)$, with $a$ the cosmic scale factor.
The adopted fiducial values for cosmological parameters are
$\Omega_{\rm m}=0.3$,
$\Omega_{\rm b}=0.045$,
$\Omega_{\rm de}=0.7$,
$w_0=-0.95$,
$w_a=0.0$,
$h=0.7$,
$n_s=1.0$, and
$\sigma_8=0.8$. Here, $\Omega_{\rm m}$, $\Omega_{\rm b}$ and $\Omega_{\rm de}$ are the density parameters of the matter (including cold dark matter and baryons), baryons and the dark energy at present time, $n_s$ is the spectral index of the primordial power spectrum of scalar perturbations, $h$ is the dimensionless Hubble parameter defined by $H_0=100\, h\, {\textrm{km/s/Mpc}}$, and $\sigma_8$ is the rms mass fluctuation in spheres of radius $8h^{-1}$Mpc.
\section{The nulling technique applied to three-point shear tomography}
\label{sec:theory}
\subsection{Principle of the nulling technique}
\label{sec:nulling}
The shear on the image of a distant galaxy is a result of gravitational distortion of light caused by the inhomogeneous three-dimensional matter distribution in the foreground of that galaxy. For notational simplicity, we will use the dimensionless surface mass density (the convergence) $\kappa$ instead of the shear ${\gamma}$ as a measure for the lensing signal throughout the paper, although in reality the signal is based on the measurement of the shear. This will not affect our results since $\kappa$ and ${\gamma}$ are linearly related on each redshift plane while our method is dealing with the redshift dependence of them (the same reason justifies the turning to the Fourier domain in the next subsection).
When one measures the shear ${\gamma}$, the direct observable is the galaxy ellipticity $\epsilon_{\rm obs}=\epsilon_{\rm I}+{\gamma}$. The shear ${\gamma}$ is a signal caused by gravitational distortion which is a deterministic process, where the intrinsic ellipticity $\epsilon_{\rm I}$ can be further written as the sum of a deterministic part $\epsilon_{\rm I}^{\rm det}$ which is caused by intrinsic alignment, and a stochastic part $\epsilon_{\rm I}^{\rm ran}$ which does not correlate with any other quantity. There is no correlation between $\epsilon_{\rm I}^{\rm ran}$ of different galaxies either.
We define $\kappa_{\rm obs}$ and $\kappa_{\rm I}$ which are the correspondences of $\epsilon_{\rm I}^{\rm det}+{\gamma}$ and $\epsilon_{\rm I}^{\rm det}$. We remove the stochastic part since $\kappa$ is deterministic. Note that $\kappa_{\rm obs}$ and $\kappa_{\rm I}$ are analogs of the dimensionless surface mass density $\kappa$ but do not have any direct physical meaning as $\kappa$ does. They are complex quantities in general and can lead to a B-mode signal. To better distinguish the real measurable $\kappa$ from them, we denote it as $\kappa_{\rm G}$ in the rest of the paper since it is the physical quantity which is related to the gravitational lensing signal. Keeping the dominating linear term, the convergence $\kappa_{\rm G}$ can be written as \citep[details see e.g.][]{s06}:
\eq{
\label{eq:d2k}
\kappa_{\rm G}(\vek{\theta},\chi_{\rm s}) = \frac{3 \, \Omega_{\rm m} H_0^2 }{2 c^2} \int_0^{\chi_{\rm s}} d\chi\, \frac{\chi (\chi_{\rm s}-\chi)}{\chi_{\rm s}} ~\frac{\delta\br{\chi \vek{\theta},\chi}}{a(\chi)}\;,
}
\noindent where $\delta$ is the three-dimensional matter density contrast, $\chi_{\rm s}$ is the comoving distance of the background galaxy which is acting as a source, and $a(\chi)$ is the cosmic scale factor at the comoving distance $\chi$ of $\delta$ which is acting as a lens.
Equation (\ref{eq:d2k}) clearly shows that the contribution of the matter inhomogeneity $\delta$ at comoving distance $\chi_{i}$ to the cosmic shear signal of background galaxies can be considered as a function of the source distance $\chi_{\rm s}$, and this function is proportional to $1-{\chi_{i}}/{\chi_{\rm s}}$. The nulling technique takes advantage of this characteristic dependence on source distance $\chi_{\rm s}$ by constructing a weight function $T(\chi_i, \chi_{\rm s})$ such that the product of $T(\chi_i, \chi_{\rm s})$ and $1-{\chi_{i}}/{\chi_{\rm s}}$ has an average of zero on the range between $\chi_{i}$ and the comoving distance to the horizon $\chi_{\textrm{hor}}$:
\eq{
\label{eq:nc}
\int_{\chi_{i}}^{\chi_{\textrm{hor}}} d\chi_{\rm s} ~T(\chi_i, \chi_{\rm s})\, \br{1 - \frac{\chi_{i}}{\chi_{\rm s}}} = 0\;.
}
\noindent One then uses this weight function as a weight for integrating over the source distance:
\eq{
\label{eq:projectedkappanew}
\hat{\kappa}_{\rm G}(\chi_i, \vek{\theta}) := \int_{\chi_{i}}^{\chi_{\textrm{hor}}} d\chi_{\rm s}\; T(\chi_i, \chi_{\rm s})\; \kappa_{\rm G}(\vek{\theta},\chi_{\rm s})\;.
}
\noindent The resulting new measure of shear signal $\hat{\kappa}_{\rm G}(\chi_i, \vek{\theta}) $ is then free of contributions from the matter inhomogeneity at distance $\chi_{i}$. Note that although the weight function $T$ has two arguments $\chi_i$ and $\chi_{\rm s}$ here, we consider it as a function of $\chi_{\rm s}$ for a particular $\chi_{i}$.
Consider a correlator $\ba{{\kappa}_{\rm obs}^{i}{\kappa}_{\rm obs}^{j}}$ with comoving distances $\chi_{i}< \chi_{j}$. With a similar decomposition as (\ref{eq:split}), it is straightforward to see that the GI term in it is $\ba{{\kappa}_{\rm I}^{i}{\kappa}_{\rm G}^{j}}$. The term $\ba{{\kappa}_{\rm G}^{i}{\kappa}_{\rm I}^{j}}$ vanishes since the lensing signal at $\chi_{i}$ is correlated only with matter with $\chi \le \chi_{i}$, whereas ${\kappa}_{\rm I}^{j}$ originates solely from physical processes happening at $\chi_{j}$. If we integrate $\ba{{\kappa}_{\rm I}^{i}{\kappa}_{\rm G}^{j}}$ over $\chi_j$ with a weight function that eliminates the contributions to ${\kappa}_{\rm G}^{j}$ by the matter inhomogeneity at distance $\chi_{i}$, this correlator will also vanish,
\eq{
\int_{\chi_{i}}^{\chi_{\textrm{hor}}} d\chi_j \; T(\chi_i, \chi_j)\; {\ba{{\kappa}_{\rm I}^{i}{\kappa}_{\rm G}^{j}}} = 0,}
\noindent since it is just the matter inhomogeneity at distance $\chi_{i}$ that gives rise to the correlation between ${\kappa}_{\rm I}^{i}$ and ${\kappa}_{\rm G}^{j}$. Thus, when we integrate over $\ba{{\kappa}_{\rm obs}^{i}{\kappa}_{\rm obs}^{j}}$ with the same weight function, the GI contamination in it will be ``nulled out''. Equation (\ref{eq:nc}) is the condition that the weight function $T$ should satisfy in order to ``null'' the intrinsic-shear alignment terms, so we call it ``the nulling condition''.
The same applies to three-point statistics. Consider a correlator $\ba{{\kappa}_{\rm obs}^{i}{\kappa}_{\rm obs}^{j}{\kappa}_{\rm obs}^{k}}$ with $\chi_{i}$ being the smallest comoving distance of the three. Both GII and GGI systematics contained in it also originate from the matter inhomogeneity at distance ${\chi}_{i}$. Typically, the generation of GII systematics requires that $\chi_{i} \approx \chi_{j} < \chi_{k}$, while the generation of GGI requires $\chi_{i} < \chi_{j}$ and $\chi_{i} < \chi_{k}$. For both cases, the dependence of GII or GGI systematics on $\chi_{k}$ is also just $1-{\chi_{i}}/{\chi_{k}}$. So new measures built as $\int_{\chi_{i}}^{\chi_{\textrm{hor}}} d\chi_{k} \; T(\chi_i, \chi_j, \chi_{k})\ba{{\kappa}_{\rm obs}^{i}{\kappa}_{\rm obs}^{j}{\kappa}_{\rm obs}^{k}}$ with $T$ satisfying the nulling condition for three-point statistics
\eq{
\int_{\chi_{i}}^{\chi_{\textrm{hor}}} d\chi_k ~T(\chi_i, \chi_j, \chi_k)\, \br{1 - \frac{\chi_{i}}{\chi_k}} = 0
}
\noindent will be free of both GII and GGI contamination. Again, $T(\chi_i, \chi_j, \chi_{k})$ here should be seen as a function of $\chi_{k}$ whose form depends on $\chi_i$ and $\chi_j$.
Note that this method only depends on the characteristic redshift dependencies of the lensing signal and intrinsic-alignment signals, and is not limited to E-mode fields. This is a reassuring feature since while the $\kappa_{\rm G}$ field is a pure E-mode field to first order, the $\kappa_{\rm I}$ field can have a B-mode component. However, if parity-invariance is assumed, any correlation function which contains an odd number of B-mode shear components vanishes \citep{s03}, thus there should be no B-mode component in the GGI signal.
\subsection{Nulling formalism for lensing bispectrum tomography}
\label{sec:bispectomo}
Since the nulling technique relies on the distinct redshift dependence of the intrinsic-alignment signal, redshift information is crucial for it. With the help of infrared bands, forthcoming multicolor imaging surveys can provide rather accurate photometric redshift information for the galaxies \citep[e.g.][]{abdalla08, bordoloi09}, allowing tomographic studies of cosmic shear statistics. We base our study on cosmic shear bispectrum tomography, and outline the corresponding formalism of the nulling technique in the following.
Given the galaxy redshift probability distribution of redshift bin $i$ which we denote as $p^{(i)}_{\textrm{s}}(z)=p^{(i)}_{\textrm{s}}(\chi_{\rm s})\,d\chi_{\rm s}/dz\,$, one can define the average convergence field in redshift bin $i$ by integrating $\kappa(\vek{\theta},\chi_{\rm s})$ in (\ref{eq:d2k}) over $p^{(i)}_{\textrm{s}}(\chi_{\rm s})$. We turn to angular frequency space now and define
\eq{
\label{eq:kappai}
\tilde{\kappa}^{(i)}_{\rm G}(\vek{\ell}):=\int_0^{\chi_{\rm hor}}~d\chi_{\rm s}\, p_{\rm s}^{(i)}(\chi_{\textrm{s}}) ~\tilde{\kappa}_{\rm G}(\vek{\ell},\chi_{\rm s}) \;,
}
\noindent where $\tilde{\kappa}_{\rm G}(\vek{\ell},\chi_{\rm s})$ is the Fourier transform of $\kappa_{\rm G}(\vek{\theta},\chi_{\rm s})$. To better show the relation between $\tilde{\kappa}^{(i)}_{\rm G}(\vek{\ell})$ and three-dimensional matter inhomogeneity in Fourier space $\tilde{\delta}\br{k,\chi}$, one can write, basing on (\ref{eq:d2k}) and (\ref{eq:kappai}),
\eq{
\label{eq:d2kappa}
\tilde{\kappa}^{(i)}_{\rm G}(\vek{\ell})=\int_0^{\chi_{\rm hor}}\!\!d\chi~ W^{(i)}(\chi)~ \tilde{\delta}\br{\vek{\ell}/ \chi,\chi}\;,
}
\noindent by defining a lensing weight function $W^{(i)}(\chi)$ as
\eq{
\label{eq:weight}
W^{(i)}(\chi) := \frac{3\, \Omega_{\textrm{m}}H_0^2 \,\chi}{2\, a(\chi)\, c^2} \int_{\chi}^{\chi_{\rm hor}}\!\!d\chi_{\textrm{s}}~ p_{\rm s}^{(i)}(\chi_{\textrm{s}}) \frac{\chi_{\textrm{s}}-\chi}{\chi_{\textrm{s}}}\;.
}
The tomographic lensing bispectrum is defined via
\eq
\label{eq:fbisp}
\ba{\tilde{\kappa}_{\rm G}^{(i)}(\vek{\ell}_1)\tilde{\kappa}_{\rm G}^{(j)}(\vek{\ell}_2)
\tilde{\kappa}_{\rm G}^{(k)}(\vek{\ell}_3)}=(2\pi)^2
B^{(ijk)}_{\rm GGG}({\ell}_1,{\ell}_2,{\ell}_3)\;\delta_D(\vek{\ell}_1+\vek{\ell}_2+\vek{\ell}_3)\;,
}
\noindent where the Dirac delta function ensures that the bispectrum is defined only when $\vek{\ell}_1$, $\vek{\ell}_2$, and $\vek{\ell}_3$ form a triangle. This fact arises from statistical homogeneity, while that the bispectrum can be defined as a function independent of the directions of the angular frequency vectors arises from statistical isotropy.
In a survey, the convergence field $\tilde{\kappa}_{\rm obs}$ is determined from the observed galaxy ellipticities, and the corresponding bispectrum $B_{\textrm{obs}}$ suffers from intrinsic-shear alignments. As we did with the three-point correlator in Sect.$\,$\ref{sec:introduction}, we separate the observed lensing bispectrum into the four terms:
\eq{
B_{\textrm{obs}}=B_{\textrm{GGG}}+B_{\textrm{GGI}}+B_{\textrm{GII}}+B_{\textrm{III}}\;.
}
\noindent Among them, $B_{\textrm {GGI}}$, $B_{\textrm{GII}}$ and $B_{\textrm{III}}$ can be linked to the convergence in a similar way as (\ref{eq:fbisp}), for example
\eq{
\label{eq:bGGI}
\ba{\tilde{\kappa}_{\rm I}^{(i)}(\vek{\ell}_1)\tilde{\kappa}_{\rm G}^{(j)}(\vek{\ell}_2)
\tilde{\kappa}_{\rm G}^{(k)}(\vek{\ell}_3)}=(2\pi)^2
B^{(ijk)}_{\rm GGI}({\ell}_1,{\ell}_2,{\ell}_3)\;\delta_D(\vek{\ell}_1+\vek{\ell}_2+\vek{\ell}_3)\;.
}
\noindent Here we assume disjunct redshift bins and let $i$ to be the redshift bin with the lowest redshift, so $\ba{\tilde{\kappa}_{\rm G}^{(i)}(\vek{\ell}_1)\tilde{\kappa}_{\rm I}^{(j)}(\vek{\ell}_2)\tilde{\kappa}_{\rm G}^{(k)}(\vek{\ell}_3)}$ and $\ba{\tilde{\kappa}_{\rm G}^{(i)}(\vek{\ell}_1)\tilde{\kappa}_{\rm G}^{(j)}(\vek{\ell}_2)\tilde{\kappa}_{\rm I}^{(k)}(\vek{\ell}_3)}$ both vanish due to the same reason as explained in Sect.$\,$\ref{sec:nulling} for the two-point statistics.
The purpose of the nulling technique is to filter $B_{\textrm{obs}}$ in such a way that the GGI term is strongly suppressed in comparison with the GGG term. The GII and III terms can be removed by ignoring the signal coming from bispectrum $B^{(ijk)}_{\rm obs}({\ell}_1,{\ell}_2,{\ell}_3)$ with two or three equal redshift bins.
To fulfill this purpose, we construct our new measures as
\eq{
\label{eq:Yij}
Y^{(ij)}({\ell}_1,{\ell}_2,{\ell}_3):=\sum_{k=i+1}^{N_z} T^{(ij)}(\chi_k)~B^{(ijk)}_{\textrm{obs}}({\ell}_1,{\ell}_2,{\ell}_3)\;\chi'_k\;\Delta z_k \;,
}
\noindent where $N_z$ is the total number of redshift bins, $\chi'_k$ is the derivative of comoving distance with respect to redshift, and $\Delta z_k$ is the width of redshift bin $k$. The weight function is written now as $T^{(ij)}(\chi_k)$ since $i$ and $j$ indicate two redshift bins, i.e. two populations of galaxies, rather than two comoving distances as in the previous subsection. The weight $T^{(ij)}$ is required to satisfy the nulling condition (\ref{eq:nc}) in its discretized form,
\eq{
\label{eq:nc2}
O^{(ij)}:=\sum_{k=i+1}^{N_z} T^{(ij)}(\chi_k) ~ \br{1-\frac{\chi_i}{\chi_k}} ~ \chi'_k \; \Delta z_k = 0\;,
}
\noindent for all $j>i$. Here, $\chi_i$ and $\chi_k$ should be chosen such that they represent well the distance to redshift bins $i$ and $k$. In this paper we choose them to be the distances corresponding to the median redshift of the bin. The summation over index $k$ runs from $i+1$ rather than $i$ since we consider only bispectrum measures with $j > i$ and $k > i$ to avoid III and GII systematics. In this case $B^{(ijk)}_{\textrm{obs}}$ in (\ref{eq:Yij}) can be written as a sum of $B^{(ijk)}_{\textrm{GGG}}$ and $B^{(ijk)}_{\textrm{GGI}}$, and $Y^{(ij)}$ can be expressed as
\eq{
\label{eq:shownulling}
\begin{split}
& Y^{(ij)}({\ell}_1,{\ell}_2,{\ell}_3)=\sum_{k=i+1}^{N_z} T^{(ij)}(\chi_k)~B^{(ijk)}_{\textrm{GGG}}({\ell}_1,{\ell}_2,{\ell}_3)\;\chi'_k\;\Delta z_k \\
& + \sum_{k=i+1}^{N_z} T^{(ij)}(\chi_k)~B^{(ijk)}_{\textrm{GGI}}({\ell}_1,{\ell}_2,{\ell}_3)\;\chi'_k\;\Delta z_k\;.
\end{split}
}
Suppose one has infinitely many redshift bins, then the lensing signal in bin $k$ caused by the matter inhomogeneity in bin $i$ is exactly proportional to $1-\chi_i/\chi_k$, which means $B^{(ijk)}_{\textrm{GGI}}({\ell}_1,{\ell}_2,{\ell}_3)$ can be written as a product of $1-\chi_i/\chi_k$ and some function of the parameters other than $\chi_k$:
\eq{
\label{eq:GGIsplit}
B^{(ijk)}_{\textrm{GGI}}({\ell}_1,{\ell}_2,{\ell}_3) = {\cal F}(\chi_i, \chi_j, {\ell}_1,{\ell}_2,{\ell}_3)\, \br{1-\frac{\chi_i}{\chi_k}}\;.
}
\noindent Then we have
\eq{
\label{eq:tomonull}
\begin{split}
& \sum_{k=i+1}^{N_z} T^{(ij)}(\chi_k)~B^{(ijk)}_{\textrm{GGI}}({\ell}_1,{\ell}_2,{\ell}_3)\;\chi'_k\;\Delta z_k = \\
& {\cal F}(\chi_i, \chi_j, {\ell}_1,{\ell}_2,{\ell}_3) \, \sum_{k=i+1}^{N_z} T^{(ij)}(\chi_k)~\br{1-\frac{\chi_i}{\chi_k}} \;\chi'_k\;\Delta z_k = 0\;.
\end{split}
}
\noindent This suggests that only the GGG contribution is left in the nulled measure $Y^{(ij)}$, the GGI contribution has been ``nulled out'' due to the nulling condition. If only a limited number of redshift bins is available, (\ref{eq:GGIsplit}) holds only approximately, leading to a residual in (\ref{eq:tomonull}).
Since the nulling condition is the only condition that the weight $T^{(ij)}$ must satisfy in order to ``null'', there is much freedom in choosing the form of it. We would like to further specify its form such that it preserves as much Fisher information in $Y^{(ij)}$ as possible. The method we have adopted for the nulling weight construction will be detailed in Sect.$\,$\ref{sec:weight}.
Note that for each $(i,j)$ combination, one can in principle apply more than one nulling weight to the original bispectrum, and obtain more nulled measures. If one retains the condition of maximizing the Fisher information and demands that all the weight functions built for one $(i,j)$ combination are orthogonal to each other, one arrives at higher-order modes that have the second-most, third-most, etc., information content (higher-order weights, see JS08). The total number of such linearly independent nulled measures for a certain $(i,j)$ equals the possible values of $k \ge i+1$. In this schematic study we will only use the optimum, i.e. the first-order nulling weights. We will assess the information loss due to this limitation in Sect.$\,$\ref{sec:compression}.
\section{Modeling}
\label{sec:model}
\subsection{Survey characteristics}
\label{sec:setup}
We set up a fictitious survey with a survey size of $A$ = 4000 deg$^2$ which is similar to the survey size of DES. This can be easily scaled to any survey size using the proportionality of statistical errors to $A^{-1/2}$. We assume a galaxy intrinsic ellipticity dispersion $\sigma_{\epsilon}=\sigma(\epsilon_{\rm I}^{\rm ran})=0.35$. As galaxy redshift probability distribution we adopt the frequently used parameterization \citep{smail94},
\eq{
p_{\textrm{s}}(z) \propto \br{\frac{z}{z_0}}^{\alpha} \exp \bc{ -\br{\frac{z}{z_0}}^\beta},
}
\noindent and use $z_0 = 0.64$, $\alpha = 2$, $\beta = 1.5$. The distribution is cut at $z_{\textrm{max}} = 3$ and normalized to 1. The corresponding median redshift of this fictitious survey is $z_{\rm m}=0.9$, which is compatible to a survey like EUCLID. We adopt an average galaxy number density $\bar{n}_{\textrm{g}}=40~ {\textrm{arcmin}}^{-2}$ which is again EUCLID-like.
Disjunct redshift bins without photo-z error are assumed, which means that the galaxy redshift probability distribution in redshift bin $i$ takes the form $p^{(i)}_{\textrm{s}}(z) \propto p_{\textrm{s}}(z)$ if and only if the redshift that corresponds to comoving distance $\chi_{\rm s}$ is within the boundaries of redshift bin $i$. A number of 10 redshift bins are used by default. The boundaries of the redshift bins are set such that each bin contains the same number of galaxies.
We adopt 20 angular frequency bins spaced logarithmically between $\ell_{\textrm{min}}=50$ and $\ell_{\textrm{max}}=3000$, and denote the characteristic angular frequency of a bin as $\bar{\ell}$. Within this range the noise properties of the cosmic shear field are still not too far in the non-Gaussian regime, allowing a more realistic theoretical estimation of the bispectrum and its covariance. Whether this number of angular frequency bins can reconstruct the angular frequency dependence of the bispectrum is tested, and 20 bins are found to be sufficient for our requirements on precision. This is also expected since the bispectrum is rather featureless as a function of angular frequency.
\subsection{Bispectrum and its covariance}
We show the modeling of $B_{\textrm{GGG}}$ and its covariance in this section. We will only consider the tomographic bispectrum at redshift bins satisfying $z_i<z_j$ and $z_i<z_k$, which already ensures an elimination of $B_{\textrm{III}}$ and $B_{\textrm{GII}}$ systematics in our case.
Applying Limber's equation, it can be shown that the tomographic convergence bispectrum can be written as a projection of the three-dimensional matter bispectrum $B_\delta\!\left({k_1},{k_2},{k_3}; \chi\right)$ (see e.g. TJ04):
\eq{
\label{eq:B_GGG}
\begin{split}
& B^{(ijk)}_{\rm GGG}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3)= \\
& \int^{\chi_{\rm hor}}_0\!\!d\chi ~ \frac{W^{(i)}(\chi)W^{(j)}(\chi)W^{(k)}(\chi)}{\chi^{4}}~ B_\delta\!\left(\frac{\bar{\ell}_1}{\chi},\frac{\bar{\ell}_2}{\chi},\frac{\bar{\ell}_3}{\chi};
\chi\right)\;.
\end{split}
}
To compute $B_{\delta}$, we employ the fitting formula by \citet{sco01}, which is based on hyper-extended perturbation theory \citep{sco99}. A comparison of this formula with the halo model results can be found in \citet{TJ03b,TJ03c}.
Estimating the bispectrum covariance is often done within a flat-sky spherical harmonic formalism \citep[Hu00 hereafter]{hu00}, which suffers - at least formally - from drawbacks since its basis functions $Y_{lm}$ are only defined for discrete angular frequency values and full sky coverage. In \citet{joachimi09b} another approach exclusively based on the two-dimensional Fourier formalism was constructed, which we use here.
The bispectrum covariance is a six-point correlation function which can be expanded into its connected parts as outlined in e.g. \citet{ber02a}. As argued in TJ04, for angular scales $\ell \le 3000$, the term which is a triple of two-point functions still dominates. Keeping only this term, the bispectrum covariance reads
\eq{
\label{eq:bicov}
\begin{split}
& {\textrm{Cov}} \br{B^{(ijk)}_{\rm GGG}({\bar{\ell}}_1,{\bar{\ell}}_2,{\bar{\ell}}_3),\;B^{(lmn)}_{\rm GGG}({\bar{\ell}}_4,{\bar{\ell}}_5,{\bar{\ell}}_6) } = \\
& \; \frac{(2\,\pi)^3}{A\; {\bar{\ell}_1}{\bar{\ell}_2}{\bar{\ell}_3} \; {\Delta \bar{\ell}_1}{\Delta \bar{\ell}_2}{\Delta \bar{\ell}_3}} \; \Lambda^{-1}\br{\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3} \; \\
& \times \br{ \bar{P}^{(il)}(\bar{\ell}_1)\bar{P}^{(jm)}(\bar{\ell}_2)\bar{P}^{(kn)}(\bar{\ell}_3)\;\delta_{\bar{\ell}_1 \bar{\ell}_4}\;\delta_{\bar{\ell}_2 \bar{\ell}_5}\;\delta_{\bar{\ell}_3 \bar{\ell}_6} + 5\, \mbox{perms.}},
\end{split}
}
\noindent in which $\Delta \bar{\ell}_i$ is the bin width of the angular frequency bin with typical value $\bar{\ell}_i$, and $\bar{P}^{(ij)}(\bar{\ell})$ is the observed power spectrum which contains the intrinsic ellipticity noise \citep[e.g.][]{kaiser92,hu99,JS08a}:
\eq{
\label{eq:psshapenoise}
\bar{P}^{(ij)}(\bar{\ell}) = P^{(ij)}(\bar{\ell}) + \delta_{ij}\frac{\sigma^2_\epsilon}{2 \bar{n}_i}\;,
}
\noindent where $\bar{n}_i$ is the galaxy number density in redshift bin $i$. The term $\Lambda \br{\ell_1,\ell_2,\ell_3}$ is defined as
\eqa{
\label{eq:deflambda}
\begin{split}
\Lambda \br{\ell_1,\ell_2,\ell_3} \equiv \left\{ \begin{array}{ll} & \bc {\frac{1}{4} \sqrt{2 \ell_1^2 \ell_2^2 + 2 \ell_1^2 \ell_3^2 + 2 \ell_2^2 \ell_3^2 - \ell_1^4 - \ell_2^4 - \ell_3^4} }^{-1} \\
& \hspace{2cm} \,\, \mbox{if}~~~ |\ell_1-\ell_2| < \ell_3 < \ell_1 + \ell_2\,, \\
& 0 \hspace{2cm} \mbox{else}.\\ \end{array} \right.
\end{split}
}
\noindent When $|\ell_1-\ell_2| < \ell_3 < \ell_1 + \ell_2$ is satisfied, $\Lambda^{-1} \br{\ell_1,\ell_2,\ell_3}$ is the area of a triangle with side lengths $\ell_1,\ell_2$ and $\ell_3$. We use the \citet{EH} transfer function to evaluate the linear three-dimensional matter power spectrum, and the \citet{SMITH} fitting function for the nonlinear power spectrum.
Note that there is no intrinsic ellipticity noise in the observed bispectrum, since the galaxy intrinsic ellipticity distribution is assumed to be skewless.
Under the assumptions of a compact survey geometry and scales much smaller than the extent of the survey area, (\ref{eq:bicov}) provides a bispectrum covariance that naturally incorporates the scaling with survey size, is not restricted to integer angular frequencies, and allows for any appropriate binning. In terms of Fisher information, the result given by this approach and the Hu00 one agree to high accuracy \citep{joachimi09b}.
\subsection{Toy intrinsic-shear alignment model}
In this section we present a toy model for generating GGI systematics. Since the physical generation of intrinsic-shear alignments concerns nonlinear growth of structure and complex astrophysical processes which are not easy to quantify, a realistic model is not yet available. Current simulations involving baryonic matter also have some way to go before they can simulate the generation of the GGI systematics reliably.
Up to now there has not been any attempt to measure GGI and GII in galaxy surveys. \citet{sembo08} studied these systematics using ray-tracing simulations. They provided fits in real space to projected GII and GGI signals, but the results are still too crude to lead to sufficient constraints on an intrinsic-shear alignment model.
This situation emphasizes the importance of a method intended to control intrinsic-shear alignment to be model-independent, especially at the three-point level. Since this is the case for the nulling technique, for this work we only require a simple model for $B_{\textrm {GGI}}^{(ijk)}$ which satisfies the characteristic redshift dependence and leads to a reasonable bias.
Based on the observation that the lensing bispectrum expression (\ref{eq:B_GGG}) comes directly from (\ref{eq:d2kappa}) and the definition of the tomography bispectrum (\ref{eq:fbisp}), we link $B_{\textrm {GGI}}^{(ijk)}$ also to a three-dimensional bispectrum $B_{\delta_I\delta\delta}$ via
\eq{
\label{eq:ggiia}
\begin{split}
& B_{\textrm {GGI}}^{(ijk)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3)= \\
& \int^{\chi_{\rm hor}}_0\!\!d\chi ~\frac{p_{\rm s}^{(i)}(\chi) W^{(j)}(\chi)W^{(k)}(\chi)}{\chi^{4}}~ B_{\delta_I\delta\delta}\br{\frac{\bar{\ell}_1}{\chi},\frac{\bar{\ell}_2}{\chi},\frac{\bar{\ell}_3}{\chi};\chi}.
\end{split}
}
\noindent Similar to $B_{\delta}(k_1,k_2,k_3)$ which is given by
\eq{
\begin{split}
&\ba{\tilde{\delta}(\vek{k_1},\chi)\tilde{\delta}(\vek{k_2},\chi)\tilde{\delta}(\vek{k_3},\chi)}=\\
&\hspace{2cm} (2\pi)^3\,\delta_D(\vek{k_1}+\vek{k_2}+\vek{k_3})\,B_{\delta}(k_1,k_2,k_3;\chi),
\end{split}
}
\noindent $B_{\delta_I\delta\delta}$ is defined via
\eq{
\begin{split}
& \ba{\tilde{\delta}_I(\vek{k_1},\chi)\tilde{\delta}(\vek{k_2},\chi)\tilde{\delta}(\vek{k_3},\chi)}=\\
&\hspace{2cm} (2\pi)^3\,\delta_D(\vek{k_1}+\vek{k_2}+\vek{k_3})\,B_{\delta_I\delta\delta}(k_1,k_2,k_3;\chi),
\end{split}
}
\noindent where $\tilde{\delta}_{\rm I}(\vek{k})$ is the three-dimensional density field which is responsible for the intrinsic alignment, and it satisfies
\eq{
\label{eq:d2kappa_I}
\tilde{\kappa}^{(i)}_{\rm I}(\vek{\ell})=\int_0^{\chi_{\rm hor}}\!\!d\chi~ p_{\rm s}^{(i)}(\chi)~ \tilde{\delta}_{\rm I}\br{\frac{\vek{\ell}}{\chi},\chi}\;,
}
\noindent The definition of both $\tilde{\kappa}^{(i)}_{\rm I}$ and $\tilde{\delta}_{\rm I}$ originates from the deterministic part of galaxy intrinsic ellipticity $\epsilon_{\rm I}^{\rm det}$. We have assume the existence of these underlying smooth fields. Similar quantities have been defined in \citet{joachimi09c}, see also \citet{hirata04} and \citet{schneider09}. We would like to point out again that, although we introduce these quantities for the clarity of our model, we do not need them for the main purpose of this paper. What we need to model is the projected GGI bispectrum $B_{\textrm {GGI}}^{(ijk)}$.
Note that in (\ref{eq:ggiia}), the weight for the lowest redshift bin $i$ is the source distribution function $p_{\rm s}^{(i)}$ which is zero outside redshift bin $i$, rather than the lensing weight $W^{(i)}$ which is a much broader function. Since $\tilde{\kappa}_{\rm I}^{(i)}$ depends only on physical processes at redshift bin $i$ and is inferred from ellipticity measurements in this bin, and $\tilde{\kappa}_{\rm G}^{(j)}$ is linked to the three-dimensional matter density through the lensing weight $W^{(j)}$, this assignment of weight functions will ensure the correct redshift dependence of $B_{\textrm {GGI}}^{(ijk)}$.
When the redshift bins are not disjunct, however, the intrinsic alignment signal can no longer be associated with bin $i$. There will be two permutations in both the left-hand side of (\ref{eq:bGGI}) and the right-hand side of (\ref{eq:ggiia}), similar to the two-point case, e.g. Eq.\,11 in \citet{hirata04}.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{loptimal.eps}
\caption{Distribution of the nulled Fisher information as defined in (\ref{eq:foij}) per $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ bin and per redshift bin combination among different angular frequency triangle shapes and sizes. Results for four redshift bin combinations $(i,j)$ are presented.
\textit{Left panel}: Distribution of the nulled Fisher information among different triangle configurations. We consider triangles with the common shortest side length $\bar{\ell}_1=171$ which corresponds to the 7th angular frequency bin. Due to our logarithmic binning and the constraint that the three side lengths must be able to form a triangle, only 8 such triangle configurations exist. Plotted is the nulled Fisher information contained in these 8 triangles against $\alpha$, which is the angle opposite to the shortest side length in that triangle. Smaller $\alpha$ correspond to more elongated triangles, and larger $\alpha$ correspond to almost equilateral triangles.
\textit{Right panel}: Distribution of the nulled Fisher information contributed by each $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ bin over different triangle sizes. A fixed triangle shape with $\bar{\ell}_1:\bar{\ell}_2:\bar{\ell}_3=1:3.64:4.52$ (corresponds to the leftmost points in the left panel) is chosen. The nulled Fisher information contained in one $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ bin is plotted against the shortest side length $\bar{\ell}_1$ of each triangle.}
\label{fig: loptimal}
\end{figure*}
The modeling of $B_{\delta_I\delta\delta}$ is then a pure matter of choice. We build a simple three-dimensional GGI bispectrum with power-law dependence on both redshift $z$ and spatial frequency $k$:
\begin{equation}
\label{eq:IAmodel}
\begin{split}
& B_{\delta_I\delta\delta}\br{{k_1},{k_2},{k_3};\chi} := - {\cal A} \, B_{\delta}\,(k_{\textrm {ref}},k_{\textrm {ref}},k_{\textrm {ref}};\chi(z_{\textrm {med}}))\, \br{\frac{1+z}{1+z_{\textrm{med}}}}^{r-2}\\
& \times \;\bc{{\br{\frac{k_1}{k_{\textrm {ref}}}}}^{2(s-2)} + {\br{\frac{k_2}{k_{\textrm {ref}}}}}^{2(s-2)} + {\br{\frac{k_3}{k_{\textrm {ref}}}}}^{2(s-2)}}\;,
\end{split}
\end{equation}
\noindent where $z_{\textrm {med}}$ is the median redshift of the whole survey, and ${\cal A}$, $k_{\textrm {ref}}$, $r$, $s$ are free parameters. Among them the parameter $k_{\textrm {ref}}$ is designed to be a characteristic wave number, whose value we set to be a weakly nonlinear scale of $10\;h{\rm Mpc}^{-1}$ here. The minus sign ensures that the contamination of GGI systematics leads to an underestimation of the GGG signal, as found by \citet{sembo08}.
Little is known about the redshift and angular scale dependence of $B_{\delta_I\delta\delta}$. However one can roughly estimate how it compares to the $B_{\delta\delta\delta}$ signal.
A linear alignment model suggests $\delta_I \propto \delta_{\rm lin}\,\bar{\rho}(z)/((1+z)\,D_{+}(z))$ \citep[see e.g.][]{hirata04}, in which $\bar{\rho}(z)$ is the mean density of the universe, ${D_{+}}(z)$ is the growth factor, and $\delta_{\rm{lin}}$ is the linear matter density contrast. Thus we have, very roughly, $\delta_I \propto (1+z)^3\,\delta_{\rm lin}$ which suggests $B_{\delta_I\delta\delta} \propto (1+z)^3\,B_{\delta\delta\delta}$.
The linear alignment model assumes that the intrinsic alignment is linearly related to the local tidal gravitational field \citep[e.g.][]{catelan01,hirata04}. If this holds true, we also expect $B_{\delta_I\delta\delta}$ to have a stronger angular scale dependence than $B_{\delta\delta\delta}$ since tidal gravitational interaction follows the inverse cube law rather than the inverse square law which gravity itself follows.
For a $\Lambda$CDM model, in the weakly nonlinear regime where perturbation theory holds, the dependence of $B_{\delta\delta\delta}$ on $(1+z)$ has a negative power shallower than $-4$, and the dependence on $k$ has a power of around $-2$. In this paper we choose $r=0$, $s=1$ as default. We also study the cases of $r=-2$, $r=2$, and $s=0$ whose results will be shown in Fig.$\,$\ref{fig:nz} below.
As for the amplitude ${\cal A}$ of the GGI signal, the only direct study up to now is \citet{sembo08}, which suggests an overall GGI/GGG ratio of 10$\,\%$ for a $z_{\rm m} = 0.7$ survey for elliptical galaxies and few percent for a mixed sample of elliptical and spiral galaxies. In this paper we adjust ${\cal A}$ such that the amplitude of the tomographic GGI bispectrum is limited to be within 10$\,\%$ of the amplitude of the lensing GGG signal, i.e. GGI/GGG$\lesssim 10\,\%$ at redshift bin combinations with $z_i \ll z_j$ and $z_i \ll z_k$ where the GGI signal is expected to be most significant. This leads to a relatively modest overall GGI/GGG ratio at percent level. We will show examples of the generated GGI and GGG signals in Fig.$\,$\ref{fig:ggiggg}.
As an order-of-magnitude estimate, one can also relate the GGI/GGG ratio to that of GI/GG by expanding three-point signals to couples of two-point signals using perturbation theory, in analogy to the \citet{sco01} fitting formula. For the case of $z_i\ll z_j \approx z_k$, the leading order terms would give that the GGI/GGG ratio approximates that of GI/GG evaluated at redshifts $z_i$ and $z_j$. This suggests that our adopted GGI/GGG ratio is also consistent with available observational studies of the GI signal \citep{mandel06, mandel09, hirata07, fu08, okumura09a, okumura09b}, although the results of these studies vary a lot according to different median redshift, color and luminosity of the selected galaxy sample.
\section{Construction of nulling weights }
\label{sec:weight}
As mentioned in Sect.$\,$\ref{sec:bispectomo}, we would like to construct a single first-order weight function $T^{(ij)}(\chi)$ for each $(i,j)$ combination which preserves the maximum of information. This can be seen as a constrained optimization problem. The constraining condition here is the nulling condition and the quantity to be optimized is the Fisher information after nulling. In JS08, several practical methods were developed to solve this optimization problem at the two-point level, and very good agreement was found among the different methods.
We adopt the simplified analytical approach as described in JS08, and reformulate it for three-point statistics here. For convenience we introduce the following notations:
\noindent the bispectrum covariance matrix ${\rm \textbf CovB}$, whose elements are
\eq{
{\rm CovB}({^{ijk}_{\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3}};{^{lmn}_{\bar{\ell}_4,\bar{\ell}_5,\bar{\ell}_6}}) := {\textrm{Cov}} \br{B_{\rm GGG}^{(ijk)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3),\;B_{\rm GGG}^{(lmn)}(\bar{\ell}_4,\bar{\ell}_5,\bar{\ell}_6) }\;;
}
\noindent the covariance matrix ${\rm \textbf CovY}$ of the nulled bispectra $Y$, whose elements are
\eq{
\label{eq:cov_Y}
\begin{split}
& {\rm CovY}({^{ij}_{\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3}};{^{lm}_{\bar{\ell}_4,\bar{\ell}_5,\bar{\ell}_6}}) := \textrm{Cov}\br{Y^{(ij)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3),Y^{(lm)}(\bar{\ell}_4,\bar{\ell}_5,\bar{\ell}_6)} \\
& = \sum_{k= \atop i+1}^{N_z} \sum_{n= \atop l+1}^{N_z} \textrm{Cov} \br{B_{\rm GGG}^{(ijk)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3),B_{\rm GGG}^{(lmn)}(\bar{\ell}_4,\bar{\ell}_5,\bar{\ell}_6)} \\
& \times \, T^{(ij)}\br{\chi_{k}} T^{(lm)} \br{\chi_{n}} \chi'_{k} \;\chi'_{n} \;\Delta z_{k} \;\Delta z_{n}\; ;
\end{split}
}
\noindent a vector ${\rm \textbf B_{\textbf ,\,{\mu} }}$ whose elements are partial derivatives of the bispectrum with respect to the cosmological parameter $p_{\mu}$
\eq{
{ B_{ ,\,\mu}}({^{ijk}_{\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3}}) := \frac{\partial B_{\rm GGG}^{(ijk)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3)}{\partial p_{\mu}}\;;
}
\noindent and a corresponding vector ${\rm \textbf Y_{\textbf ,\,\mu}}$ for nulled bispectra $Y$, whose elements are
\eq{
{Y_{ ,\,\mu}}({^{ij}_{\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3}}) := \frac{\partial Y^{(ij)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3)}{\partial p_{\mu}}\;.
}
Then the Fisher information matrix from the original bispectra can be written as (following TJ04)
\eq{
\label{eq:F_i}
{\rm F^{\rm i}_{\mu\nu}} = {\rm \textbf B_{\textbf ,\,\mu}}\, {\rm \textbf CovB}^{-1}\, {\rm \textbf B_{\textbf ,\,\nu}}\;\;,
}
\noindent and that from the nulled bispectra can be written as
\eq{
\label{eq:F_f}
{\rm F^{\rm f}_{\mu\nu}} = {\rm \textbf Y_{\textbf ,\,\mu}}\, {\rm \textbf CovY}^{-1}\, {\rm \textbf Y_{\textbf ,\,\nu}}\;\;.
}
\noindent Here the matrix multiplication is a summation of all possible angular frequency combinations $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ and redshift bin combinations, ${(ijk)}$ for the original bispectra and ${(ij)}$ for the nulled bispectra. In (\ref{eq:F_i}) and (\ref{eq:F_f}), ${\rm \textbf CovB}^{-1}$ and ${\rm \textbf CovY}^{-1}$ indicate the inverse of the covariance matrix. When the covariance is approximated by triples of power spectra, the covariance between two different angular frequency combinations $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)\neq(\bar{\ell}_4,\bar{\ell}_5,\bar{\ell}_6)$ is zero, see (\ref{eq:bicov}), which means that the covariance matrix is block diagonal. In this case the matrix inversion can be done separately for each block specified by an angular frequency combination $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$.
According to the idea of the simplified analytical approach, we consider the Fisher information on one cosmological parameter contained in bispectrum measures $B_{\rm{GGG}}^{(ijk)}(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ with a single $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ combination and with redshift bin $(i,j,k)$ combinations having common $(i,j)$ indices. For every $(i,j)$ combination we build nulling weights $T^{(ij)}$ which maximizes the nulled Fisher matrix using the method of Lagrange multipliers. Since here the nulled Fisher matrix receives contribution only from certain angular frequency and redshift combinations, we denote it as $F_{\rm o}^{(ij)}$ to avoid ambiguity. $F_{\rm o}^{(ij)}$ has only one component since only one cosmological parameter is taken into consideration. As only a single $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ combination is involved, we will omit the $\bar{\ell}$-dependence in all variables in the rest of this subsection to keep a compact form.
Again for notational simplicity, we follow JS08 and introduce a vector notation as follows. For each $(i,j)$ in consideration, let the values of the weights $T^{(ij)}(\chi_k)$ form a vector $\vek{T}=T_k$, and define another vector $\vek{\rho}$ and a matrix $\bar{\bf{C}}$ with elements
\eq{
\rho_{k}:= { B_{,\,\mu}}({^{ijk}}) \, \chi'_{k}\;\Delta z_{k} \;,
}
\eq{
\bar{C}_{kn} := {\rm CovB}({^{ijk}};{^{ijn}}) \;\chi'_{k} \;\chi'_{n} \;\Delta z_{k} \;\Delta z_{n}\;.
}
\noindent Thus $F_{\textrm{o}}^{(ij)}$ can be expressed, according to (\ref{eq:F_f}), as
\eq{
\label{eq:foij}
F_{\textrm{o}}^{(ij)} := {\rm \textbf Y_{\textbf ,\,\mu}}({^{ij}})\,\;{{\rm \textbf CovY}^{-1}({^{ij}};{^{ij}})}\,\;{\rm \textbf Y_{\textbf ,\,\mu}}({^{ij}}) = \frac{\br{\vek{T} \cdot \vek{\rho}}^2} {\vek{T}^\tau \bar{\bf{C}} \vek{T}}.
}
\noindent We further define a vector $\vek{f}$ with elements
\eq{
f_{k} = \br{1-\frac{\chi_{i}}{\chi_{k}}}\;\chi'_{k}\;\Delta z_{k}
}
\noindent to write the nulling condition (\ref{eq:nc2}) as
\eq{
O^{(ij)} = \vek{T} \cdot \vek{f} = 0 \;.
}
The problem of finding nulling weights $\vek{T}$ which maximize $F^{(ij)}_o$ under the constraint given by the nulling condition can be solved with the method of Lagrange multipliers by defining a function
\eq{
\label{eq:maxi}
G := F_{\textrm{o}}^{(ij)} + \lambda O^{(ij)} = \frac{\br{\vek{T} \cdot \vek{\rho}}^2} {\vek{T}^\tau \bar{\bf{C}} \vek{T}} + \lambda \vek{T}\cdot\vek{f}\;
}
\noindent with $\lambda$ being the Lagrange multiplier, and setting the gradient of $G$ with respect to $\vek{T}$ to zero,
\eq{
\nabla_T\; G = 2 \vek{\rho}\; \frac{\br{\vek{T} \cdot \vek{\rho}}}{\vek{T}^\tau \bar{\bf{C}} \vek{T}} - 2\; \bar{\bf{C}} \vek{T} \;\br{ \frac{\br{\vek{T} \cdot \vek{\rho}}}{\vek{T}^\tau \bar{\bf{C}} \vek{T}} }^2 + \lambda \vek{f} = 0\;.
}
\noindent The solution to this equation is (for more details see JS08)
\eq{
\vek{T} = {\cal N} \bc{\bar{\bf{C}}^{-1} \vek{\rho} - \frac{\vek{f}^\tau \bar{\bf{C}}^{-1} \vek{\rho}}{\vek{f}^\tau \bar{\bf{C}}^{-1} \vek{f}}\; \bar{\bf{C}}^{-1} \vek{f} }\;,
}
\noindent with the normalization $\cal N$ adjusted to give ${|T|}^2=1$.
Apparently the thus constructed nulling weights depend on which $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ combination is considered and with respect to which cosmological parameter we optimize the information content. In this paper the default cosmological parameter to optimize is $\Omega_{\rm m}$, and we choose for each $(i,j)$ combination the $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ combination which maximizes $F_{\rm o}^{(ij)}$. However one needs to be aware that this serves only as a clear choice of a $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ combination and is not necessarily the best in terms of information preservation considering all angular frequency bins and all cosmological parameters.
To show which triangle shapes and sizes contain more information, we plot $F_{\rm o}^{(ij)}$ against the $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ triangle shape and size for four typical $(i,j)$ combinations in Fig.$\,$\ref{fig: loptimal}. In the left panel, the nulled information $F_{\rm o}^{(ij)}$ contained in different triangles with a common shortest side length $\bar{\ell}_1=171$ is plotted against $\alpha$, which is the angle opposite to $\bar{\ell}_1$. Due to our logarithmic binning in angular frequency, only eight $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ combinations with $\bar{\ell}_1=171$ can form triangles. One sees that the more elongated triangles (small $\alpha$) contain much more Fisher information than the almost equilateral triangles (large $\alpha$). The small separation between the 3rd and the 4th points from the left is caused by the degeneracy of different triangle shapes with respect to $\alpha$, e.g. two equal and very long side lengths can result in the same value of $\alpha$ as two shorter side lengths with a length difference close to the length of the shortest side length. The right panel shows the distribution of the Fisher information contained in one $(\bar{\ell}_1, \bar{\ell}_2, \bar{\ell}_3)$ bin over the triangle size. When the redshift in consideration is higher, the peak of the information distribution moves to higher angular frequencies. The figure suggests that most information comes from high redshifts and small angular scales.
\begin{figure}[t]
\centering
\includegraphics[width=9cm]{poptimal.eps}
\caption{Nulling weights $T^{(ij)}$ for redshift bins $i=1, j=2$ are plotted against the redshift value of the third redshift index $k$. Remarkable consistency is found between nulling weights optimized on different parameters, shown with different line styles.}
\label{fig: poptimal}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{ggiggg.eps}
\caption{Tomographic convergence bispectrum (GGG, solid curves) and intrinsic-shear alignment (GGI, dashed curves) for equilateral triangles are plotted against triangle side length. Measures both before (left panel) and after (right panel) applying the nulling technique ($B$ and $Y$ respectively) are shown for three typical redshift bin $(i,j)$ combinations in the three rows.}
\label{fig:ggiggg}
\end{figure*}
To explore the sensitivity of nulling weights on the choice of the cosmological parameter, we construct seven sets of weight functions, each optimizing the information content in terms of one parameter. For all $(i,j)$ combinations we find that the nulling weights are not very sensitive to the choice of parameter. As an example, the weights for $(i,j) = (1,2)$ are shown in Fig.$\,$\ref{fig: poptimal}. This result is rather surprising at first sight, since for different parameters the distribution of information (contained in the bispectrum) over redshift bins is quite different. However, such insensitivity suggests that the shapes of nulling weights are already strongly constrained under our construction scheme. One constraint is, evidently, the nulling condition. Moreover, considering the fact that we optimize the nulling weights for each $(i,j)$ combination with respect to the information content they preserve, we have already required the shapes of these first order nulling weights to be as smooth as possible.
The fact that these two conditions have already imposed strong constraints on the nulling weights also suggests that nulling weights can be robustly and efficiently constructed, i.e. it is not critical to construct the ``best'' nulling weights.
\section{Performance of the nulling technique}
\subsection{GGI/GGG ratio}
\label{sec:ratio}
What the nulling technique ``nulls'' is the GGI signal $B_{\rm GGI}$, so the GGI/GGG ratio is the most direct quantification of its performance. We plot the modeled GGI and GGG bispectra before and after nulling in Fig.$\,$\ref{fig:ggiggg}. The original GGI signal is shown in the left panels by dashed lines. For comparison the GGG signals are shown as solid curves. The results are shown for equilateral triangle configurations for the convenience of presenting. One sees that when the redshift bin number $j$ and/or $k$ increase, the changes in GGG and GGI signals are different, which shows the expected different redshift dependence. For all redshift bin combinations the GGI signal is modeled to be subdominant to the GGG signal. In the nulled measures shown in the right panels, the GGI/GGG ratio is suppressed by a factor of 10 over all angular scales, which reflects the success of the nulling technique.
\subsection{Information loss and downweighting of systematics}
We further evaluate the performance of the nulling technique by looking at the constraining power of cosmic shear bispectrum tomography on cosmological parameters, as well as the biases caused by the GGI systematics before and after nulling.
The full characterization of the bispectrum involves three angular frequency vectors which form a triangle. In some works concerning three-point statistics, only equilateral triangle configurations i.e. $\ell_1=\ell_2=\ell_3=\ell$ are used for simplicity reasons \citep[e.g.][]{pires09}. But as several authors have pointed out \citep[e.g.][]{kilbinger05,berge09}, only a low percentage of information is contained in equilateral triangles. Thus, to calculate the full information content, we use general triangle configurations but limit our calculation to triangles with three different side lengths, again for reasons of simplicity (for details see Appendix\,\ref{app:1} ).
We will use the figure of merit (FoM, \citealp{TaskForce06}) to quantify the goodness of parameter constraints. Here the FoM for constraints in the parameter plane $p_\alpha-p_\beta$ is defined to be proportional to the inverse of the area of the parameter constraint ellipses:
\eq{
\label{eq:fom}
\rm{FoM}(p_\alpha,p_\beta) \equiv \br{(\rm \textbf F^{-1})_{\alpha \alpha}(\rm \textbf F^{-1})_{\beta \beta}-(\rm \textbf F^{-1})_{\alpha \beta}^{2}}^{-\frac{1}{2}}.
}
To compute biases, we adopt a method based on a simple extension of the Fisher matrix formalism \citep[e.g.][]{hut06,amara08}. Then one needs to define a bias vector $\mathbb{B}^{\textrm{GGI}}$ which in our case reads:
\eq{
\label{eq:Bsys_i}
{\mathbb B}_{\nu, \textrm{i}}^{\textrm{GGI}}={\rm \textbf B_{\rm GGI }}\, {\rm \textbf CovB}^{-1}\, {\rm \textbf B_{\textbf ,\,\nu}}\;\;,
}
\eq{
\label{eq:Bsys_o}
{\mathbb B}_{\nu, \textrm{f}}^{\textrm{GGI}}={\rm \textbf Y_{\rm GGI }}\, {\rm \textbf CovY}^{-1}\, {\rm \textbf Y_{\textbf ,\,\nu}}\;\;,
}
\noindent with
\eq{
{\rm \textbf B_{\rm GGI }}({^{ijk}_{\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3}}) := B_{\rm GGI}^{(ijk)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3)\;\;,
}
\eq{
{\rm \textbf Y_{\rm GGI }}({^{ij}_{\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3}}) := Y_{\rm GGI}^{(ij)}(\bar{\ell}_1,\bar{\ell}_2,\bar{\ell}_3)\;\;.
}
\noindent The bias of the parameter estimator $\hat{p}_\mu$ is given by the difference between its ensemble average and the fiducial value of the parameter $p_{\mu}^{\rm fid}$:
\eq{
\label{eq:bias}
b_{\mu}= \langle \hat{p}_\mu \rangle - p_{\mu}^{\rm fid} =\sum_{\nu} \br{\rm \textbf F^{-1}}_{\mu\nu} \mathbb{B}_{\nu}^{\textrm{GGI}} \;.
}
The information content before and after nulling can be seen in Fig.$\,$\ref{fig:para}. On the cost of increasing the error on each parameter to about twice its original value, GGI systematics are reduced to be within the original statistical error. The relative information loss in terms of FoM can be found in Table$\,$\ref{tab:FoM}. The constraints shown in Fig.$\,$\ref{fig:para} do not represent the best constraints obtainable from a cosmic shear bispectrum analysis since we consider only the triangles with angular scale $\bar{\ell}_1 \ne \bar{\ell}_2 \ne \bar{\ell}_3$. Also note that the nulling technique can in principle remove the GGI systematics completely. But as shown in Fig.$\,$\ref{fig:para}, the systematics still cause some residual biases on cosmological parameters after nulling, due to the finite number of redshift bins. The GGI systematics will be reduced to a lower level when more redshift bins are available. We will discuss this further in the following subsection.
\subsection{How many redshift bins are needed?}
Analyzing the cosmic shear signal in a tomographic way was originally meant to maximize the information. For this purpose alone, a crude redshift binning will suffice \citep{hu99}. However, to control intrinsic-shear alignment, which is a redshift-dependent effect, much more detailed redshift information is required \citep[e.g.][]{king02,bridle07b,JS08b}. Thus, for a method intended to eliminate intrinsic-shear alignment, it is necessary to show its requirement on the redshift precision. In the case of nulling, detailed redshift information is not only needed for the method to be able to eliminate the bias, but also for the preservation of a reasonably large amount of information through the nulling process. JS08 examined the number of redshift bins required for the nulling technique in the two-point case, and showed that 10 redshift bins already ensure that parameters are still well-constrained after nulling.
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{nz3.eps}
\caption{ Ratio of the nulled and the original biases for cosmological parameter $\Omega_{\textrm m}$ as a function of number of redshift bins $N_z$. Results for different GGI models are shown. Parameter $s$ and $r$ are the slopes of angular frequency and redshift dependence of our power-law model (\ref{eq:IAmodel}). }
\label{fig:nz}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{fom_nz.eps}
\caption{ Figure of merit (FoM) as defined in (\ref{eq:fom}) in the $\Omega_{\rm m}$ - $\sigma_{\rm 8}$ plane as a function of number of redshift bins $N_z$. FoM from two-point (2p) measures, three-point measures (3p) and combined (2p+3p) are shown both before nulling (original) and after nulling (nulled). }
\label{fig:fom_nz}
\end{figure}
\begin{figure*}[t]
\begin{minipage}[c]{.75\textwidth}
\includegraphics[width=13cm]{para.eps}
\end{minipage}
\begin{minipage}[c]{.22\textwidth}
\caption{Projected 1-sigma (68$\,\%$ CL) parameter constraints from cosmic shear bispectrum tomography. Hidden parameters are marginalized over. The black solid and blue dash-dotted ellipses correspond to the original constraints and those after nulling, respectively. The black cross in the center of each panel represents the fiducial values adopted for the parameters, and the distance from the center of one ellipse to the black cross reflects the bias caused by intrinsic-alignment GGI systematics on the corresponding parameter. As nulling can be seen as a linear data compression under the constraint of the nulling condition, we also plot the constraints and biases after an unconditioned linear data compression as magenta dashed ellipses for comparison (see Sect.$\,$\ref{sec:compression}). }
\label{fig:para}
\end{minipage}
\end{figure*}
To re-assess this problem at the three-point level, we consider two different situations to address the requirements coming from control of the intrinsic-shear alignment and preservation of the information content separately. In both cases we split the redshift range between $z=0$ and $z=3$ into 5, 10, 15, and 20 (only in the first situation) redshift bins, with the redshift bins split in a way that there is an equal number of galaxies in each bin.
First we consider a single cosmological parameter, $\Omega_{\rm m}$, to be free and study the biases introduced by the GGI signal on $\Omega_{\rm m}$ both before and after nulling. We use only equilateral triangle configurations to reduce the amount of calculation.
The results are shown in Fig.$\,$\ref{fig:nz}. Within the range of consideration, the ratio of the nulled and the original biases drops quickly with the increase of the number of redshift bins for all GGI models. For most of the models, 5 redshift bins seem to be not sufficient for the nulling technique to control the bias induced by GGI down to a percent level. Going from 5 redshift bins to 10 redshift bins is very rewarding in terms of bias reduction.
However, we note that a decrease $|\rm{b}_f/\rm{b}_i|$ doesn't neccessarily indicate a better performance of the nulling method, or generally speaking, of any method intended to control the intrinsic-shear alignment. One can see the reason for this by noticing that, it is the original unbinned GGI/GGG signal that is directly controlled by any of these methods. Between $|\rm{b}_f/\rm{b}_i|$ and the original unbinned GGI/GGG signal lies the binning process as well as the summation over angular frequency bins and redshift bins. Since the signs of the biases contributed by different angular frequencies and redshifts can be different, there can be bias cancellation during these processes. In another word, $|\rm{b}_f/\rm{b}_i|$ can depend on binning choices.
\begin{figure*}[t]
\begin{minipage}[c]{.75\textwidth}
\centering
\includegraphics[width=13cm]{combine.eps}
\end{minipage}
\begin{minipage}[c]{.22\textwidth}
\caption{The thick green (gray) solid, thick blue (black) solid and thin black dashed ellipses indicate 1-sigma (68$\,\%$ CL) parameter constraints from the nulled power spectrum measures, bispectrum measures, and combined. Hidden parameters are marginalized over. The distance from the center of an ellipse to the black cross reflects the nulled bias on the corresponding parameter. The original biases from bispectrum measures can be seen in Fig.\ref{fig:para}. The thin black solid ellipses over-plotted on to the centers of the nulled combined constraint ellipses indicate the statistical power (68$\%$ CL) of combined constraints before nulling. Note the different ranges of parameters compared to Fig.\ref{fig:para}.
}
\label{fig:combined}
\end{minipage}
\end{figure*}
We then vary two cosmological parameters ($\Omega_{\rm m}$ and $\sigma_{\rm 8}$) and investigate how the original and the nulled parameter constraints change with respect to the number of redshift bins available. For this case we use all triangle shapes to enable a comparison with results for two-point statistics.
Our result (Fig.$\,$\ref{fig:fom_nz}) shows that a further increase of the number of redshift bins beyond 10 is not very rewarding in terms of information preservation as characterized by the FoM, in either 2p, 3p, or 2p+3p cases. This suggests, when the possibility of more redshift bins exists, the choice of redshift bin number should be based mainly on the requirement of bias reduction level in case of negligible photometric errors. When there are non-negligible photometric errors, however, the information loss will probably be more severe, as found by \citet{joachimi09a} for the two-point case.
\subsection{The nulling technique as a conditioned compression of data}
\label{sec:compression}
The necessity of carrying out data compression in cosmology has long been recognized \citep[e.g.][]{tegmark97} and has been ever increasing due to the increasing size of the data sets. In cosmic shear studies the survey area of next generation multicolor imaging surveys will be an order of magnitude larger than the current ones. The study of three-point statistics also implies a huge increase in the amount of data directly entering the Fisher-matrix/ likelihood analysis, compared to the two-point case.
\begin{table}[h]
\begin{minipage}[t]{\columnwidth}
\caption{Change of cosmic shear bispectrum statistical power after nulling (null) and linear data compression (compress).}
\label{tab:FoM}
\begin{tabular}[t]{c|c|c|c|c|c}
& i & null & null/i & compress & compress/i \\
\hline\hline
$\Omega_{\rm m}$-$\sigma_8$ & 21455 & 4609 &21.5$\,\%$ & 12242 &57.1$\,\%$\\
$\Omega_{\rm m}$-$w_0$ & 637 & 123 &19.3$\,\%$ & 428 &67.2$\,\%$\\
$\Omega_{\rm m}$-$w_a$ & 145 & 33 &23.0$\,\%$ & 110 &75.9$\,\%$\\
$\sigma_8$- $w_0$ & 434 & 87 &20.0$\,\%$ & 299 &68.9$\,\%$\\
$\sigma_8$- $w_a$ & 101 & 26 &25.4$\,\%$ & 72 &71.3$\,\%$\\
$w_0$- $w_a$ & 11.4 & 2.3 &20.2$\,\%$ & 8.0 &70.2$\,\%$\\
\end{tabular}
\tablefoot{Presented are FoM on two-dimensional parameter planes between cosmological parameters $\Omega_{\rm m}$, $\sigma_8$, $w_0$ and $w_a$. The cosmological parameters $h$, $\Omega_{\rm b}$ and $n_{\rm s}$ are marginalized over. The second column is the FoM from the original bispectrum; the third and fifth columns are FoM from the nulled and the compressed measures, respectively; the fourth (sixth) column shows the percentage of the third (fifth) column compared to the first column, which reflects the relative information loss through the nulling (the unconditioned compression) procedure.}
\end{minipage}
\end{table}
The basic principle of data compression is to reduce the amount of data while preserving most of the information. This is already naturally encoded in the nulling technique. If one keeps only the first-order weights for nulling, as we do in this paper, the nulling procedure reduces the number of data entries in each angular frequency bin from the number of redshift bin $(i,j,k)$ combinations, to the number of $(i,j)$ combinations, which means roughly from ${N_z}^3$ to ${N_z}^2$. The nulling transformation is linear since the resulting nulled entry is a linear combination of $k$ original entries weighted by the nulling weight (\ref{eq:Yij}). In the sense that an ``optimum'' set of nulling weights is constructed, the nulling technique also intends to preserve as much information as possible. But there is yet another additional constraining condition in the nulling procedure: the nulling condition (\ref{eq:nc}), which largely confines the shape of the nulling weights by requiring the existence of at least one zero-crossing (see Fig.\,\ref{fig: poptimal}). In short, the nulling technique can be seen as a conditioned linear compression of data.
\begin{table*
\centering
\caption{FoM before ($`i'$) and after ($`f'$) nulling and their ratio, using the cosmic shear power spectrum (2pt), bispectrum (3pt), and combined (2pt+3pt) analysis.}
\label{tab:FOM23null}
\begin{tabular}[t]{c|c|c|c|c|c|c|c|c|c}
& 2pt,\,i & 3pt,\,i & 2pt+3pt,\,i & 2pt,\,f & 2pt,\,i/f & 3pt,\,f & 3pt,\,i/f & 2pt+3pt,\,f & 2pt+3pt,\,i/f\\
\hline\hline
$\Omega_{\rm m}$-$\sigma_8$ & 21774 & 21455& 86851 & 3297 & 15.1$\,\%$ & 4609 &21.5$\,\%$ & 18555 & 21.4$\,\%$\\
$\Omega_{\rm m}$-$w_0$ & 1590 & 637 & 3806 & 236 & 14.8$\,\%$ & 123 &19.3$\,\%$ & 600 & 15.8$\,\%$\\
$\Omega_{\rm m}$-$w_a$ & 517 & 145 & 872 & 69 & 13.3$\,\%$ & 33 &23.0$\,\%$ & 121 & 13.9$\,\%$\\
$\sigma_8$- $w_0$ & 864 & 434 & 3832 & 132 & 15.2$\,\%$ & 87 &20.0$\,\%$ & 488 & 17.2$\,\%$\\
$\sigma_8$- $w_a$ & 326 & 101 & 709 & 47 & 14.4$\,\%$ & 26 &25.4$\,\%$ & 107 & 15.1$\,\%$\\
$w_0$- $w_a$ & 45 & 11 & 184 & 7.4 & 16.4$\,\%$ & 2.3 &20.2$\,\%$ & 27 & 14.5$\,\%$\\
\end{tabular}
\renewcommand{\footnoterule}{}
\end{table*}
It is then interesting to know how much of the information loss during the nulling process actually comes from the nulling condition, and how much just comes from the fact that a data compression process is naturally involved in nulling. To explore this, we perform an unconditioned linear data compression, by simply ignoring the nulling condition in the whole nulling procedure i.e. dropping the Lagrange multiplier term in (\ref{eq:maxi}), but otherwise keeping the simplifications inherent to the analytical approach. The results are shown in Fig.$\,$\ref{fig:para}. A summary of the FoM from the original and the nulled bispectrum measures as well as the compressed measures is shown in Table$\,$\ref{tab:FoM}.
In contrast to nulling, an unconditioned linear compression does not eliminate the parameter bias, but increases or reduces some of them marginally. Regarding the parameter constraints, although the increase in the size of the ellipses is much less than in the case of nulling, around one third of the information in terms of FoM is lost through compression, which means that the amount of degradation in parameter constraints after compression is not negligible. This suggests that keeping only the first-order terms contributes to non-negligible information loss. To regain part of this information, one could add higher-order weights to the nulling procedure. But the difference between the nulled and the compressed FoM serves as an indication for the inevitable information loss through the nulling process, which is imposed by the nulling condition.
\subsection{Two-point and three-point constraints combined}
Besides constraining cosmological parameters using three-point cosmic shear alone, we investigate the combined constraints from both two-point and three-point cosmic shear measures. The performance of the nulling technique on cosmic shear power spectrum tomography alone and the resulting constraints on cosmological parameters were presented in JS08. For consistency, we use the same setting for the cosmic shear power spectrum as described for the bispectrum in Sect.\,\ref{sec:model}. In particular, we neglect photometric redshift errors, use only a limited range and number of $\ell$-bins, and adopt a power-law intrinsic-shear alignment model with a form described by (36) in JS08 and a slope of 0.4. We have confirmed the consistency between our power spectrum and bispectrum codes with those used in \citet{berge09}. Our power spectrum code agrees also with iCosmo \citep{refregier08}.
Figure \ref{fig:combined} shows the resulting constraint ellipses after nulling from the cosmic shear power spectrum analysis, the bispectrum analysis, and the two combined. To show how much information is lost during the nulling process, we overplot the original two- and three-point combined constraints on top of the nulled constraint ellipses in Fig.\,\ref{fig:combined}, but center them on the corresponding nulled constraints by subtracting the bias difference before and after nulling. The information content in terms of FoM for each parameter pair is presented in Table\,\ref{tab:FOM23null}.
One sees that the amount of information contained in bispectrum measures and power spectrum measures are indeed comparable. With bispectrum information added, typically three times better constraints in terms of FoM are achieved, both before and after nulling. This factor is smaller than the result in TJ04, although the same angular frequency range and the same set of 7 cosmological parameters are chosen for both studies. However a direct comparison is prohibited by different fiducial values adopted and different survey specification.
Through the nulling procedure, around 15$\,\%$ of the original information in terms of FoM is preserved in the two-point case, and around 20$\,\%$ in the three-point case. It is a bit higher in the three-point case, in accordance to the fact that a roughly ${N_z}^3 \rightarrow {N_z}^2$ compression is involved in the three-point case and a ${N_z}^2 \rightarrow {N_z}^1$ one in the two-point case, while this fact is due to the summation over one redshift bin index during the nulling procedure (the same trend is evident in Fig.$\,$\ref{fig:fom_nz}). The information loss is considerable, but it is a price to pay for a model-independent method. As we have discussed in the previous subsection, the difference between the information loss through the nulling and the unconditioned compression procedures represents the inevitable loss of information through nulling. However, this difference is less than 50$\,\%$ in the considered three-point case. The other information loss is due to the simplifications we adopted in this study, including using only the first-order weights, and discarding the measures with two or three equal redshift bins. A further detailed consideration of these aspects can regain part of the lost information. Another simplification we have made in the three-point case is to use only triangles with three different angular frequencies. This reduces both the original and the nulled information contained in the three-point measures. However, this simplification can be easily removed with a careful distinction of all cases.
Also notice that, the dependence of number of possible bispectrum modes, i.e. triangles, on the maximum angular frequency $\ell_{\rm max}$ is roughly $\ell_{\rm max}^3$, while that of power spectrum modes is roughly $\ell_{\rm max}^1$. For this study $\ell_{\rm max}=3000$ is chosen. If reliable information on smaller angular scales can be obtained, the three-point statistics will possibly give us more information than the two-point statistics.
\section{Conclusion}
In this study we developed a method to control the intrinsic-shear alignment in three-point cosmic shear statistics by generalizing the nulling technique. We showed that the generalization of the nulling technique to three-point statistics is quite natural, providing a model-independent method to reduce the intrinsic-shear alignment signals (GGI and GII) in comparison to the lensing GGG signal.
To test the performance of the nulling technique, we assumed a fictitious survey with a setup typical of future multicolor imaging surveys, and applied the nulling technique to the modeled bispectra with intrinsic-shear alignment contamination. The lensing bispectra (GGG) was computed based on perturbation theory, while the GGI signal was modeled by a simple power-law toy model. We focused on the reduction of the GGI contaminant, since GII can be removed simply by not considering tomographic bispectra with two or three equal redshift bins.
The reduction of the intrinsic-shear alignment contamination at the three-point level by the nulling technique was demonstrated both in terms of the GGI/GGG ratio, and in terms of biases on cosmological parameters in the context of an extended Fisher matrix study. In terms of the GGI/GGG ratio, a factor of 10 suppression is achieved after nulling over all angular scales. Correspondingly, the biases on cosmological parameters are reduced to be less than or comparable to the original statistical errors. We studied the performance of the nulling technique when 5, 10, 15, or 20 redshift bins are available, and found that the performance on bias reduction, rather than how much information is preserved during the nulling procedure, depends more significantly on the number of redshift bins. In case one requires better control of intrinsic-shear alignment, more detailed redshift information allowing more redshift bins is the most direct way to go.
When dealing with real data, there is one further source of complication which we did not consider in this paper, that is the photometric redshift uncertainty. The photometric redshift uncertainty can be characterized by a redshift-dependent photometric redshift scatter and catastrophic outliers. \citet{joachimi09a} studied the influence of photometric redshift uncertainty on the performance of the nulling technique at the two-point level. They found that the photometric scatter places strong bounds on the remaining power to constrain cosmological parameters after nulling. The existence of catastrophic outliers, on the other hand, can lead to an incomplete removal of the intrinsic (II, III) alignments as well as the intrinsic-shear alignments (GI, GII, GGI). However, methods to control the photometric redshift uncertainty have been proposed. For example, recent studies concerning the problem of catastrophic outliers point to the solutions of either limiting the lensing analysis to $z<2.5$ or by conducting an additional small-scale spectroscopic survey \citep{sun09, bernstein09, bordoloi09}.
As already demonstrated by JS08 in the two-point case, some information loss is inherent to the nulling procedure.
For the setup of this paper we found that, in terms of FoM about 20$\,\%$ of the original information is preserved through the nulling procedure in the three-point case, and 15$\,\%$ in the two-point case. We further studied the source of such information loss by comparing the nulling technique to an unconditioned linear compression of the data, since the nulling procedure can be seen as a linear compression of data under the constraint of the nulling condition (\ref{eq:nc}). We found that around one third of the original information is lost through an unconditioned compression of the data, suggesting that this situation can be improved by considering higher-order terms in the nulling and compression processes.
Results on parameter constraints from the two- and three-point cosmic shear statistics combined are also presented. The amount of nulled information contained in bispectrum measures and power spectrum measures are comparable. With bispectrum information added, typically three-times better constraints are achieved both before and after nulling, in terms of FoM.
Again, due to the large amount of information existing in the three-point cosmic shear field, one would certainly like to exploit it in the future. The nulling method we developed in this work solves a potentially severe problem hampering the use of three-point information, namely the intrinsic-shear alignment systematic. Our method works at the cost of a large information loss, which can hopefully be avoided by a future method of removing the intrinsic-shear alignment contaminants. But as the only completely model-independent method so far, the nulling technique can serve as a working method now and can provide a valuable cross check even with the availability of better methods.
\begin{acknowledgements}
We thank Joel Berg{\'e} and Masahiro Takada for the help in comparing power spectrum and bispectrum codes. We are also grateful to the anonymous referee for the helpful comments and suggestions. BJ acknowledges support by the Deutsche Telekom Stiftung and the Bonn-Cologne Graduate School of Physics and Astronomy. This work was supported by the RTN-Network `DUEL' of the European Commission, and the Deutsche Forschungsgemeinschaft under the Transregional Research Center TR33 `The Dark Universe'.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{First section}
Contemporary terrestrial experiments,
and extra--terrestrial observatories, indicate that the
Standard $SU(3)_C\times SU(2)_L\times U(1)_Y$ Model of the
gauge interactions, augmented with three families of quarks and
leptons, correctly accounts for the extracted data. The remaining
piece, the scalar Higgs field and the mechanism of electroweak
symmetry breaking, awaits the data and final judgement of the
LHC experiment. Post 1998 data from astrophysical and reactor
neutrino experiments also indicate that neutrinos are
massive, which suggests augmentation of each Standard Model generation
with a Standard Model singlet.
Perhaps its most tantalising property, is the fact that the Standard Model
matter and gauge sectors can be embedded into Grand Unified Theories.
Most appealing in this context is the embedding into $SO(10)$, that
accomodates each matter generation into a single spinorial
16 representation. The gauge charges of the matter states are experimental
observables and the embedding into $SO(10)$ reduces these number
of observables from $54=3\times 3\times 6$, to just one, namely the
number of generations. This could of course be a cruel accident,
but it is the best clue we have for the fundamental origins of the
Standard Model.
Additional evidence for the validity of grand unification exists in the
logarithmic running of the Standard Model parameters, which is compatible
with the data in the gauge and heavy matter sectors. The
Higgs sector receives quadratic corrections from the cutoff scale, and
restoration of its logarithmic evolution requires the introduction of a
new symmetry -- {\it e.g.} supersymmetry. Additionally, proton longevity
and neutrino masses suggest that the unification scale is vastly separated
from the electroweak scale. Indeed, proton stability is an accidental feature
of the Sandard Model at the renormalizable level. Non--renormalizable
operators, suppressed by an effective cutoff scale, which must exist because
the Standard Model is incomplete, will, in general, induce fast proton
decay. The proton lifetime is therefore another fact, set in our bones,
that hints at a large unification scale.
Gran unification, however, cannot be the end of the story. The flavour
data is not explained in this context. The choice of gauge and matter
sectors is also ad hoc. Most importantly, gravity, is not included in
this framework. Furhermore, non--conformal quantum field theories and
general relativity are incompatible at a fundamental level. String theory
is a unique extension of the Standard Model in the sense that
while providing a consistent theory of quantum gravity, it also
gives rise to the gauge and matter structures that appear in the
Standard Model. It provides a unifying framework for all
the basic particle constituents that appear in nature,
and enables the development of a phenomenological approach
to quantum gravity. Furthermore, string theory is predictive.
It predicts that the rank of the gauge interactions is extended,
and should be, for example, below 22 in the heterotic string.
Given that there are five different theories in ten dimensions,
which are all believed to be connected by various duality
transformations, a natural question is which of these theories
should be used as a starting point for phenomenology. The Standard
Model data motivates the hypothesis that the perturbative
string vacuum should accommodate the canonical $SO(10)$ embedding
of its spectrum. Additionally, a perturbative string vacuum
should contain three chiral generations and a pair
of Higgs doublets with potentially viable Yukawa couplings.
Among the five perturbative string theories the only one that
which is compatible with these requirements is the $E_8\times
E_8$ heterotic--string, as it is the only one that produces
spinorial representations in the perturbative massless spectrum.
It may well be that exploring other properties of the
fundamental vacuum will necessitate utilising a different
perturbative limit of the underlying non--perturbative theory.
The ultimate question in this regard is how are the different
limits connected, and how can we be guaranteed that we are exploring
different limits of a single nonperturbative vacuum.
The quasi--realistic free fermionic heterotic--string models
were constructed in the late eighties and early nineties
\cite{fsu5,fny,alr,nahe,eu,top,cslm,cfn,cfs,su421,fmt,cfmt}.
They provide a
concrete framework to study many of the issues that pertain to the
phenomenology of the Standard Model
and grand unification. A few highlights of studies in the free
fermionic standard--like models are listed below:
\begin{itemize}
\item[{$\bullet$}] { Top quark mass $\sim$ 175--180GeV \cite{top}}
\item[{$\bullet$}] { Generation mass hierarchy \cite{fmm}}
\item[{$\bullet$}] { CKM mixing \cite{ckm,fh}}
\item[{$\bullet$}] { Stringy seesaw mechanism \cite{seesaw,fh2,seesawII}}
\item[{$\bullet$}] { Gauge coupling unification \cite{gcu,df}}
\item[{$\bullet$}] { Proton stability \cite{ps}}
\item[{$\bullet$}] { Squark degeneracy \cite{fp2,fv}}
\item[{$\bullet$}] { Minimal Standard Heterotic String Model (MSHSM)
\cite{cfn}}
\item[{$\bullet$}] { Moduli fixing \cite{modulifixing}}
\item[{$\bullet$}] {
Classification \cite{nooij, classification} \&
spinor--vector duality \cite{spduality,tristan}}
\end{itemize}
The fermionic formulation was developed in the mid--eighties \cite{abk,klt}.
The equivalence of bosons and fermions is two dimensions
entails that a model constructed
using the fermionic approach corresponds to a model
constructed using the bosonic approach in which the
target--space is compactified on a six dimensional internal manifold.
The free fermionic formalism corresponds to using
a free bosonic formalism in which the radii of the internal dimensions
are fixed at a special point in the compact space.
Deformation from the special point in the moduli space are parametrized
in terms of world--sheet Thirring interactions among the world--sheet
fermions \cite{egrs}.
The simplicity of the free fermionic formalism entails that
the string consistency constraints are solved in terms of
the world--sheet free fermion transformation properties on the string
world--sheet, which are encoded in sets of basis vectors and
one--loop GSO projection coefficients among the basis vectors.
The formalism to extract the physical spectrum and superpotential
interaction terms are also straightforward. The simplest free fermionic
constructions correspond to a $Z_2\times Z_2$ orbifold of a six dimensional
toroidal manifold, augmented with discrete Wilson lines that are needed
to break the $SO(10)$ GUT symmetry.
Perhaps, the most tantalising achievement is the successful calculation of the
top quark mass \cite{top},
which was obtained several years prior to the experimental
discovery. This calculation demonstrats how string theory enables the
calculation of the fermion--scalar Yukawa couplings in terms of the
unified gauge coupling. Furthermore, the string models offer
an explanation for the hierarchical mass splitting between the
top and bottom quarks.
The top quark Yukawa coupling is obtained at the cubic level
of the superpotential and is of order one, whereas the Yukawa
couplings of the lighter quarks and leptons are obtained from
nonrenormalizable operators that are suppressed relative to the
leading cubic level term. Thus, only the top quark mass is
characterised by the electroweak scale and the masses of the lighter
quarks and leptons are naturally suppressed compared to it.
As the heavy generation Yukawa couplings
are obtained at low orders in the superpotential, the calculation of these
Yukawa couplings is robust and is common to a large class of models.
Further analysis of fermion masses shows that
quasi--realistic fermion mass textures arise for reasonable
choices of supersymmetric flat directions \cite{ckm}. Issues like left--handed
neutrino masses \cite{fh2}, gauge coupling unification \cite{gcu,df},
proton stability \cite{ps} and squark
degeneracy \cite{fp2,fv}
were studied in concrete quasi--realistic free fermionic string
models and for detailed solutions of the supersymmetric flat direction
constraints.
It was also demonstrated in ref. \cite{cfn} that the free fermionic
heterotic string vacua give rise to models that produce in the
observable charged sector below the string
unification scale solely the matter
spectrum of the minimal supersymmetric standard model.
Such models are dubbed Minimal Standard Heterotic
String Models (MSHSM). The free fermionic models also provide important
clues to the problem of moduli fixing in string theory \cite{modulifixing}.
They
highlight the fact that string theory may utilize geometrical structures
that do not have a classical correspondence. Primarily, they allow boundary
conditions that distinguish between the left-- and right--moving
coordinates of the six dimensional compactified space.
Such boundary conditions necessarily lead to the projection
of the moduli fields associated with the extra internal coordinates.
The free fermionic models have also been instrumental in recent years to
unravel a new duality symmetry under the exchange of spinor and vector
representations of the GUT group \cite{spduality, tristan}.
The free fermionic formalism is also adaptable to a computerised
classification, of $Z_2\times Z_2$ orbifolds with symmetric
shifts \cite{nooij, classification}.
A generic property of string vacua is the existence of fractionally charged
states, due to the breaking of the $SO(10)$ GUT symmetry by discrete Wilson
lines. The lowest state among those is stable due to electric charge
conservation. Experimental constriants severely constrain the mass
and abundance of such such particles. Using the free fermionic
classification method we recently constructed examples of Pati--Salam
string models that are completely free of massless exotics \cite{exophobic}.
The quasi--realistic free fermion models were constructed in the
late eighties and early nineties, and continue to be represent
viable string vacua that merit deeper exploration into their properties.
In particular, their relation to $Z_2\times Z_2$ orbifolds is of
vital importance. One should note that the different formalisms of
string compactifications are related and provide complementary insight
to the properties of the underlying vacua. So, for example,
while the free fermionic approach provides easy access to the
massless spectra and potential, and is adaptable to a computerised
classification, orbifold representation may be more amenable to study
moduli dynamics and vacuum selection. In this vein, it is vital to
establish a detailed dictionary between the two laguages.
While each language may be well understood on its own,
writing a detailed dictionary between the two is often non--trivial.
In this talk I highlight how this correspondence can be established
for two key properties of the models, and is indeed found to be highly
non--trivial. The first is with respect to the $Z_2\times Z_2$
orbifold action on the $SO(12)$ lattice at the free fermionic point.
It is well known that a $Z_2\times Z_2$ orbifold of a generic six
dimensional torus produces 48 fixed points, whereas $Z_2\times Z_2$
of the $SO(12)$ lattice produces 24 fixed points. The 48 points can
be reduced to 24 by adding a freely acting shift, under which
pairs of fixed points are identified. Now one can add different
freely acting shifts to this effect,
and the question is which one reproduces the free fermionic model.
Namely, it is not sufficient
to examine the massless spectrum for this purpose, but one must also
establish the
correspondence of the massive spectrum. String theory affords even more
interesting possibilities
because, in addition to the ordinary Kaluza--Klein momentum modes, it
also gives rise to
space--like winding modes. Therefore, one can add freely acting shifts
that shift the momenta,
winding, as well as asymmetric shifts that act on both the momenta and
winding and are therefore
intrinsicallly stringy! It turms out that the shifts that reproduce
the partition function at the
free fermionic point are uniquely of the third kind.
The second property, related to the free
fermion--orbifold correspondence, illustrates the utility of the free fermion
method in accessing classes of string vacua that are
obscure from the point of view of the orbifold contruction.
In the orbifold models one typically starts with the
$E_8\times E_8$ heterotic--string compactified to
four dimensions and subsequently break one $E_8$ gauge factor
by the orbifold twistings and discrete Wilson lines.
In contrast the quasi--realistic free fermionic models
start with $SO(16)\times SO(16)$, where the reduction from
$E_8\times E_8$ to $SO(16)\times SO(16)$ is realised in
terms of a Generalised GSO (GGSO) phase in the partition function.
One can then show that the partition function at the free fermionic
point is generated by starting with the partition function of the
$E_8\times E_8$ and acting on it with a freely acting $Z_2\times
Z_2$ twist given by
$Z_2^a\times~Z_2^b~:~=
(-1)^{F_{\xi^1}}\delta_1 ~\times~ (-1)^{F_{\xi^2}}\delta_1 \,$
with
$\delta_1 X^9 = X_9 +\pi R_9~,$
where $F_{\xi^{1,2}}$ are fermion numbers acting on the first and
second $E_8$ factors,
and $\delta_9$ is a mod 2 shift in one internal direction
\cite{parti}.
The action of the $Z_2\times Z_2$
freely acting twist reduces the $E_8\times E_8$ symmetry to
$SO(16)\times SO(16)$.
The next step in the construction employs a $Z_2^c$ twist
that acts on the internal
coordinates and produces 16 fixed points. The gauge symmetry is broken to
$E_7\times SU(2)$, and $SO(12)\times SO(4)$, in each case, respectively.
The matter and Higgs states in the first case are in the 56
representation of $E_7$,
which breaks as $56 = (32,1) + (12,2)$ under $SO(12)\times SU(2)$, where
the $32$ spinorial, and the $12$ vectorial, representations of $SO(12)$
contain the matter and Higgs representations of the Standard Model,
respectively.
Hence, the freely acting $Z_2^a\times Z_2^b$ twists induce a string
Higgs--matter
splitting mechanism, similar to the string doublet--triplet splitting
mechanism \cite{ps}.
However, it turns out that simply adding the $Z_2^c$ twist, {\it i.e.} taking
$[Z_+/(Z_2^a\times Z_2^b)]/Z_2^c$, where $Z_+$ is the $E_8\times E_8$
partition function, projects the twisted massless spinorial representations
and keeps the vectorials.
To restore the matter representations one needs to analyse
the full $Z_+/(Z_2^a\times Z_2^b\times Z_2^c)$ partition and take into account
the seven independent modular orbits (seven discrete torsions) \cite{cft}. A
non--trivial proposition indeed! Now, one should note that this is merely
one example. The free fermionic classification accesses $2^5$ vacua at this
level, and by breaking the hidden $SO(16)$ factor to $SO(8)\times SO(8)$
this is increased to $2^9$. We therefore note that the free fermionic
approach provides a superior tool to easily access
large classes of string vacua.
The bosonic formalism provides a more intuitive approach to the string moduli
and hopefully to the dynamics that fix the string vacuum.
\begin{acknowledgement}
This work is supported in part by the
STFC under grant ST/G00062X/1 and the EU under
contract MRTN-CT-2006-035863.
\end{acknowledgement}
|
\section{Introduction}
\label{intro}
Mounting evidence suggests that a dominant fraction of bulges in
fairly massive ($M_{\ast}\sim 10^{10} - 10^{11}$~M$_{\odot}$)
$z\sim$~0 galaxies (e.g., Weinzirl et al. 2009) as well as the bulk of
the cosmic star formation rate density since $z < $~1 in intermediate
and high mass galaxies, are not triggered by ongoing major mergers
\citep{Hammer05, Wolf05, Bell05, JogeeIAU, Jogee08} but are likely
related to a combination of minor mergers and internally-driven
secular processes
Barred disks have been studied extensively in the nearby
universe. Quantitative methods show that approximately 45\% of disk
galaxies are barred in the optical at $z\sim$~0 \citep{MJ07, Reese07,
BJM08, Aguerri08}. Bars drive gas to the centers of galaxies,
where powerful starbursts can be ignited \citep{Schwarz81, KK04,
Jogee05, Sheth05}. These starbursts caused by bar-driven inflow may
help build up central, high $v/\sigma$, stellar concentrations called
disky bulges or pseudobulges \citep{Kormendy82, Kormendy93, Jogee05,
Weinzirl08}. Bars can also create boxy/peanut bulges through
vertical buckling \citep{Bureau99, Ath05, MVShlos06} and resonance
scattering \citep{Combes81, Combes90}.
But what is the relationship between bar-driven secular evolution and
environmental effects? Little is known about barred galaxies in dense
environments, as the situation is complicated by the relative
importance of processes such as ram pressure stripping, galaxy tidal
interactions, mergers, and galaxy harassment. In addition, further
complications are introduced by the fact that the bar fraction and
properties in clusters depend on the epoch of bar formation and the
evolutionary history of clusters. We explore these questions using the
Space Telescope Abell 901/902 Galaxy Evolution Survey (STAGES;
\citealp{Gray08}).
\section{Data and sample}
\label{data}
The main data for this study comes from the STAGES survey
$0\fdg5\times0\fdg5$ degree HST/ACS F606W mosaic of the Abell 901/902
cluster system at $z\sim$~0.165. Spectro-photometric redshifts for all
galaxies are available from the COMBO-17 survey \citep{Wolf04}, with
an accuracy of $\delta z/(1+z)\sim$~0.01 for the sample used in this
study. In addition, X-ray maps of the intra-cluster medium (ICM)
density from XMM-Newton, UV from GALEX, Spitzer 24$\mu$m coverage,
dark matter maps from weak lensing \citep{Heymans08}, and stellar
masses \citep{Borch06} are available for this field.
We use the spectro-photometric redshifts to select a sample of 785
bright ($M_{V} \le -18$), cluster galaxies. The point spread function
(PSF) for the ACS images is $0\farcs1$, corresponding to $\sim$~282 pc
at $z\sim$~0.165\footnote{We assume in this paper a flat cosmology
with $\Omega_{\rm m} = 1 - \Omega_{\Lambda} = 0.3$ and $H_{\rm 0}$
=70~km~s$^{-1}$~Mpc$^{-1}$.}. While this resolution is sufficient
for detecting primary bars in spiral galaxies, bars in dwarf galaxies
may often be smaller than the detection limit
(Sect.~\ref{method}). Therefore, in this study we focus on galaxies
brighter than $M_{V} = -18$ to avoid dwarf galaxies, to keep field
contamination low ($\sim$~10\%; \citealp{WGM05}), and to maintain
adequate resolution for bar detection through ellipse-fitting. From
the 785 bright cluster galaxies we are able to ellipse-fit 762 (97\%;
Sect.~\ref{barmeth}). This is the total sample from which all other
subsamples are derived in the subsequent analysis.
\section{Method}
\label{method}
\subsection{Selecting disk galaxies}
\label{diskselmeth}
Because bars are a disk phenomenon, the bar fraction $f_{\rm bar}$ is
defined as the number of barred disk galaxies over the total number of
disk galaxies:
\begin{equation}
f_{\rm bar} = \frac{N_{\rm barred}}{N_{\rm disk}} = \frac{N_{\rm
barred}}{N_{\rm barred} + N_{\rm unbarred}}.
\end{equation}
For this reason, it is necessary to select a sample of disk galaxies
from our total cluster sample.
We compare three widely used methods of selecting disk galaxies:
visual classification, $U-V$ blue-cloud cut, and S{\'e}rsic cut. We
visually classify all galaxies in the sample into broad bins of $B/T$:
`pure bulge', `bulge+disk', and `pure disk'. For the blue-cloud cut,
only galaxies on the blue cloud in color-magnitude space are selected
as disks. For the S{\'e}rsic cut, only galaxies with single-component
S{\'e}rsic index $n < 2.5$ are selected as disks.
\subsection{Identifying and characterizing bars}
\label{barmeth}
We use ellipse-fitting to identify and characterize bars (e.g.,
\citealp{Wozniak95, Friedli96, Jogee04, Knapen00, Sheth03, MJ07}). We
use the standard IRAF task `ellipse' to fit isophotes to each galaxy
out to a radius $a_{\rm max}$ where the surface brightness reaches sky
level. We generate radial profiles of the surface brightness,
ellipticity ($e$), and position angle (PA), as well as overlays of the
fitted ellipses onto the galaxy images. We use the radial profiles and
overlays (as an extra check) to classify galaxies as highly inclined
($i > 60^{\circ}$), unbarred, or barred. A galaxy is classified as
barred if (a)~the $e$ profile rises to a global maximum ($> 0.25$)
while the PA stays constant and (b)~after the global maximum near the
bar end, $e$ drops and the PA changes indicating the outer disk
region. We do not attempt to classify highly inclined ($i >
60^{\circ}$) galaxies as barred or unbarred with
ellipse-fitting. However, morphological features such as bars and
spiral arms are classified regardless of inclination during visual
classification of the sample.
After excluding highly inclined galaxies (226 or 37\%) and those with
poor fits (32 or 5\%), we are left with 350 moderately inclined ($i <
60^{\circ}$), bright ($M_{V} \le -18$), ellipse-fitted,
visually-identified disk galaxies. To ensure adequate resolution for
bar identification, we make a cut in disk size $a_{\rm disk}$ of
3~kpc. This reduces our sample by 10 galaxies, leaving us with 340
bright, moderately inclined, ellipse-fitted, visually-identified disk
galaxies.
\section{Results}
\subsection{Disk selection}
We compare three widely used methods of disk selection: S{\'e}rsic
cut, blue-cloud cut, and visual classification. The three methods pick
out 467, 341, and 608 ellipse-fitted disk galaxies,
respectively. Fig.~\ref{diskmiss}a shows that when a S{\'e}rsic cut is
used, 31\% of disk galaxies are missed compared to visual
classification. This underscores the difficulty in selecting disks
using S{\'e}rsic and color cuts, because of the prevalence of S0s and
red spirals in low-z cluster environments. When a blue-cloud cut is
used (Fig.~\ref{diskmiss}b), 51\% of visually-identified disk galaxies
are missed. There is inherent difficulty in distinguishing unbarred
disk galaxies without spiral arms from spheroids using visual
classification. We can obtain a lower limit to how many galaxies are
disks in the red sample or with $n > 2.5$, by only considering
galaxies exhibiting unambiguous disk signatures such as bars and
spiral arms. We find that at least 25\% (67/267) of galaxies with
$n>2.5$ and 22\% (84/390) of galaxies on the red sample exhibit such
features.
\begin{figure*}[t!]
\resizebox{\hsize}{!}{\includegraphics[clip=true]{marinova_f01.eps}}
\vskip -2.8 in
\caption{\footnotesize
(a)~Rest-frame $U-V$ color vs. $M_{V}$ absolute magnitude.
The solid line shows the separation between the red sample and blue
cloud. (b)~Rest-frame $U-V$ color vs. S{\'e}rsic index $n$. The
solid line shows the cut at $n = 2.5$. In both panels,
moderately-inclined disks with ellipse-fit identified bars are shown
as green triangles, while for inclined galaxies ($i>60^{\circ}$),
barred galaxies identified through visual classification are plotted
as blue triangles. Unbarred disks with spiral arms (from visual
classification) are shown as pink squares for all inclinations. Disk
galaxies identified with visual classification. Unbarred disks
without spiral arms are plotted as black points. A large fraction of
visually-identified disk galaxies are missed by the color cut and
S{\'e}rsic methods.}
\label{diskmiss}
\end{figure*}
\subsection{Optical bar fraction}
For moderately inclined disks, the three methods of disk selection
yield a similar \textit{global} optical bar fraction ($f_{\rm
bar-opt}$) of 34\%$^{+10\%}_{-3\%}$, 31\%$^{+10\%}_{-3\%}$, and
30\%$^{+10\%}_{-3\%}$, respectively. However, the \textit{global}
optical bar fraction may not reveal the full picture, as there is
evidence that in the field, the optical bar fraction is a function of
$B/T$, and possibly color \citep{BJM08, Aguerri08}. We therefore
explore how the optical bar fraction varies as a function of $M_{V}$
and qualitative measures of $B/T$ in the Abell 901/902 cluster
system. We note that while we did not quantitatively measure $B/T$ in
the sample, we visually classified the extreme ends of the
distribution into the broad classes described
above. Table~\ref{fbar_BT} shows that in the cluster, the optical bar
fraction is a strong function of $B/T$ and absolute magnitude. At a
given morphological type, $f_{\rm bar-opt}$ rises for brighter
galaxies. At a given $M_{V}$, $f_{\rm bar-opt}$ rises for `pure disk'
galaxies.
\begin{table*}
\caption{Optical bar fraction as a function of host absolute
magnitude and morphological class}
\label{fbar_BT}
\begin{center}
\begin{tabular}{lcc}
\hline
\\
$M_{V}$ range & Bulge+Disk & Pure Disk\\
\hline
\\
$-18 \ge M_{V} > -19$& 13\%$\pm4\%$ (8/63) & 39\%$\pm8\%$ (16/41) \\
$-19 \ge M_{V} > -20$& 19\%$\pm4\%$ (17/88)& 59\%$\pm9\%$ (16/27) \\
$-20 \ge M_{V} > -21$& 40\%$\pm6\%$ (29/73) & 63\%$\pm17\%$ (5/8) \\
$-21 \ge M_{V} > -22$& 63\%$\pm8\%$ (22/35) & --- \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{figure*}[t!]
\vskip -1 in
\resizebox{\hsize}{!}{\includegraphics[clip=true]{marinova_f02.eps}}
\vskip -0.7 in
\caption{\footnotesize
The fraction of barred galaxies a function of:~ (a) distance from
nearest cluster center, (b) log$\Sigma_{10}$, (c) $\kappa$, and (d)
ICM density. Bar classifications are from ellipse fits and disks
are identified by visual classification. The vertical dashed lines
denote the core radius at 0.25~Mpc and the virial radius at
1.2~Mpc. Between the core and the virial radius of the cluster
$f_{\rm bar-opt}$ varies at most by a factor of $\sim$~1.3, allowed
by the error bars.}
\label{densityall}
\end{figure*}
The result that $f_{\rm bar-opt}$ rises toward disk-dominated galaxies
may be interpreted in the context of bar formation models. If bars
form and grow through the swing amplification, a large bulge that
leads to an inner Lindblad resonance (ILR) may inhibit bar formation
by cutting off the feedback loop. Early simulations with rigid halos
suggested that a hot DM halo component can inhibit the formation of a
bar (e.g., \citealp{OP73}). However, more recent simulations with live
halos show that bars grow through the exchange of angular momentum
with the DM halo \citep{DebSell98, DebSell00, Ath03}. These studies
suggest that the evolution of the bar is intimately tied to the mass
distribution of the DM halo within the radius of the galaxy disk. If
fainter galaxies have more dominant DM halos (e.g., \citealp{Persic96,
Kassin06}), bars may develop and grow more slowly in these galaxies,
however in the end, they may be stronger.
\subsection{Bars as a function of local density}
How does the optical bar fraction vary as a function of local density
in the cluster? We use four available measures of local density:
projected distance to nearest cluster center, $\kappa$, $\Sigma_{10}$,
and ICM density. $\kappa$ is line-of-sight projected surface mass
density, and $\Sigma_{10}$ is the local galaxy number
density. Figure~\ref{densityall} shows that the optical bar fraction
varies at most by a factor of 1.3 between the core and virial radius
of the clusters. Our results agree with Aguerri et al. (2009) who
find no dependence of the optical bar fraction on local environment
density, over a wide range of environments comparable to the outer
region and outskirt of our cluster sample. Theoretical models of bar
formation in a cosmological context, suggest that interaction with the
halo sub-structure induces bars \citep{Shlos08}. Because this
substructure is present in all environments, these models imply a
similar bar fraction across a large range of environment densities,
which is consistent with our results.
In the core, our data does not allow a firm conclusion, because the
four tracers of local density show opposing trends, projection effects
become severe, and number statistics are low. Previous studies using
visual identification of bars have argued that the optical bar
fraction increases in the Coma and Virgo cluster cores
\citep{Thompson81, Andersen96}. If this trend is real, it could be
caused by competing effects. Our results suggest that because
bulge-dominated galaxies, which are prevalent in cluster cores, host a
lower fraction of bars, the bar fraction should decrease
there. However, the high velocity dispersion in the core regions
favors numerous, short-lived, tidal interactions, which may increase
the bar fraction, while preserving intact the galaxy disk.
\begin{acknowledgements}
IM, SJ, and AH acknowledge support from NSF grant AST 06-07748, NASA
LTSA grant NAG5-13063, as well as HST G0-10395.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\paragraph{Motivation.}
From the viewpoint of \emph{reasoning techniques}, two main trends
in modeling in Service Oriented Computing (SOC) can be singled out.
On the one hand, an \emph{operational approach} focuses on how process interactions can lead to correct configurations.
Typical
representatives of this approach
are based on process calculi and Petri nets (see, e.g., \cite{vanderaalst1998apn,boreale2006ssc,lanese2007doa,lapadula7cows}), and count with behavioral equivalences and type disciplines as main analytic tools.
On the other hand, in a \emph{declarative approach}
the focus is on the set of conditions components should fulfill in order to be considered correct,
rather than on the complete specification of the control flows within process activities (see, e.g., \cite{vanderaalst2006dtt,pesic2006daf}).
Even if these two trends address similar concerns, we find that they have evolved rather independently from each other.
The quest for a unified approach in which
operational and declarative techniques can harmoniously converge is therefore a legitimate research direction.
In this paper we shall argue that Concurrent Constraint Programming (CCP) \cite{cp-book} can serve as a foundation for such an approach.
Indeed, the unified framework for operational and logic techniques
that CCP provides can be
fruitfully exploited for analysis in SOC, possibly in conjunction with other techniques such as type systems.
Below we briefly introduce the \ccp model and then elaborate
on how it can shed light on a particular issue: the analysis of
structured communications.
\ccp\ \cite{cp-book} is a well-established model for concurrency where processes interact with each other by \emph{telling} and \emph{asking} for pieces of information (\emph{constraints}) in a shared medium, the \emph{store}.
While the former operation simply adds a given constraint to the store (thus making it available for other processes),
the latter allows for rich, parameterizable forms of process synchronization.
Interaction is thus inherently \emph{asynchronous}, and can be related to a broadcast-like communication discipline, as opposed to the point-to-point discipline enforced by formalisms such as the $\pi$-calculus \cite{SaWabook}.
In \ccp, the information in the store grows monotonically, as constraints cannot be removed.
This condition is relaxed in \emph{timed} extensions of \ccp (e.g., \cite{tcc-lics94,NPV02}), where processes evolve along a series of \emph{discrete time units}.
Although each unit contains its own store, information is not automatically transferred from one unit to another.
In this paper we shall adopt a \ccp process language that is timed in this sense.
In addition to the traditional operational view of process calculi, \ccp enjoys a
\emph{declarative} nature that distinguishes it from other models of concurrency: \ccp programs can be seen, at the same time, as computing agents and as logic formulas \cite{cp-book,NPV02,Olarte:08:PPDP}, i.e., they can be read and understood as logical specifications.
Hence,
CCP-based languages are
suitable for \emph{both} the specification and
verification
of programs. In the \ccp\ language used in this paper, processes can be interpreted as linear-time temporal logic formulas;
we shall exploit this correspondence to verify properties of our models.
\paragraph{This Work.}
We describe initial results on the definition of a formal framework for the declarative analysis of structured communications.
We shall exploit \utcc \cite{Olarte:08:SAC}, a timed CCP process calculus, to give
a declarative
interpretation to the language defined by Honda, Vasconcelos, and Kubo in \cite{honda1998lpa} (henceforth referred to as \hvk).
This way, structured communications can be analyzed in a declarative framework where
time is defined explicitly.
We begin by proposing an encoding of the \hvk language into \utcc
and studying its correctness.
We then move to the timed setting, and propose \hvkplus, a timed extension of \hvk.
The extended language explicitly includes information on session duration,
allows for declarative preconditions within session establishment constructs,
and features a construct for session abortion.
We then discuss how the encoding of \hvk into \utcc straightforwardly extends to \hvkplus.
\paragraph{A Compelling Example.}
We now give intuitions on how a declarative approach could be useful in the analysis of structured communications.
Consider the ATM example from \cite[Sect. 4.1]{honda1998lpa}.
There, an ATM has established two sessions:
the first one with a user, sharing session $k$ over service $a$,
and the second one with the bank, sharing session $h$ over service $b$.
The ATM offers \texttt{deposit}, \texttt{balance}, and \texttt{withdraw} operations.
When executing a withdraw, if there is not enough money in the account, then an \emph{overdraft} message appears to the user.
It is interesting to analyze what occurs when
this scenario is extended
to consider a
card reader that acts as a malicious interface between the user and the ATM. The user communicates his personal data with the reader using the service $r$, which will be kept by the reader after the first withdraw operation
to continue withdrawing money without the authorization of the user.
A greedy card reader could even withdraw repeatedly until causing an overdraft, as expressed below:
{\footnotesize \centering
\begin{tabular}{lcl} \\
${\it Reader}$ & = & $ \accept{r}{k'} \receive{k'}{id}$\\
& & \quad $ \request{a}{k} \send{k}{id} $
\qquad $\branching{k'}{ \begin{array}{l}
withdraw: \receive{k'}{amt} \\
~~\selection{k}{withdraw} \send{k}{amt} \\
\quad \branching{k}{dispense : \selection{k'}{dispense} \send{k}{amt} R(k,amt) || overdraft: Q}
\end{array}
}$\\
${\it R(j,x)}$ & = & $\recursionH{R'} \selection{k}{withdraw} \send{j}{x} \branching{j}{dispense : \receive{j}{amt} R' || overdraft: Q}$\\
${\it User}$ & = & $ \request{r}{k'} \send{k'}{myId}$\\
& & \quad $\selection{k'}{withdraw} \send{k'}{58}
\tab $\branching{k'}{dispense : \receive{k'}{amt} P || overdraft: Q}$ \\ \\
\end{tabular}
}
By creating sessions between them, the card reader $Reader$ is able to receive the
user's information, and to use
it later by attempting a session establishment with the bank.
Following authentication steps (not modeled above),
the card reader allows the user to obtain the requested amount.
Additional withdrawing transactions between the reader and the bank are defined by the recursive process $R$.
In the specification above, the process $Q$ can be assumed to send a message (through a session with the bank)
representing the fact that
the account has run out of money: $Q = \send{k_{bank}}{\underline{\mathsf{0}}} \inact$.
Even in this simple scenario,
the combination of operational and declarative reasoning techniques
may come in handy to reason about
the possible states of the system.
Indeed, while an operational approach
can be used to describe an operational description of the compromised ATM above,
the declarative approach can complement such a description by
offering declarative insights regarding its evolution.
For instance, assuming $Q$ as above,
one could show that a \utcc specification of the ATM example
satisfies the linear temporal logic formula $\sometime \outp(k_{bank}, 0)$,
which intuitively means that in presence of a malicious card reader
the user's bank account will eventually reach an overdraft status.
\paragraph{Related Work.}
One approach to combine the declarative flavor of constraints and process calculi techniques is represented by a number of works that
have extended name-passing calculi with some form of partial information (see, e.g., \cite{VictorP98,DiazRV98}).
The crucial difference between such a strand of work
and \ccp-based calculi
is that the latter offer a tight correspondence with logic, which greatly broadens the spectrum of reasoning techniques at one's disposal.
Recent works similar to ours include CC-Pi \cite{BuscemiM07} and the calculus for structured communications in \cite{CD08}.
Such languages feature elements that resemble much ideas underlying CCP (especially \cite{BuscemiM07}).
The main difference between our approach and such works is that
we adhere to the use of declarative reasoning techniques based on temporal logic
as an effective way of complementing operational reasoning techniques.
In \cite{BuscemiM07},
the reasoning techniques associated to CC-Pi are essentially operational, and used to reason about service-level agreement protocols.
In \cite{CD08},
the key for analysis is represented by a type system which provides consistency for session execution, much as in the original approach in \cite{honda1998lpa}.
\section{Preliminaries}
\label{s:preli}
\subsection{A Language for Structured Communication}
We begin by introducing \hvk, a language for structured communication proposed in \cite{honda1998lpa}. We assume the following conventions:
\emph{names} are ranged over by $a, b, \dots$;
\emph{channels} are ranged over by $k, k'$;
\emph{variables} are ranged over by $x, y, \dots$;
\emph{constants} (names, integers, booleans) are ranged over by $c, c' ,\dots$;
\emph{expressions} (including constants) are ranged over by $e, e', \dots$;
\emph{labels} are ranged over by $l, l', \dots$;
\emph{process variables} are ranged over by $X, Y , \dots$.
Finally, $ u, u', \dots$ denote names and channels. We shall use $\vec{x}$
to denote a sequence (tuple) of variables $x_1...x_n$ of length $n=|\vec{x}|$.
Notation $\vec{x}$ will be similarly applied to other syntactic entities. The sets of free names/channels/variables/process variables of $P$, is defined in the standard way, and are respectively denoted by $\fn(\cdot)$, $\fc(\cdot)$, $\fv(\cdot)$, and $\fpv(\cdot)$. Processes without free variables or free channels are called \emph{programs}.
\begin{definition}[The \hvk language \cite{honda1998lpa}] \label{def-hondasLanguage}
Processes in \hvk are built from:
{\small
\begin{longtable}{rrllrrll}
P,Q &$::=$ & \request{a}{k}P & Session Request
&$|$ & \accept{a}{k}P & Session Acceptance\\
&$|$ & \send{k}{\vec e}P & Data Sending
&$|$ & \receive{k}{\vec x}P & Data Reception\\
&$|$ & \selection{k}{l}P & Label Selection
&$|$ & \branching{k}{l_1:P_1 \parallel \cdots || l_n:P_n} & Label Branching \\
&$|$ & \throw{k}{k'}P & Channel Sending
&$|$ & \catch{k}{k'}P & Channel Reception \\
&$|$ & \itn{e}{P}{Q} & Conditional Statement
&$|$ & P $|$ Q & Parallel Composition\\
&$|$ & \inact & Inaction
&$|$ & $(\nu u)$P & Hiding\\
&$|$ & \recursionH{D}P & Recursion
&$|$ & $X[\vec{e} \, \vec{k} ]$ & Process Variables\\
D &$::=$ & \multicolumn{2}{l}{$X_1(x_1 k_1) = P_1$ $\mathbf{and} \cdots \mathbf{and}$ $X_n(x_n k_n) = P_n$}\\
& &\multicolumn{2}{r}{Declaration for Recursion}
\end{longtable}
}
\end{definition}
\paragraph{Operational Semantics of \hvk.}
The operational semantics of \hvk\ is given by the reduction relation $\redihvk$ which is the smallest relation on processes generated by the rules in Figure \ref{tab:sos-hvk}. In Rule $\textsc{Str}$, the structural congruence $\equiv _h$ is the smallest relation satisfying :
1) $P \equiv _h Q$ if they differ only by a renaming of bound variables (alpha-conversion). 2) $P \ |\ \inact \equiv _h P$, $P \ |\ Q \equiv _h Q \ |\ P$, $(P \ |\ Q) \ |\ R \equiv _h P \ |\ (Q \ |\ R)$. 3) $(\nu u) \inact \equiv _h \inact$, $(\nu u u')P \equiv _h (\nu u' u)P$, $(\nu u )(P \ |\ Q) \equiv _h (\nu u)P \ |\ Q$ if $x \notin \fv(Q)$, $(\nu u)(\recursionH{D}P) \equiv _h (\recursionH{D}((\nu u)P)) $ if $u \notin \fv(D)$. 4) $(\recursionH{D}P) \ |\ Q \equiv _h \recursionH{D}(P \ |\ Q) $ if $\fpv(D)\cap \fpv(Q) = \emptyset$. 5) $\recursionH{D}{(\recursionH{D'}P)} \equiv _h \recursionH{D \mbox{ and } D'}P$ if $\fpv(D)\cap \fpv(D') = \emptyset$.
\begin{figure}[h]
{\small
\[\begin{array}{ll}
{\textsc{Link}} & \request{a}{k}Q \ |\ \accept{a}{k}P \redihvk (\nu k)(P \ |\ Q ) \\
\textsc{Com} & (\send{k}{\vec e}P) \ |\ (\receive{k}{\vec{x}}Q ) \redihvk P \ |\ Q [\vec{c}/\vec{x}] \mbox{\quad if } e\downarrow \vec{c}\\
\textsc{Label} & \selection{k}{l_i}P \ |\ \branching{k}{l_1:P_1 \parallel \cdots \ \parallel l_n:P_n} \redihvk P \ |\ Pi \ (1 \leq i \leq n) \\
\textsc{Pass} & \throw{k}{k'}P \ |\ \catch{k}{k'}Q \redihvk P \ |\ Q \\
\textsc{If1} & \itn{e}{P}{Q} \redihvk P \ (e \downarrow \true) \\
\textsc{If2} & \itn{e}{P}{Q} \redihvk Q\ (e \downarrow \false) \\
\textsc{Def} & \recursionH{D}{(X[\vec{e} \, \vec{k} ] \ |\ Q )} \redihvk
\recursionH{D}{(P[\vec{c}/\vec{x}] \ |\ Q) } \ (e \downarrow \vec{c}, X(\vec{x}\vec{k}) = P \in D)\\
\textsc{Scop} & P \redihvk P' \mbox{ implies } (\nu u)P \redihvk (\nu u)P' \\
\textsc{Par} & P \redihvk P' \mbox{ implies } P \ |\ Q \redihvk P' \ |\ Q \\
\textsc{Str} & \mbox{If } P \equiv _h P' \mbox{ and } P' \redihvk Q' \mbox{ and } Q' \equiv _h Q \mbox{ then } P \redihvk Q
\end{array}\]
}
\caption{Reduction Relation for \hvk\ ($\redihvk$)\cite{honda1998lpa}. \label{tab:sos-hvk}}
\end{figure}
Let us give some intuitions about the language constructs and the rules in Figure \ref{tab:sos-hvk}.
The central idea in \hvk\ is the notion of a \emph{session}, i.e., a series of reciprocal interactions between two parties, possibly with
branching, delegation and recursion, which serves as an abstraction unit for describing structured communication.
Each session has associated a specific port, or \emph{channel}. Channels are generated at session initialization; communications inside the session take place on the same channel.
More precisely, sessions are initialized by
a process of the form
$ \request{a}{k}Q \ |\ \accept{a}{k}P$.
In this case, there is a request, on name $a$, for the initiation of a session and the
generation of a fresh channel.
This request is matched by an accepting process on $a$, which
generates a new channel $k$, thus allowing $P$ and $Q$ to communicate each other.
This is the intuition behind rule $\textsc{Link}$.
Three kinds of atomic interactions are available in the language:
sending (including name passing), branching, and channel passing (also referred to as delegation).
Those actions are described by rules $\textsc{Com}$, $\textsc{Label}$, and $\textsc{Pass}$, respectively.
In the case of $\textsc{Com}$,
the expression $\vec{e}$ is sent on the port (session channel) $k$.
Process $\receive{k}{\vec{x}}Q $ then receives such a data and executes $Q[\vec{c}/\vec{x}]$,
where $\vec{c}$ is the result of evaluating the expression $\vec{e}$.
The case of $\textsc{Pass}$ is similar but considering that in the constructs $\throw{k}{k'}P$ and $\catch{k}{k'}Q$, only session names can be transmitted.
In the case of
$\textsc{Label}$, the process $\selection{k}{l_i}P$ selects one label and then the corresponding process $P_i$ is executed.
The other rules are self-explanatory.
For the sake of simplicity, and without loss of generality (due to rule 5 of $\equiv _h$), in the sequel we shall assume programs of the form $\recursionH{D}P$ where there are not procedure definitions in $P$.
\subsection{Timed Concurrent Constraint Programming}\label{sec:utcc}
Timed concurrent constraint programming (\tcc) \cite{tcc-lics94} extends CCP
for modeling reactive systems. In \tcc, time is conceptually divided into
\emph{time units} (or \emph{time intervals}). In a particular time
unit, a \tcc process $P$ gets an input (i.e. a constraint) $c$
from the environment, it executes with this input as the initial
\emph{store}, and when it reaches
its resting point, it \emph{outputs} the resulting store $d$ to the
environment. The resting point determines also a residual process $Q$
which is then executed in the next time unit.
It is worth noticing that the final store is not automatically transferred to the next time unit.
The \utcc calculus \cite{Olarte:08:SAC} extends \tcc for reactive systems featuring mobility.
Here \emph{mobility} is understood as
the dynamic reconfiguration of system linkage through communication,
much like in the $\pi$-calculus \cite{SaWabook}.
\utcc generalizes \tcc by considering
a \emph{parametric} ask operator of the form $\absp{\vec{x}}{c}{P}$, with the following intuitive meaning:
process $P[\vec{t}/\vec{x}]$ is executed for every term $\vec{t}$ such that the current store entails an admissible substitution $c[\vec{t}/\vec{x}]$.
This process can be seen as an \emph{abstraction} of the process $P$ on the variables $\vec{x}$ under the constraint (or with the \emph{guard}) $c$.
\utcc\ provides a number of reasoning techniques:
First, \utcc processes can be represented as partial closure operators (i.e. idempotent and extensive functions). Also, for
a significant fragment of the calculus,
the input-output behavior of a process $P$ can be retrieved from the set of fixed points of its associated closure operator \cite{Olarte:08:PPDP}. Second, \utcc\ processes can be characterized as
First-order Linear-time Temporal Logic (FLTL)
formulas \cite{mp91}. This declarative view of the processes allows for the use of the well-established verification techniques from FLTL to reason about \utcc\ processes. \\
\noindent{\bf Syntax}.
Processes in \utcc\ are parametric in a \emph{constraint system} \cite{cp-book} which specifies
the basic constraints
that agents can tell or ask during execution.
It also defines an \emph{entailment} relation ``$\entails$'' specifying interdependencies among constraints.
Intuitively, $c \entails d$ means that the information in $d$ can be deduced from that in $c$ (as in, e.g., $x > 42 \entails x > 0$).
The notion of constraint system can be set up by using first-order logic (see e.g., \cite{NPV02}). We assume a first-order signature $\Sigma$ and a (possibly empty) first-order theory $\Delta$, i.e., a set of sentences over $\Sigma$ having at least one model. Constraints are then first-order formulas over
$\Sigma$. Consequently, the entailment relation is defined as follows: $c\entails d$ if the implication $c\Rightarrow d$ is valid in $\Delta$.
The syntax of the language is as follows:
\[
{
\begin{array}{ll}
P,Q := & \skipp \; \; \mbox{\Large{$\mid$}}\;\; \tellp{c} \; \; \mbox{\Large{$\mid$}}\;\; \absp{\vec{x}}{c}{P} \; \; \mbox{\Large{$\mid$}}\;\; P\parallel Q \; \; \mbox{\Large{$\mid$}}\;\; \localp{\vec{x} ; c}{P} \; \; \mbox{\Large{$\mid$}}\;\; \nextp{P} \; \; \mbox{\Large{$\mid$}}\;\; \unlessp{c}{P} \; \; \mbox{\Large{$\mid$}}\;\; \bangp{P}
\end{array}
}
\] with the variables in $\vec{x}$ being pairwise distinct.
A process $\skipp$ does nothing; a process $\tellp{c}$ adds $c$ to the store in the current time interval. A process $Q=\absp{\vec{x}}{c}{P}$ binds the variables $\vec{x}$ in $P$ and $c$. It executes $P[\vec{t}/\vec{x}]$ for every term $\vec{t}$ s.t. the current store entails an admissible substitution over $c[\vec{t}/\vec{x}]$. The substitution $[\vec{t}/\vec{x}]$ is admissible if $|\vec{x}| = |\vec{t}|$ and no $x_i$ in $\vec{x}$ occurs in $\vec{t}$.
Furthermore, $Q$ evolves into $\skipp$ at the end of the time unit, i.e., abstractions are not persistent when passing from one time unit to the next one.
$\parp{P}{Q}$ denotes $P$ and $Q$ running in parallel during the current time unit. A process \(\localp{\vec{x};c}{P} \) \emph{binds} the variables $\vec{x}$ in $P$ by declaring them private to $P$ under a constraint $c$. If $c=\true$, we write $\localp{\vec{x}}{P}$ instead of $\localp{\vec{x};\true}{P}$.
The \emph{unit delay} \( \nextp{P} \) executes $P$ in the next time unit. The \emph{time-out} \( \unlessp{c}{P} \) is also a unit delay, but \( P \) is executed in the next time unit iff \( c \) is not entailed by the final store at the current time unit. Finally, the \emph{replication} \( !\,P \) means \( P\parallel \mathbf{next}\,P\parallel \mathbf{next}^{2}P\parallel
\dots\), i.e.,
an unbounded number of
copies of \( P \) but one at a time.
We shall use $\bangp_{[n]}P$ to denote
\emph{bounded replication}, i.e., $P\parallel \nextp{P}\parallel ... \parallel \nextp^{n-1}P$.
From a programming language perspective, variables $\vec{x}$ in $\absp{\vec{x}}{c}{P}$ can be seen as the formal parameters of $P$. This way, \emph{recursive definitions} of the form $X(\vec{x}) \defsymbol P$ can be encoded in \utcc as
\begin{equation}\label{eq:def-encoding}
\cR\os X(\vec{x}) \defsymbol P \cs = \bangp\absp{\vec{x}}{{\it call_x}(\vec{x})}{\widehat{P}}
\end{equation}
where ${\it call_x}$ is an uninterpreted predicate (a constraint) of arity $|\vec{x}|$.
Process $\widehat{P}$ is obtained from $P$ by replacing recursive calls of the form $X(\vec{t})$ with $\tellp{{\it call_x}(\vec{t})}$. Similarly, calls of the form $X(\vec{t})$ in other processes are replaced with $\tellp{{\it call_x}(\vec{t})}$.
\paragraph{Operational Semantics.}
The operational semantics considers \emph{transitions} between process-store \emph{configurations} \(
\left\langle P,c\right\rangle \) with stores represented as constraints and processes quotiented by the structural congruence $\equiv_u$ defined below. We shall use $\gamma,\gamma',\ldots$ to range over configurations.
The semantics is given in terms of an \emph{internal} and an \emph{observable} transition relation; both are given in Figure \ref{opersem}.
The \emph{internal transition} { \( \left\langle P,d\right\rangle \redi \left\langle P',d' \right\rangle \)} informally means ``\( P \) with store \( d \) reduces, in one internal
step, to \( P' \) with store \( d'\)\,''.
We sometimes abuse of notation by writing $P \redi P'$ when $d,d'$ are unimportant.
The \emph{observable transition} { \( P\rede{(c,d)} R \)} means ``\( P \)
on input \( c \), reduces \emph{in one time unit} to \( R \) and outputs \( d \)''. The latter is obtained from a finite sequence of internal transitions.
In rule $R_S $, the structural congruence \( \equiv _u \)
is the smallest congruence satisfying:
1) \( P \equiv _u Q \) if they differ only by a renaming of bound variables. 2) \( P \parallel \skipp \equiv _u P \). 3) \( P \parallel Q \equiv _u Q \parallel P \), \( P \parallel (Q \parallel R) \equiv _u (P \parallel Q) \parallel
R \). 4) \( P\parallel \localp{\vec{x};c}{Q} \equiv _u \localp{\vec{x};c}{(P \parallel Q)} \) if $\vec{x} \not\in \fv(P)$. 5) $\localp{\vec{x};c}{}\localp{\vec{y};d}{P} \equiv _u
\localp{\vec{x};\vec{y}\ ; c\wedge d}{P}$ \ \ if $\vec{x}\cap\vec{y}=\emptyset $ and $\vec{y} \notin \fv(c)$.
Extend $\equiv _u$ by decreeing that $\mconf{P,c}\equiv _u \mconf{Q,c}$ iff $P\equiv _u Q$.
\begin{definition}[Output Behavior]\label{def:out-behav}
Let $s=c_1.c_2....c_n$ be a sequence of constraints. If $P=P_1\rede{(\true,c_1)}P_2 \rede{(\true,c_2)} \dots P_n \rede{(\true,c_n)} P_{n+1} \equiv_u Q$ we shall write $P\rede{s}^{*} Q$. If $s=c_1.c_2.c_3...$ is an infinite sequence, we omit $Q$ in $P\rede{s}^{*}Q$.
The \emph{output behavior} of $P$ is defined as $\obehav{P} = \{s \ | \ P\rede{s} ^{*}\}$. If $\obehav{P} = \obehav{Q}$ we shall write $P \oequiv{} Q$.
Furthermore, if $P\rede{s} Q$ and $s$ is unimportant we simply write $P\rede{}^{*}Q$.
\end{definition}
\begin{figure}
\centering
\begin{tabular}{c}
\\
\scriptsize
$ \begin{array}{l}
\hline\\
\rightinfer[\rTell]{\left\langle
\tellp{c},d\right\rangle \: \redi_{{}} \: \left\langle
\skipp,d \wedge c\right\rangle}{ }
\tab\tab\tab
\rightinfer[\rPar]{\left\langle P\parallel
Q,c\right\rangle \: \redi_{{}} \: \left\langle P'\parallel
Q,d\right\rangle }{\left\langle P,c\right\rangle \redi_{{}}
\left\langle P',d\right\rangle}
\tab\tab\tab
\rightinfer[\rUnless]{\left\langle \mathbf{unless}\textrm{
}c\textrm{ }\mathbf{next}\textrm{ }P,d\right\rangle \: \redi_{{}} \:
\left\langle \mathbf{skip},d\right\rangle }{d\entails c}
\\\\
\rightinfer[\rLocal]{\left\langle \localp{\vec{x};c}{P},d \right\rangle \: \redi_{{}} \:
\left\langle \localp{\vec{x};c'}{P'},d\wedge \exists {\vec{x}}c'\right\rangle }
{\left\langle P, c \wedge (\exists {\vec{x}} d) \right\rangle \:
\redi \:
\left\langle P', c' \wedge (\exists{\vec{x}} d) \right\rangle \ \ {} }
\tab\tab\tab
\rightinfer[\rAbs]
{\tuple{\absp{\vec{x}}{c}{P}}{d}
\redi \tuple{ P[\vec{t}/\vec{x}] || \absp{\vec{x}}{c \wedge \vec{x} \not\doteq \vec{t}\ }{P}}{d}
}
{ d \entails c[\vec{t}/\vec{x}] \ \ \ \ \ \ \ [\vec{t}/\vec{x}] \mbox{ is admissible }}
\\\\
\ \rStruct \ \frac{\raisebox{.1cm}{$\gamma_1 \redi \gamma_2$}}
{\raisebox{-.2cm}{$\gamma_1' \redi \gamma_2'$}} \ \mbox{ if }
{\gamma_1 \equiv_u \gamma_1'} \mbox{ and } {\gamma_2\equiv_u \gamma_2'}
\tab\tab\tab
\rightinfer[\rBang]{\left\langle
\bangp{P},d\right\rangle \: \redi_{{}} \: \left\langle
P || \nextp{\bangp{P}},d \right\rangle}{ }
\\\\
\hline \\
\ \rObserv \ \frac{\raisebox{.1cm}{$\left\langle
P,c\right\rangle \: \redi_{{}}^{*}\, \left\langle Q,d\right\rangle \:
\not\redi$}}
{\raisebox{-.2cm}{$P\rede{(c,d)} F(Q)$}} \ \mbox{ where \ \ }
{F}(P)=\left\{
\begin{array}{ll}
\skipp & \mbox{ if $ P=\skipp $ or $P=\absp{\vec{x}}{c}{Q} $}
\\ {F}(P_{1})\parallel {F}(P_{2}) & \mbox{ if $P=P_{1}\parallel P_{2} $}
\\ \localp{\vec{x}}{F(Q)} & \mbox{ if $ P=\localp{\vec{x};c}{Q} $ }
\\ Q & \mbox{ if $ P=\nextp{Q} $ or $ P=\unlessp{c}{Q} $}
\end{array} \right.
\\\\
\hline
\end{array}$
\\
\end{tabular}
\caption{\label{opersem} {
Operational Semantics for \utcc. In $\rAbs$, $\vec{x} \not\doteq \vec{t}$ ($\vec{x}$ syntactically different from $\vec{t}$)
denotes $\bigvee_{1\leq i \leq |\vec{x}|}x_i\not\doteq t_i$. If $|\vec{x}| =0$,
$\vec{x} \not\doteq \vec{t}$ is defined as $\false$.}
}
\end{figure}
\paragraph{Logic Correspondence.}
Remarkably, in addition to this operational view,
\utcc processes admit a declarative interpretation based on
Pnueli's first-order linear-time temporal logic (FLTL) \cite{mp91}.
This is formalized by the encoding below, which maps \utcc\ processes into FLTL formulas.
\begin{definition}\label{proc2ltl:def} Let $\Logic\os \cdot \cs$ a map from $\utcc$ processes to FLTL formulas given by:
{\footnotesize
\[
\begin{array}{lllllllllll}
\Logic\os \skipp \cs \!\!&=&\!\! \true & & \Logic\os \tellp{c} \cs \!\!&=&\!\! c \\%& &
\Logic\os P \parallel Q \cs \!\!&=&\!\! \Logic\os P \cs \wedge \Logic\os Q \cs & &
\Logic\os \absp{\vec{y}}{c}{P} \cs \!\!&=&\!\! \forall \vec{y}(c \imply \Logic\os P
\cs) \\% && &
\Logic\os \localp{\vec{x};c}{P} \cs \!\!&=&\!\! \exists \vec{x} (c \wedge \Logic\os P \cs) & & \Logic\os \nextp{P} \cs \!\!&=&\!\!\nextt \Logic\os P \cs \\
\Logic\os \unlessp{c}{P} \cs \!\!&=&\!\! c \lor \nextt \Logic \os P \cs & & \Logic\os \bangp{P}\cs \!\!&=&\!\! \Box \Logic\os P \cs \\
\end{array}
\]
}
\end{definition}
Modalities $\nextt F$ and $\always F$ represent that $F$ holds \emph{next} and \emph{always}, respectively.
We use the \emph{eventual} modality $\sometime F$ as an abbreviation of $\neg \always \neg F$.
The following theorem
relates the operational view of processes with their logic interpretation.
\begin{theorem}[Logic correspondence \cite{Olarte:08:SAC}] \label{theo:FLTL}
Let $\Logic\os \cdot \cs$ be as in Definition \ref{proc2ltl:def}, $P$ a \utcc\ process and $s=c_1.c_2.c_3...$ an infinite sequence of constraints s.t. $P\rede{s}^{*}$. For every constraint $d$, it holds that:
$ \ \Logic\os P \cs \entails \sometime d \ \mbox{ iff there exists $i\geq 1$ s.t. $c_i \entails d$}\,$.
\end{theorem}
Recall that an observable transition $P\rede{(c,c')}Q$ is obtained from a finite sequence of internal transitions (rule $\rObserv$).
We notice that there exist processes that may produce infinitely many internal transitions and as such, they cannot exhibit an observable transition; an example is $\absp{x}{c(x)}{\tellp{c(x+1)}}$. The \utcc\ processes considered in this paper are \emph{well-terminated}, i.e., they never produce an infinite number of internal transitions during a time unit. Notice also that in the Theorem \ref{theo:FLTL} the process $P$ is assumed to be able to output a constraint $c_i$ for all time-unit $i\geq 1$. Therefore, $P$ must be a well-terminated process.
\paragraph{Derived Constructs.}
Let $\outp$ be an uninterpreted predicate.
One could attempt at representing the actions of sending and receiving as in a name-passing calculus (say, $k\bangp[\vec{e}]$ and $\receive{k}{\vec{x}}P$, resp.) with
the \utcc processes $\tellp{\outp(k,\vec{e})}$ and $\absp{\vec{x}}{\outp(k,\vec{x})}{P}$, respectively.
Nevertheless, since these processes are not automatically transferred from one time unit to the next one, they will disappear right after the current time unit, even if they do not interact.
To
cope with this kind of behavior,
we shall define versions of
$\absp{\vec{x}}{c}{P}$ and $\tellp{c}$ processes that are \emph{persistent in time}.
More precisely,
we shall use the process $\waitpp{\vec{x}}{c}{P}$,
which transfers itself from one time unit to the next one until, for some $\vec{t}$,
$c[\vec{t}/\vec{x}]$ is entailed by the current store.
Intuitively, the process behaves like an input that is active until interacting with an output.
When this occurs, the process outputs the constraint $\overline{c}[\vec{t}/\vec{x}]$,
as a way of acknowledging the successful read of $c$.
When $|\vec{x}| = 0$, we shall write $\wheneverp{c}{P}$ instead of
$\waitpp{\vec{x}}{c}{P}$.
Similarly, we define $\tellpp{c}$ for the persistent output of $c$ until some process ``reads" $c$.
These processes can be expressed in the basic \utcc syntax as follows (in all cases, we assume $stop, go \notin \fv(c)$):
{\footnotesize
\[
\begin{array}{llll}
\tellpp{c} & \defsymbol & \localp{go,stop}{} (&
\tellp{\outp'(go)} \parallel \bangp\whenp{\outp'(go)}{\tellp{c}} \parallel \\
& & & \bangp \unlessp{\outp'(stop)}{\tellp{\outp'(go)}} \parallel \\
& & & \bangp \whenp{\overline{c}}{\bangp{\tellp{\outp'(stop)}}}) \\
\waitp{\vec{x}}{c}{P} & \defsymbol &
\localp{stop,go}{}(& \tellp{\outp'(go)} \parallel \bangp{}\unlessp{\outp'(stop)}{\tellp{\outp'(go)}} \\
& & & \parallel \bangp \absp{\vec{x}}{c\wedge \outp'(go)}{(P \parallel\bangp{\tellp{\outp'(stop)}}} ) \\
\waitpp{\vec{x}}{c}{P} & \defsymbol & \multicolumn{2}{l}{ \waitp{\vec{x}}{c}{(P \parallel \tellp{\overline{c}})} }
\end{array}
\]
}
Notice that once a pair of processes $\underline{\mathbf{tell}}$ and $\underline{\mathbf{wait}}$ interact, their continuation in the next time unit is a process able to output only a constraint of the form
$\exists_{x} \outp'(x)$ (e.g., $\exists_{stop}(\outp'(stop))$).
We define the following equivalence relation that allows us to abstract from
these processes.
\index{$\obsequiv{}$}
\begin{definition}[Observables]\label{def:obs:equivalence}
Let $\oequiv{}$ be the output equivalent relation in Definition \ref{def:out-behav}.
We say that $P$ and $Q$ are observable equivalent, notation
$P\obsequiv{} Q$, if $P \parallel \bangp\tellp{\exists_x{\outp'(x)}} \oequiv{} Q \parallel \bangp\tellp{\exists_x{\outp'(x)}} $.
\end{definition}
Using the previous equivalence relation, we can show the following.
\begin{proposition}\label{prop-wait-tell}
Assume that $c(\vec{x})$ is a predicate symbol of arity $|\vec{x}|$.
\begin{enumerate}
\item If $d \notentails c[\vec{t}/\vec{x}]$ for any $\vec{t}$ then $\waitpp{\vec{x}}{c}{P} \rede{(d,d)} \waitpp{\vec{x}}{c}{P} $.
\item If $P \equiv_u \tellpp{c(\vec{t})} \parallel \waitpp{\vec{x}}{c(\vec{x})}{\nextp Q}$ then $P \rede{} \obsequiv{}Q[\vec{t}/\vec{x}]$.
\end{enumerate}
\end{proposition}
\section{A Declarative Interpretation for Structured Communications}
\label{s:decla}
The encoding $[|\cdot|] $ from \hvk into \utcc is defined in Table \ref{table:enc}.
Two noteworthy aspects when considering such a translation are \emph{determinacy} and \emph{timed behavior}.
Concerning determinacy, it is of uttermost importance to
recall that
while \utcc\ is a deterministic language,
\hvk\ processes may exhibit non-deterministic behavior.
Moreover, while
\hvk is a synchronous language, whereas \utcc is asynchronous.
Consider, for instance, the \hvk process:
\[
P =\send{k}{\vec{e}}Q_1 \ | \ \send{k}{\vec{e'}}Q_2 \ | \ \receive{k}{\vec{x}}Q_3
\]
Process $P$ can have two possible transitions,
and evolve into $\send{k}{\vec{e'}}Q_2 \ | \ \ Q_3[\vec{e}/\vec{x}]$
or into $\send{k}{\vec{e}}Q_1 \ | \ \ Q_3[\vec{e'}/\vec{x}]$.
In both cases, there is an output that cannot interact with the input $\receive{k}{\vec{x}}Q_3$.
In \utcc, inputs are represented by abstractions which are persistent during a time unit.
As a result, in the encoding of $P$ we shall observe that \emph{both} outputs react with the same input, i.e. that
$\os P \cs \rede{} \os Q_3[\vec{e}/\vec{x}]\cs \parallel \os Q_3[\vec{e'}/\vec{x}]\cs$.
As for timed behavior,
it is crucial to observe that while
\hvk is an untimed calculus,
\utcc provides constructs for explicit time.
In the
encoding we
shall advocate
a timed interpretation of \hvk in which all available synchronizations between processes occur at a given
time unit,
and the continuations of synchronized processes will be executed in the next
time unit.
This will prove convenient when showing the operational correspondence between both calculi, as we can relate the observable behavior in \utcc and the reduction semantics in \hvk.
Let us briefly provide some intuitions on $[|\cdot|]$.
Consider \hvk processes $P=\request{a}{k}P'$ and $Q=\accept{a}{x}Q'$.
The encoding of $P$ declares a new variable session $k$ and sends it through the channel $a$ by posting the constraint $\reqp(a,k)$. Upon reception of the session key (local variable) generated by $\os P\cs$, process $\os Q\cs$ adds the constraint $\accp(a,k)$
to notify the acceptance of $k$.
They
can then synchronize on this constraint, and execute their continuations in the next time unit.
The encoding of label selection and branching
is similar, and uses constraint $\selp(k,l)$ for synchronization.
We use the parallel composition
$\prod\limits_{1\leq i \leq n} \whenp{l=l_i}{\nextp{\os P_i\cs}}$ to execute the selected choice. Notice that we do not require a non-deterministic choice since the constraints $l=l_i$ are mutually exclusive.
As in \cite{honda1998lpa}, in the encoding of $\itn{e}{P}{Q}$ we assume an evaluation function on expressions. Once $e$ is evaluated, $\downarrow e$ is a \emph{constant} boolean value.
The encoding of $\recursionH{D}P$
exploits the scheme described in Equation \ref{eq:def-encoding}.
\begin{table}[t]
\centering
{\footnotesize
\begin{longtable}{c}
\\
$
\begin{array}{r c l}
[| \request{a}{k}P |] & = &
\localp{k}{}( \tellpp{\reqp(a,k)} \parallel \wheneverp{\accp(a,k)}{\nextp} \os P \cs) \\
[|\accept{a}{k}P |] &= & \waitpp{k}{\reqp(a,k)}{}(\tellp{\accp(a,k)} \parallel \nextp \os P \cs) \\ \\
[| \send{k}{\vec{e}}P |] & = & \tellpp{\outp(k,\vec{e})} \parallel \wheneverp{\overline{\outp(k,\vec{e})}}\nextp{\os P \cs} \\
[| \receive{k}{\vec{x}}P |] &=& \waitpp{\vec{x}}{\outp(k,\vec{x})}{\nextp{\os P \cs}} \\ \\
[| \selection{k}{l}P |] &= & \tellpp{\selp(k,l)} \parallel \wheneverp{\overline{\selp(k,l)}}{\nextp{\os P \cs}} \\
[| \branching{k}{l_1:P_1 || \dots || l_n:P_n} |] & = & \waitpp{l}{\selp(k,l)}{
\prod\limits_{1\leq i \leq n} \whenp{l=l_i}{\nextp{\os P_i\cs}}
} \\
[| \throw{k}{k'}P |] & = & \tellpp{\outkp(k,k')} \parallel \wheneverp{\overline{\outkp(k,k')}}{\nextp{\os P \cs}} \\
[| \catch{k}{k'}P |] & = & \wheneverpp{\outkp(k,k')}{\nextp{\os P \cs}} \\ \\
[| \itn{e}{P}{Q} |] & = & \whenp{e\downarrow\true}{\nextp{\os P \cs}} \parallel \whenp{e\downarrow\false}{\nextp{\os Q \cs}} \\
[| P | Q |] & = & \os P \cs \parallel \os Q \cs \\
[| \inact|] & = & \sskip \\
[| (\nu u) P|] & = & \localp{u}{ \os P \cs} \\
\os \recursionH{D}P \cs & =& \prod\limits_{X_i(x_ik_i) \in D } \cR\os X_i(x_ik_i)\cs {\widehat{P}} \\
\end{array}
\end{longtable}
}
\caption{{An Encoding from \hvk into \utcc. $\cR \os \cdot \cs$ and $\widehat{P}$ are defined in Equation \ref{eq:def-encoding}. \label{table:enc}}}
\end{table}
\paragraph{Operational Correspondence.}
Here we study an operational correspondence property for our encoding.
The differences with respect to (a)synchrony and determinacy discussed above
will have a direct influence on the correspondence.
Intuitively, the encoding falls short for \hvk programs featuring the kind of non-determinism that results from ``uneven pairings'' between session requesters/providers, label selection/branching, and inputs/outputs as in the example above.
We thus find it convenient to appeal to the type system of \hvk\ to obtain some basic determinacy
of the source terms.
Roughly speaking, the type discipline in \cite{honda1998lpa} ensures a correct pairing
between actions and co-actions once a session is established.
Although the type system guarantees a correct match between (the types of) session requesters and providers,
it does not rule out the
kind of non-determinism induced by different orders in the pairing of requesters and providers.
We shall then require session providers to be always willing to engage into a session.
This is, given a channel $a$, we require that there is at most one $\mathbf{accept}$ process (possibly replicated) on $a$
that is able to synchronize with every process requesting a session on $a$.
Notice that this requirement is in line with a meaningful class of programs, namely those
described by the type discipline developed in \cite{Berger2008Completeness-an,BergerHY01}.
Before presenting the operational correspondence, we introduce some
auxiliary notions.
\begin{definition}[Processes in normal form]\label{def:normal-form:hvk}
We say that a \hvk\ process $P$ is in \emph{normal form} if takes the form $\inact$ or $\recursionH{D}\nu \vec{u}(Q_1 \ |\ \cdots \ |\ Q_n)$ where neither
the operators ``$\nu$'' and ``$|$'' nor process variables occur in the top level of $Q_1,\ldots,Q_n$.
\end{definition}
The following proposition states that given a process $P$ we can find a process $P'$ in normal form,
such that: either $P'$ is structurally congruent to $P$, or it results from
replacing the process variables at the top level of $P$ with their corresponding definition (using rule $\textsc{Def}$).
\begin{proposition}
For all \hvk\ process $P$ there exists $P'$ in normal form s.t. $P \redihvk^{*}\equiv _h P'$ only using the rules $\textsc{Def}$ and $\textsc{Str}$ in Figure \ref{tab:sos-hvk}.
\end{proposition}
\begin{proof}
Let $P$ be a process of the form $\recursionH{D}Q$ where there are no procedure definitions in $Q$. By repeated applications of the rule $\textsc{Def}$, we can show that $P\redihvk^{*} P'$ where $P'$ does not have occurrences of processes variables in the top level. Then, we use the rules of the structural congruence to move the local variables to the outermost position and find $P''\equiv _h P'$ in the desired normal form.
\end{proof}
Notice that the rules of the operational semantics of \hvk\ are given for pairs of processes that can interact with each other.
We shall refer to each of those pairs as a \emph{redex}.
\begin{definition}[Redex]\label{d:redex}
A \emph{redex} is a pair of complementary processes composed in parallel as in:
{\footnotesize \begin{align*}
(1) &\request{a}{k}P \ |\ \accept{a}{k}Q & (3) & \throw{k}{k'}P \ | \ \catch{k}{k'}Q.\\
(2) &\send{k}{\vec{e}}P \ |\ \receive{k}{\vec{x}}Q & (4) & \selection{k}{l}P \ |\ \branching{k}{l_1:P_1 \parallel \cdots || l_n:P_n}
\end{align*} }
\end{definition}
Notice that a redex in \hvk\ synchronizes and reduces in a single transition as in $ (\send{k}{\vec e}P) \ |\ (\receive{k}{\vec{x}}Q )$ $\redihvk P \ |\ Q [\vec{e}/\vec{x}] $.
Nevertheless, in \utcc, the encoding of the processes above requires several internal transitions for adding the constraint $\outp(k,\vec{e})$ to the current store, and for ``reading" that constraint by means of $\waitpp{\vec{x}}{\outp(k,\vec{x})}{\nextp \os Q \cs}$ to later execute $\nextp \os Q[\vec{e}/\vec{x}]\cs$.
We shall then establish the operational correspondence between an observable transition of \utcc (obtained from a finite number of internal transitions) and the following subset of reduction relations over \hvk\ processes:
\begin{definition}[Outermost Reductions]
Let $P\equiv_h \recursionH{D}\nu \vec{x}(Q_1 \,|\,\cdots \, | \,Q_n)$ be an \hvk\ program in normal form.
We define the \emph{outermost reduction relation} $P\rede{}_{h}P'$ as the maximal sequence of reductions
$ P \redihvk^{*} P' \equiv _h \recursionH{D}\nu\vec{x'}(Q_1' \, |\, \cdots \, | \, Q_n')$ such that for every $i\in\{1,..n\}$, either
\begin{enumerate}
\item $Q_i = \itn{e}{R_1}{R_2} \redihvk R_{1/2} = Q_i'$;
\item for some $j\in\{1,..n\}$, $Q_i | Q_j$ is a redex
such that $Q_i | Q_j \redihvk \nu\vec{y}(Q_i' | Q_j')$, with $\vec{y} \subseteq \vec{x'}$;
\item there is no $k\in\{1,..n\}$ such that $Q_i \, | \, Q_k$ is a redex and $Q_i \equiv _h Q_i'$.
\end{enumerate}
\end{definition}
One may argue that the above-presented definition may rule out some possible reductions in \hvk. Returning to the concerns about determinacy, an outermost reduction filters out cases where there are more than one possible reduction for a set of parallel processes (i.e.: the parallel composition of two outputs and one input with the same session key). The use of outermost reductions gives us a subset of possible reductions in \hvk that keeps synchronous processes and discard processes that are not going to interact in any way (recall that in the typing discipline of \hvk the composition of an input and an output with the same session key will consume the channel used; hence, every other process sending information over the same session will not have any complementary process to synchronize with).
In the sequel we shall thus consider only \hvk\ processes $P$ where for $n\geq 1$, if $P\equiv_h P_1\rede{}_{h}P_2 \rede{}_{h} \cdots \rede{}_{h} P_n$ and $P\equiv_h P_1'\rede{}_{h}P_2' \rede{}_{h}\cdots \rede{}_{h} P_n'$ then $P_i \equiv _h P_i'$ for all $i\in \{1,..,n\}$, i.e., $P$ is a \emph{deterministic} process.
\begin{theorem}[Operational Correspondence] \label{thm:oper-corr}
Let $P,Q$ be deterministic \hvk\ processes in normal form and $R$,$S$ be \utcc\ processes. It holds: \\
1) {\em Soundness}: If $P \rede{}_{h} Q$ then, for some $R$, $\os P \cs \rede{} R \obsequiv{} \os Q \cs$; \\
2) {\em Completeness}: If $ \os P\cs \rede{} S$ then, for some $Q$, $P \rede{}_{h} Q$ and $\os Q\cs \obsequiv{} S$.
\end{theorem}
\begin{proof}
Assume that $P\equiv _h \recursionH{D}\nu \vec{x}(Q_1 \,|\,\cdots \, | \,Q_n)$
and $Q\equiv _h \recursionH{D}\nu \vec{x'}(Q_1' \,|\,\cdots \, | \,Q_n')$.
\begin{enumerate}
\item {\em Soundness}.
Since $P \rede{}_{h} Q$ there must exist a sequence of derivations of the form
$P \equiv _h P_1 \redihvk P_2 \redihvk ... \redihvk P_n \equiv _h Q$. The proof
proceeds by induction on the length of this derivation, with a case analysis on the last applied rule. We then have the following cases:
\begin{enumerate}
\item {\bf Using the rule \textsc{If1}}. It must be the case that
there exists $Q_i \equiv _h \itn{e}{R_1}{R_2}$ and $Q_i \redihvk R_{1} \equiv _h Q_i'$ and $e\downarrow \true$. One can easily show that $\whenp{e \downarrow \true}\nextp$ $\os Q_i'\cs \rede{}\os Q_i'\cs$.
\item {\bf Using the rule \textsc{If2}} Similarly as for \textsc{If1}.
\item {\bf Using the rule \textsc{Link}}. It must be the case that
there exist $i,j$ such that $Q_i \equiv _h \request{a}{k}Q_i'$ and
$Q_j \equiv _h \accept{a}{x}Q_j'$ and then
$Q_i \ | \ Q_j \redihvk (\nu k)(Q_i' \ | \ Q_j')$. We then have a derivation
\[\footnotesize
\begin{array}{lll}
\os Q_i\cs \parallel \os Q_ k\cs & \redi^{*} &
\localp{k;c}{}( R_i' \parallel \wheneverp{\accp(a,k)}{\nextp} \os Q_i' \cs \parallel \\
& & \tab\tab\tab\tab\ \ \waitpp{k'}{\reqp(a,k')}{}(\tellp{\accp(a,k')} \parallel \nextp(\os Q_j' \cs)) \\
& \redi^{*} &
\localp{k;c'}{}( R_i' \parallel \wheneverp{\accp(a,k)}{\nextp} \os Q_i' \cs \parallel \\
& & \tab\tab\tab\tab\ R_j' \parallel \tellp{\accp(a,k)} \parallel \nextp(\os Q_j'[k/k'] \cs) \\
& \redi^{*} &
\localp{k;c''}{}( R_i' \parallel R_j' \parallel \nextp \os Q_i' \cs \parallel \nextp(\os Q_j'[k/k'] \cs) \not\redi\\
\end{array}
\]
where
${\small
c = \reqp(a,k), \
c' = c \wedge \overline{\reqp(a,k)}, \
c'' = c' \wedge \accp(a,k) \wedge
\overline{\accp(a,k)}
}
$
and
$R_i'$, $ R_j'$ are the processes resulting after the interaction of the processes in the parallel composition
$\tellpp{\reqp(a,k)} \parallel \waitpp{k'}{\reqp(a,k')}{\cdots}$, i.e.:\[\footnotesize
\begin{array}{rll}
R_i'& \equiv_u & \localp{go,stop ; \outp'(go)\wedge \outp'(stop) \wedge c(\vec{t})}{}\\
& & \tab\tab\tab\tab\tab \nextp\bangp \unlessp{\outp'(stop)}{\tellp{\outp'(go)}} \parallel \nextp {\bangp{\tellp{\outp'(stop)}}}\\
R_j' & \equiv_u & \localp{stop',go' ; \outp'(go') \wedge \overline{c}(\vec{t}) \wedge \outp'(stop') }{} \nextp\bangp\tellp{\outp'(stop')} \\
& & \ \ \parallel \nextp \bangp{}\unlessp{\outp'(stop')}{\tellp{\outp'(go')}} \\
& & \ \ \parallel \absp{\vec{x}}{c\wedge \outp'(go') \wedge \vec{x} \not\doteq \vec{t}}{(Q \parallel \tellp{\overline{c}(\vec{t})} \parallel \bangp{\tellp{\outp'(stop')}}} \\
& & \ \ \parallel \nextp\bangp \absp{\vec{x}}{c\wedge \outp'(go') }{(Q \parallel \tellp{\overline{c}(\vec{t})} \parallel \bangp{\tellp{\outp'(stop')}}}
\end{array}\]
We notice that $R_i' \parallel R_j' \notredi$ and it is a process that can only output the constraint $\outp'(x)$ where $x$ is a local variable. By appealing to Proposition \ref{prop-wait-tell} we conclude
$\os Q_i\cs \parallel \os Q_ j\cs \rede{}\obsequiv{} \localp{k}{}(\os Q_i'\cs \parallel \os Q_ j'\cs$).
\item The cases using the rules $\textsc{Label}$ and $\textsc{Pass}$ can be proven similarly as the case for
$\textsc{link}$.
\end{enumerate}
\item {\em Completeness}. Given the encoding and the structure of $P$, we have a \utcc process $R = \os P \cs$ s.t.
\[
R \equiv _u \localp{\vec{x}}(\os Q_1 \cs \parallel ... \parallel \os Q_n \cs) \, .
\]
Let $R_i = \os Q_i \cs$ for $1\leq i \leq n$.
By an analysis on the structure of $R$, if $R_i \redi R_i'$ then it must be the case that either (a)
$R_i= \whenp{e}\nextp{\os Q_i'\cs}$ and $R_i' = \nextp \os Q_i'\cs$ or (b) $\langle R_i , c\rangle \redi \langle R_i',c \wedge d \rangle$
where $d$ is a constraint of the form $\reqp(\cdot)$, $\selp(\cdot)$, $\outp(\cdot)$, or $\outkp(\cdot)$.
In both cases we shall show that there exists a $R_i''$ such that $R_i \redi^{*} R_i'' \notredi$ such that
$Q_i \redihvk Q_i' $ and $R_i'' = \nextp \os Q_i' \cs$.
\begin{enumerate}
\item Assume that $R_i = \whenp{e\downarrow \true} \nextp \os Q_i' \cs$ for some $Q_i'$.
Then it must be the case that
$Q_i=\itn{e}{Q_i'}{Q_i''}$.
If $e\downarrow\true$ we then have
$R_i'' = \nextp \os Q_i'\cs$.
The case when $e \downarrow \false$ is similar by considering $R_i = \whenp{e\downarrow \false}{Q_i'}$.
\item Assume now that $\langle R_i,c \rangle \redi \langle R_i' , c \wedge d \rangle$ where
$d$ is of the form $\reqp(\cdot)$, $\selp(\cdot)$, $\outp(\cdot)$ or $\outkp(\cdot)$.
We proceed by case analysis of the constraint
$d$. Let us consider only the case $d = \exists_k(\reqp(a,k))$; the cases in which $d$ takes the form $\selp(\cdot)$, $\outp(\cdot)$, or $\outkp(\cdot)$ are handled similarly.
If $d = \exists_k(\reqp(a,k))$ for some $a$, then we must have that $Q_i \equiv _h \request{a}{k}Q_i'$ for some $i$.
If there exists $j$ such that
$Q_j \equiv _h \accept{a}{x}Q_j'$, one can show a derivation similar to the case of the rule $\textsc{Link}$ in soundness to prove that $R_i \parallel R_j \redi^{*} \oequiv{} \localp{k}(\nextp \os Q_i' \cs \parallel \nextp \os Q_j' \cs)$.
If there is no $Q_j$ such that $Q_i \ | Q_j$ forms a redex,
then one can show by using (1) in Proposition \ref{prop-wait-tell} that $R_i \rede{} \obsequiv{} R_i$ .
\end{enumerate}
\end{enumerate}
\end{proof}
\section{A Timed Extension of \hvk}
We now propose an extension to \hvk in which
a bundled treatment of time is explicit and session closure is considered.
More precisely, the \hvkplus language arises as the extension of \hvk processes (Def. \ref{def-hondasLanguage})
with refined constructs for session request and acceptance, as well as with a construct for
session abortion:
\begin{definition}[A timed language for sessions]
\hvkplus processes are given by the following syntax:
\begin{center}
{\footnotesize
\begin{tabular}{rrll}
P &$::=$ & \timedRequest{a}{k}{m}P & Timed Session Request\\
&$|$ & \timedAccept{a}{k}{c}P & Declarative Session Acceptance\\
&$|$ & $\cdots$ & \{~~the other constructs, as in Def. \ref{def-hondasLanguage} ~~\} \\
&$|$ & \killS{c_k} & Session Abortion\\
\end{tabular}
}
\end{center}
\end{definition}
The intuition behind these three operators is the following: $\timedRequest{a}{k}{m}P$ will request a session $k$ over the service name $a$ during $m$ time units. Its dual construct is $\timedAccept{a}{k}{c}P$:
it will grant the session key $k$ when requested over the service name $a$
provided by a session and a successful check over the constraint $c$.
Notice that $c$ stands for a precondition for agreement between session request and acceptance. In $c$,
the duration $m$ of the corresponding session key $k$
can be referenced by means of the variable ${\it dur_k}$. In the encoding we syntactically replace it by the variable corresponding to $m$.
Finally, \killS{c_k} will remove $c_k$ from the valid set of sessions.
\begin{table}[h]
\centering
{\footnotesize
\begin{tabular}{c}
\\
$
\begin{array}{r c l}
[| \timedRequest{a}{k}{m}P |] & = &
\localp{k}{}\tellpp{\reqp(a,k,m)} \parallel \\
& & \ \ \tab\tab\tab \wheneverp{\accp(a,k)}{\nextp}( \tellp{\actp(k)} \parallel \Guarded_{\actp(k)} (\os P \cs ) \parallel \\
& & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \tab \ \ \bangp_{[m]}\unlessp{\killp(k)} \tellp{\actp(k)}) \\
[|\timedAccept{a}{k}{c}P |] &= & \waitpp{k}{\reqp(a,k,m) \land c[m/{\it dur_k}]}{}(\tellp{\accp(a,k)} \parallel \nextp\Guarded_{\actp(k)}(\os P \cs)) \\
\os \killS{k} \cs & = & \bangp\tellp{\killp(k)} \\ \\
\end{array}
$\\
\end{tabular}
}
\caption{{\label{table:enc2} The Extended Encoding. $\Guarded_d(P)$ is in Definition \ref{def:guarded}.}}
\label{enc:ext}
\end{table}
Adapting the encoding in Table \ref{table:enc} to consider \hvkplus processes
is remarkably simple (see Table \ref{enc:ext}).
Indeed, modifications to the encoding of session request and acceptance are straightforward.
The most evident change is the addition of the parameter $m$ within the constraint $\reqp(a,k,m)$.
The duration of the requested session is suitably represented as a bounded replication of the process defining the activation of
the session $k$ represented as the constraint $\actp(k)$. The execution of the continuation $\os P \cs$ is guarded by the constraint $\actp(k)$ (i.e. $P$ can be executed only when the session $k$ is valid).
Thus, in the encoding we use the function $\Guarded_d(P)$ to denote the process behaving as $P$ when the constraint $d$ can be entailed from the current store, doing nothing otherwise. More precisely:
\begin{definition}\label{def:guarded}
Let $\Guarded: \cC \to Procs \to Procs$ be defined as
{ \footnotesize
\begin{align*}
&\Guarded_d(\skipp) &=& \skipp & & \Guarded_d( P_1 \parallel P_2) &=& \Guarded_d(P_1) \parallel \Guarded_d(P_2) \\
&\Guarded_d(\tellp{c}) &=& \whenp{d}{\tellp{c}} & & \Guarded_d( \bangp{Q}) &=& \bangp\Guarded_d(Q) \\
&\Guarded_d( \nextp{Q}) &=& \whenp{d}{\nextp\Guarded_d(Q)} & & \Guarded_d(\absp{\vec{x}}{c}{Q}) &=& {\absp{\vec{x}}{c}{\Guarded_d(Q)}} \quad \text{ if $\vec{x} \notin \fv(d)$} \\
&\Guarded_d( \unlessp{c}{Q}) &=& \whenp{d}{\unlessp{c}{\Guarded_d(Q)}} & & \Guarded_d( \localp{\vec{x};c}{Q}) &=& \localp{\vec{x};c}{\Guarded_d(Q)} \quad \text{ if $\vec{x} \notin \fv(d)$}
\end{align*}}
\end{definition}
On the side of session acceptance, the main novelty is the introduction of $c[m/{\it dur_k}]$. As explained before, we syntactically replace the variable ${\it dur_k}$ by the corresponding duration of the session $m$.
This is a generic way to represent the agreement that should exist between a service provider and a client; for instance, it could be the case that the
client is requesting a session longer than what the service provider can or want to grant.
\subsection{Case Study: Electronic booking
Here we present
an example that
makes use of the constructs introduced in \hvkplus.
Let us consider an electronic booking scenario.
On one side, consider
a company AC which offers flights directly from its website. On the other side, there is a customer looking for the best offers.
In this scenario, the customer establishes a timed session with AC and asks for a flight proposal given a set of constraints (dates allowed, destination, etc.). After receiving an offer from AC, the customer can refine the selection further (e.g. by checking that the prices are below a given threshold) and loops until finding a suitable option, that he will accept by starting the booking phase.
One possible \hvkplus specification of this scenario is described in Table \ref{t:ex1}.
\begin{table}[h]
\footnotesize{
\centering
\begin{tabular}{lcl}
\\
Customer & = & $\timedRequest{ob}{k}{m} (\send{k}{booking data} Select(k))$\\
Select(k) & = &~ $\receive{k}{offer} ( \itn{(offer.price \leq 1500) }{\selection{k}{Contract}}{Select(k)})$ \\
AC & = & $\timedAccept{ob}{k}{dur_k \leq MAX\_TIME} ( \receive{k}{bookingData} $ \\
& & ~~ $(\nu u) \send{k}{u} \branching{k}{Contract: \overline{Accept} || Reject : \killS{k}})$\\
\end{tabular} \caption{{Online booking example with two agents. \label{t:ex1}}}
}
\end{table}
In a second stage, the customer uses an online broker to mediate between him and a set of airlines acting as service providers. Let $n$ be the number of service providers, and consider two vectors of fixed length: ${\it Offers}$, which contains the list ${\it [Offers_0,\dots, Offers_i,\dots, Offers_n]}$ of offers received by a customer, and $SP$, which contains the list of trusted services. First, the customer establishes a session with the broker for a given period $m$; later on, he/she starts requesting for a flight by providing the details of his/her trip to the broker. On the other side,
the broker
will look into his pool of trusted service providers for the ones
that can supply
flights
that suit the customer's requirements.
All possible offers are transferred back to the customer, who will invoke a local procedure $Sel$ (not specified here) that selects one of the offers by performing an output on name $a$.
Once an offer is selected, the broker will allow a final interaction between the customer and the selected service. He does so by delegating to the customer the session key used previously between him and the chosen service provider.
Finally, the broker proceeds to cancel all those sessions concerning the discarded services.
An \hvkplus specification of this scenario is given in
Table \ref{t:ex2}
where, for the sake of readability, processes denoting post-processing activities are abstracted from the specification.
A notable advantage in using \hvkplus as a modeling language is
the possibility of exploiting timed constructs in the specification of
service enactment and service cancellation.
In the above scenario it is possible to see how \hvkplus allows (i) to effectively take explicit account on the maximal times accepted by the customer: the composition of nested services can take different speeds but the service broker will ensure that customers with low speeds are ruled out of the communication; and (ii) to have a more efficient use of the available resources: since there is not need to maintain interactions with discarded services, the service broker will free those resources by sending kill signals.
\begin{table}[h]
\footnotesize{
\hspace{-0.3cm}
\subfigure[Customer and Service Provider]{
\begin{tabular}{ll}
Customer = &\hspace{-0.4cm} $\timedRequest{ob}{k}{m} (\send{k}{booking data} $\\
&\hspace{-0.5cm} ~$\receive{k}{ n } ($\\
&\hspace{-0.5cm} ~$\prod\limits_{i \in n} (\receive{k}{\it Offers_i} ( $\\
&\hspace{-0.5cm} ~$Sel({\it Offers});\receive{a}{x}\send{k}{\it x}$ \\
&\hspace{-0.5cm} ~~$\catch{k}{k'}$\\
&\hspace{-0.5cm} ~~$\send{k'}{PaymentDetails} \inact))))$\\\\
SP = &\hspace{-0.4cm} $\timedAccept{SP_i}{k_i'}{N \leq 300ms} ($\\
&\hspace{-0.5cm} ~$\receive{k_i'}{bookingData} $\\
&\hspace{-0.5cm} ~~$\send{k_i'}{\it offer}$\\
&\hspace{-0.5cm} ~~$\receive{k_i'}{paymentDetails} \inact)$\\
\end{tabular}
}
\subfigure[Online Broker]{ \hspace{-0.3cm}
\begin{tabular}{ll}
Broker = & \hspace{-0.4cm} $\timedAccept{ob}{k}{m \leq 500ms} ( $ \\
&\hspace{-0.5cm} ~ $\receive{k}{bookingData} \send{k}{ |SP| }$ \\
&\hspace{-0.5cm} ~ $(\nu u) \prod\limits_{ i \in |SP|} (\timedRequest{SP_i}{k_i'}{N} $\\
&\hspace{-0.5cm} \quad$\send{k_i'}{bookingData}$\\
&\hspace{-0.5cm} \quad $\receive{k_i'}{\it offer_i} (\send{u}{\it offer_i} \inact || S(u,k)))$\\
&\hspace{-0.5cm} \quad $\receive{k}{y} \recursionH{ X({\it Offers}, k_1', \dots, k_n') = P }$\\
&\hspace{-0.5cm} \quad \quad $ \prod\limits_{i \in |SP|} (\itn{(\it y = offers_i)}{(\throw{k}{k_i'} PostProc )}{}$\\
&\hspace{-0.5cm} \quad \quad \quad \quad ${\killS{k_i'}||P(X-\{{\it offers_i}, k_i'\})}))$\\\\
S(u,k) = &\hspace{-0.5cm} $\prod\limits_{ i \in |SP|} (\receive{u}{\it offer_i} \inact || \send{k}{\it offer_i} \inact )$\\
\end{tabular}
}
\caption{Online booking example with online broker.\label{t:ex2}}
}
\end{table}
\subsection{Exploiting the Logic Correspondence}
To exploit the logic correspondence we can draw inspiration from the \emph{constraint templates} put forward in
\cite{pesic2006daf},
a set of LTL formulas that represent desirable/undesirable situations in service management. Such templates are divided in three types: \emph{existence constraints}, that specify the number of executions of an activity; \emph{relation constraints}, that define the relation between two activities to be present in the system; and \emph{negation constraints}, which are essentially the negated versions of relation constraints.
By appealing to Theorem \ref{theo:FLTL}, our framework allows for the verification of existence and relation constraints over \hvkplus programs. Assume a \hvkplus program $P$ and let $F = \Logic\os \os P \cs \cs$ (i.e., the FLTL formula associated to the \utcc\ representation of $P$). For existence constraints, assume that $P$ defines a service accepting requests on channel $a$. If the service is eventually active, then it must be the case that $F \entails \sometime \exists_k(\accp(a,k))$ (recall that the encoding of $\mathbf{accept}$ adds the constraint $\accp(a,k)$ when the session $k$ is accepted). A slight modification to the encoding of $\mathbf{accept}$ would allow us to take into account the number of accepted sessions and then support the verification of properties such as $F \entails \sometime (N_{sessions}(a) = N)$, informally meaning that the service $a$ has accepted $N$ sessions. This kind of formulas correspond to the existence constraints in \cite[Figure 3.1.a--3.1.c]{pesic2006daf}. Furthermore, making use of the guards associated to ask statements, we can verify relation constraints as eventual consequences over the system. Take for instance the specification in Table \ref{t:ex1}. Let $\overline{Accept}$ be a process that outputs ``${\it ok}$" through a session $h$. We then may verify the formula $F \entails \exists_u (u.price < 1.500 => \outp(h, {\it ok}))$. This is a responded existence constraint describing how the presence of an offer with price less or equal than $1.500$ would lead to an acceptance state.
\section{Concluding Remarks}
\label{s:concl}
We have argued for a timed CCP language as a suitable foundation for analyzing structured communications.
We have presented
an encoding of the language for structured communication in \cite{honda1998lpa} into \utcc, as well as an extension of such a language that considers explicitly elements of partial information and session duration.
To the best of our knowledge, a unified framework where behavioral and declarative techniques converge for the analysis of structured communications has not been proposed before.
Languages for structured communication and CCP process calculi are conceptually very different.
We have dealt with some of these differences (notably, determinacy)
when stating an operational correspondence property for the declarative interpretation of \hvk processes.
We believe there are at least two ways of achieving more
satisfactory
notions of operational correspondence.
The first one involves considering
extensions of \utcc
with (forms of) non-determinism.
This would allow to capture some scenarios of session establishment in which the operational correspondence presented here falls short.
The main consequence of adding non-determinism to \utcc is that
the correspondence with FLTL as stated in Theorem \ref{theo:FLTL} would not longer hold. This is mainly because non-deterministic choices cannot be faithfully represented as logical disjunctions (see, e.g., \cite{NPV02}).
While a non-deterministic extension to \tcc with a tight connection with temporal logic
has been developed
(\ntcc \cite{NPV02}),
it does not provide for representations of mobile links.
Exploring whether there exists a CCP language between \ntcc and \utcc combining both non-determinism and mobility
while providing logic-based reasoning techniques is interesting on its own and appears challenging.
The second approach consists in defining a type system for \hvk and \hvkplus processes better suited to the nature of \utcc processes.
This would imply enriching the original type system in \cite{honda1998lpa} with
e.g., stronger typing rules for dealing with session establishment.
The definition of such a type system is delicate and needs care, as one would not like to rule out
too many processes as a result of too stringent typing rules.
An advantage of a type system ``tuned" in this way is that one could aim at obtaining a correspondence between
well-typed processes and logic formulas, similarly as the given by Theorem \ref{theo:FLTL}.
In these lines, plans for future work include the investigation of effective mechanisms for
the seamless integration of new type disciplines and reasoning techniques based on temporal logic within the elegant framework
provided by
(timed) CCP languages.
The timed extension to \hvk presented here includes notions of time that involve only session engagement processes.
A further extension could involve the inclusion of time constraints over input/output actions.
Such an extension might be useful to realistically specify scenarios in which factors such as, e.g,
network traffic and long-lived transactions,
prevent interactions between services from occurring instantaneously.
Properties of interest in this case could include, for instance,
the guarantee that a given interaction has been fired at a valid time, or
that the nested composition of services does not violate a certain time frame.
We plan to explore case studies of structured communications involving this kind of timed behavior, and
extend/adjust \hvkplus accordingly.
\paragraph{Acknowledgments.}
We are grateful to Marco Carbone and Thomas Hildebrandt for insightful discussions on the topics of this paper.
We also grateful to Roberto Zunino who provided useful remarks on a previous version of this document.
The contribution of Olarte and P\'{e}rez was initiated during short research visits to the IT University of Copenhagen.
They are most grateful to the IT University and to the FIRST PhD Graduate School for funding such visits.
\bibliographystyle{abbrv}
|
\section{Introduction}
The rate at which a bar rotates, its pattern speed, $\omp$, is the
principle parameter controlling a barred (SB) galaxy's dynamics and
morphology. Most determinations of this parameter are indirect. An
often used method is to match hydrodynamical simulations of SB
galaxies to observed velocity fields. The only direct and
model-independent technique to measure $\omp$ is the Tremaine \&
Weinberg (1984, hereafter TW) method. They show that for any tracer
that satisfies the continuity equation, $\kin = \pin\,\omp\,\sin{i}$.
Here, $\pin \equiv \int h(Y)\,X\,\Sigma\,dX\,dY$ and $\kin \equiv \int
h(Y)\,V_{\rm los}\,\Sigma\,dX\,dY$ are the luminosity-weighted mean
positions and velocities respectively.
The number of successful applications of the TW method is rather
limited and confined mostly to early type SB galaxies (e.g.,
\citealt{agu02}; \citealt{ger03}). Late-type barred galaxies often
display prominent dust lanes along the leading edges of the bar which
can invalidate the assumption of tracer-continuity. Our motivation for
this study is to explore how important dust is for TW measurements of
both early and especially late-type galaxies. In the latter, star
formation can be a potential problem as it violates the basic
assumption of the TW method. We therefore also explore the ability to
measure $\omp$ in a hydrodynamical simulation of a barred galaxy that
includes star formation \citep{deb06}.
\begin{figure*}[t!]
\resizebox{\hsize}{!}{\includegraphics[clip=true]{gerssen_f01.ps}}
\caption{\footnotesize
Examples of the projected distribution of particle weights in our
models. The models are published in \citet[][TW1-TW4]{deb03},
\citet[][TW5]{deb00}, and \citet[][TW6]{deb06}. They are shown here
face-on to highlight their morphological features. Dust lanes have
been included to illustrate how dust attenuation is implemented in
our simulations. For clarity the dust lanes have an exaggerated
extinction ($A_V = 15$) and width.}
\label{f:maps}
\end{figure*}
\section{Method}
In a particle implementation of the TW method, the integrals over
surface brightness and velocities are replaced by sums over the
particles. Five out of our six N-body models all have particles that
are coeval and have equal mass. We therefore start by assigning the
same intrinsic weights to all particles in these models. Dust lanes
on the leading edges of the bars are implemented by lowering the
weights of particles that reside in these areas. We assume that the
dust distribution within the dust lanes is described by a double
exponential model. That is, $D_{0,{\rm lane}} = e^{-R/h_{R,{\rm
lane}}}\,e^{-|z|/h_{z,{\rm lane}}}$ when a particle resides in the dust
lane and is $0$ otherwise. Particles that are shadowed by dust lanes
contribute to the TW integrals with lower weights. Unlike a
foreground screen model, the dust extinction in our models varies with
position within the disc. For each particle we need to calculate the
amount of intervening dust by integrating the projected dust
distribution along the line-of-sight. The weight for each particle is
then given by $w_i = e^{-\tau_i}$, where the optical depth $\tau_i$ is
obtained from $\tau_i = \int^{s_i}_{-\infty} D(s')\,ds'$. Examples
of the weight distribution in each of our six models are shown in
Fig.~\ref{f:maps}. They are shown here face-on to better illustrate
their morphological features.
\begin{figure*}[]
\resizebox{\hsize}{!}{\includegraphics[clip=true]{gerssen_f02.ps}}
\caption{ \footnotesize
Left panel: The ratio $\omtw / \omint$ of the bar pattern speed
observed with the TW method to the intrinsic bar pattern speed as a
function of dust lane extinction $A_V$. The different curves show
the behaviour of this ratio for the different N-body models. Two
curves are shown for each model. Curves with a ratio $> 1$ are at
$\rm PA_{\rm bar} = +45^\circ$, while curves with a ratio $< 1$ are
at $\rm PA_{\rm bar} = -45^\circ$. The dust lane geometry and dust
distribution are the same for every model: \DLwidth/$a_{\rm B} =
0.1$, \DLslope\ = $-0.3$, \DLxmin/$a_{\rm B} = -0.1$, $h_{R,{\rm
lane}} = 10$ and $h_{z,{\rm lane}} = 0.1$. For clarity, the
$1\sigma$ errors are shown for model TW1 only. Right panel: The bar
pattern speed ratio as a function of dust extinction for the model
that includes star formation (TW6). The result is not very
different from the models shown in the left panel. For comparison
model TW1 is overplotted as the solid line with error bars.}
\label{f:omega}
\end{figure*}
\section{Results}
\subsection{Dust extinction}
In these proceedings we explore the behaviour of the measured pattern
speed as a function of dust extinction, expressed in $A_V$ magnitude
($A_V = 1.086\,\tau_0$). The effects of adding a global dust disc and
varying the dust lane geometry are small and are detailed in
\citep{ger07}. Fig.~\ref{f:omega} (left panel) shows the ratio
$\Omega_{\rm p}/\Omega_0$ of the bar pattern speed observed with the
TW method to the intrinsic bar pattern speed as a function of dust
lane extinction $A_V$ (defined face-on). The different curves show the
behaviour of this ratio for the different N-body models (TW1 to TW5).
Two curves are shown for each model. Curves with a ratio $>1$ are at
$\rm PA_{\rm bar} = +45^{\circ}$, while curves with a ratio $<1$ are
at $\rm PA_{\rm bar} = -45^{\circ}$. The dust lane geometry and dust
distribution are the same for every model.
\subsection{Star formation}
In late type systems, vigorous star formation may invalidate the TW
assumption that the tracer population satisfies the continuum
condition. We use an N-body + SPH model (TW6) to explore these
effects. The results, shown on the left, are qualitatively similar to
the results derived from the N-body models TW1 to TW5. For a
quantitatively comparison model TW1 is overplotted as the solid line
with error bars in Fig.~\ref{f:omega} (right panel).
\section{Conclusions}
With realistic amounts of absorption in the dust lanes ($A_V \sim 3$)
the models predict observationally insignificant errors (20--40\%) in
the observed pattern speed. Our results suggest that the application
of the TW method can be extended to later-type barred galaxies and
facilitate direct comparisons with pattern speeds measured by
hydrodynamical modeling.
\bibliographystyle{aa}
|
\section*{Introduction}
\L ukasiewicz logic is a non-classical logic whose synctactic apparatus was introduced by \L ukasiewicz and Tarski in \cite{luktar}. A first, synctactic, proof of the completeness of \L ukasiewicz's infinite-valued sentential calculus was given by Rose and Rosser in \cite{ros}. Later on~---~in \cite{chang,chang2}~---~Chang introduced the class of MV-algebras and showed that the Lindenbaum algebra of \L ukasiewicz logic is an MV-algebra. Since then, MV-algebras were deeply investigated by many authors and they received a great impulse from the work \cite{mun}, by Mundici, where a categorical equivalence between MV-algebras and lattice-ordered Abelian groups with strong unit was established.
We recall that an MV-algebra is a structure $\mathbf{A} = \langle A, \oplus , ^*, 0\rangle$ of type (2,1,0) satisfying the equations
\begin{enumerate}
\item[-]$(x\oplus y) \oplus z = x \oplus (y\oplus z)$,
\item[-]$x \oplus y = y \oplus x$,
\item[-]$x \oplus 0 = x$,
\item[-]$(x^*)^* = x$,
\item[-]$x \oplus 0^*=0^*$,
\item[-]$(x^* \oplus y)^* \oplus y = (y^* \oplus x)^* \oplus x$.
\end{enumerate}
The element $0^*$ is denoted by $1$. The prototypical example of MV-algebra is the real unit interval $[0,1]$ equipped with the truncated sum $x \oplus y := \min\{1, x+y\}$ and the involution $x^* := 1 - x$. It is well-known that the variety of MV-algebras is generated by this algebra and that the free MV-algebra over $n$ generators is the algebra of \emph{McNaughton functions}~---~i.e. piecewise linear functions with integer coefficients~---~from $[0,1]^n$ to $[0,1]$, equipped with the operations naturally inherited from the MV-algebra over $[0,1]$.
Since the free MV-algebra is an algebra of functions, it is naturally equipped with the further operation of map composition. In \cite{diflge} the authors defined an extension of MV-algebras obtained by adding another binary operation whose behaviour with respect to the MV-algebraic structure mirrors the composition of McNaughton functions in the free algebras: the Composition MV-algebras, \emph{CMV-algebras} for short. Many results that we present in this paper can also be found in \cite{diflge}.
The introduction of this class was stimulated by the observation of a connection between the behaviour of McNaughton functions and chaotic deterministic dynamical systems. In this paper we recall some properties of CMV-algebras and introduce an expansion of the one-variable fragment of \L ukasiewicz propositional logic whose algebraic semantics is the variety of CMV-algebras.
The paper is organized as follows. In the first section we recall the definition and basic properties of CMV-algebras, showing many examples of such structures. Among them, an important role is played by the free MV-algebra in one generator, $\mathbf{M}_1$, endowed with a monoid operation that is essentially map composition and that is also the example of CMV-algebra that inspired the aforementioned paper \cite{diflge}.
In Section~\ref{cmvend} we show the link between composition MV-algebras and MV-algebra endomorphisms: the monoid reduct of any CMV-algebra $\mathbf{A}$ is isomorphic to a submonoid of the endomorphism monoid of the MV-algebra reduct of $\mathbf{A}$.
In Section~\ref{cmvcong} we treat CMV-ideals and congruences, presenting several examples of them. We also present simple algebras and point out a wide class of such algebras.
Modules over CMV-algebras are introduced in Section~\ref{mod}. Apart from many examples presented, we will see that every MV-algebra is canonically a module over the CMV-algebra $\mathbf{M}_1$.
Finally, in Section~\ref{logic}, we present the logic $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$~---~that is an expansion of the one-variable fragment of \L ukasiewicz propositional calculus~---~whose associated Tarski-Lindenbaum algebra is a CMV-algebra.
\section{Composition MV-algebras}
In this section we introduce the definition of CMV-algebras as a special case of composition algebras.
\begin{definition}\label{compop}
Let $f: A^n \longrightarrow A$ be a n-ary operation on a set $A$ and suppose that a monoid structure $\langle A, \diamond, i \rangle$ is defined on the same set. We say that $f$ and the monoid structure are compatible if $f(x_1, x_2, \ldots, x_n) \diamond y = f(x_1 \diamond y, x_2 \diamond y, \ldots, x_n \diamond y)$. In the special case $n = 0$ we have an element $a \in A$ and the condition reduces to $a \diamond y = a$.
\end{definition}
\begin{definition}\label{compalg}
Let $A$ be an algebra of a given signature and $\langle A, \diamond, i\rangle$ a monoid on the same universe. If all the operations of $A$ are compatible with the monoid structure we say that $A$ is a \emph{composition algebra} (of the corresponding type) with respect to the given monoid.
A \emph{morphism} of composition algebras (of the same type) is a morphism of the same algebras that is also a morphism of the associated compatible monoid structures.
\end{definition}
Clearly a composition algebra is an algebra itself but the previous definition is more useful for the scope of the present work.
\begin{definition}\label{comprel}
Let $\Phi \subseteq A^n$ be an $n$-ary relation on $A$ and $\langle A, \diamond, i\rangle$ a monoid. We say that $\Phi$ and the monoid structure are compatible if from $(x_1, x_2, \ldots, x_n) \in \Phi$ follows $(x_1 \diamond y, x_2 \diamond y, \ldots, x_n \diamond y) \in \Phi$, for every $y \in A$.
\end{definition}
An important example of composition algebras is the following. Consider a group $\langle G,+\rangle$ and the set of all the functions $f: G \longrightarrow G$ equipped with the operation $+$ defined pointwise. The composition of functions $\circ$ makes the algebra $\langle G^G,+ \rangle$ a composition algebra. Note that the composition algebra $\langle G^G, +, \circ\rangle$ is a near-ring (see \cite{pilz}).
We already recalled the definition of MV-algebra in the introduction. MV-algebra morphisms are defined in the usual way according to the signature, namely they are maps that preserve $\oplus$, $^*$ and $0$; the same holds for congruences, i.e. a congruence over an MV-algebra is an equivalence relation that preserves the operations. Let us recall also some basic facts about MV-algebras.
First of all we observe that Boolean algebras are precisely those MV-algebras satisfying the additional equation $x \oplus x = x$. Every MV-algebra contains as a subalgebra the two-element boolean algebra $\{0,1\}$; the set $B(\mathbf{A})$ of all idempotent elements of an MV-algebra $\mathbf{\mathbf{A}}$ is the largest boolean algebra contained in $A$ and is called the \emph{Boolean skeleton} of $\mathbf{A}$.
On every MV-algebra $\mathbf{A}$ we define also further (derived) operations and a structure of bounded lattice. In particular, $\odot$ is the operation defined by $x \odot y = (x^* \oplus y^*)^*$, $1$ (i.e. $0^*$) is the neutral element for $\odot$ and $\to$ is defined by $x \to y = x^* \oplus y$; moreover we define $x \ominus y := x \odot y^*$ and denote by $nx$ the element $\underbrace{x \oplus \ldots \oplus x}_{\textrm{$n$-times}}$ and by $x^n$ the element $\underbrace{x \odot \ldots \odot x}_{\textrm{$n$-times}}$. The order relation on $\mathbf{A}$ is defined by $x \leq y$ if and only if $x^* \oplus y = 1$. In this way we obtain a bounded lattice (with 0 as the bottom element and 1 as the top), with $x \vee y := x \oplus (x^* \odot y)$ and $x \wedge y := x \odot (x^* \oplus y)$. Last, the structure $\langle A, \vee, \wedge, \odot, 0, 1\rangle$ is a bounded integral commutative residuated lattice and the residuum of $\odot$ is $\to$.
\begin{exm}\label{exampleMV}
\begin{enumerate}
\item The real unit interval $[0,1]$, equipped with operations
$$x \oplus y = \min\{1,x+y\}, \quad x \odot y = \max\{0,x+y-1\}, \quad x^* = 1 - x,$$
is an MV-algebra, often referred to as the \emph{standard MV-algebra}; in this algebra the order relation is the one among real numbers. For any $n \in \mathbb N$, the set $\operatorname{\textnormal{\L}}_{n+1} = \left\{0, \frac{1}{n}, \ldots, \frac{n-1}{n}, 1\right\}$, equipped with the same operations, is a finite linearly ordered MV-algebra.
\item If $X$ is any set and $\mathbf{A}$ is an MV-algebra, $\mathbf{A}^X = \langle A^X, \oplus, ^*, 0\rangle$ is an MV-algebra with the operations defined pointwisely from those in $\mathbf{A}$.
\item The set $M_n$ of all the functions from $[0,1]^n$ to $[0,1]$ that are continuous, piecewise linear and such that each linear piece has integer coefficients~---~with the operations defined, again, pointwisely from the ones in $[0,1]$~---~is an MV-algebra. Such functions are called \emph{McNaughton functions} and the MV-algebra of McNaughton functions with $n$ variables is known to be the free MV-algebra over $n$ generators.
\end{enumerate}
\end{exm}
\begin{definition}\label{mvid}
Let $\mathbf{A}$ be an MV-algebra. An \emph{MV-ideal} of $\mathbf{A}$ is a non-empty subset $I$ of $A$ such that
\begin{itemize}
\item[-] if $x \in I$ and $x \geq y$ then $y \in I$,
\item[-] if $x, y \in I$ then $x \oplus y \in I$.
\end{itemize}
Let $\mathbf{A}$ be an MV-algebra and $I \subseteq A$ an MV-ideal of $\mathbf{A}$. The equivalence relation $\sim_I$ defined by $x \sim_I y$ if and only if $(x \odot y^*) \oplus (y \odot x^*) \in I$, is a congruence of the MV-algebra $\mathbf{A}$, called the \emph{congruence generated by $I$}. Conversely, if $\sim$ is a congruence of $\mathbf{A}$ then the set $I_\sim = \{x \in A \mid x \sim 0\}$ is an MV-ideal of $\mathbf{A}$.
\end{definition}
An ideal $I$ is called \emph{prime} if it is proper, that is $I \neq A$ and, for all $x, y \in A$, either $x \ominus y \in I$ or $y \ominus x \in
I$; $I$ is called \emph{maximal} if it is proper and no proper ideal of $\mathbf{A}$ strictly contains $I$. The radical $\operatorname{Rad}(\mathbf{A})$ of $\mathbf{A}$ is the intersection of all maximal ideals of $\mathbf{A}$ and consists of $0$ and all the infinitesimals of the algebra, where an element $a$ is called an \emph{infinitesimal} if $a \neq 0$ and $na \leq a^*$ for all $n \in \mathbb N$. An MV-algebra $\mathbf{A}$ is called \emph{perfect} if $A = \operatorname{Rad}(\mathbf{A}) \cup \operatorname{Rad}(\mathbf{A})^*$, where $\operatorname{Rad}(\mathbf{A})^* := \{x \mid x^* \in \operatorname{Rad}(\mathbf{A})\}$. An MV-algebra $\mathbf{A}$ is \emph{simple} if it has only one proper ideal.
The following representation theorem holds:
\begin{theorem}\label{chang}
An equation holds in every MV-algebra if and only if it holds in the standard MV-algebra $[0,1]$.
\end{theorem}
An important result by Mundici \cite{mun} establishes an equivalence between the category of MV-algebras and the one of lattice-ordered abelian groups ($\ell$-groups) with strong unit. A standard reference for MV-algebras is \cite{cigdotmun}.
\begin{definition}\label{cmv}
A \emph{composition MV-algebra}~---~\emph{CMV-algebra} for short~---~is a structure $\mathbf{A} = \langle A, \oplus, *, 0, \diamond, i\rangle$ such that $\langle A, \oplus, ^*, 0\rangle$ is an MV-algebra, $\langle A, \diamond, i\rangle$ is a monoid and these two structures are compatible in the sense of Definition~\ref{compalg}, i.e. for every $x, y, z \in A$, the following conditions hold:
\begin{enumerate}
\item[$(i)$] $(y \oplus z) \diamond x = (y \diamond x) \oplus (z \diamond x)$,
\item[$(ii)$] $x^* \diamond y = (x \diamond y)^*$,
\item[$(iii)$] $0 \diamond x = 0$.
\end{enumerate}
\end{definition}
Observe that $(ii)$ and $(iii)$ yield also $1 \diamond x = 1$.
Since CMV-algebras are an equational class, they form a variety that we denote by $\mathcal{CMV}$. We shall denote by $\mathbf{A}_{\mathcal{MV}}$ the MV-reduct of the CMV-algebra $\mathbf{A}$, and by $\mathbf{A}_\mathcal{M}$ the monoid reduct of $\mathbf{A}$. Further, when there is no danger of confusion, we shall denote a structure by its domain.
\begin{remark}
If $0=1$ the algebra is reduced to $\{0\}$; we call that algebra the \emph{trivial} CMV-algebra. In a non trivial CMV-algebra the operation $\diamond$ is not commutative: $1 \diamond 0 = 1$ but $0 \diamond 1 = 0$.
\end{remark}
\begin{lemma}\emph{\cite{diflge}}\label{lemma:proprts}
Let $\mathbf{A}$ be a non trivial CMV-algebra. Then the following properties hold:
\begin{enumerate}
\item[$(i)$] $0 < i < 1$;
\item[$(ii)$] $i^* \diamond i^* = i$;
\item[$(iii)$] $i^* \neq i$;
\item[$(iv)$] $a \diamond x = b \diamond x$ \ for all $x$ \ if and only if \ $a = b$.
\end{enumerate}
\end{lemma}
\begin{corollary}
Every non trivial CMV-algebra has at least four elements.
\end{corollary}
\begin{proposition}\emph{\cite{diflge}}\label{prop:compa}
Let $\mathbf{A}$ be a CMV-algebra. The operations $\odot, \vee, \wedge, \ominus$ are compatible with the monoid structure. Moreover the order relation is compatible with the monoid structure.
\end{proposition}
\begin{theorem}
There are no totally ordered non trivial CMV-algebras.
\end{theorem}
\begin{proof}
If $i \leq i^*$, using Proposition~\ref{prop:compa}, we get $i \diamond i^* \leq i^* \diamond i^*$ so $i^* \leq i$ and so $i = i^*$ which is absurd by Lemma~\ref{lemma:proprts}. If $i^* \leq i$ then $i^* \diamond i^* \leq i \diamond i^*$ and we get $i \leq i^*$ so, again, $i = i^*$.
\end{proof}
\begin{corollary}
If $\mathbf{A}$ is a CMV-algebra, then the MV-reduct $\mathbf{A}_{\mathcal{MV}}$ of $\mathbf{A}$ is neither simple nor perfect.
\end{corollary}
\begin{proof}
It is known that simple MV-algebras are totally ordered; thus $\mathbf{A}_{\mathcal{MV}}$ cannot be simple. On the other hand, it is self-evident that $i, i^* \notin \operatorname{Rad}(\mathbf{A})$ for any CMV-algebra $\mathbf{A}$, hence $\mathbf{A}$ cannot be perfect.
\end{proof}
Now we come to some examples of CMV-algebras.
\begin{exm}\label{ex:CMV}
\begin{enumerate}
\item Let $\mathbf{M}$ be any MV-algebra and let $\langle M^M, \circ, i\rangle$ be the monoid of functions from $M$ to itself, with the operation $\circ$ of composition and the identity function $i$. $\mathbf{M}^M = \langle M^M, \oplus, *, 0, 1\rangle$ is an MV-algebra with pointwise defined operations. It is easy to check that $\mathbf{A} = \langle M^M, \oplus, *, 0, \circ, i\rangle$ is a CMV-algebra.
\item Let $\mathbf B_2 = \{0,1\}$ be the Boolean algebra with two elements. We refer to the previous example and set $\mathbf{M} = \mathbf B_2$ and denote by $\mathbf{A}_4$ the corresponding CMV-algebra. It is easy to show that $\mathbf{A}_4$ is, up to isomorphisms, the unique CMV-algebra with four elements.
\item Let $\operatorname{\textnormal{\L}}_3 = \{0, 1/2, 1\}$ be the standard MV-algebra with three elements and $A$ be the set of functions $f: \operatorname{\textnormal{\L}}_3 \longrightarrow \operatorname{\textnormal{\L}}_3$ such that $f[\{0,1\}] \subseteq \{0,1\}$. With pointwise operations and map composition, $A$ becomes a composition MV-algebra. Indeed it is a subalgebra of a CMV-algebra like in Example 1. In general if $\mathbf{A} \subseteq \mathbf{B} \subseteq \mathbf{C}$ are MV-algebras, then the set $\{f \mid f(x) \in A \textrm{ for all } x \in B\}$ is the domain of a subalgebra of $\mathbf{C}^C$.
\item Consider the MV-algebra $\mathbf{M}_1$ of the McNaughton functions of one variable. The composition of two McNaughton functions (in one variable) is still a McNaughton function, so $\mathbf{M}^\mathcal C_1 = \langle M_1, \oplus, *, 0, \circ, i\rangle$ is a CMV-algebra, where $i$ is the identity map of $[0,1]$.
\item The set $\Lambda$ of all the continuous maps $f: [0,1] \longrightarrow [0,1]$ can be endowed with a CMV-algebra structure as in the previous example.
\end{enumerate}
\end{exm}
An example of CMV-algebra of functions with more than one variable is given after Proposition~\ref{prop:MVtoCMV}.
\begin{definition}
A CMV-algebra $\mathbf{A}$ is called \emph{functional} if it is a subalgebra of a CMV-algebra $\mathbf{M}^M = \langle M^M, \oplus, *, 0, \circ, i\rangle$ where $\mathbf{M}$ is an MV-algebra, $\langle M^M, \oplus, *, 0, 1\rangle$ is the MV-algebra obtained by $M$ with pointwise operations and $\circ$ is the composition of functions (see Example~\ref{ex:CMV}.1).
\end{definition}
\begin{theorem}[Cayley-type theorem]
All CMV-algebras are functional.
\end{theorem}
\begin{proof}
Let $\mathbf{A} = \langle A, \oplus, *, 0, \diamond, i \rangle$ be a CMV-algebra, $\mathbf{A}_{\mathcal{MV}}$ its MV-reduct and let $\mathbf{A}^A = \langle A^A, \oplus, *, 0, \circ, i\rangle$ be the CMV-algebra of all the functions from $A$ to $A$.
Consider, for any $a \in A$, the function $f_a \in A^A$ defined by $f_a(x) = a \diamond x$ for every $x \in A$. Then the map $\mu: A \longrightarrow A^A$ defined by $\mu \left(a\right) = f_a$ is a CMV-monomorphism of $\mathbf{A}$ onto $\mathbf{A}^A$. Hence $\mathbf{A}$ is functional.
\end{proof}
There is another way of considering the relation between an MV-algebra $\mathbf{M}$ and the CMV-algebra $\mathbf{M}^M$. We can identify the elements of $M$ with the constant functions of $M^M$. This means to consider the map $\tau: a \in M \longmapsto c_a \in M^M$, where $c_a$ are constant functions of $M$, i.e. $c_a(x) = a$ for all $x \in M$. It is easy to show that $\tau$ is an MV-monomorphism and so $\mathbf{M}$ is isomorphic to the MV-subalgebra $\tau[\mathbf{M}]$ of $\mathbf{M}^M_{\mathcal{MV}}$ consisting of constant functions. If $\mathbf{M}$ is not trivial, i.e. $0 \neq 1$, then $\tau[\mathbf{M}]$ never contains the identity map of $M^M$, so it will never be a CMV-subalgebra of $\mathbf{M}^M$, although it is closed under composition (indeed $\tau[\mathbf{M}]$ is a semigroup of ``left zeros'' with respect to composition). We can consider the CMV-subalgebra $\widetilde{\mathbf{M}}$ generated by $\tau[\mathbf{M}]$ in $\mathbf{M}^M$. In this context we have
\begin{proposition}\label{tilde}
$\widetilde{\operatorname{\textnormal{\L}}_{n+1}} = \operatorname{\textnormal{\L}}_{n+1}^{\operatorname{\textnormal{\L}}_{n+1}}$.
\end{proposition}
\begin{proof}
Since $\widetilde{\operatorname{\textnormal{\L}}_{n+1}}$ is a subalgebra of $\operatorname{\textnormal{\L}}_{n+1}^{\operatorname{\textnormal{\L}}_{n+1}}$, by definition, we just need to prove that $\operatorname{\textnormal{\L}}_{n+1}^{\operatorname{\textnormal{\L}}_{n+1}} \subseteq \widetilde{\operatorname{\textnormal{\L}}_{n+1}}$. Let us denote, for simplicity, $\operatorname{\textnormal{\L}}_{n+1} = \{0, 1, \ldots, n\}$ and a function $f: \operatorname{\textnormal{\L}}_{n+1} \longrightarrow \operatorname{\textnormal{\L}}_{n+1}$ by $(f(0), f(1), \ldots, f(n))$. We obviously have that the functions $i = (0, 1, \ldots, n)$, $i^* = (n, n-1, \ldots, 0)$, $f_c = (c, \ldots, c)$ belong to $\widetilde{\operatorname{\textnormal{\L}}_{n+1}}$ for all $c \in \operatorname{\textnormal{\L}}_{n+1}$.
Then
$$(n, n-1, \ldots, 0) \ominus (n-1, n-1, \ldots, n-1) = (1, 0, \ldots, 0) \in \widetilde{\operatorname{\textnormal{\L}}_{n+1}}$$
and
$$(0, 1, \ldots, n) \ominus (n-1, n-1, \ldots, n-1) = (0, \ldots, 0, 1) \in \widetilde{\operatorname{\textnormal{\L}}_{n+1}}.$$
So $(n, 0, \ldots, 0) \in \widetilde{\operatorname{\textnormal{\L}}_{n+1}}$ and, from
$$(n, n-1, \ldots, 0) \ominus (n, 0, \ldots, 0) = (0, n-1, n-2, \ldots, 0) \in \widetilde{\operatorname{\textnormal{\L}}_{n+1}},$$
we have
$$(0, n-1, n-2, \ldots, 0) \ominus (n-2, n-2, \ldots, n-2) = (0, 1, 0, \ldots, 0) \in \widetilde{\operatorname{\textnormal{\L}}_{n+1}}.$$
With a similar argument we get
$$
(1, 0, \ldots, 0), (0, 1, 0, \ldots, 0), (0, 0, 1, \ldots, 0), \ldots, (0, \ldots, 0, 1) \in \widetilde{\operatorname{\textnormal{\L}}_{n+1}}.
$$
Since every function $f$ can be written as
$$f = (f(0), 0, \ldots, 0) \oplus (0, f(1), 0, \ldots, 0) \oplus \ldots \oplus (0, 0, \ldots, 0, f(n)),$$
we have $\operatorname{\textnormal{\L}}_{n+1}^{\operatorname{\textnormal{\L}}_{n+1}} \subseteq \widetilde{\operatorname{\textnormal{\L}}_{n+1}}$.
\end{proof}
\begin{proposition}\label{01tilde}
The CMV-algebra $\widetilde{[0,1]}$ consists of the continuous piecewise linear functions from $[0,1]$ to $[0,1]$ where each piece has the form $x \longmapsto mx + \alpha$, with $m \in \mathbb{Z}$ and $\alpha \in \mathbb{R}$.
\end{proposition}
\begin{proof}
Let us denote by $A$ the set of continuous piecewise linear functions of the form above. It is easy to see that $\mathbf{A} = \langle A, \oplus, *, 0, \circ, i \rangle$ with pointwise defined operations and with composition of functions is a CMV-subalgebra of $[0,1]^{[0,1]}$ containing the constant functions. Note that the MV-subalgebra of $[0,1]^{[0,1]}_{\mathcal{MV}}$ generated by constant functions $c_r$, with $r \in [0,1]$, and the identity is precisely $\mathbf{A}_{\mathcal{MV}} = \langle A, \oplus, *, 0, 1\rangle$, hence $\mathbf{A}$ is the smallest CMV-algebra containing $c_r$ for any $r \in [0,1]$.
\end{proof}
\begin{definition}\label{def:constant}
An element $a$ of a CMV-algebra $A$ is a \emph{constant} if $a \diamond x = a$ for all $x \in A$.
\end{definition}
For example $0$ and $1$ are constants. Constant functions in the CMV-algebra $\mathbf{M}^M$ are constants.
\begin{lemma}\emph{\cite{diflge}}
The following properties of constants hold.
\begin{itemize}
\item[(i)] An element $a$ of a CMV-algebra $\mathbf{A}$ is a constant if and only if $a \diamond 1 = a$ if and only if $a \diamond 0 = a$.
\item[(ii)] If $f: \mathbf{A} \longrightarrow \mathbf{B}$ is a morphism of CMV-algebras and $a \in A$ is a constant, then $f(a) \in B$ is a constant. \item[(iii)] The set $K$ of all the constants of the CMV-algebra $\mathbf{A}$ is an MV-subalgebra of $\mathbf{A}_{\mathcal{MV}}$ and an ideal of the monoid $\langle A, \diamond, i\rangle$.
\end{itemize}
\end{lemma}
\section{CMV-algebras and MV-endomorphisms}\label{cmvend}
If $\mathbf{A} = \langle A, \oplus, ^*, 0\rangle$ is an MV-algebra, then $E(\mathbf{A})$ shall denote the monoid of MV-endomorphisms with composition operation $\boxdot$ given for every $a \in A$ by $(f \boxdot g)(a) = g(f(a))$, and the identity map, $\operatorname{id}_A$, as neutral element. With the above notations we have
\begin{proposition}\label{endcomp}
Let $\mathbf{A} = \langle A, \oplus, ^*, 0, \diamond, i\rangle$ be a CMV-algebra. Then there exists an injective homomorphism of monoids from $\mathbf{A}_\mathcal{M}$ to $E(\mathbf{A}_{\mathcal{MV}})$.
\end{proposition}
\begin{proof}
For all $x, y \in A$ we set $\mu_y(x) = x \diamond y$. It is easy to see that for any $y \in A$, $\mu_y$ is an MV-endomorphism, i.e. $\mu_y \in E(\mathbf{A}_{\mathcal{MV}})$. Then we can define a map $\mu: \mathbf{A}_\mathcal{M} \longrightarrow E(\mathbf{A}_{\mathcal{MV}})$, by setting $\mu(y) = \mu_y$, for all $y \in A$. It is easy to check that $\mu$ is a homomorphism of monoids. To show that $\mu$ is injective we observe that, assuming $\mu(y) = \mu(z)$, we have $x \diamond y = x \diamond z$ for all $x\in A$. So, for $x = i$, we have $y = z$.
\end{proof}
\begin{proposition}\emph{\cite{diflge}}\label{prop226}
Let $\mathbf{A} = \langle A, \oplus, ^*, 0\rangle$ be an MV-algebra and $M_A = \langle A, \diamond, i\rangle$ be any monoid over $A$ such that the map $\Psi: M_A \longrightarrow E(\mathbf{A})$, given by $\Psi(y)(x) = x \diamond y$ for all $x, y \in A$ is a monoid homomorphism. Then we have:
\begin{enumerate}
\item[$(i)$] $\mathbf{A} = \langle A, \oplus, ^*, 0, \diamond, i\rangle$ is a CMV-algebra;
\item[$(ii)$] $\mathbf{A}_{\mathcal{MV}} = \mathbf{A}$;
\item[$(iii)$] $M_A = \mathbf{A}_\mathcal{M}$;
\item[$(iv)$] $\Psi$ is injective.
\end{enumerate}
\end{proposition}
\begin{proposition}\emph{\cite{diflge}}\label{prop:MVtoCMV}
Let $\mathbf{A} = \langle A, \oplus, ^*, 0\rangle$ be an MV-algebra such that there is a mapping $\Phi: A \longrightarrow E(\mathbf{A})$ satisfying the following conditions:
\begin{enumerate}
\item[$(i)$] $\Phi$ is injective;
\item[$(ii)$] $\Phi[A]$ is a submonoid of $E(\mathbf{A})$;
\item[$(iii)$] $\Phi(\Phi(z)(y))(x) = (\Phi(z) \boxdot \Phi(y))(x)$, for all $x, y, z \in A$;
\item[$(iv)$] $\Phi(x)(i) = x$, for any $x\in A$, where $i = \Phi^{-1}(\operatorname{id}_A)$.
\end{enumerate}
Then there exists a CMV-algebra $\mathbf{B}$ such that $\mathbf{A} = \mathbf{B}_{\mathcal{MV}}$.
\end{proposition}
Consider the free MV-algebra $\mathbf{M}_n$ over $n$ generators $p_1, \ldots, p_n$. For any $f \in M_n$ let $\sigma_f: M_n \longrightarrow M_n$ be the map defined by setting $\sigma_f(p_1) = f$ and $\sigma_f(p_i) = p_i$ for $i = 2, \ldots, n$. Then $\sigma_f$ is an endomorphism of $\mathbf{M}_n$ mapping each element $g(p_1, \ldots, p_n)$ of $M_n$ in the element $g(f, p_2, \ldots, p_n)$ of $M_n$ obtained by substituting each occurrence of $p_1$ in $g$ with $f$. We will call the endomorphisms $\sigma_f$ \emph{partial substitutions}.
The map $\Phi: f \in M_n \longmapsto \sigma_f \in E(M_n)$ satisfies the assumptions of Proposition~\ref{prop:MVtoCMV}. Indeed $\Phi$ is clearly an injective function and, for all $f, g \in M_n$, $\sigma_f \circ \sigma_g = \sigma_{\sigma_f(g)}$. Since $\operatorname{id}_{M_n} = \sigma_{p_1}$, then $\Phi[M_n]$ is a submonoid of $E(M_n)$ and $\Phi(\Phi(h)(g))(f) = \Phi(\sigma_h(g))(f) = \sigma_{\sigma_h(g)}(f) = (\sigma_h \circ \sigma_g)(f) (\Phi(h)\circ \Phi(g))(f)$. Further, $\Phi^{-1}(\operatorname{id}_{M_n}) = p_1$ and $\sigma_f(p_1) = f$ by definition. Hence $M_n$ can be equipped with a structure of CMV-algebra by setting $f \diamond g = \sigma_g(f)$ and $i = p_1$.
\section{Congruences, CMV-ideals and simple algebras}\label{cmvcong}
\begin{definition}
An MV-ideal $I$ of the MV-reduct $\mathbf{A}_{\mathcal{MV}}$ of a CMV-algebra $\mathbf{A}$ is a \emph{CMV-ideal} if it is also a right ideal of the monoid $\langle A, \diamond, i\rangle$ and $x \sim_I y$ implies $a \diamond x \sim_I a \diamond y$ for all $a, x, y \in A$, where $\sim_I$ is the MV-congruence generated by $I$ (see Definition~\ref{mvid}).
Alternatively, we can say that a non-empty subset of a CMV-algebra $\mathbf{A}$ is a CMV-ideal if the following conditions are satisfied:
\begin{enumerate}
\item[$(i)$] if $x \in I$ and $x \geq y$ then $y \in I$,
\item[$(ii)$] if $x, y \in I$ then $x \oplus y \in I$,
\item[$(iii)$] $x \diamond y \in I$ for all $x \in I$ and $y \in A$,
\item[$(iv)$] for all $a, x, y \in A$, if $x \sim_I y$ then $a \diamond x \sim_I a \diamond y$.
\end{enumerate}
A subset of $\mathbf{A}$ satisfying conditions $(i$--$iii)$ is called a \emph{$\diamond$-ideal}.
\end{definition}
\begin{lemma}\label{idcong}
Let $\mathbf{A}$ be a CMV-algebra and $I$ a subset of $A$. Then $I$ is a CMV-ideal if and only if the relation $\sim_{I}$ is a CMV-algebra congruence.
\end{lemma}
\begin{proof}
From the definition of CMV-ideal it follows that, if $x \sim_I y$ then $a \diamond x \sim_I a \diamond y$ for all $a \in A$, hence $\sim_I$ is a CMV-congruence.
Conversely, if $\sim_I$ is a congruence of CMV-algebras then, for $x \in I$ and $y \in A$, we have $x \sim_I 0$ and $(x \diamond y) \sim_I (0 \diamond y) = 0$, i.e. $(x \diamond y) \sim_I 0$; hence $(x \diamond y) \in I$. Moreover, to show that $I$ is a CMV-ideal, assume $x \sim_I y$. Since $\sim_I$ is a congruence, we have, for any $a \in A$, $a \diamond x \sim_I a \diamond y$. Hence $I$ is a CMV-ideal.
\end{proof}
The following result has a standard algebraic proof.
\begin{proposition}
Let $\mathbf{A}$, $\mathbf{B}$ be CMV-algebras and $h: A \longrightarrow B$ a CMV-homomorphism. Then for any CMV-ideal $J$ of $\mathbf{B}$ the set $h^{-1}(J) = \{x \in A \ \mid \ h(x) \in J\}$ is a CMV-ideal of $\mathbf{A}$.
\end{proposition}
\begin{exm}\label{ex:3.9}
\begin{enumerate}
\item Consider the CMV-algebra $\mathbf{M}^\mathcal C_1$ of McNaughton functions of one variable. Let $I$ be the set of all the functions $f \in M_1$ such that $f(0) = f(1) = 0$. Then $I$ is a CMV-ideal.
\emph{Indeed, it is clear that $I$ is an MV-ideal; now if $f \in I$ and $g \in M_1$ then $g[\{0,1\}] \subseteq \{0,1\}$ and therefore $f \circ g \in I$. It is easy to see that $f \sim_I g$ iff $f(0) = g(0)$ and $f(1) = g(1)$. Then we have that $h \circ f \sim h \circ g$ for any $h \in M_1$.}
\item To get another example of CMV-ideal on $\mathbf{M}^\mathcal C_1$, consider the MV-ideal $J_0$ of the functions that are constantly equal to zero in a
neighbourhood of $0$ (the neighbourhood depending on the function) and the MV-ideal $J_1$ constructed in the same way for neighbourhoods of $1$. They are prime MV-ideals. Let $J$ be $J_0 \cap J_1$. Then $J$ is a CMV-ideal.
\emph{Clearly $J$ is an MV-ideal; let $f \in J$ and suppose $f = 0$ on $[0,c) \cup (d,1]$. Take $h \in M_1$ and assume, for example, that $h(0) = 0$ and $h(1) = 0$. Since $h$ is continuous, it is possible to find a neighbourhood $U$ of $0$ and a neighbourhood $V$ of $1$ such that $h[U], h[V] \subseteq [0,c)$ and, then, $f \circ h \in J$. The other cases are treated in the same way. Now if $\sim_J$ is the equivalence defined by $J$ we have that $f \sim_J g$ iff $f = g$ in a neighbourhood of $0$ and in a neighbourhood of $1$. It is obvious that, in this case, $h \circ f \sim_J h \circ g$ for any $h \in M_1$.}
\end{enumerate}
\end{exm}
We generalize Example~\ref{ex:3.9} as follows. Let $\mathbf{A}$ be a CMV-algebra and $\mathbf{B}$ an MV-subalgebra of $\mathbf{A}$. We set $S_B = \{a \in A \ \mid \ a \diamond x \in B \ \forall x \in B\}$. It is immediate to verify that $S_B$ is a CMV-subalgebra of $A$. The interpretation in ``functional'' terms is: $S_B$ is the CMV-algebra of the functions that fix the subalgebra $B$. Now we set $J = \{a \in A \ \mid \ a \diamond x = 0 \ \forall x \in B\}$. Obviously $J \subseteq S_B$. Moreover $J$ is a CMV-ideal of $S_B$.
As concrete examples we can consider the CMV-algebra $\operatorname{\textnormal{\L}}_{n+1}^{\operatorname{\textnormal{\L}}_{n+1}}$ and take the subalgebra $B$ of the constant functions. In this case the quotient CMV-algebra $S_B/J$ can be identified with the CMV-subalgebra of $\operatorname{\textnormal{\L}}_{n+1}^{\operatorname{\textnormal{\L}}_{n+1}}$ consisting of the functions $f$ satisfying $f[\{0,1\}] \subseteq \{0,1\}$. It is a CMV-algebra with $4(n+1)^{n-1}$ elements.
According to Universal Algebra, we define \emph{simple} CMV-algebras as those algebras having $\{0\}$ and the whole algebra as the only CMV-ideals.
\begin{lemma}\label{lem:id}
Let $\mathbf{M}$ be an MV-algebra and $\mathbf{A} = \mathbf{M}^M$ the corresponding functional CMV-algebra. Let $I$ be a $\diamond$-ideal of $A$, $a \in M$ and $I(a) = \{f(a) \ \mid \ f \in I\} \subseteq M$. Then $I(a)$ is an MV-ideal of $\mathbf{M}$.
\end{lemma}
\begin{proof}
If $x, y \in I(a)$ then $x = f(a), y = g(a)$ with $f, g \in I$; $f \oplus g \in I$ so $x \oplus y = (f \oplus g)(a) \in I(a)$ and if $z \leq f(a)$ then the function $h: M \longrightarrow M$ defined by $h(a) = z, h(b) = f(b)$ for $b \neq a$ is in $I$ and so $z \in I(a)$. We have more: $I(a) = \bigcup_{f \in I} f[M]$ (hence it is independent from $a$). This is due to the fact that, denoting by $c_b$ the function constantly equal to $b$, if $I$ is a $\diamond$-ideal, then $f \diamond c_b = c_{f(b)} \in I$.
\end{proof}
\begin{theorem}\emph{\cite{diflge}}\label{th:simple}
If $\mathbf{M}$ is an MV-algebra then $\mathbf{A} = \mathbf{M}^M$ is a simple CMV-algebra.
\end{theorem}
Let $\mathbf{A} = \langle A, \oplus, ^*, 0\rangle$ be an MV-algebra and $\mathbf{B}$ a CMV-subalgebra of the CMV-algebra $\mathbf{A}^\mathbf{A}$. The zero element of $B$ will be denoted by $f_{0}$. For any element $f$ of $B$, we set $\operatorname{zero}(f) = \{a \in A \ \mid \ f(a) = 0\}$ and we denote by $\operatorname{supp}(f)$ the complement of $\operatorname{zero}(f)$. For a $\diamond$-ideal $I$ of $\mathbf{B}$ we set $\Sigma(I) = \bigcap_{f \in I} \operatorname{zero}(f)$. Let $S$ be any subset of $A$, then we set $Z(S) = \{f \in B \ \mid \ f(s) = 0 \quad \forall s \in S\}$. $S$ will be called $B$-\emph{stable} if and only if, for every $f \in B$, $f[S] \subseteq S$. With the above notations and definitions we have
\begin{lemma}\label{lem1}
Let $S$ be a $B$-stable non-empty subset of $A$. Then $Z(S)$ is a proper CMV-ideal of $\mathbf{B}$.
\end{lemma}
\begin{proof}
It is easy to check that $Z(S)$ is a proper MV-ideal. Moreover, for $f \in Z(S)$ and $g \in B$ we have, for every $s \in S$, $g(s) \in S$ and $f(g(s)) = 0$. Thus $f \diamond g \in Z(S)$, i.e. $Z(S)$ is a $\diamond$-ideal. Finally, for $f, g \in B$, assume $f \sim_{Z(S)} g$. Then $d(f,g) \in Z(S)$, that is $f(s) = g(s)$ for all $s \in S$. Hence, for any $h \in B$ and $s \in S$, $d(h \diamond f, h \diamond g)(s) = 0$. Therefore $h \diamond f \sim_{Z(S)} h \diamond g$ and $Z(S)$ is a CMV-ideal.
\end{proof}
\begin{lemma}\label{lem2}
Let $J$ be a proper $\diamond$-ideal of $\mathbf{B}$ such that $\Sigma(J) \neq \varnothing$. Then $\Sigma(J)$ is a $B$-stable subset of $A$.
\end{lemma}
\begin{proof}
Let $g\in J$, then for an arbitrary element $h$ of $B$ we have $g \diamond h \in J$ and, for any $a \in \Sigma(J)$, $g(h(a)) = 0$. Hence $h(a) \in \operatorname{zero}(g)$. So $h(a) \in \Sigma(J)$, i.e. $h(\Sigma(J)) \subseteq \Sigma(J)$.
\end{proof}
\section{Modules}\label{mod}
In the present section we are going to give the definition of CMV-module. The main result of the section states that every MV-algebra $\mathbf{B}$ is an
$\mathbf{M}^\mathcal C_1$-module, where $\mathbf{M}^\mathcal C_1$ is the CMV-algebra of one-variable McNaughton functions. In particular, if $\mathbf{B}$ is the MV-algebra $\mathbf{M}_1$, we can associate, to each element $f$ of the MV-algebra $\mathbf{M}_1$, an $\mathbf{M}^\mathcal C_1$-module endomorphism $\Phi_f$ via the mapping $\Phi_f(h) = f \circ h$. In this way the algebraic nature of substitutions is displayed in the suitable framework of $\mathbf{M}^\mathcal C_1$-module.
\begin{definition}\label{def:5.1}
Let us consider an MV-algebra $\mathbf{M} = \langle M, \oplus_M, ^{*_M}, 0_M, 1_M\rangle$ and a CMV-algebra $\mathbf{A} = \langle A, \oplus_A, ^{*_A}, 0_A, \diamond, i\rangle$. $\mathbf{M}$ is an \emph{$\mathbf{A}$-module} if there is an external law
$$\begin{array}{ccc}
A \times M & \longrightarrow & M \\
(a,x) & \longmapsto & ax
\end{array}$$
such that
\begin{enumerate}
\item[$(i)$] $(a \oplus_A b) x = ax \oplus_M bx$,
\item[$(ii)$] $a^{*_A} x = (ax)^{*_M }$,
\item[$(iii)$] $0_A x = 0_M$,
\item[$(iv)$] $ix = x$,
\item[$(v)$] $(a \diamond b) x = a(bx)$,
\end{enumerate}
for all $a, b \in A$ and $x \in M$.
\end{definition}
If $\mathbf{M}$ and $\mathbf{P}$ are $\mathbf{A}$-modules then a module morphism from $\mathbf{M}$ to $\mathbf{P}$ is a morphism of MV-algebras $\varphi: M \longrightarrow P$ such that $\varphi(ax) = a \varphi(x)$ for all $a \in A$, $x \in M$. It is a trivial remark that $\mathbf{A}$-modules and their morphisms form a category in a canonical way.
\begin{exm}
\begin{enumerate}
\item The MV-algebra reduct $\mathbf{A}_{\mathcal{MV}} = \langle A, \oplus, ^*, 0\rangle$ of a CMV-algebra $\mathbf{A}$ is an $\mathbf{A}$-module by setting $a x = a \diamond x$.
\item Let $\mathbf{M}$ be an MV-algebra and $\mathbf{A} = \mathbf{M}^M$. We define $fx = f(x)$ for all $f \in A$ and $x \in M$. This clearly makes $\mathbf{M}$ an $\mathbf{A}$-module.
\item One can generalize the previous example by taking a CMV-subalgebra $\mathbf{B}$ of $\mathbf{A}$. So, for example, $[0,1]$ is a module over the CMV-algebra $\mathbf{M}^\mathcal C_1$ of McNaughton functions.
\item An MV-algebra is an $A_4$-module if and only if it is a Boolean algebra. Indeed it is obvious that the external law can be uniquely defined and satisfies all the conditions needed.
\item Let $\mathbf{A}$ be a CMV-algebra and $\mathbf{K}$ be the MV-algebra of the constants of $A$ (see Definition~\ref{def:constant}), it is easy to prove that $K$ is an MV-subalgebra of $A$. Then $\mathbf{K}$ is, in a natural way, an $\mathbf{A}$-module.
Indeed we define $a k = a \diamond k$ for all $a \in A$ and $k \in K$; we have $a k \diamond x = (a \diamond k) \diamond x = a \diamond (k \diamond x) = a \diamond k = ak$ and the definition is correct. Conditions $(i$---$v)$ of Definition~\ref{def:5.1} are trivially satisfied.
\item If $\mathbf{A}$ is a CMV-subalgebra of $\mathbf{B}$ then $\mathbf{B}$ is in a canonical way an $\mathbf{A}$-module.
\item By combining examples 5 and 6, if $\mathbf{K}_B$ is the algebra of constants of $\mathbf{B}$ then $\mathbf{K}_B$ is an $\mathbf{A}$-module.
\item We can generalize Example 6 to the case of a morphism of CMV-algebras $\Phi: A \longrightarrow B$ and consider $B$ as an $\mathbf{A}$-module defining $ab = \Phi(a) \diamond b$ for all $a \in A$ and $b \in B$. It is easy to check conditions of Definition~\ref{def:5.1}. Even more generally, the morphism $\Phi$ turns naturally every $\mathbf{B}$-module into an $\mathbf{A}$-module; this operation is functorial and may be called ``restriction of
scalars''.
\item Let $\mathbf{M}$ be an $\mathbf{A}$-module, $X$ a given (non-empty) set and consider the set $P$ of all the functions $f: X \longrightarrow M$. $P$ is an MV-algebra in an obvious way (pointwise defined operations). Define $af$, for $a \in A$ and $f\in P$, by $(af)(x) = af(x)$, for all $x \in X$. We obtain a structure of $\mathbf{A}$-module on $P$. In particular the MV-algebra $\mathbf{A}^n$ is an $\mathbf{A}$-module.
\item Let $\mathbf{M}$ be an MV-algebra and $X,P$ as in the previous example; let $\mathbf{A}$ be the CMV-algebra $\mathbf{M}^M$. Then $P$ is an $\mathbf{A}$-module by defining $af = a \circ f$ where $\circ$ is the composition of functions.
\end{enumerate}
\end{exm}
Let $\mathbf{M}^\mathcal C_1$ be the CMV-algebra of McNaughton functions of one
variable where the composition is the usual composition of functions
(see Example~\ref{ex:CMV}.4).
If $f \in M_1$ we denote by $\Phi_f: M_1 \longrightarrow M_1$ the function $\Phi_f(h) = h \circ f$. It is clear that $\Phi_f$ is an endomorphism of the $\mathbf{M}^\mathcal C_1$-module $\mathbf{M}_1$. We call \emph{substitutions} the morphisms of type $\Phi_f$. Note that $\Phi_f(i) = f$.
\begin{lemma}\label{lem:5.3}
$\Phi_{f \circ g} = \Phi_g \circ \Phi _f$.
\end{lemma}
\begin{proof}
$\Phi_{f \circ g}(h) = h \circ (f \circ g) = (h \circ f) \circ g = \Phi_g (\Phi_f(h))$.
\end{proof}
Let $\mathbf{B}$ be any MV-algebra. If $x \in B$ then there is a unique morphism of MV-algebras $\varphi_x: \mathbf{M}_1 \longrightarrow \mathbf{B}$ such that $\varphi_x(i) = x$. This is due to the fact that the MV-algebra $\mathbf{M}_1$ is free over the generator $i$, hence elements of $M_1$ are MV-terms modulo provable equality in $i$ and, by the universal property of free algebras, for any $x \in B$ the map $\varphi_x$ sends any MV-term $\tau(i)$ to the element $\tau(x) \in B$. Denote also $fx = \varphi_x(f)$ for all $f \in M_1$ and $x \in B$.
\begin{proposition}\label{prop5.4}
$\varphi_x \circ \Phi_f = \varphi_{fx}$.
\end{proposition}
\begin{proof}
Elements of the free algebra $M_1$ are MV-polynomials in $i$. By the universal property of free algebras, for any $x \in B$ the map $\varphi_x$ sends any polynomial $\tau(i)$ into the element $\tau(x) \in B$. Hence, for all $g \in M_1$, $(\varphi_x \circ \Phi_f)(g) = \varphi_x(g \circ f) = g(f(x)) = \varphi_{fx}(g)$.
\end{proof}
\begin{theorem}
Every MV-algebra $\mathbf{B}$ is, in a canonical way, a module over the CMV-algebra $\mathbf{M}^\mathcal C_1$ of McNaughton functions.
\end{theorem}
\begin{proof}
We observe that the function $(f,x) \longmapsto fx$ gives to $\mathbf{B}$ a structure of $\mathbf{M}^\mathcal C_1$-module. Then it is enough to prove that $(f \circ g)x = f(gx)$ for all $f, g \in M_1$ and $x \in B$. We have, by Proposition~\ref{prop5.4}, $\varphi_{(f \circ g)x} = \varphi_x \circ \Phi_{f \circ g}$ and, using Lemma~\ref{lem:5.3}, $\varphi_{(f \circ g)x} = \varphi_x \circ (\Phi_g \circ \Phi_f) = (\varphi_x \circ \Phi_g) \circ \Phi_f = \varphi_{gx} \circ \Phi_f = \varphi_{f(gx)}$. Now we can apply all this to $i$ and get the result.
\end{proof}
\section{The logic $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$}\label{logic}
We shall extend the fragment of \L ukasiewicz propositional calculus of one variable $v$, here denoted by $\operatorname{\textnormal{\L}}_\omega^1$. The extension of $\operatorname{\textnormal{\L}}_\omega^1$ that we are going to define will be denoted by $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$; formulas of $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$ are built from the connectives of negation ($\neg $) and implication $\to$ in the usual way (the conjunction $\odot$ and the disjunction $\oplus$ being considered as derived from $\neg$ and $\to$), and by a binary connective $\blacktriangleleft$ defined as follows:
$\varphi \blacktriangleleft \psi $ is the formula obtained by the formula $\varphi$ where every occurrence of the variable $v$ is substituted by the formula $\psi$. Actually the set $\operatorname{Form}$ of formulas of $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$ is obtained inductively by the following stipulations:
\begin{enumerate}
\item The single variable $v$ is a (atomic) formula;
\item If $\alpha$ is a formula, then $\neg \alpha$ is a formula;
\item If $\alpha$ and $\beta$ are formulas, then $\alpha \rightarrow \beta$ and $\alpha \blacktriangleleft \beta$ are formulas.
\end{enumerate}
In the light of next results, the formula $\alpha \blacktriangleleft \beta$ can be thought of as the formula obtained from $\alpha$ by the substitution of each occurrence of the variable $v$ with the formula $\beta$.
Interpretation of connectives of $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$ is given by the following definition.
\begin{definition}
Let $\mathbf{A} = \langle A, \oplus, ^*, 0, \diamond, i \rangle$ be a CMV-algebra. Then an \emph{$\mathbf{A}$-valuation} is a function $\operatorname{val}: \operatorname{Form} \longrightarrow A$ satisfying the following properties, where $\alpha$ and $\beta$ denote arbitrary formulas:
\begin{itemize}
\item[-] $\operatorname{val}(v) = i$,
\item[-] $\operatorname{val}(\neg \alpha) = (\operatorname{val}(\alpha))^*$,
\item[-] $\operatorname{val}(\alpha \rightarrow \beta) = \operatorname{val}(\alpha) \rightarrow \operatorname{val}(\beta))$,
\item[-] $\operatorname{val}(\alpha \blacktriangleleft \beta) = \operatorname{val}(\alpha) \diamond \operatorname{val}(\beta)$.
\end{itemize}
\end{definition}
An $\mathbf{A}$-valuation $\operatorname{val}$ is said to $\mathbf{A}$-satisfy a formula $\alpha$ iff $\operatorname{val}(\alpha) = 1$; $\alpha$ is an $\mathbf{A}$-tautology iff $\alpha$ is satisfied by all $\mathbf{A}$-valuations. Formulas $\alpha$ and $\beta$ are semantically $\mathbf{A}$-equivalent iff $\operatorname{val}(\alpha) = \operatorname{val}(\beta)$ for all $\mathbf{A}$-valuations $\operatorname{val}$. The syntactic approach to many-valued logics is the same as for propositional classical logic. A set of formulas, called {\em axioms}, and a set of inference rules are fixed. In $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$ we have
Axioms:
\begin{enumerate}
\item[$\operatorname{Ax}1$] \quad $\varphi \to (\psi \to \varphi)$
\item[$\operatorname{Ax}2$] \quad $(\varphi \to \psi) \to ((\psi \to \chi)\to (\varphi\to \chi))$
\item[$\operatorname{Ax}3$] \quad $((\varphi \to \psi)\to \psi)\to ((\psi \to \varphi)\to \varphi)$
\item[$\operatorname{Ax}4$] \quad $(\neg \varphi\to \neg \psi)\to (\psi \to \varphi)$
\item[$\operatorname{Ax}5$] \quad $(\varphi \blacktriangleleft v)\to \varphi$
\item[$\operatorname{Ax}6$] \quad $(v\blacktriangleleft \varphi)\to \varphi$
\item[$\operatorname{Ax}7$] \quad $\varphi \to (v \blacktriangleleft \varphi)$
\item[$\operatorname{Ax}8$] \quad $\varphi \to (\varphi \blacktriangleleft v)$
\item[$\operatorname{Ax}9$] \quad $((\varphi \to \psi)\blacktriangleleft \gamma) \to ((\varphi\blacktriangleleft \gamma) \to (\psi \blacktriangleleft \gamma))$
\item[$\operatorname{Ax}10$] \quad $((\neg \varphi \to \psi)\blacktriangleleft \gamma)\to (\neg(\varphi\blacktriangleleft \gamma) \to (\psi \blacktriangleleft \gamma))$
\item[$\operatorname{Ax}11$] \quad $(\neg(\varphi\blacktriangleleft \gamma) \to (\psi \blacktriangleleft \gamma))\to ((\neg \varphi \to \psi)\blacktriangleleft \gamma)$
\item[$\operatorname{Ax}12$] \quad $(\varphi \blacktriangleleft (\psi\blacktriangleleft \gamma))\to ((\varphi\blacktriangleleft \psi)\blacktriangleleft \gamma)$
\item[$\operatorname{Ax}13$] \quad $((\varphi\blacktriangleleft \psi)\blacktriangleleft \gamma)\to (\varphi \blacktriangleleft (\psi\blacktriangleleft \gamma))$
\item[$\operatorname{Ax}14$] \quad $\neg(\varphi\blacktriangleleft \psi)\to (\neg\varphi\blacktriangleleft \psi)$
\item[$\operatorname{Ax}15$] \quad $(\neg\varphi\blacktriangleleft \psi)\to \neg(\varphi\blacktriangleleft \psi)$
\end{enumerate}
Inference rules:
$$\frac{\alpha \quad \alpha \to \beta}{\beta} \qquad \textrm{\emph{Modus Ponens}},$$
$$\frac{\alpha}{\alpha \blacktriangleleft \beta} \qquad \textrm{$\blacktriangleleft$-\emph{rule}},$$
$$\frac{\alpha}{\alpha \to (\alpha \blacktriangleleft \beta)} \qquad \textrm{\emph{Arrow $\blacktriangleleft$-rule}}.$$
By $\vdash \alpha $ we mean that $\alpha$ is a theorem of $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$, and the set of all the theorems of $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$ shall be denoted by $\operatorname{Th}$. Let now $\sim_{\blacktriangleleft}$ be the binary relation on $\operatorname{Form}$ defined by:
$$\alpha \sim_{\blacktriangleleft} \beta \qquad \text{iff} \qquad \vdash \alpha \to \beta \ \text{ and } \ \vdash \beta \to \alpha.$$
With the above notations we have:
\begin{proposition}\label{simequiv}
$\sim_{\blacktriangleleft}$ is an equivalence relation such that if $\alpha \sim_{\blacktriangleleft} \beta$ then $((\alpha \blacktriangleleft \gamma) \sim_{\blacktriangleleft} (\beta \blacktriangleleft \gamma))$, for every formula $\gamma$.
\end{proposition}
\begin{proof}
$\sim_{\blacktriangleleft}$ is trivially an equivalence relation. The fact that $\alpha \sim_{\blacktriangleleft} \beta$ implies $((\alpha \blacktriangleleft \gamma) \sim_{\blacktriangleleft} (\beta \blacktriangleleft \gamma))$, for all $\gamma$, is as an easy application of $\blacktriangleleft$-rule, $\operatorname{Ax}9$ and Modus Ponens.
\end{proof}
The equivalence classes of $\sim_{\blacktriangleleft}$ will be denoted by
$$\mid \alpha \mid \ := \ \{\beta \in \operatorname{Form} \mid \beta \sim_{\blacktriangleleft}\alpha\}.$$
\begin{lemma}\label{thms}
For each formula $\alpha$, $\vdash \alpha$ if and only if $\mid \alpha \mid \ = \ \operatorname{Th}$.
\end{lemma}
\begin{proof}
If $\mid \alpha \mid = \operatorname{Th} $, then $\vdash \alpha$, because $\alpha \in \mid \alpha \mid$.
Conversely, suppose that $\vdash \alpha$ and $\vdash \beta$. Then, by $\operatorname{Ax}1$ and Modus Ponens, we get $\vdash \beta \to \alpha$ and $\vdash \alpha \to \beta$, hence $\beta \sim_{\blacktriangleleft} \alpha$. Then $\operatorname{Th} \subseteq \mid\alpha\mid$. Now take $\beta \in \mid\alpha\mid$. Then $\vdash \alpha \to \beta$ and therefore, via Modus Ponens, we get $\vdash \beta$, which implies $\mid \alpha \mid \subseteq \operatorname{Th}$.
\end{proof}
\begin{theorem} The quotient set $\frac{\operatorname{Form}}{\sim_{\blacktriangleleft}}$ becomes a CMV-algebra,
once equipped with the operations $\to, \neg, \diamond$ and the
constants $1$ and $i$ given by the following stipulations:
\begin{enumerate}
\item[$(i)$] $\mid \alpha \mid \to \mid \beta \mid \ := \ \mid \alpha \to \beta \mid$
\item[$(ii)$] $(\mid \alpha \mid )^* \ := \ \mid \neg \alpha \mid$
\item[$(iii)$] $\mid \alpha \mid \diamond \mid \beta \mid \ := \ \mid \alpha \blacktriangleleft \beta \mid$
\item[$(iv)$] $1 \ := \ \operatorname{Th}$
\item[$(v)$] $i \ := \ \mid v \mid$.
\end{enumerate}
\end{theorem}
\begin{proof}
By Proposition~\ref{simequiv} and Lemma~\ref{thms} $(i$--$iii)$ are well defined operations on $\frac{\operatorname{Form}}{\sim_{\blacktriangleleft}}$. It remains to prove that the axioms of CMV-algebras are satisfied. Here we limit ourselves to check the proper axioms of CMV-algebra. Let us prove the associativity of $\diamond$. Indeed, to prove that $(\mid \alpha \mid \diamond \mid \beta \mid)\diamond \mid \gamma \mid = \mid \alpha \mid \diamond(\mid \beta \mid \diamond \mid \gamma \mid)$ we have to prove that $\alpha \blacktriangleleft (\beta \blacktriangleleft \gamma) \sim_{\blacktriangleleft} (\alpha \blacktriangleleft \beta) \blacktriangleleft \gamma$. Equivalently we should show that
\begin{equation}\label{(j)}
\vdash (\alpha \blacktriangleleft (\beta \blacktriangleleft \gamma)) \to ((\alpha \blacktriangleleft \beta) \blacktriangleleft \gamma),
\end{equation}
\begin{equation}\label{(jj)}
\vdash ((\alpha \blacktriangleleft \beta) \blacktriangleleft \gamma) \to (\alpha \blacktriangleleft (\beta \blacktriangleleft \gamma)),
\end{equation}
but (\ref{(j)}) is precisely $\operatorname{Ax}12$ and (\ref{(jj)}) is $\operatorname{Ax}13$. To prove the distributivity of $\diamond$ with respect to the $\oplus$ operation we need to prove that
$$(\neg \alpha \to \beta) \blacktriangleleft \gamma \quad \sim_{\blacktriangleleft} \quad \neg (\alpha \blacktriangleleft \gamma) \to (\beta \blacktriangleleft \gamma),$$
i.e. we need to show that the following hold:
\begin{equation}\label{(jjj)}
\vdash ((\neg \alpha \to \beta) \blacktriangleleft \gamma) \to (\neg(\alpha \blacktriangleleft \gamma) \to (\beta \blacktriangleleft \gamma)),
\end{equation}
\begin{equation}\label{(jv)}
\vdash (\neg(\alpha \blacktriangleleft \gamma) \to (\beta \blacktriangleleft \gamma)) \to ((\neg \alpha \to \beta) \blacktriangleleft \gamma).
\end{equation}
Again, (\ref{(jjj)}) and (\ref{(jv)}) are, respectively, $\operatorname{Ax}10$ and $\operatorname{Ax}11$.
To prove that $1 \diamond \mid \alpha \mid = 1$ we have to show that for some $\beta \in \operatorname{Th}$ we have $\beta \sim_{\blacktriangleleft} (\beta \blacktriangleleft \alpha)$, which is absolutely trivial, as well as the fact that $i \diamond \mid \alpha \mid = \mid \alpha \mid$ and $\mid \alpha \mid \diamond i = \mid \alpha \mid$.
Last, to verify that $\mid \alpha \mid^* \diamond \mid \beta \mid = (\mid \alpha \mid \diamond \mid \beta \mid)^*$, we just need to apply $\operatorname{Ax}14$ and $\operatorname{Ax}15$.
\end{proof}
\section{Conclusion}
In this last section we would like to stress some facts about CMV-algebras and $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$.
We presented plenty of examples of CMV-algebras and we have seen that the one generated free MV-algebra is in a canonical way a CMV-algebra. Moreover, CMV-algebra have many interesting algebraic properties and connections with MV-algebras. For example, Propositions~\ref{prop226} and~\ref{prop:MVtoCMV} tell us how to extend an MV-algebra $\mathbf{A}$ to a CMV-algebra $\mathbf{A}^\mathcal C$ using the endomorphism monoid of $\mathbf{A}$, and Theorem~\ref{th:simple} highlights a wide class of simple CMV-algebras.
On the other hand, the system $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$, introduced in Section~\ref{logic}, looks like the single variable fragment of some interesting~---~yet unknown~---~expansion of \L ukasiewicz logic. Hence, it could be fruitful to find a suitable $\operatorname{S}\operatorname{\textnormal{\L}}_\omega$ system (i.e. a system in a numerable set of variables) containing $\operatorname{S}\operatorname{\textnormal{\L}}_\omega^1$ as a fragment.
So the theory of CMV-algebras is still at its initial stage, but many signs indicate that it is worth to keep on investigating such structures, their connection with logic and their possible applications (e.g. at the theory of chaotic dynamical systems, as pointed out in~\cite{diflge}).
|
\section{The Provenance of Astrophysical Data}
The ever increasing datasets archived and dealt with by
astrophysical institutions and researchers are moving
astrophysics into one of the fastest growing, and more public,
e-Science disciplines.
Within e-Science, Data Provenance is the tracking of the origin of
any piece of information that has been recorded, with knowledge
of all processing steps undergone. In the astronomical realm,
we need to track data Provenance in order to answer questions like
\emph{what was done to the photons captured by this archived
entity}, and \emph{what entities were responsible for it?}
We need to be able to answer those questions because for typical
astronomical workflows across large datasets Provenance must be
recorded, and later used throughout the workflow, in order to get
physically meaningful and accurate results.
Fig.~\ref{fig:P64.santander.fig1} illustrates the role of
Data Provenance in astronomical workflows.
\begin{figure}[tb]
\centering
\includegraphics[width=\textwidth]
{P64.santander.fig1.eps}
\caption{
Data
reduction software are
convert properties measured with the telescope
detectors into files ($F_n$), and later into
physical parameters and data products ($P_m$).
Knowledge and models of telescope, detectors
settings, astronomical source, and
absorbing media is needed for the reduction.
}
\label{fig:P64.santander.fig1}
\end{figure}
In particular, relevant questions for Data Provenance are
\emph{which
were the files used to create a particular data product}
(Associations), and \emph{which are the data products created
from it}? (Inverse Association).
Additionally, we need to know
\emph{are we using the same source data} (photons) \emph{for
any two given data products}? (Photon Accountancy), as the use
of the same photons in different data products which are later
combined can produce unreliable science.
Finally, we need to be
able to \emph{track parameters related to the models for
emission, absorption and detection} used by the reduction
software, as these encode implicit knowledge that needs to be
made explicit in order to allow the reproduction of
measurements and processes by other scientists.
\section{Data Provenance Systems' Requirements:
High-Level Use Cases}
\label{sec:use_cases}
The use cases for a Data Provenance system within the ESO
archive (see Fig.~\ref{fig:P64.santander.fig2}) have been
classified in three kinds of services:
\begin{description}
\item[Data Discovery services] which query on the
Provenance records to find scientific data fulfilling
different criteria. These data discovery encompasses
the Provenance of observing projects at all stages:
Organization (observing proposals, PIs and COIs),
Technical (instrument setup), Ambient, and Publications.
In addition, Exposure Time Calculators (ETCs) must rely on
the
historical information of telescope elements, such as
filters or detectors, whose properties (or actual
elements) can change with time.
\item[Dataset Documentation services] which allow scientists
to retrieve the full record of observation origin and
further processing. They are the inverse of the services
above, as the dataset is already known, and the inverse
association must be constructed. ETCs can be used in this
family of use cases to assess data quality for a particular
dataset.
\item[Observation Preparation services] where both
observations and projects with given data acquisition
characteristics are retrieved.
\end{description}
\begin{figure}[tb]
\centering
\includegraphics[width=\textwidth]
{P64.santander.fig2.eps}
\caption{
Classification of use cases compiled for the ESO archive.
The main group corresponds to Data Discovery using
Provenance metadata (mostly through inverse
association queries), while similar queries are
desirable for dataset documentation (using direct
association queries).
}
\label{fig:P64.santander.fig2}
\end{figure}
\section{Implementing Data Provenance at ESO}
\subsection{Challenges}
Following the taxonomy of Data Provenance systems by Simmhan,
Plale, \& Gannon (2005), the main elements to be specified for
Data Provenance systems are Storage, Granularity,
Representation, and Dissemination.
\begin{description}
\item[Storage] deals with the actual way in which
Data Provenance is made persistent. While Storage
implementation is not relevant from an interoperability
point of view, it provides the foundation for all
Provenance services. Insertion and retrieval speeds
must be able to cope with the expected data generation
and query rate.
\item[Granularity] is the degree of detail for which
Provenance metadata is collected. It must be chosen
to serve all ESO use cases, and also imposes
restrictions on Storage (to be able to cope with the
chosen Granularity). We can also include the selection
of which particular metadata (from FITS headers to
observation logs) are committed to Storage as part of
the Granularity. From Fig.~\ref{fig:P64.santander.fig2}
we can see that other metadata, such as Observation logs,
and ADP processing logs, must be taken into account,
and provide a time dimension to the problem.
\item[Representation] the way Provenance is expressed
when shared with requesting entities (as opposed to
Storage). Hierarchical trees, series of keyword/value
pairs, or even plain-text files for informational use
cases. However, due to the rich nature of the objects
described by Provenance ---a detailed description of the
light-path in ESO instruments is shown by Delgado
(2009)---, a hierarchical structure is preferred.
Representation directly affects interoperability, so
the generalization of a Provenance representation as
an IVOA Data Model would be desirable.
\item[Dissemination] the way Provenance metadata is
to be available for shared datasets. Decisions in this
respect include creating Provenance services to satisfy
the use cases outlined in Section~\ref{sec:use_cases}
\end{description}
The challenges, then, lie in successfully managing the huge
datasets involved, do so maintaining system responsiveness, and
standardize the way to retrieve and disseminate Provenance
metadata.
\subsection{Jump-Starts}
In order to find the best solution for some of the issues
outlined above, we can count on several already started
initiatives.
There are active discussions on the Provenance of
astrophysical datasets within the IVOA Working Group
for Data Modelling (DM WG). A Representation, in the form of a
complete IVOA Observation data model, has already been proposed
by Santander-Vela (2009), and is currently under study.
In addition, ESO maintains a complete FITS keyword/metadata
database (Vuong et al. 2008), which will be one of the basis
for establishing Storage and Granularity of the future ESO
Provenance services. Currently, that facility stores more than
five thousand million keyword-value pairs, for more than ten
million raw file entries, and already includes metadata
versioning support (a requirement for Provenance audit trails
and historical dataset documentation), while at the same time
providing excellent query speed. There is also an ongoing
project for using that same database to ingest keywords for
highly processed datasets.
\section{Conclusions}
By identifying the elements needed for a successful data
Provenance management system, and the selection of already in
place systems, the ESO archive will be able to provide even
more advanced scientific-oriented data products and services,
while at the same time benefiting from existing IVOA standards
for the Virtual Observatory, and helping in their development,
leading by example.
\acknowledgements JDSV is grateful to the support
on his Ph.D. thesis by Spanish MICINN grants
AYA2005-07516-C02-00 and AYA 2008-06181-C02-02; to the ESO
Archive Department for support to this conference; and to Paul
Eglitis for valuable comments. AD work is part of the Spanish
in-kind contribution to ESO, and was supported by MICINN grant
CAC-2006-47.
|
\section{Introduction}
In his classic paper \cite{MR885560}, Witten unraveled a beautiful
connection between the elliptic genus discovered by Ochanine
\cite{MR895567} and a certain two-dimensional quantum field theory,
known as the supersymmetric sigma model. Extending his idea to the
case with no left-moving fermions, he was naturally led to discover a
new genus. The Witten genus, as it is now called, associates to every
closed string manifold $M$ of dimension $4k$ an integral modular form
$W(M)$ of weight $2k$.
About a decade after the discovery of the Witten genus, H\"ohn and
Stolz independently made the following
\begin{conjecture}[H\"ohn, Stolz \cite{MR1380455}]
Let $M$ be a closed string manifold. If $M$ admits a Riemannian
metric of positive Ricci curvature, then the Witten genus $W(M)$
vanishes.
\end{conjecture}
The aim of this paper is to explain how recent developments in the
study of two-dimensional supersymmetric sigma models provide a new
approach to the K\"ahler case of the H\"ohn--Stolz conjecture. This
approach is based on a remarkable phenomenon that certain sigma models
with $(0,2)$ supersymmetry exhibit, which was first predicted by
Witten \cite{MR2320663} and subsequently confirmed by Tan and the
author \cite{MR2415553}---specifically, the phenomenon that the chiral
algebras associated to these models vanish due to instanton effects.
The emergence of a chiral algebra (or vertex algebra) is a
characteristic property of two-dimensional quantum field theories with
$(0,2)$ supersymmetry. Although this fact had been realized
\cite{Vafa:1989pa} as early as a few years after the appearance of
Witten's paper \cite{MR885560}, it had not been paid much attention by
physicists until relatively recently. During the last several years,
however, there have been significant advances in our understanding of
the chiral algebras of $(0,2)$ supersymmetric sigma models, thanks
largely to the pioneering works of Kapustin \cite{Kapustin:2005pt} and
Witten \cite{MR2320663}. What inspired these advances is the theory
of chiral differential operators developed earlier in mathematics.
The author hopes that the results presented here from physics give
some inspirations to mathematicians.
We begin our discussion in the next section by reviewing how the
Witten genus arises in two-dimensional supersymmetric sigma models.
Then, in Section 3, we describe how to construct a chiral algebra
using $(0,2)$ supersymmetry in the case when the target space is
K\"ahler. In Section 4, we outline the argument for the vanishing
phenomenon of the chiral algebra and explain its relation to the
H\"ohn--Stolz conjecture. In the last section, we discuss what this
phenomenon implies for the geometry of loop spaces.
\section{Supersymmetry and the Witten genus}
Let $M$ be a closed spin manifold of dimension $d = 4k$. We will
consider the supersymmetric sigma model with no left-moving fermions
that has target space $M$. This theory has a hermitian operator $Q_+$
satisfying the $(0,1)$ supersymmetry algebra
\begin{equation}
\{Q_+,Q_+\} = H - P, \qquad
\{Q_+, (-1)^{F_R}\} = 0,
\end{equation}
where $H$ and $P$ are the generators of translations respectively in
time and space directions, and $(-1)^{F_R}$ is the operator that
counts the number of right-moving fermions modulo two. With a
suitable choice of the complex structure on the two-dimensional
spacetime manifold, the operator $H - P$ generates antiholomorphic
translations, $H - P \propto \partial_\zb$. States annihilated by $Q_+$
are called \emph{supersymmetric}. Supersymmetric states are those
states that have $H - P = 0$.
On the torus $\mathbb{C}/2\pi(\mathbb{Z} +\tau\mathbb{Z})$ with $\tau$ in the upper
half-plane, the partition function of the theory is computed by
\begin{equation}
Z(M; q) = \mathop{\mathrm{Tr}}\nolimits\bigl((-1)^{F_R} q^{(H + P)/2} \qb^{(H - P)/2}\bigr).
\end{equation}
Here $q = e^{2\pi i\tau}$ and the trace is taken in the Hilbert space
of states. Since the supercharge $Q_+$ pairs bosonic ($F_R$ even) and
fermionic ($F_R$ odd) states not in its kernel, $Z(M; q)$ receives
contributions only from supersymmetric states. The operator $P$ takes
values in $-d/24 + \mathbb{Z}$; the shift by $-d/24$ comes from the
regularization of an infinite sum. Thus we can write $Z(M; q)$ in a
$q$-series as
\begin{equation}
\label{Z}
Z(M; q) = q^{-d/24} \sum_{n = 0}^\infty q^n \mathop{\mathrm{Tr}}\nolimits_{\CH_n} (-1)^{F_R}.
\end{equation}
The coefficients are the indices of $Q_+$ in the spaces $\CH_n$ of
supersymmetric states with $P = -d/24 + n$. The Witten genus of $M$
is then given by
\begin{equation}
\label{WG}
W(M) = \eta(q)^d Z(M;q),
\end{equation}
with $\eta(q) = q^{1/24} \prod_{n=1}^\infty (1-q^n)$ the Dedekind eta
function. Since the Witten genus and the partition function differ
only by a factor, we will mainly talk about the latter which is a
physically more natural object to consider.
The geometric meaning of the partition function becomes clear if we
canonically quantize the theory. Quantization identifies states with
spinors on the free loop space $\CL M$ and $(-1)^{F_R}$ with the
chirality operator. The supercharge $Q_+$, on the other hand, is
identified with the Dirac operator $\slashed{D}_{\CL M}$ on $\CL M$ coupled to
the Killing vector field $V$ generating rotations of loops:
\begin{equation}
Q_+ = \slashed{D}_{\CL M} + t \Gamma \cdot V.
\end{equation}
Here $\Gamma$ denotes Clifford multiplication and $t \in \mathbb{R}^+$ is a
parameter of the theory.%
\footnote{There are two ways to think about this parameter. The
volume of the target space vs. the radius of the loop.}
Taking the limit $t \to 0$ in the expression \eqref{Z}, we learn
that the partition function is the $S^1$-equivariant index of the
Dirac operator on the loop space with respect to the natural circle
action \cite{MR970288}.
Taking the limit $t \to \infty$, we obtain a formula that expresses
$Z(M;q)$ in terms of characteristic classes of $M$. Since $H - P$
contains the potential $t^2 |V|^2$, supersymmetric states localize in
this limit to the zeros of $V$. The zeros are the constant loops
forming a copy of $M$ inside $\CL M$. Such localized states can be
thought of as spinors on $M$, taking values in various vector bundles;
an approximate supersymmetric state with $P = -d/24 + n$ is a spinor
with values in the bundle given by the coefficient of $q^n$ in the
series
\begin{equation}
\bigotimes_{m = 1}^\infty S_{q^m}(TM)
= 1 + qTM + q^2\bigl(TM + S^2(TM \otimes TM)\bigr) + \dotsb,
\end{equation}
where the symmetric power operation $S$ of a vector bundle is defined
by $S_q(E) = 1 + qS(E) + q^2S^2(E) + \dotsb$. On these localized
states, $Q_+$ acts as the Dirac operator $\slashed{D}_M$ on $M$. Thus we
obtain
\begin{equation}
\begin{split}
Z(M;q)
&= q^{-d/24} \mathop{\mathrm{index}}\Bigl(\slashed{D}_M
\otimes \bigotimes_{m = 1}^\infty S_{q^m}(TM \otimes \mathbb{C})\Bigr) \\
&= q^{-d/24} \Bigl\langle \hat A(M)
\mathop{\mathrm{ch}}\nolimits\Bigl(\bigotimes_{m = 1}^\infty S_{q^m}(TM \otimes \mathbb{C})\Bigr),
[M]\Bigr\rangle.
\end{split}
\end{equation}
Mathematically, this formula defines the Witten genus via the relation
\eqref{WG}.
Although the above definition of the Witten genus $W(M)$ works for any
closed spin manifold $M$, the underlying quantum field theory does
not. For the theory to be well defined, $M$ must be string (that is,
spin and has $p_1(M)/2 = 0$). In addition, the theory suffers from
the usual problem of infinities of quantum field theory. To ensure
that these infinities can be removed by renormalization (at least to
one-loop level), one must impose that the Ricci curvature of $M$ be
positive or else vanish. Hence, one expects that these conditions are
necessary for the partition function $Z(M;q)$ to be interpreted as the
$S^1$-equivariant index of the Dirac operator on $\CL M$. The
modularity of $W(M)$ for $M$ a string manifold follows from the
diffeomorphism invariance of the theory. Actually, the weaker
condition that $p_1(M) = 0$ rationally is sufficient for this since
the Witten genus can be computed in the perturbative approximation, by
which one loses the torsion information of anomalies.
\section{Chiral Algebras}
Now suppose that $M$ is K\"ahler. In this case, the Dirac operator on
$M$ splits into two pieces as $\slashed{D}_M = \CD_M + \CDb_M$
\cite{MR1031992}. Each of these pieces squares to zero, $\CD_M^2 =
\CDb_M^2 = 0$, and is the adjoint of the other, $\CD_M = \CDb_M^*$.
Furthermore, the spinor bundle has a $\mathbb{Z}$-grading under which $\CD_M$
has degree one and $\CDb_M$ degree minus one. The $r$th
$\CD_M$-cohomology group of the spinor bundle is isomorphic to the
space of harmonic spinors of degree $r$.
Similarly, in the K\"ahler case the supercharge splits as $Q_+ = Q +
Q^*$, with $Q$ satisfying the $(0,2)$ supersymmetry algebra
\begin{equation}
Q^2 = 0, \qquad
\{Q, Q^*\} = \frac{1}{2} (H - P).
\end{equation}
There is also a grading given by \emph{R-charge}, which coincides
modulo two with $(-1)^{F_R}$ and under which $Q$ has charge one and
$Q^*$ charge minus one. This is a $\mathbb{Z}$-grading at the level of
perturbation theory, but reduced to a $\mathbb{Z}/2n\mathbb{Z}$-grading by instanton
effects for $c_1(M) \neq 0$, where $2n$ is the greatest divisor of
$c_1(M)$. (Recall that $M$ is spin, thus $c_1(M) \equiv 0$ (mod
$2$).) The $r$th $Q$-cohomology group computed in the Hilbert space
of states is isomorphic to the space of supersymmetric states of
R-charge $r$.
What makes a quantum field theory with $(0,2)$ supersymmetry
interesting for us is that one can construct another $\mathbb{Z}/2n\mathbb{Z}$-graded
$Q$-cohomology, namely the $Q$-cohomology in the space of local
operators. The action of $Q$ on a local operator $\CO$ is given by
supercommutator, $\{Q, \CO\} = Q\CO \mp \CO Q$ depending on whether
$\CO$ is bosonic or fermionic. The $Q$-cohomology of states and local
operators are isomorphic when the theory is conformally invariant via
the state-operator correspondence. This is not the case in general.
An important property of the $Q$-cohomology of local operators is that
its elements vary holomorphically on the spacetime. By this we mean
the following. Let $\CO$ be a $Q$-closed local operator, $\{Q,\CO\} =
0$, and $[\CO]$ its $Q$-cohomology class. Recalling that $H - P
\propto \partial_\zb$, we have
\begin{equation}
\partial_\zb\CO
\propto \{H - P, \CO\}
= \{\{Q, Q^*\}, \CO\}
= \{Q, \{Q^*,\CO\}\}.
\end{equation}
Thus $\partial_\zb\CO$ is $Q$-exact, and $\partial_\zb$ annihilate
$Q$-cohomology classes.
Another important property is that the $Q$-cohomology of local
operators has natural operator product expansions (OPE) inherited from
those of the underlying theory:
\begin{equation}
\label{OPECA}
[\CO_i(z)] \cdot [\CO_j(z')]
= [\CO_i(z) \cdot \CO_j(z')]
\sim \sum_k c_{ij}{}^k(z - z') [\CO_k(z')].
\end{equation}
The coefficient functions $c_{ij}{}^k(z - z')$ are holomorphic for $z
\neq z'$, but can have poles at $z = z'$.
The holomorphic $Q$-cohomology of local operators, equipped with this
OPE structure, is the \emph{chiral algebra} of the $(0,2)$
supersymmetric theory. The chiral algebra forms a holomorphic sector
of the theory, in which correlation functions are holomorphic in the
insertion points as long as no two points coincide. For the
topological A-model (considered as a $(0,2)$ supersymmetric theory),
these correlation functions generalize Gromov--Witten invariants.
The chiral algebra of a quantum field theory with $(0,2)$
supersymmetry has the same structure as the chiral algebras (vertex
algebras) of conformal field theories do, except that the grading by
conformal weight is missing. In fact, the chiral algebra is in
general not conformally invariant, and certainly not for our sigma
model if $c_1(M) \neq 0$. Though our model is classically conformally
invariant, quantum mechanically the conformal invariance is broken if
the Ricci curvature of $M$ is nonzero. Correspondingly, though the
energy-momentum tensor $T_{zz}$ lies in the $Q$-cohomology
classically, there exists a $Q$-closed local operator $\theta$
associated to $c_1(M)$, which, under perturbative quantum corrections,
satisfies the relation
\begin{equation}
\label{QTt}
\{Q, T_{zz}\} = \partial_z\theta.
\end{equation}
If $c_1(M) = 0$, then $\theta$ is $Q$-exact and one can find quantum
corrections for $T_{zz}$ to make it $Q$-closed again. However, this
is not possible if $c_1(M) \neq 0$. Hence, the chiral algebra lacks
an energy-momentum tensor in this case.
In view of this, one may be surprised by the fact that even though the
chiral algebra has in general no conformal invariance, in a slightly
different variant of the theory---obtained by the so-called
``twisting'' procedure---there \emph{is} a grading by conformal weight
at the perturbative level. It was shown by Witten \cite{MR2320663}
that with this additional grading, the perturbative chiral algebra of
the twisted model is isomorphic to the cohomology of a sheaf
$\CD^{\mathit{ch}}_M$ of chiral differential operators, introduced by Malikov
et al.\ \cite{MR1704283}. In particular, when $c_1(M) = 0$, the
perturbative chiral algebra is conformally invariant and, by the
state-operator correspondence, its character computes the partition
function:%
\footnote{On the torus (or any Riemann surface with trivial canonical
bundle), the twisted and untwisted models are equivalent and have
the same partition function.}
\begin{equation}
\mathop{\mathrm{ch}}\nolimits H^*(M; \CD^{\mathit{ch}}_M) = Z(M;q).
\end{equation}
By the same token, Kapustin \cite{Kapustin:2005pt} showed that the
perturbative chiral algebra of the A-model is isomorphic to the
cohomology of a sheaf of chiral de Rham complex. The case in which
the theory is coupled to a general ``gauge bundle'' $E$ was considered
by Tan \cite{MR2302273} and found related to the construction of
Gorbounov et al.\ \cite{MR1995376}. The A-model corresponds to $E =
T_M$, the holomorphic tangent bundle of $M$.
Beyond perturbation theory, the gradings by R-charge and conformal
weight can be violated by instantons, which are holomorphic maps from
the spacetime manifold to $M$. Thus, these perturbative relations
between the chiral algebras and chiral differential operators may be
altered drastically.
\section{The Vanishing ``Theorem''}
In \cite{MR2320663}, Witten made a remarkable prediction. He claimed
that the chiral algebras of sigma models of $(0,2)$ supersymmetry,
though perturbatively infinite-dimensional, can nevertheless vanish
nonperturbatively in the presence of instantons. More specifically,
he predicted that this occurs for $M = \mathbb{CP}^1$ via an instanton-induced
relation
\begin{equation}
\label{Qt1}
\{Q, \theta\} \sim 1.
\end{equation}
In the case of $M = \mathbb{CP}^1$, the chiral algebra is $\mathbb{Z}/2\mathbb{Z}$-graded and
the operator $\{Q,\theta\}$, which has R-charge two perturbatively,
has chance to be equal to the identity operator $1$ nonperturbatively.
This relation shows that $1 = 0$ in the $Q$-cohomology and therefore
the chiral algebra is identically zero.
Witten's prediction was confirmed by an explicit computation in
\cite{MR2415553}, where this result was also extended to all complete
flag manifolds. In fact, one has a more general vanishing ``theorem''
which has been ``proved'' in \cite{Yagi:2010tp} at the physicists'
level of rigor:
\begin{"theorem"}[Yagi \cite{Yagi:2010tp}]
Let $M$ be a closed K\"ahler string manifold with $c_1(M) > 0$. If
$M$ contains a rational curve with trivial normal bundle, then the
chiral algebra of a $(0,2)$ supersymmetric sigma model with no
left-moving fermions vanishes for target space $M$.
\end{"theorem"}
The normal bundle $N_{C/M}$ of a rational curve $C \subset M$ is a
holomorphic vector bundle defined by the exact sequence
\begin{equation}
0 \longrightarrow T_C \longrightarrow T_M|_C \longrightarrow N_{C/M} \longrightarrow 0.
\end{equation}
The condition on the normal bundle says that
\begin{equation}
N_{C/M} \cong \CO^{\oplus d/2 - 1}.
\end{equation}
This ensures that the chiral algebra is $\mathbb{Z}/2\mathbb{Z}$-graded by R-charge.
We can only give the outline of the ``proof'' here. Suppose that we
have an instanton-induced relation $\{Q, \theta\} = \CO$. The local
operator $\CO$ is $Q$-closed, but need not be $Q$-exact in
perturbation theory, since we are assuming that this relation arises
only after instantons are taken into account. Let us look at the
R-charge zero, weight zero part of $\CO$. A local operator of
R-charge zero and weight zero is a function on $M$. Since $[\CO]$
varies holomorphically on the compact complex manifold $M$, this
function must be constant modulo $Q$-exact terms. Applying the
computation from the case of $M = \mathbb{CP}^1$, one sees that this constant
is nonzero on the rational curve $C$, hence on the whole manifold $M$.
The positive weight part of $\CO$, on the other hand, can be shown to
be all $Q$-exact. Thus, we again have the relation $\{Q,\theta\} \sim
1$, the equation $1 = 0$ holds in the $Q$-cohomology, and the chiral
algebra vanishes in the presence of instantons.
The intuitive picture behind the vanishing ``theorem'' is as follows.
When the normal bundle of $C$ is trivial, $C$ can be translated in
every direction in $M$ and we obtain a family of rational curves.
This family must sweep out the whole manifold $M$, and the
contributions to $\{Q,\theta\}$ from its members add up to a nonzero
constant. For example, in the case where $M$ is a flag manifold the
transitive group action generates such a family.
We are now ready to explain the relation of the vanishing phenomenon
to the H\"ohn--Stolz conjecture. It is based on the following
observation: the $Q$-cohomology of states vanishes if the chiral
algebra does.
The argument is very simple. Notice that the $Q$-cohomology of states
is naturally a module over the chiral algebra, with the action of
$[\CO]$ on $[\ket{\Psi}]$ defined by $[\CO] \cdot [\ket{\Psi}] =
[\CO\ket{\Psi}]$. If the chiral algebra vanishes, then $[1] = 0$ and
for any $Q$-closed state $\ket{\Psi}$ we have
\begin{equation}
[\ket{\Psi}]
= [1] \cdot [\ket{\Psi}]
= 0.
\end{equation}
Therefore, any $Q$-closed state is $Q$-exact; in other words, the
$Q$-cohomology of states vanishes.
Now, recall that the $Q$-cohomology of states is isomorphic to the
space of supersymmetric states. Therefore, when this vanishes the
partition function, which counts the number of bosonic supersymmetric
states minus that of fermionic ones at each level of $P$, vanishes as
well.
It would be interesting to see to what extent our vanishing
``theorem'' generalizes to other K\"ahler manifolds.
\section{Loop-Space Lichnerowicz}
A natural geometric framework for understanding the two-dimensional
supersymmetric sigma model with target space $M$ is the loop space
$\CL M$. Hence, the vanishing of the chiral algebra should have
implications for the geometry of this space. One such implication is
that $\CL M$ has no harmonic spinors. This is in accordance with the
original argument of Stolz, who interpreted the conjecture as arsing
from a loop-space analog of the Lichnerowicz theorem for
finite-dimensional spin manifolds with positive scalar curvature.
We have already explained that the partition function of the sigma
model with target space $M$ is the $S^1$-equivariant index of the
supercharge $Q_+$, and this is equal to the $S^1$-equivariant index of
the Dirac operator $\slashed{D}_{\CL M}$ on $\CL M$. When $M$ is K\"ahler,
one can actually say more: the space of supersymmetric states is
isomorphic to the space of harmonic spinors on $\CL M$. The reason is
that $Q$ is related to the degree one part $\CD_{\CL M}$ of $\slashed{D}_{\CL
M}$ by the conjugation
\begin{equation}
Q = e^{\CA_0} \CD_{\CL M} e^{-\CA_0},
\end{equation}
where $\CA_0$ is the symplectic action functional with Hamiltonian set
to zero. Thus, the two operators have an isomorphic cohomology.
The $Q$-cohomology of states vanishes if the chiral algebra does.
Therefore, the vanishing of the chiral algebra implies that there
exist no harmonic spinors on $\CL M$.
|
\section{Introduction}
\label{intro}
The perturbation of point spectra for classical vibration and quantum mechanical problems has a long history
\cite{rayleigh,rellich}. The more difficult problem of assessing the structural stability of the continuous spectrum
in e.g.\ scattering problems has also been widely investigated \cite{friedrichs,kato}. Because general linear Hamiltonian
systems are not governed by Hermitian or symmetric operators, the spectrum need not be stable and a transition to
instability is possible. For finite degree-of-freedom Hamiltonian systems, the situation is described by Krein's theorem
\cite{krein50,KJ80,moser58}, which states that a necessary condition for a bifurcation to instability
under perturbation is to have a collision between eigenvalues of opposite signature. The purpose of the present paper is to
investigate Krein-like phenomena in Hamiltonian systems with a continuous spectrum. Of interest are systems that describe
continuous media which are Hamiltonian in terms of noncanonical Poisson brackets (see e.g.\ \cite{morrison98,morrison05}).
Our study differs from that of \cite{grillakis}, which considered canonical Hamiltonian systems with continuous spectra in
a Hilbert space where the time evolution operator is self-adjoint. The effects of relatively compact
perturbations on such a system were studied and it was proved that the existence of a negative energy mode in the continuous
spectrum caused the system to be structurally unstable. It was also proved that such systems are otherwise structurally
stable. Also, our study differs from analyses of fluid theories concerning point spectra \cite{mackay,KM95} and point and continuous spectra \cite{hirota}, the latter using hyperfunction theory.
A representative example of the kind of Hamiltonian system of interest is the Vlasov-Poisson equation \cite{morrison80},
which when linearized about stable homogeneous equilibria gives rise to a linear Hamiltonian system with pure continuous
spectra that can be brought into action-angle form \cite{MP92,morrison94,MS94,morrison00}. A definition of signature
was given in these works for the continuous spectrum. In the present paper we concentrate on the Vlasov-Poisson equation,
but the same structure is possessed by Euler's equation for the two-dimensional fluid, where signature for shear flow
continuous spectra was defined \cite{BM98,BM02} and, indeed, a large class of systems \cite{morrison03}. Thus, modulo
technicalities, the behavior treated here is expected to cover a large class of systems.
In Sec.~\ref{ncham} we review on a formal level the noncanonical Hamiltonian structure for a class of systems that includes
the Vlasov-Poisson equation as a special case. Linearization about equilibria is described, the concept of dynamical
accessibility, and the linear Hamiltonian operator $T$, the main subject of the remainder of the paper, are defined.
In the remainder of the paper we sketch proofs, in varying levels of detail, pertaining to properties of this linear operator for various equilibria. In
Sec.~\ref{specstab} we describe spectral stability in general terms and analyze the spectrum of $T$ for the Vlasov case.
The existence of a continuous component to the spectrum is demonstrated and Penrose plots are used to describe the point component. In Sec.~\ref{structstab} we describe structural stability and, in particular, consider the structural stability
of $T$ under perturbation of the equilibrium state. We show that any equilibrium is unstable under perturbation of an
arbitrarily small function in $W^{1,1}$. In Sec.~\ref{secKM} we introduce the Krein-Moser theorem and restrict to dynamically
accessible perturbations. We prove that
equilibria without signature changes are structurally stable and those with changes are structurally unstable. In Sec.~\ref{kreinvp}
we define critical states of the linearized Vlasov equation that are structurally unstable under perturbations that are
further restricted. We prove that a mode with the opposite signature of the continuum is structurally unstable and that
the opposite combination cannot exist unless the system is already unstable. Finally, in Sec.~\ref{conclu}, we conclude.
\section{Noncanonical Hamiltonian form}
\label{ncham}
The class of equations of interest have a single dependent variable $\zeta(x,v,t)$, such that for each time $t$, $\zeta\colon\mathcal{D}\rightarrow\mathbb{R}$, where the particle phase space $\mathcal{D}$ is some two-dimensional domain with coordinates $(x,v)$. The dynamics is assumed to be Hamiltonian in terms of a noncanonical Poisson bracket of the form
\begin{equation}
\{F,G\}=\int_{\mathcal{D}}\!dxdv \, \zeta\,\left[\frac{\delta F}{\delta \zeta},\frac{\delta G}{\delta \zeta}\right]\,,
\label{ncze}
\end{equation}
where $[f,g]:=f_xg_v-f_vg_x$ is the usual Poisson bracket, where the subscripts denote partial differentiation, and $\delta F/\delta \zeta$ denotes the functional derivative of a functional $F[\zeta]$. The equation of motion is generated from a Hamiltonian functional $\mathcal{H}[\zeta]$ as follows:
\begin{equation}
\zeta_{t}=\{\zeta,\mathcal{H}\}=-[\zeta,\mathcal{E}]\,,
\end{equation}
where $\mathcal{E}:=\delta \mathcal{H}/\delta \zeta$. The Poisson bracket of (\ref{ncze}) is noncanonical: it uses only a single noncanonical variable $\zeta$, instead of the usual canonically conjugate pair; it possesses degeneracy reflected in the existence of Casimir invariants, $C=\int_{\mathcal{D}}\!dxdv\, \mathcal{C}(\zeta)$, that satisfy $\{F,C\}=0$ for all functionals $F$; but it does satisfy the Lie-algebraic properties of usual Poisson brackets. For further details see \cite{morrison82,holm85,morrison98,morrison03}.
For the Vlasov-Poisson equation we assume $\mathcal{D}=X\times\mathbb{R}$, where $X\subset\mathbb{R}$ or $X=S$, the circle, the distinction will not be important. The dependent variable is the particle phase space density $f(x,v,t)$ and the Hamiltonian is given by
\begin{equation}
\mathcal{H}[f]=\frac1{2}\int_{X}\!dx\!\int_{\mathbb{R}}\!dv\, v^2f + \frac1{2}\int_X \!dx\, |\phi_x|^2\,,
\end{equation}
where $\phi$ is a shorthand for the functional dependence on $f$ obtained through solution of Poisson's equation,
$\phi_{xx}=1-\int_{\mathbb{R}}\!fdv$, for a positive charge species with a neutralizing background. Using $\delta \mathcal{H}/\delta f=\mathcal{E}=v^2/2 +\phi$, we obtain
\begin{equation}
f_t=\{f,\mathcal{H}\}=-[f,\mathcal{E}]=-vf_x+\phi_x f_v \,,
\end{equation}
where, as usual, the plasma frequency and Debye length have been used to nondimensionalize all variables.
This Hamiltonian form for the Vlasov-Poisson equation was first published in \cite{morrison80}.
For a discussion of a general class of systems with this Hamiltonian form, to which the ideas of the present
analysis can be applied, see \cite{morrison03}. In a sequence of papers \cite{morrison87,MP92,MS94,SM94,morrison00,MS08}
various ramifications of the Hamiltonian form have been explored -- notably, canonization and diagonalization of the
linear dynamics to which we now turn.
Because of the noncanonical form, linearization requires expansion of the Poisson bracket as well as the Hamiltonian.
Equilibria, $\zeta_0$, are obtained by extremization of a free energy functional, $\mathcal{F}=\mathcal{H} + C$, as was first done for
Vlasov-like equilibria in \cite{KO58}. Writing $\zeta=\zeta_0 + \zeta_1$ and expanding gives the Hamiltonian form for the linear
dynamics
\begin{equation}
{\zeta_1}_t =\{\zeta_1,H_L\}_L
\end{equation}
where the linear Hamiltoinian, $H_L=\frac1{2}\int_{\mathcal{D}}\!dxdv\,\zeta_1\mathcal{O}\,\zeta_1$, is the second variation of $\mathcal{F}$, a
quadratic from in $\zeta_1$ defined by the symmetric operator $\mathcal{O}$, and $\{F,G\}_L=\int_{\mathcal{D}}\!dxdv\, \zeta_0[F_1,G_1]$
with $F_1:=\delta F/\delta \zeta_1$. Thus the linear dynamics is governed by the time evolution operator
$T\, \cdot :=-\{\ \cdot\ ,H_L\}_L=[\zeta_0,\mathcal{O}\, \cdot \, ]$.
Linearizing the Vlasov-Poisson equation about an homogeneous equilibrium,
$f_0(v)$, gives rise to the system,
\begin{eqnarray}
{f_1}_t&=&- v{f_1}_{x} + {\phi_1}_xf_0'
\label{f1t} \\
{\phi_1}_{xx} &=& - \int_{\mathbb{R}}\!dv\, f_1\,,
\label{ph1xx}
\end{eqnarray}
for the unknown $f_1(x,v,t)$. Here $f_0':=df_0/dv$. This is an infinite-dimensional linear Hamiltonian system
generated by the Hamiltonian functional:
\begin{align}
H_L[f_1] &= -\frac{1}{2}\int_X\!dx \int_{\mathbb{R}}\!dv \, \frac{v}{f_0'}\, |f_1|^2+\frac{1}{2}
\int_X\!dx\, |{\phi_1}_{x}|^2\,.
\end{align}
We concentrate on systems where $x$ is an ignorable coordinate, and either Fourier expand or transform. For Vlasov-Poisson this gives the system
\begin{equation}
{f_k}_t=- i kvf_k+\frac{if_0'}{k} \int_{\mathbb{R}}\!d\bar{v}\, f_k(\bar{v},t)=:-T_k f_k\,,
\label{fk}
\end{equation}
where $f_k(v,t)$ is the Fourier dual to $f_1(x,v,t)$. Perturbation of the specturm of the operator defined by Eq.~(\ref{fk}) is the primary subject of this paper. The operator $T_k$ is a Hamiltonian operator generated by the Hamiltonian functional
\begin{align}
H_L[f_k,f_{-k}] &= \frac{1}{2}\sum_k\left(-\int_{\mathbb{R}}\!dv\, \frac{v}{f_0'}\, |f_k|^2+
|\phi_k|^2\right)\,,
\end{align}
with the Poisson bracket
\begin{align}
\{F,G\}_L &= \sum_{k=1}^{\infty}ik\int_{\mathbb{R}}\!dv\, f_0'\left(\frac
{\delta F}{\delta f_k}\frac{\delta G}{\delta f_{-k}}-\frac{\delta F}{\delta f_{-k}}
\frac{\delta G}{\delta f_k}\right)\,.
\label{pbk}
\end{align}
Observe from (\ref{pbk}) that $k\in\mathbb{N}$ and thus $f_k$ and $f_{-k}$ are independent variables that are almost canonically conjugate. Thus the complete system is
\begin{equation}
{f_k}_t=-T_k f_k\qquad{\rm and}\qquad {f_{-k}}_t=-T_{-k} f_{-k}\,,
\label{Tsym}
\end{equation}
from which we conclude the spectrum is Hamiltonian.
\begin{lemma}
If $\lambda$ is an eigenvalue of the Vlasov equation linearized about the equilibrium $f_0'(v)$, then so are $-\lambda$ and $\overline{\lambda}$ (complex conjugate). Thus if $\lambda=\gamma +i\omega$, then eigenvalues occur in the pairs, $\pm\gamma$ and $\pm i\omega$, for purely real and imaginary cases, respectively, or quartets, $\lambda=\pm \gamma \pm i\omega$, for complex eigenvalues.
\end{lemma}
\label{specsym}
\begin{proof}
That $-\lambda$ is an eigenvalue follows immediately from the symmetry $T_{-k}= -T_{k}$, and that $\overline{\lambda}$ is an eigenvalue follows from $T_{k}f_k= -\overline{(T_{k}f_k)}$. $\Box$
\end{proof}
In \cite{MP92,MS94,morrison00,MS08} it was shown how to scale $f_k$ and $f_{-k}$ to make them canonically conjugate variables. In order to do this requires the following definition of dynamically accessibility, a terminology introduced in \cite{MP89,MP90}.
\begin{definition}
\label{da}
A particle phase space function $k$ is {\it dynamically accessible} from a particle phase space function $h$, if $g$ is an area-preserving rearrangement of $h$; i.e., in coordinates $k(x,v)=h(X(x,v),V(x,v))$, where $[X,V]=1$. A peturbation $\delta h$ is {\it linearly dynamically accessible} from $h$ if $\delta h=[G,h]$, where $G$ is the infinitesimal generator of the canonical transformation $(x,v)\leftrightarrow (X,V)$.
\end{definition}
Dynamically accessible perturbations come about by perturbing the particle orbits under the action of some Hamiltonian. Since electrostatic charged particle dynamics is Hamiltonian, one can make the case that these are the only perturbations allowable within the confines of Vlasov-Poisson theory.
Given an equilibrium state $f_0$, linear dynamically accessible perturbations away from this equilibrium state satisfy $\delta f_0=[G,f_0]=G_xf_0'$. Therefore assuming the initial condition for the linear dynamics is linearly dynamically accessible, we can define
\begin{equation}
q_k(v,t)=f_k\qquad{\rm and}\qquad p_k(v,t)=-i{f_{-k}}/(k{f_0'})
\label{pq}
\end{equation}
without worrying about a singularity at the zeros of $f_0'$ and $k=0$. With the definitions of (\ref{pq}), the Poisson bracket of (\ref{pbk}) achieves canonical from
\begin{equation}
\{F,G\}_L = \sum_{k=1}^{\infty} \int_{\mathbb{R}}\!dv \,
\left(\frac
{\delta F}{\delta q_k}\frac{\delta G}{\delta p_k}-\frac{\delta F}{\delta p_k}
\frac{\delta G}{\delta q_k}\right)\,.
\label{cpbk}
\end{equation}
The full system has the new Hamiltonian $\bar \mathcal{H} = \mathcal{H} + U\mathcal{P}$ in frame moving with speed $U$, where $\mathcal{P}= \int_{\mathcal{D}}\!dxdv\, vf$. Linearizing in this frame yields the linear Hamiltonian $\bar H_L= H_L + P_L$, from which we identify the linear momentum
\begin{equation}
P_L[f_k,f_{-k}] = \frac{1}{2}\sum^{\infty}_{k=1}\int_{\mathbb{R}}\!dv\, \frac{k}{f_0'}\, |f_k|^2\,,
\label{pl}
\end{equation}
which must be conserved by the linear dynamics. It is easy to show directly that this is the case.
\begin{lemma}
The momentum $P_L$ defined by (\ref{pl}) is a constant of motion, i.e., $\{P_L,H_L\}=0$.
\end{lemma}
\begin{proof}
This follows immediately from (\ref{Tsym}): $\int_{\mathbb{R}}\!dv\,\left( f_kT_{-k} + f_{-k}T_k\right)=0$.
$\Box$
\end{proof}
Observe, that like the Hamiltonian, $H_L$, the momentum $P_L$ is conserved for each $k$, which in all respects appears
only as a parameter in our system. Assuming the system size to be $L$ yields $k=2\pi n/L$ with $n\in \mathbb{N}$, and, thus,
this parameter can be taken to be in $\mathbb{R}^+/\{0\}$. Alternatively, we could suppose $X=\mathbb{R}$, Fourier transform, and split the
Fourier integral to obtain an expression similar to (\ref{pbk}) with the sum replaced by an integral over positive values
of $k$. For the present analysis we will not be concerned with issues of convergence for reconstructing the spatial
variation of $f_1(x,v,t)$, but only consider $k\in \mathbb{R}^+/\{0\}$ to be a parameter in our operator.
We will see in Sec.~\ref{specstab} that the operator $T_k$ possesses a continuous component to its spectrum.
But, we emphasize that this continuous spectrum of interest arises form the multiplicative nature of the velocity
operator, i.e.\ the term $vf_k$ of $T_k$, not from having an infinite spatial domain, as is the case for free
particle or scattering states in quantum mechanics. In the remainder of the paper, $f$ will refer to either $f_1$
or $f_k$, which will be clear from context, and the dependence on $k$ will be suppressed, e.g. in $T_k$, unless $k$
dependence is being specifically addressed.
\section{Spectral stability}
\label{specstab}
Now we consider properties of the evolution operator $T$ defined by (\ref{fk}).
We define spectral stability in general terms, record some properties of $T$, and describe
the tools necessary to characterize the spectrum of $T$. We suppose $f_k$ varies as $\exp(-i\omega t)$,
where $\omega$ is the frequency and $i\omega$ is the eigenvalue. For convenience we also use $u:=\omega/k$, where
recall $k\in \mathbb{R}^+$. The system is spectrally stable if the spectrum of $T$ is less than
or equal to zero or the frequency is always in the closed lower half plane. Since the system is Hamiltonian,
the question of stability reduces to deciding if the spectrum is confined to the imaginary axis.
\begin{definition}
\label{SpectralStability}
The linearized dynamics of a Hamiltonian system around some equilibrium solution, with the
phase space of solutions in some Banach space $\mathcal{B}$, is {\it spectrally stable} if
the spectrum $\sigma(T)$ of the time evolution operator $T$ is purely imaginary.
\end{definition}
Spectral stability does not guarantee that the system is stable, or that the equilibrium $f_0$ is linearly stable. (See e.g.\ \cite{morrison98} for general discussion). The solutions of a spectrally stable system are guaranteed to grow at most sub-exponentially and one can construct a spectrally stable system with polynomial temporal growth for certain initial conditions. (See e.g.\ \cite{degond86} for analysis of the Vlasov system.)
Spectral stability relies on functional analysis for its definition, since the spectrum of the operator $T$ may depend on the choice of function space $\mathcal{B}$. The time evolution operators arising from the types of noncanonical Hamiltonian systems that are of interest here generally contain a continuous spectrum \cite{morrison03} and the effects of perturbations that we study can be categorized by properties of the continuous spectrum of these operators. In general for the operators of \cite{morrison03}, the operator $T$ is the sum of a multiplication operator and an integral operator. In the Vlasov case, the multiplicative operator is i$v\cdot$ and the integral operator is $f_0'\int\!dv\, \cdot$. As we will see, the multiplication
operator causes the continuous spectrum to be composed of the entire imaginary axis except possibly for some discrete points.
Instability comes from the point spectrum. In particular, the linearized Vlasov Poisson equation is not spectrally stable when the time evolution operator has a spectrum that
includes a point away from the imaginary axis, with the necessary counterparts implied by Lemma \ref{ncham}.\ref{specsym}. For the operator $T$ this will always be a discrete mode; i.e.\ an eigenmode associated with an eigenvalue in the point spectrum.
\begin{theorem}
\label{SpecTheorem}
The one-dimensional linearized Vlasov-Poisson system with homogeneous equilibrium $f_0$ is spectrally unstable
if for some $k\in\mathbb{R}^+$ and $u$ in the upper half plane, the plasma dispersion relation
\[
\varepsilon(k,u):=1-k^{-2}\int_{\mathbb{R}}\!dv
\frac{f_0'}{v-u}=0\,.
\]
Otherwise it is spectrally stable.
\end{theorem}
\begin{proof}
The details of this proof are given in plasma textbooks. It follows directly from (\ref{f1t}) and (\ref{ph1xx}), and the assumption $f_1\sim\exp(ikx-i\omega t)$.
$\Box$
\end{proof}
Using the Nyquist method that relies on the argument principle of complex analysis, Penrose \cite{penrose} was able to relate the vanishing of $\varepsilon(k,u)$ to the winding number of the closed curve determined by the real and imaginary parts of $\varepsilon$ as $u$ runs along the real axis. Such closed curves are called Penrose plots. The crucial quantity is the integral part of $\varepsilon$ as $u$ approaches the real axis from above:
\[
\lim_{u\rightarrow 0^+}\frac{1}{\pi}\int_{\mathbb{R}}\!dv\, \frac{f_0'}{v-u}=H[f_0'](u)-if_0'(u)\,,
\]
where $H[f_0']$ denotes the Hilbert transform, $H[f_0']=\frac{1}{\pi}\Xint-\!dv\,{f_0'}/({v-u})$, where $\Xint-:=PV\int_{\mathbb{R}}$ indicates the Cauchy principle value. (See \cite{king} for an in depth treatment of Hilbert transforms.) The graph of the real line under this mapping is the essence of the Penrose plot, and so we will refer to these closed curves as Penrose plots as well. When necessary to avoid ambiguity we will refer to the former as $\varepsilon$-plots.
For example, Fig.~\ref{gaus1} shows the derivative of the distribution function, $f_0'$, for the case of a Maxwellian distribution
and Fig.~\ref{penroseGaus} shows the contour $H[f_0']-if_0'(u)$ that emerges from the origin in the complex plane at $u=-\infty$, descends, and then wraps around to return to the origin at $u=\infty$. From this figure it is evident that the winding number of the $\varepsilon(k,u)$-plot is zero for any fixed $k\in\mathbb{R}$, and as a result there are no unstable modes.
Making use of the argument principle as described above, Penrose obtained the following criterion:
\begin{theorem}
\label{PenroseCriterion}
The linearized Vlasov-Poisson system with homogeneous equilibrium $f_0$ is spectrally unstable if there exists a point $u$ such that
\[
f_0'(u)=0 \qquad {\rm and}\qquad \Xint-\!dv\, \frac{f_0'(v)}{v-u}>0\,,
\]
with $f_0'$ traversing zero at $u$. Otherwise it is spectrally stable.
\end{theorem}
\begin{figure}[htbp]
\includegraphics[scale=.66]{s1gaussian.pdf}
\caption{$f_0'$ for a Maxwellian distribution.}
\label{gaus1}
\end{figure}
\begin{figure}[htbp]
\includegraphics[scale=.6]{s1Hgaussian.pdf}
\caption{Stable Penrose plot for a Maxwellian distribution.}
\label{penroseGaus}
\end{figure}
Penrose plots can be used to visually determine spectral stability. As described above, the Maxwellian distribution
$f_0=e^{-v^2}$ is stable, as the resulting $\varepsilon$-plot does not encircle the origin. However, it is not difficult to
construct unstable distribution functions. The superposition of two displaced Maxwellian distributions,
$f_0=e^{-(v+c)^2}+e^{-(v-c)^2}$, is such a case.
As $c$ increases the distribution goes from stable to unstable. Figures \ref{2max} and \ref{uspen} demonstrate how the transition from stability to instability is manifested in a Penrose plot. The two examples are $c=3/4$ and $c=1$.
(Note, the normalization of $f_0$ only affects the overall scale of the Penrose plots and so is ignored for convenience.) It is evident from Fig.~\ref{uspen} that for some $k\in\mathbb{R}$ the $\varepsilon$-plot (which is a displacement of the curve shown by multiplying by $-k^{-2}$ and adding unity) will encircle the origin, and thus will be unstable for such $k$-values.
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.6]{s2stablepenrose.pdf}
\end{center}
\caption{Penrose plot for a stable superposition of Maxwellian distributions}
\label{2max}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.6]{s2unstablepenrose.pdf}
\end{center}
\caption{The unstable Penrose plot corresponding to two separated Maxwell distributions.}
\label{uspen}
\end{figure}
We now are positioned to completely determine the spectrum. For convenience we set $k=1$ when it does not affect the essence of our arguments, and consider the operator $T\colon f\mapsto ivf- if_0'\int f$ in the space
$W^{1,1}(\mathbb{R})$, but we also discuss the space $L^1(\mathbb{R})$. The space $W^{1,1}(\mathbb{R})$ is the Sobolev space containing
the closure of functions under the
norm $\|f\|_{1,1}=\|f\|_1+\|f'\|_1$. Thus it contains all functions that are in $L^1(\mathbb{R})$ whose weak
derivatives are also in $L^1(\mathbb{R})$. First we establish the expected facts that $T$ is densely defined and closed.
In $W^{1,1}$ the operator $T$ is the sum of the multiplication operator and a bounded operator -- that it is densely defined and
closed follows from the fact that the multiplication operator is densely defined and closed in these spaces, where
\[
D_1(T):= \{f|vf\in W^{1,1}(\mathbb{R})\}.
\]
\begin{theorem}
The operator $T\colon W^{1,1}(\mathbb{R})\rightarrow W^{1,1}(\mathbb{R})$ with domain $D_1(T)$ is both
(i) densely defined and (ii) closable.
\end{theorem}
\begin{proof}
(i) The set of all smooth functions with compact support, $C_c^{\infty}(\mathbb{R})$ is a subset of $D_1$.
This set is dense in $W^{1,1}(\mathbb{R})$ so $D_1$ is dense and $T$ is densely defined.
(ii) The operator $T$ is closable if the operator $v$ is closable because $T$ and $v$ differ by a bounded operator.
The multiplication operator $v$ is closed if for each sequence $f_n\subset W^{1,1}(\mathbb{R})$ that converges to
$0$ either $vf_n$ converges to $0$ or $vf_n$ does not converge. Suppose $vf_n$ converges. At each point
$f_n$ converges to $0$. Therefore $vf_n$ converges to $0$ at each point, so $vf_n$ converges to $0$ if it converges.
$\Box$
\end{proof}
Therefore there exists some domain $D$ such that the graph $(D,TD)$ is closed.
\medskip
In determining the spectrum of the operator $T$, denoted $\sigma(T)$, we split
the spectrum into point, residual, and continuous components as follows:
\begin{definition}
\label{spectra}
For $\lambda\in\sigma(T)$ the resolvent of $T$ is $R(T,\lambda)=(T-\lambda I)^{-1}$, where $I$ is the identity operator. We say $\lambda$
is (i) in the point spectrum, $\sigma_p(T)$, if $T-\lambda I$ fails to be injective, (ii) in the
residual spectrum, $\sigma_r(T)$, if $R(T,\lambda)$ is not densely defined, and (iii) in the continuous spectrum,
$\sigma_c(T)$, if $R(T,\lambda)$ is densely defined but unbounded.
\end{definition}
Using this definition we characterize the spectrum of the operator $T$.
\begin{theorem}
\label{spectrumT}
The component $\sigma_p(T)$ consists of all points $\lambda=iu\in \mathbb{C}$ where $1-k^{-2}\int_{\mathbb{R}}\!dv\,{f_0'}/({v-u})=0$, $\sigma_c(T)$
consists of all $\lambda=i u$ with $u\in\mathbb{R}/(-i
\sigma_p(T)\cap\mathbb{R})$, and $\sigma_r(T)$ contains all points $\lambda=iu$ in the complement of $\sigma_p(T)\cup \sigma_c(T)$ that satisfy $f_0'(u)=0$.
\end{theorem}
\begin{proof}
By the Penrose criterion we can identify all the points in the
point spectrum. If
$1-k^{-2}\int_{\mathbb{R}} \,dv{f_0'}/({v-u})=0$ then $iu=\lambda\in\sigma_p(T)$.
Because the system is Hamiltonian these modes will occur for the linearized Vlasov-Poisson system in quartets (two for $T_k$ and two for $T_{-k}$), as follows from Lemma \ref{ncham}.\ref{specsym}.
It is possible for there to be discrete modes with real frequencies and these will occur in pairs. If
for real $u$ the map $u\mapsto \varepsilon$ passes through the origin then there will be such an embedded mode.
For convenience we drop the wavenumber subscript $k$ on $f_k$ and add the subscript $n$ to identify $f_n$ as an element of a sequence of functions that converges to zero with, for each $n$, support contained
in an interval of length $2\epsilon(n)$ surrounding the point $u$ and zero average value. Let $u\in\mathbb{R}$ and choose the sequence $\{f_n\}$ so that $\epsilon(n)\rightarrow 0$. Then for each $n$
\begin{align}
\|R(T, iu)\| &\geq \frac{\|f_n\|_{1,1}}{\|(v-u)f_n\|_{1,1}}\nonumber \\
&\geq \frac{\|f_n\|_{1,1}}{\|v-u\|_{W^{1,1}(u-\epsilon,u+\epsilon)}\|f_n\|_{1,1}} \nonumber\\
&=\frac{1}{\|v-u\|_{W^{1,1}(u-\epsilon,u+\epsilon)}}
\nonumber
\end{align}
In the above expression, $W^{1,1}(u-\epsilon,u+\epsilon)$ refers to the integral of $|f|+|f'|$ over the interval $(u-\epsilon,u+\epsilon)$.
Therefore the resolvent is an unbounded operator and $iu=\lambda$ is in the spectrum. If the frequency $u$ has an imaginary component $i\gamma$ then
$\|R(T,iu)\|<1/\gamma$ so unless $iu=\lambda$ is part of the point spectrum it is part of the resolvent set.
The residual spectrum of $T$ is contained in the point spectrum of $T^*$. The dual of $W^{1,1}$ is the space $W^{-1,1}$ defined by pairs $(g,h)\in W^{-1,1}$ with
$\|(g,h)\|_{{-1,1}}<\infty$ (cf.\ \cite{adams}). The operator $T^*(g,h)= i(vg-h+\int (gf_0'-hf_0'')dv,-vh)$ is the adjoint of $T$. If we search for a
member $iu=\lambda$ of the point spectrum we get two equations, one of which is $(v-u)h=0$. This forces $h=0$ because $h$ cannot be a $\delta$-function in $W^{-1,1}$.
The other equation is then $(v- u)g+\int gf_0'dv=0$ which
can only be true if the integral is zero or if $(v- u)g$ is a constant. For this $g=\frac{1}{v- u}$ and the resulting equation for $u$ is the
same equation as that for the frequency of the point modes of $T$. If the integral is zero then $g=\delta(v-u)$ is a solution when $f_0'(u)=0$. Therefore the residual
spectrum contains the points $\lambda=iu$ satisfying $f_0'(u)=0$.
$\Box$
\end{proof}
This characterization of the spectrum fails in Banach spaces with less regularity than $W^{-1,1}$, such as $L^p$ spaces, because the Dirac $\delta$ is not contained in the dual space. In this case the
residual spectrum vanishes because $\sigma_p(T^*)=\sigma_p(T)$. This calculation is nearly identical to that of Degond \cite{degond86}, who characterizes the residual spectrum slightly differently than
we do. In any event, the result is that the Penrose criterion determines whether $T$ is spectrally stable or not. If the winding number of the $\varepsilon$-plot is positive, then there is spectral instability and if it
is zero there is spectral stability.
\section{Structural stability}
\label{structstab}
Spectral stability characterizes the linear dynamics of a nonlinear Hamiltonian system
in a neighborhood of an equilibrium. The main question now is to determine when a spectrally
stable system can be made spectrally unstable with a small perturbation. When this is impossible for our choice of allowed perturbations,
we say the equilibrium is structurally stable, and when there is an infinitesimal perturbation
that makes the system spectrally unstable we say that the equilibrium is spectrally unstable.
We can make this more precise by stating it in terms of operators on a Banach space.
\begin{definition}
\label{genstructuralstability}
Consider an equilibrium solution of a Hamiltonian system and the corresponding
time evolution operator $T$ for the linearized dynamics, with a phase space some
Banach space $\mathcal{B}$. Suppose that $T$ is spectrally stable. Consider perturbations
$\delta T$ of $T$ and define a norm on the space of such perturbations. Then we say that the
equilibrium is {\it structurally stable} under this norm if there
is some $\delta>0$ such that for every $\|\delta T\|<\delta$ the operator $T+\delta T$ is spectrally stable.
Otherwise the system is {\it structurally unstable}.
\end{definition}
Because we are dealing with physical systems
it makes sense to have physical motivation for the choice of norm on the space of
perturbations. In this paper we
are interested in perturbations of the Vlasov equation through changes in the equilibrium. This
choice is motivated by the Hamiltonian structure of the equations and Krein's theorem for finite-dimensional systems. In general the space
of possible perturbations is quite large, but perturbations of equilibria give rise to operators in certain Banach spaces and motivate the definition of norm. Even in the case of unbounded perturbations there may exist such a norm
(see Kato \cite{kato}, for instance).
Consider a stable equilibrium function $f_0$. We will consider perturbations of the
equilibrium function and the resulting perturbation of the time evolution operator. Suppose that the time
evolution operator of the perturbed system is $T+\delta T$.
In the function space that we will consider these perturbations are bounded operators
and their size can be measured by the norm $\|\delta T\|$. This norm will be proportional
to the norm of $\|\delta f_0'\|$, where $\delta f_0$ is the
perturbation of the equilibrium.
\begin{definition}
\label{StructuralStability}
Consider the formulation of the linearized Vlasov-Poisson equation in the Banach
space $W^{1,1}(\mathbb{R})$ with a spectrally stable homogeneous equilibrium function $f_0$. Let $T_{f_0+\delta f_0}$
be the time evolution operator corresponding to the linearized dynamics around the distribution
function $f_0+\delta f_0$. If there exists some $\delta$ depending only on $f_0$ such
that $T_{f_0+\delta f_0}$ is spectrally stable whenever $\|\delta T_{\delta f_0}\|=\|T_{f_0}-T_{f_0+\delta f_0}\|<\delta$, then the equilibrium $f_0$
is structurally stable under perturbations of $f_0$.
\end{definition}
The aim of this work is to characterize the structural stability of the linearized
Vlasov-Poisson equation.
We will prove that if the perturbation function is some homogeneous $\delta f_0$ and the norm
is $W^{1,1}$ (and $L^1$ as a consequence) every equilibrium distribution function is structurally unstable
to an infinitesimal perturbation in this space. This fact will force us to consider more restricted sets of
perturbations.
\subsection{Winding number}
We need to compute the winding number of Penrose plots and the change in winding number
under a perturbation, both in this section and the rest of the paper.
We use the fact that one way to compute the winding number is to draw a ray from the origin to infinity and
to count the number of intersections with the contour accounting for orientation.
\begin{lemma}
\label{winding}
Consider an equilibrium distribution function $f_0'$. The winding number of the
Penrose $\varepsilon$-plot around the origin is equal to $\sum_u {\rm sgn}(f_0''(u))$ for all
$u\in\mathbb{R}^-$, satisfying $f_0'(u)=0$.
\end{lemma}
To calculate the winding number of the Penrose $\varepsilon$-plot using this lemma one counts the number of zeros of
$f_0'$ on the negative real line and adds them with a positive sign if $f_0''$ is positive,
a Penrose crossing from the upper half plane to the lower half plane, a negative sign if
$f_0''$ is negative, a crossing from the lower half plane to the upper
half plane, and zero if $u$ is not a crossing of the x-axis, a tangency.
This lemma comes from the following equivalent characterization of the
winding number from differential topology \cite{guillemin}.
\begin{definition}
\label{diffwinding}
If $X$ is a compact, oriented, $l$-dimensional manifold and $f\colon X\rightarrow\mathbb{R}^{l+1}$
is a smooth map, the winding number of $f$ around any point $z\in\mathbb{R}^{l+1}-f(X)$
is the degree of the direction map $u\colon X\rightarrow S^l$ given by $u(x)=\frac{f(x)-z}{|f(x)-z|}$.
\end{definition}
In our case the compact manifold is the real line plus the point at $\infty$ and $l=1$.
The degree of $u$ is the intersection number of $u$ with any point on the circle taken
with a plus sign if the differential preserves orientation and a minus sign if it reverses
it. The lemma is just a specialization of this definition to the negative $x$-direction on the circle. If more than one derivative of $f_0$ vanishes at a zero of $f_0'$ there is a standard procedure for calculating the winding number by determining if there is a sign change in
$f_0'$ at the zero.
\subsection{Structural instability of general $f_0$}
In a large class of function spaces it is possible to
create infinitesimal perturbations that make any equilibrium distribution function unstable. This can happen
in any space where the Hilbert transform is an unbounded operator. In these spaces
there will be an infinitesimal $\delta f_0$ such that $H[\delta f_0']$ is order one at a zero of $f_0'$. Such a
perturbation can turn any point where $f_0'=0$ into a point where $H[f_0'+\delta f_0']>0$
as well. Because $\delta f_0'$ is small and the region where $H[\delta f_0']$ is not small is also small, the only
effect on the Penrose plot will be to move the location of the zero. Thus, such a perturbation will increase the winding number and cause instability.
We will explicitly demonstrate this for the Banach space $W^{1,1}(\mathbb{R})$ and,
by extension, the Banach space $L^1\cap C_0$. This will imply that any distribution
function is infinitesimally close to instability when the problem is set in one
of these spaces, implying the structural instability of every distribution function.
Suppose we perturb $f_0$ by a function $\delta f_0$. The resulting perturbation to
the operator $T$ is the operator mapping $f$ to $\delta f_0'\int dv f$. In the space
$W^{1,1}$ this is a bounded operator and thus we take the norm of the perturbing operator
to be $\|\delta f_0'\|_{1,1}$. Now we introduce a class of perturbations that can be
made infinitesimal, but have Hilbert transform of order unity.
Consider the function
$\chi(v,h,d,\epsilon)$ defined by
\[
\chi =
\left\{
\begin{array}{ll}
{hv}/{\epsilon} &\quad |v|<\epsilon \\
h\, {\rm sgn}(v) &\quad \epsilon<|v|<d+\epsilon \\
h+d/2+\epsilon/2-v/2 &\quad 2h+d+\epsilon>v>d+\epsilon \\
-h-d/2-\epsilon/2-v/2 &\quad 2h+d+\epsilon>-v>d+\epsilon \\
0 &\quad |v|>2h+d+\epsilon
\end{array}\,.
\right.
\]
Figures \ref{chi} and \ref{Hchi} show the graph of $\chi$ and its Hilbert transform, $H[\chi]$, respectively.
\bigskip
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.65]{chi}
\end{center}
\caption{The perturbation $\chi$ for $\epsilon=e^{-10}$, $h=d=.1$.}
\label{chi}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.85]{Hchi1}
\end{center}
\caption{The Hilbert transform of $\chi$.}
\label{Hchi}
\end{figure}
\begin{lemma}
\label{chi_lemma}
If we choose $d=h$ and $\epsilon=e^{-(1/h)}$, then for any $\delta,\gamma>0$ we can choose
an $h$ such that $\|\chi\|_{1,1}<\delta$ and $\Xint-\!dv\, {\chi}/{v}>1-O(h)$,
and $|\Xint-\!dv\, {\chi}/({u-v})|<|\gamma/u|$ for $|u|>|2h+d+\epsilon|$.
\end{lemma}
\begin{proof}
In the space $W^{1,1}$ the function $\chi$ has norm $2h^2+2hd+h\epsilon + 4h$, which is less than any $\delta$ for small enough $h$.
We can compute the value of the Hilbert transform of this
function at a given point $u$ by calculating the principal values:
\begin{eqnarray}
\Xint-\!dv\, \frac{\chi}{v-u}&=& \frac{hu}{\epsilon}\log\left(\frac{|u-\epsilon|}{|u+\epsilon|}\right)
+h\log\left(\frac{|d+\epsilon-u||u+d+\epsilon|}{|\epsilon+u||\epsilon-u|}\right)\nonumber\\
&+& \frac{1}{2}(d+\epsilon+2h-u)\log\left(\frac{|d+\epsilon+2h-u|}{|d+\epsilon-u|}\right)
\nonumber\\
&+&
\frac{1}{2}(d+\epsilon+2h+u)\log\left(\frac{|d+\epsilon+2h+u|}{|d+\epsilon+u|}\right)\,.
\label{chiint}
\end{eqnarray}
We analyze the asymptotics of this function as $h$, $d$, and
$\epsilon$ go to zero, with the desiderata that i) the norm of $\chi$
goes to zero, ii) the maximum of the Hilbert transform of $\chi$ is $O(1)$, and iii) there
is a band of vanishing width around the origin outside of which the Hilbert
transform can be made arbitrarily close to zero.
Note that (\ref{chiint}) can be written as a linear combination of translates of
the function $x\log x$:
\begin{eqnarray}
\Xint-\!dv\, \frac{\chi}{v-u} &=&
\frac{h}{\epsilon}((u-\epsilon)\log(|u-\epsilon|)
- (u+\epsilon)\log(|u+\epsilon|))
\nonumber\\
&-& \frac{1}{2}(d+u+\epsilon)\log(|d+u+\epsilon|) -\frac{1}{2}(d-u+\epsilon) \log(|d-u+\epsilon|)
\nonumber\\
&+&\frac{1}{2}(d+u+\epsilon+ 2h)\log(|d+u+\epsilon+2h|)
\nonumber\\
&+&\frac{1}{2}(d-u+\epsilon+2h) \log(|d-u+\epsilon+2h|)\,.
\label{chiagain}
\end{eqnarray}
The function $x\log x$ has a local minimum for positive $x$ at $x=1/e$.
This is the point at which the function is most negative. It has zeros at $x=0$
and $x=1$. For values of $u,d,\epsilon,h$ close to zero all of the arguments
of the $\log$ functions are less than $1/e$. Therefore, for $|u|<d+\epsilon+2h$
the $x\log x$ terms are all monotonically decreasing functions of the
argument $x$. Of the terms of (\ref{chiagain}), $\frac{h}{\epsilon}((u-\epsilon)\log(|u-\epsilon|)
-(u+\epsilon)\log(|u+\epsilon|))$ has by far the largest coefficient as long as
$\epsilon$ is much smaller than $h$. We choose $h=d$ and $\epsilon=0(e^{-1/h})$. Then the terms that
do not involve $\epsilon$ are all smaller than $(6h+\epsilon)\log (6h+\epsilon)$. With these choices $\chi$ satisfies
\begin{eqnarray}
\chi(0)&=&2-(h+e^{-1/h})\log(|h+e^{-1/h}|)+(3h+e^{-1/h})\log(|3h+e^{-1/h}|)\nonumber\\
&=&2+O(h\log h)\,.
\nonumber
\end{eqnarray}
Consider the pair of functions $-(u+c)\log(|u+c|)+(u-c)\log(|u-c|)$. The derivative
with respect to $u$ is $-\log(|u+c|)+\log(|u-c|)$. This is zero for $u=0$ and for $u>0$ it is always negative and the pair is
always decreasing, and for small values of $h$ the pair is guaranteed to be positive. Suppose that $u>\epsilon$.
Then we can bound the term with the $h/\epsilon$ coefficient:
\begin{eqnarray}
&{\ }&\frac{h}{\epsilon}\left|(u-\epsilon)\log(|u-\epsilon|)-(u+\epsilon)\log(|u+\epsilon|)\right| \nonumber\\
&{\ }& \hspace{1.0 in} =\left|\frac{h}{\epsilon}(u-\epsilon)\log\frac{|u-\epsilon|}{|u+\epsilon|}-2\epsilon\log(|u+\epsilon|)\right| \nonumber\\
&{\ }& \hspace{1.0 in} = \frac{h}{\epsilon}\left|(u-\epsilon)\log\frac{1-\frac{\epsilon}{u}}{1+\frac{\epsilon}{u}}-2\epsilon\log(|u+\epsilon|)\right|\nonumber \\
&{\ }& \hspace{1.0 in} < \frac{h}{\epsilon}\left|(u-\epsilon)\log(e^{-\epsilon/u})\right|+2\left|h\log(|u+\epsilon|)\right| \nonumber\\
&{\ }&\hspace{1.0 in} = \frac{h(u-\epsilon)}{u}+2\left|h\log(|u+\epsilon|)\right|\,.
\nonumber
\end{eqnarray}
For $u>>\epsilon$, for example if $u=O(h^2)$, this term is $O(h\log h)$. Therefore, for $|u|>h^2$ we have $\chi=O(h\log h)$ which
can be made arbitrarily small. When $|u|>3h+\epsilon$ the function $\chi$ decreases at least as fast as $O(1/u)$. With these choices of
$h$, $d$, and $\epsilon$, the norm of $\chi$ is $O(h)$, which proves the Lemma.
$\Box$
\end{proof}
Now we state the theorem that any equilibrium is strucutrally unstable in both
the spaces $W^{1,1}$ and $L^1\cap C_0$.
\begin{theorem}
\label{structuralinstability}
A stable equilibrium distribution $f_0\in C^2$ is structurally unstable under perturbations of the equilibrium in the Banach spaces $W^{1,1}$ and $L^1\cap C_0$.
\end{theorem}
\begin{proof}
If $f_0$ is stable then the Penrose $\epsilon$-plot of $f_0'$ has a winding number of zero. Because the point at $\infty$ corresponds
to a crossing where $f_0'$ goes from negative to positive there exists a point $u_0$ with $f_0'(u_0)=0$ that is an isolated zero,
$H[f_0'](u_0)<0$, and $f_0''(u_0)<0$. Let $F=\sup |f_0''|$. Choose $h$ to always be smaller than the distance from $u_0$ to
the nearest $0$ of $f_0'$. Then if $\epsilon=O(e^{-1/h})$ and $d=h$ the support of $\chi(u-u_0)$ will contain only one zero of $f_0'$.
For $h$ small enough the slope of $\chi$ at $u_0$ will be greater than $F$ so that the function $f_0'+\chi$ will be positive for $u$
in the set $(u_0,u^+)$ for some $u^+$ in the support of $\chi$. Similarly $f_0'+\chi$ will be negative for $u$ in the set $(u^-,u_0)$
for some $u^-$ in the support of $\chi$. Because $\chi$ has compact support the function $f_0'+\chi$ is positive in a neighborhood
outside of the support of $\chi$ so that the intermediate value theorem guarantees one additional zero of the function $f_0'+\chi$ for
$u>u_0$ and also for $u<u_0$. Choose $\chi$ so that this Hilbert transform of $f_0'+\chi$ is positive at the point $u_0$
and $h$ small enough that it is negative before the next zero of $f_0'+\chi$ on either side of $u_0$. Then the winding number of
$f_0'+\chi$ is positive because an additional positive crossing has been added on the negative real line.
Because the norm of $\chi$ is $O(h)$ in both $W^{1,1}$ and $L^1$ the distribution $f_0$ is unstable to an arbitrarily small
perturbation and is therefore structurally unstable.
$\Box$
\end{proof}
Thus we emphasize that we can always construct a perturbation
that makes our linearized Vlasov-Poisson system unstable.
For the special case of the Maxwellian distribution, Fig.~\ref{pert} shows the perturbed derivative of the distribution function and Fig.~\ref{pertpen} shows the Penrose plot of the unstable perturbed system. Observe the two crossings created by the perturbation on the positive axis as well as the negative crossing arising from the unboundedness of the perturbation.
In some sense Theorem \ref{structuralinstability} represents a failure
of our class of perturbations to produce any interesting structure for the Vlasov
equation. Indeed signature appears to play no role in delineating bifurcation to instability. In order to derive a nontrivial result we develop a new theory analogous to the finite-dimensional Hamiltonian perturbation theory developed by Krein
and Moser. This new theory involves a restriction to dynamically accessible perturbations of the equilibrium state. This is natural since the noncanonical Hamiltonian structure can be viewed as the union of canonical Hamiltonian motions (on symplectic leaves) labelled by the equilibrium state -- to compare with traditional finite-dimensional theory requires restriction to the given canonical Hamiltonian motion under consideration.
\bigskip
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.6]{chimaxwell}
\end{center}
\caption{$f_0'+\chi$ for a Maxwellian distribution.}
\label{pert}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.6]{pengausschi}
\end{center}
\caption{Penrose plot for perturbed Maxwellian.}
\label{pertpen}
\end{figure}
\section{The Krein-Moser theorem}
\label{secKM}
For linear finite-dimensional Hamiltonian systems, Hamilton's equations are a set
of first order linear ODEs. If the Hamiltonian is time-independent, then
the behavior of solutions is characterized by the eigenfrequencies.
If all of the eigenfrequencies are on the real axis and nondegenerate, then the system will be stable.
If there are degenerate eigenvalues the system will be stable as long as the time evolution operator
does not have any nontrivial Jordan blocks, but there will be secular growth if it does. Any complex
eigenfrequencies will lead to instability. The Hamiltonian of a linear finite-dimensional Hamiltonian system
is a quadratic form in the canonical variables. If we consider perturbations of the
coefficients of the quadratic form it is trivial to define a notion of small perturbations,
as the resulting perturbation of the Hamiltonian will be a bounded operator. Krein and Moser
independently proved a theorem characterizing the structural stability of these systems in
terms of a signature, a quantity that amounts to the sign of the energy evaluated on the eigenvector of a mode.
We consider a special case of their theorem.
\begin{theorem}{\rm (Krein-Moser)}
\label{KreinMoser}
Let $H$ define a stable linear finite-dimensional Hamiltonian system. Then $H$ is structurally
stable if all the eigenfrequencies are nondegenerate. If there are any degeneracies, $H$ is
structurally stable if the assosciated eigenmodes have energy of the same sign. Otherwise $H$ is
structurally unstable.
\end{theorem}
This Krein-Moser theorem gives a clear picture of the behavior of these systems under
small perturbations. The eigenfrequencies move around, but remain confined to the real
line unless there is a collision between a positive energy and negative energy mode, in which case they may leave the axis.
This theorem was first proved by Krein in the early 1950s and later rediscovered by
Moser in the late 1950s. Our goal is to place the perturbation theory of infinite-dimensional
Hamiltonian systems in the language of the finite-dimensional theory.
The appropriate definition of signature for the continuous spectrum of the Vlasov-Poisson
equation was introduced in \cite{MP92,morrison00} (see also \cite{MS08}), where an integral transform was also introduced for constructing a canonical transformation to action-angle variables for the infinite-dimensional system. The transformation is a generalization
of the Hilbert transform and it can be used to show that the linearized Vlasov-Poisson
equation is equivalent to the system with the following Hamiltonian functional:
\begin{align}
&H_{L}=\sum_{k=1}^{\infty}\int_{\mathbb{R}}\!du\, \sigma_k(u)\omega_k(u)J_k(u,t)\,,
\end{align}
where $\omega_k(u)=|ku|$ and $\sigma_k(u)=-{\rm sgn}(kuf_0'(u))$ is the analog of the Krein signature corresponding
to the mode labeled by $u\in\mathbb{R}$. (Note, the transformation can always be carried out in a frame where $f_0'(0)=0$. Because the Hamiltonian does not transform as a scalar for frame shifts, which are time dependent transformations, signature is frame dependent. The Hamiltonian in a shifted frame is obtained by adding a constant times the momentum $P_L$ of (\ref{pl}) to $H_L$. Later we will see that Hamiltonians that can be made sign definite is some frame are structurally stable in a sense to be defined.)
\newtheorem{signature}{Definition}
\begin{signature}
Suppose $f_0'(0)=0$. Then the signature of the point $u\in\mathbb{R}$ is
$-{\rm sgn}(uf_0'(u))$.
\end{signature}
Below is a graph that illustrates the signature for a bi-Maxwellian distribution
function.
\bigskip
\begin{figure}[htbp]
\hspace{ .2 in}
\includegraphics[scale=.9]{bimaxwelliansignature.pdf}
\caption{Signature for a bi-Maxwellian distribution function.}
\label{bigaus}
\end{figure}
\subsection{Dynamical accessibility and structural stability}
Now we discuss the effect of restricting to dynamically accessible
perturbations on the structural stability of $f_0$. In this work we only study perturbations
of $f_0$ that preserve homogeneity. Because dynamically accessible perturbations are
area preserving rearrangements of $f_0$, it is impossible to construct a dynamically accessible
perturbation for the Vlasov equation in a finite spatial domain that preserves homogeneity.
To see this we write a rearrangement as $(x,v)\leftrightarrow (X,V)$, where $V$ is a function of
$v$ alone. Because $[X,V]=1$ and $V(v)$ is not a function of $x$, we have
$V'{\partial X}/{\partial x}=1$, or $X={x}/{V'}$. If the spatial domain is finite,
this map is not an diffeomorphism unless $V'=1$.
In the infinite spatial domain case, this is not a problem and these rearrangments
exist. First we note that a rearrangement cannot change the critical points of $f_0$.
\begin{lemma}
\label{criticalda}
Let $(X,V)$ be an area preserving diffeomorphism, and let $V$ be homogeneous. Then the critical points
of $f_0(V)$ are the points $V^{-1}(v_c)$, where $v_c$ is a critical point of $f_0(v)$.
\end{lemma}
\begin{proof}
By the chain rule ${df_0(V(v))}/{dv}=V(v)'f_0'(V(v))$. The function $V'\neq 0$ because $(X,V)$ must
be a diffeomorphism. Therefore the critical points occur when $f_0'(V)=0$ or at points $v=V^{-1}(v_c)$.
$\Box$
\end{proof}
Consider the perturbation $\chi$ that was constructed earlier. If $v_c$ is a nondegenerate critical point of $f_0$
such that $f_0''(v_c)<0$,
then we want to prove that there is a rearrangement $V$ such that $f_0(V)=f_0(v)+\int_{-\infty}^v\chi(v'-v_c)dv'$
or that ${df_0(V)}/{dv}=f_0'(v)+\chi(v-v_c)$. Such a rearrangement can be constructed as long as the parameters defining $\chi$,
the numbers $h,d,\epsilon$,
are chosen such that $f_0'(v)+\chi(v-v_c)$ has the same critical points as $f_0'(v)$. Using Morse theory it is possible to construct
a $V$
so that $f_0(V)=f_0(v)+\int\chi+O((v-v_c)^3)$, where $O((v-v_c)^3)$ has compact support and is smaller than
$f_0(v)-{f_0''(v_c)}(v-v_c)^2/2$.
\begin{theorem}
\label{rearrangement}
Let $v_c$ be a nondegenerate critical point of $f_0$ with $f_0'(v_c)<0$. Then there exists a rearrangement $V$
such that $f_0(V)=f_0(v)+\int\chi+O((v-v_c)^3$, where $O$ is defined as above.
\end{theorem}
We omit the proof but it is a simple application of the Morse lemma. In order to apply the Morse lemma $f_0$ must be $C^2$. This is not restrictive
for practical applications where typically $f_0$ is smooth. The rearrangement of $f_0$ can also be made to be smooth if desired.
\bigskip
Using this result we prove a Krein-like theorem for dynamically accessible perturbations
in the $W^{1,1}$ norm.
\begin{theorem}
\label{kreinlike}
Let $f_0$ be a stable equilibrium distribution function for the Vlasov equation on an
infinite spatial domain. Then $f_0$ is structurally stable
under dynamically accessible perturbations in $W^{1,1}$, if there is only one solution
of $f_0'(v)=0$. If there are multiple solutions, $f_0$ is structurally unstable and the
unstable modes come from the zeros of $f_0'$ that satisfy $f_0''(v)<0$.
\end{theorem}
\begin{proof}
Suppose that $f_0'$ has only one zero on the real line. Because $f_0$ is an equilibrium this
zero will have $f_0''>0$. Because a dynamically accessible perturbation can never increase the number
of critical points, it will be impossible to change the winding number of the Penrose plot to a positive
number. Therefore $f_0$ is structurally stable.
Suppose that $f_0'=0$ has more than one solution on the real axis. Using the previous theorem
perturb $f_0'$ by $\chi(v-v_c)$ in a neighborhood of a critical point $v_c$ with $f_0''(v_c)<0$.
This will increase the winding number to 1 since it will add a positively oriented crossing on the
negative real axis for the correct choices of $h$, $d$, and $\epsilon$ in the definition of $\chi$.
The norm of $\chi$ can be made as small as necessary and therefore $f_0$ is structurally unstable.
Since no other critical points with $f_0''<0$ can be created the only critical points that lead to instabilities
are the ones that already exist having $f_0''<0$.
$\Box$
\end{proof}
The implication of this result is that in a Banach space where the Hilbert transform is
an unbounded operator the dynamical accessibility condition makes it so that a change in
the Krein signature of the continuous spectrum is a necessary and sufficient condition
for structural instability. The bifurcations do not occur at all points where the
signature changes, however. Only those that represent valleys of the distribution can
give birth to unstable modes.
\section{Krein bifurcations in the Vlasov equation}
\label{kreinvp}
We identify two
critical states for the Penrose plots that correspond to the transition to instability. In these
states the system may be structurally unstable under infinitesimal perturbations of $f_0'$ in the
$C^n$ norm for all $n$.
The first critical state corresponds to the existence of an embedded mode in the continuous spectrum.
If the equilibrium is stable, then such an embedded mode corresponds to a tangency of the Penrose plot to the real axis at
the origin.
If the system is
perturbed so that the tangency becomes a pair of transverse intersections, then the winding number of the Penrose plot
would jump to 1 and the system would be unstable. Considering a parametrized small perturbation, we see that the value of $k$ for the unstable mode, will correspond to some value of $k\neq 0$ for which the embedded mode exists. Figures \ref{gauscrit} and \ref{pengauscrit} illustrate a critical Penrose plot for a bifurcation at $k\neq 0$. We explore this bifurcation in Sec.~\ref{kne0}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.6]{s5embeddedpenrose}
\end{center}
\caption{Critical Penrose plot for a $k\neq 0$ bifurcation.}
\label{gauscrit}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.6]{s5embeddedpenroseclose.pdf}
\end{center}
\caption{Close up of a critical Penrose plot for a $k\neq 0$ bifurcation.}
\label{pengauscrit}
\end{figure}
Another critical state occurs when $H[f_0']=0$ at a point
where $f_0'$ transversely intersects the real axis.
If the Hilbert transform of $f_0'$ is perturbed, there
will be a crossing with a negative $H[f_0']$, and the winding number will be positive for some $k$.
This mode enters through $k=0$ because the smaller the perturbation of $H[f_0']$
the smaller $k$ must be for $T_k$ to be unstable. Figure \ref{bigaus2} is a critical Penrose
plot corresponding to the bi-Maxwellian distribution with the maximum stable separation. We explore this kind of bifurcation in Sec.~\ref{ke0}.
\bigskip
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=.6]{s4criticalpenrose.pdf}
\end{center}
\caption{Critical Penrose plot for a bi-Maxwellian distribution function.}
\label{bigaus2}
\end{figure}
\subsection{Bifurcation at $k\neq 0$}
\label{kne0}
The linearized Vlasov equation can support neutral plasma modes embedded within the continuous spectrum.
The condition for existence of a point mode is the vanishing of the plasma dispersion
relation on the real axis,
\begin{align}
\varepsilon(u) &= 1-H[f_0']+if_0'= 0\,.
\end{align}
If the spatial domain is unbounded the point modes will be analogues of
the momentum eigenstate solutions of the Schrodinger equation and have infinite energy.
Any violation of the Penrose criterion will guarantee the existence
of zeros of the plasma dispersion function on the real axis because $k$ can take any
value in this case.
If the plasma dispersion vanishes at some $u$ and $f_0''(u)=0$, there is an embedded mode in
the continuous spectrum. The signature of the continuous spectrum will not change signs at
the frequency of the mode and we will extend the definition of signature to the point $u$ even though
$f_0'(u)=0$.
The signature of an embedded mode is given by ${\rm sgn}(u\,{\partial\epsilon_R}/{\partial u})$ (see \cite{MP92,SM94}).
The signature of the continuous spectrum is $-{\rm sgn}(uf_0')$.
These signatures are the same if the value of $f_0'$ in a neighborhood of its zero is the same sign as
$H[f_0'']$.
We will prove that if $f_0$ is stable and mildly regular, it is impossible for there to be a discrete mode embedded in
the continuous spectrum with signature that is the same as the signature of the continuous
spectrum surrounding it. The proof has a simple conceptual outline. Suppose that there exists
a discrete mode with the same signature as the continuum. Then there exists some point
$u$ satisfying $f_0'(u)=0$, $-{\rm sgn}( f_0')={\rm sgn}({\partial \epsilon_R}/{\partial u})$ in a
neighborhood of $u$, and $H[f_0'](u)=0$. Perturbations of $f_0'$ centered around this
point will give the Penrose plot a negative winding number, contradicting the analyticity
of the plasma dispersion function in the upper half plane. We need $f_0'$ to be Holder continuous
so that the Penrose plot is continuous and for the plasma dispersion function to converge uniformly to
its values on the real line.
\begin{lemma}
Let $g$ be a function defined on the real line such that $g$ is Holder and
let $h=H[f]$. Then the functions $g_z$, $h_z$ that are the solutions of the Laplace equation
in the upper half plane satisfying $f_z=f$ and $g_z=g$ on the real line converge
uniformly to $f$ and $g$.
\end{lemma}
\label{firstlemma}
\begin{proof}
Because $g$ can be defined as a bounded and continuous function on $\mathbb{R}\cup\{\infty\}$
and the $g_z$ are analytic, the $g_z$
must converge uniformly to $g$. The same properties hold for $h$ and $h_z$ must converge to
$h=H[g]$.
$\Box$
\end{proof}
\begin{lemma}
Let $f_0'$ be the derivative of an equilibrium distribution function and let $f_0'$
be sufficiently regular such that the assumptions of the previous lemma are true.
Then the Penrose plot that is associated with $f_0'$ cannot have a negative winding number.
\end{lemma}
\label{previouslemma}
\begin{proof}
The Penrose plot associated with $f_0'$ is the image of the real line
under the map $\epsilon(u)=1-H[f_0']+if_0'$.
This is naturally defined as an analytic function if $u$ is in the upper half plane. By the argument principle the image of
$\mathbb{R}+it$ under
this map has a non-negative winding number. Both the
real and imaginary parts of this map converge uniformly to their values on the real line. Therefore the Penrose plot
is a
homotopy of these contours, making it possible to parametrize the contours by some $t$ such that the
distance from the Penrose plot to the contour produced by the image of $\mathbb{R}+i\delta$ is always
less than some $\eta(t)$ that goes to $0$. If the winding number of the Penrose plot were negative, there
would be some $t$ for which the winding number was negative because the winding
is a stable property under homotopy, contradicting the analyticity of the map.
$\Box$
\end{proof}
\begin{theorem}
Let $f_0'$ and $f_0''$ be Holder continuous. If $f_0$ is stable there are
no discrete modes with signature the same as the signature of the continuum.
\end{theorem}
\begin{proof}
Because $f_0$ is stable the winding number of the Penrose plot is equal to 0.
Assume that there is a discrete mode with the same signature as the contiuum surrounding it.
Then there exists a point $u$ with $f_0'(u)=0$, $f_0''(u)=0$, and ${\rm sgn}(f_0'(u+\delta)dH[f_0'](u)/{du})=1$.
Then we search for a function $g$ such that the
Penrose plot of $f_0'+g$ has a negative winding number. If such a function exists it will
contradict Lemma \ref{previouslemma}. Because $f_0''$ is Holder ${\partial\epsilon_R}/{\partial u}$ is bounded
away from zero in a neighborhood of the point. Suppose that in this neighborhood
there is only one zero of $f_0'$. Then define $g$ such that $g$ has one sign, is smooth and has compact
support, and such that the $|\partial {H[g]}/{\partial u}|<|f_0''|$ in this
neighborhood. Then for small enough $g$ the function $f_0'+g$ will have two zeros in a neighborhood of the point.
Then both of the crossings
will correspond to crossings of negative orientation and the resulting winding number
will be $-1$, a contradiction.
$\Box$
\end{proof}
\begin{corollary}
\label{otherobstruction}
If $f_0$ is stable it is impossible for there to be a point where $f_0'=0$, $f_0''<0$, and $H[f_0']>0$.
\end{corollary}
If $f_0$ is unstable the winding number is positive. In this case it may be possible for modes with
the same signature as the continuum to exist.
It is possible for a positive energy mode to be embedded in a section of negative
signature and a negative energy mode to be embedded in a section of positive signature.
This situation is structurally unstable under perturbations that are bounded by the $C^n$ norm
and remains so even when a linear dynamical accessibility constraint is enforced.
\begin{theorem}
Let $f_0'$ be the derivative of an equilibrium distribution function with a discrete mode
embedded in the continuous spectrum. Then there exists an infinitesimal function with compact
support in the
$C^{n}$ norm for each $n$ such that $f_0'+\delta f'$ is unstable.
\end{theorem}
\begin{proof}
Suppose that ${H[f_0'']}$ is non-zero in a neighborhood of the
embedded mode. Define a dynamically accessible perturbation $\delta f=hf_0'$. Then assume
that $f_0'''\neq 0$ at the mode. If we define $h$ such that it does not vanish at the
mode we find that $\delta f''=h''f_0'+h'f_0''+hf_0'''$ and therefore we can choose
$h$ such that the discrete mode becomes a crossing. This can be done with $h$
infinitesimal and smooth. The resulting perturbation will have an infinitesimal effect
on $f_0'$. The new crossings will cause a violation of the Penrose criterion, and
therefore the system with the embedded mode is structurally unstable.
$\Box$
\end{proof}
This is an analog of Krein's theorem for the Vlasov equation for the case where there is
a discrete mode. As a result of this we see that all discrete modes are either unstable
or structurally unstable.
\subsection{Little-big man theorem}
Consider a linearized equilibrium that supports three discrete modes. The signature
of each mode depends on the reference frame. There is a result that applies to a number
of Hamiltonian systems, the three-wave problem in particular \cite{CRS69,KM95}),
that gives a condition on the signature of the modes and their
frequency in some reference frame such that no frame shift can cause all the modes to
have the same signature. In a shifted frame the Hamiltonian changes by virtue of the
frequencies in the action-angle from being doppler shifted. Sometimes such shifts can render the Hamiltonian sign definite. The little-big man theorem for finite systems indicates when this cannot happen. Such a result is also possible for the point spectrum of the Vlasov equation.
\begin{theorem}
Let $f_0'$ be the derivative of an equilibrium distribution function that has three
discrete modes (elements of the point spectrum) with real frequencies. Consider a reference frame where all of the modes
have positive frequency. Then represent the energies of the three modes as a triplet
$(\pm\pm\pm)$ where the plus and minus signs correspond to the signature of each mode,
with the first mode being the one with the lowest frequency and the last the one with
the highest frequency. Then if the triplet is of the form $(+-+)$ or $(-+-)$ there
is no reference frame in which all the modes have the same signature. If the triplet has
any other form, then there is a reference frame in which all the modes have the same signature.
\end{theorem}
The formula for the energy of an embedded mode is ${\rm sgn} (u\,{\partial \epsilon_R}/{\partial u})$.
If there are three embedded modes in a frame where the frequencies are all positive the
triplet is
\[
\left(\left.{\rm sgn}\frac{\partial\epsilon_R}{\partial\omega}\right|_{\omega_1},{\rm sgn}\left.\frac{\partial\epsilon_R}{\partial\omega}\right|_{\omega_2},
{\rm sgn}\left.\frac{\partial\epsilon_R}{\partial\omega}\right|_{\omega_3}\right)
\]
\begin{proof}
If this is $(+-+)$ then as we shift frames
the possible triplets are $(0-+),(--+),(-0+),(-++),(-+0),(-+-)$. All of these are
indefinite. The other possibile initially indefinite triplet is $(--+)$. However if we
shift the two $-$ modes to negative frequency the triplet becomes $(+++)$. All other
examples are either definite or reduce to one of these two.
$\Box$
\end{proof}
\subsection{Bifurcation at $k=0$}
\label{ke0}
Assume that there are no embedded modes and that $f_0$ is stable, but that there is a point that has $f_0'=0$ and $H[f_0']=0$.
This is
the critical state for a bifurcation at $k=0$. This can be destabilized in the same way as the critical state for
$k\neq 0$. There will be a perturbation that makes $H[f_0']<0$ without changing $f_0'$ at that point.
Therefore the Penrose plot becomes unstable and the equilibrium is structurally unstable.
\begin{theorem}
\label{infinitek}
Suppose that $f_0'$ is a stable equilibrium distribution function that has a zero at $u$
of both $f_0'$ and $H[f_0']$. Then $f_0'$ is structurally unstable under perturbations
bounded by the $C^n$ norm for all $n$.
\end{theorem}
\begin{proof}
Let $\delta h$ be symmetric about the point $u$, be smooth with compact support and have its first $n$ derivatives less than
some $\epsilon$. Then let $\delta f_0'=-H[\delta h]$. The resulting
perturbation to $H[f_0']$ is $h$. If $h$ is positive at $u$, then by the symmetry of $h$
$f_0'+\delta f_0'$ has a zero at $u$ and $H[f_0']+h$ is positive there. Thus the Penrose plot has a positive
winding number and is unstable. Therefore $f_0'$ is structurally unstable.
$\Box$
\end{proof}
The previous two sections demonstrated that when the Penrose plot is critical, no amount of regularity
is sufficient to prevent $f_0$ from being structurally unstable. However, when the Penrose plot is not
critical all that is required is that a small perturbation only change the Penrose plot by a small amount
in addition to a condition to prevent perturbations near $v=\infty$. Suppose we arbitrarily restrict the support
of the perturbations so that $|v|<v_{max}$. Then if we increase the required regularity such that ${\rm sup} (H[\delta f_0'])$
is bounded there will be some $\delta$ such that for all $\delta f_0'$ with $\|\delta f'_0\|<\delta$ the distribution
$f_0+\delta f_0$ is structurally stable. This restriction can be motivated physically by restricting the particles in the
distribution function to be travelling slower than the speed of light.
\section{Conclusion}
\label{conclu}
We have considered perturbations of the linearized Vlasov-Poisson equation through
changes in the equilibrium function. The effect of these perturbations on the spectral
stability of the equations is determined by the class of allowable perturbations and the signature of
the contiuous and point spectra. Every equilibrium can be made unstable by adding an arbitrarily small
function from the space $W^{1,1}$. If we rearrange $f_0$ then only when the signature of $f_0$ changes
sign can an arbitrarily small perturbation destabilize it. When $f_0$ is stable discrete modes
always have the opposite signature of the spectrum surrounding them. The equilibria are structurally unstable
under $C^n$ small perturbations for all $n$. The signature of the spectrum and the signature of the discrete mode can
never be the same.
This generalization of Krein's theorem is more complicated than the finite-dimensional original. However the
basic ideas of Krein's theorem are still important in the infinite-dimensional case. When the perturbations
are more restricted than just belonging to $W^{1,1}$ the structural stability is determined by the signature of the spectrum.
Just as in Krein's theorem there must be a positive signature interacting with a negative signature to produce structural
instability.
This paper was devoted primarily to the Vlasov equation, but other noncanonical Hamiltonian systems admit to a similar
treatment, e.g.\ the 2D Euler equation with shear flow equilibria, and we hope to chronicle such cases in future publications.
\subsection*{Acknowledgments}
This work was supported by the U.S.\ Dept.~of Energy
Contract No.~DE-FG03-96ER-54346.
\bibliographystyle{unsrt}
\section{The Linearized Vlasov-Poisson Equation}
Consider the linearization Vlasov-Poisson equation about an homogeneous equilibrium
function $f_0$. The result is an infnite-dimensional Hamiltonian system
with a continuous spectrum. The equations of motion
are:
\begin{align}
\partial{f}{t}+v\partial{f}{x}-\frac{e}{m}\partial{\phi}{x}\de{f_0}{v} &=0 \\
\pnq{\phi}{x}{2} &= -4\pi e\int_{\mathbb{R}} fdv
\end{align}
We assume that we have a one dimensional finite spatial domain $X$ equipped
with periodic boundary conditions.
This system has a Hamiltonian structure given by the Hamiltonian functional
\begin{align}
H[f] &= -\frac{m}{2}\int_X\int_{\mathbb{R}}\frac{v|f|^2dxdv}{f_0'}+\frac{1}{8\pi}
\int_X|\partial{\phi}{x}|^2dx
\end{align}
The Poisson bracket is defined by:
\begin{align}
\{F,G\} &=\int_X\int_{\mathbb{R}}f_0\left[\frac{\delta F}{\delta f},\frac{\delta G}
{\delta f}\right]dxdv
\end{align}
Here $[g,h]=\frac{1}{m}\left(\partial{g}{x}\partial{h}{x}-\partial{g}{v}\partial{h}{x}\right)$. The time
evolution of an initial condition is given by Hamilton's equations:
\begin{equation}
\partial{p}{t}=-\{p,H\}
\end{equation}
This defines the time evolution operator $Tg=-\{g,H\}$. This formulation of Hamilton's equations
is non-canonical and as a result there are a number of Casimir invariants associated with the
degeneracies of the Poisson bracket. Any functional $C$ that satisfies
$\{C,F\}=0$ for all $F$ is a Casimir invariant. For the Vlasov-Poisson equation these are
any functional of the form $C[f]=\int \mathcal{C}(f)dxdv$.
We are interested in the spectral stability of the linearized Vlasov-Poisson system. We will say that the system is
spectrally stable if the real component of each member of the spectrum of the time evolution operator $T$ is less than or equal to zero, or that the frequency components of the spectral are in the closed lower half plane. Since this is a
Hamiltonian system this reduces to the the spectrum being confined to the imaginary axis.
\newtheorem{Spectral Stability}{Definition}
\begin{Spectral Stability}
Consider the one dimensional Vlasov-Poisson system linearized about a homogeneous equilibrium $f_0(v)$ with the phase space of solutions some Banach space $B$. The resulting sytem is said to be {\it spectrally stable} if the spectrum
$\sigma_T$ of the time evolution operator $Tf=\{f,H\}$ is purely imaginary.
\end{Spectral Stability}
Spectral stability does not guarantee that the system is stable, or that the equilibrium $f_0$ is linearly stable.
The solutions of a spectrally stable system are guaranteed to grow at most sub-exponentially, see for example the asymptotic expansions of $f$ in Degond. On the other hand one can
construct a spectrally stable system where there is polynomial growth of certain initial conditions.
The standard to determine whether or not any unstable modes exist is to use the Nyquist
method which determines the Penrose criterion.
\newtheorem{SpecTheorem}{Theorem}
\begin{SpecTheorem}
Consider the one dimensional linearized Vlasov-Poisson system with equilibrium function $f_0$.
If for some $(\omega,k)$ with $\omega$ in the upper half plane and $k\in\mathbb{Z}$ the plasma dispersion function $\epsilon(\omega,k)=1+\frac{4\pi e^2}{mk^2}\int_{\mathbb{R}}
\frac{f_0'dv}{w/k-v}$ has a zero the system is spectrally unstable. Otherwise
it is spectrally stable.
\end{SpecTheorem}
This can be calculated using
the argument principle taking the contour to be the real line and enclosing the entire upper half plane. The limit
of $\epsilon$ on for real $\omega$ is $\epsilon(w,k)=1+\frac{4\pi e^2}{mk^2}PV\int_{\mathbb{R}}\frac{f_0'dc}{w/k-v}+\frac{4\pi^2{\rm i}e^2}{mk^2}f_0'(\omega/k)$. The graph of the real line under this mapping is referred
to as the Penrose plot. The figures show the derivative of the distribution function \cite{figure 1} and the associated Penrose
plot for a Gaussian distribution \cite{figure 2}. Note that the winding number around the origin is zero
and that there are no crossings of the x-axis to the left of $x=1$ which means that
the winding number is $0$ for all values of $k$. This argument gives rise
to the Penrose criterion.
\newtheorem{PenroseCriterion}{Theorem}
\begin{PenroseCriterion}
Consider the linearized Vlasov-Poisson system with equilibrium function $f_0$. If there exists a point $u$ such that $f_0'(u)=0$ and $PV\int_{\mathbb{R}}
\frac{f_0'(v)dv}{u-v}<0$ with $f_0'$ crossing the x-axis at $u$ then the system is spectrally unstable. Otherwise it is spectrally stable.
\end{PenroseCriterion}
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph1-1.pdf}
\label{sheardomain}
\caption{$f_0'$ for a Gaussian distribution.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph1-2.pdf}
\label{sheardomain}
\caption{Penrose plot for a Gaussian distribution.}
\end{center}
\end{figure}
On the other hand it is easy to construct distribution functions where the linearized
dynamics are unstable. Consider the function $f_0=e^{(v-1)^2}+e^{(v+1)^2}$, the sum of
two Gaussians. The stability of this configuration depends on the separation of the
distribution functions, and for this separation there is instability. The graph of $f_0'$ is
figure 3 and the associated Penrose plot is figure 4.
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph2-1.pdf}
\label{sheardomain}
\caption{$f_0'$ for two separated Maxwell distributions.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph2-2.pdf}
\label{sheardomain}
\caption{The unstable Penrose plot corresponding to two separated Maxwell distributions.}
\end{center}
\end{figure}
Consider a stable equilibrium function $f_0$. We will consider perturbations of the
equilibrium function and the resulting perturbation of the time evolution operator. Suppose that the time
evolution operator of the perturbed system is $T_p$.
In the function space that we will consider these perturbations are bounded operators
and their size can be measured by their norm $\|T_-T_p\|$. This norm will be proportional
to the norm of $\|f_p'\|$ in the chosen Banach space.
\newtheorem{StructuralStability}{Definition}
\begin{StructuralStability}
Consider the formulation of the linearized Vlasov-Poisson equation in the Banach
space $\mathcal{B}$ with stable homogeneous equilibrium function $f_0$. Let $T_p$ be a class
of perturbed operators. If there exists some $\epsilon$ depending only on $f_0$ such
that $T_p$ is stable whenever $\|T-T_p\|<\epsilon$ then the Hamiltonian system defined
by $f_0$ is structurally stable under the class of perturbations $T_p$.
\end{StructuralStability}
The aim of this work will be to characterize the structural stability of the linearized
Vlasov-Poisson equation. We will choose a Banach space $\mathcal{B}$ and a class of perturbations.
We will prove that if the perturbation function is some homogeneous $f_p'$ and the norm
is $W_{1,1}$ it will be possible to make every equilibrium distribution function unstable
with an infinitesimal perturbation. Our goal will be to relate the structural stability of
$f_0$ to the Krein signature of the continuous spectrum of $f_0$. The Krein signature is a
signature that can be attached to the continuous spectrum eigenmodes or to any discrete modes.
Its definition is motivated by Krein's theorem, which classifies the structural stability of
finite dimensional Hamiltonian systems in terms of the sign of the energy and the frequency
of the discrete modes. The Krein signature that we define for the Vlasov equation is a
generalization of the definition of the sign of the energy to the conitinuous spectrum.
In order to prove a result analogous to Krein's theorem we will need to
restrict to dynamically accessible perturbations, which preserve all of the Casimir invariants.
In this case we will be able to prove that distribution functions with no signature changes will
be structurally stable while those with signature changes will have structural instabilities that
are created at points where the Penrose plot of the distribution function crosses the x-axis with
positive orientation.
\section{The Krein-Moser Theorem}
For linear finite dimensional Hamiltonian systems Hamilton's equations are a set
of first order linear homogeneous ODEs. If the Hamiltonian is time independent
the behavoir of solutions is completely characterized by the eigenfrequencies.
If all of the eigenfrequencies are on the real axis and nondegenerate the system will be stable.
If there are degenerate eigenvalues the system will be stable as long as the time evolution operator
does not have any nontrivial Jordan blocks, but there will be secular growth if it does. Any complex
eigenfrequencies will lead to instability. The Hamiltonian of a linear finite-dimensional Hamiltonian system
is a quadratic form in the coordinates and momenta. If we consider perturbations of the
coefficients of the quadratic form it is trivial to define a notion of small perturbations,
as the resulting perturbation of the Hamiltonian will be a bounded operator. Krein and Moser
independently proved a theorem characterizing the structural stability of these systems in
terms of the sign of the energy and the frequency of positive and negative energy modes.
\newtheorem{Krein-Moser}{Theorem}
\begin{Krein-Moser}
Let $H$ define a stable linear finite-dimensional Hamiltonian system. Then $H$ is structurally
stable if all the eigenfrequencies are nondegenerate. If there are any degeneracies, $H$ is
structurally stable if the assosciated eigenmodes have energy of the same sign. Otherwise $H$ is
structurally unstable.
\end{Krein-Moser}
The Krein-Moser theorem gives a clear picture of the behavoir of these systems under
small perturbations. The eigenfrequencies move around, but remain confined to the real
line unless there is a collision between a positive energy and negative energy mode.
This theorem was first proved by Krein in the early 1950s and later rediscovered by
Moser in the late 1950s. Our goal will be to place the perturbation theory of infinite dimensional
Hamiltonian systems in the language of the finite dimensional theory.
The appropriate definition of signature for the continuous spectrum of the Vlasov-Poisson
equation was introduced by Morrison in an attempt to study canonical transformations of
the Vlasov-Poisson system to action angle variables. The transformation is a generalization
of the Hilbert transform and it can be used to show that the linearized Vlasov-Poisson
equation is equivalent to the system with the following Hamiltonian functional $H_{AA}$:
\begin{align}
&H_{AA}=\sum_{k=1}^{\infty}\int_{\mathbb{R}}\sigma_k(u)\omega_k(u)J_k(u,t)du
\end{align}
where $\omega_k(u)=|ku|$ and $\sigma_k(u)=sgn(kuf_0'(u))$ is the Krein signature corresponding
to the $u$. The transformation is only value when $f_0'(0)=0$, and therefore this definition of
signature is only valid in reference frames where $f_0'(0)=0$.
\newtheorem{signature}{Definition}
\begin{signature}
Suppose $f_0'(0)=0$. Then the signature of of the point $u\in\mathbb{R}$ is
$sgn(uf_0'(u))$.
\end{signature}
We expect that in order to
define a useful generalization of Krein's theorem there must be enough structure restricting
the space of perturbations so that the systems with no signature changes do not produce structural
instabilities and those with signature changes do.
\section{A Krein-Like Theorem for the Linearized Vlasov-Poisson Equation}
Infinite-dimensional Hamiltonian systems are more complicated than finite
dimensional ones because they give rise to partial differential equations. The Krein-Moser
theorem applies to systems with discrete spectra. Systems with continuous spectra present separate
complications. There already exist some results about this problem for a class of infinite dimensional
Hamiltonian systems. Grillakis studied Hamilonian systems with continuous spectra in a Hilbert space
where the time evolution operator is self adjoint. He studied the effects of relatively compact
perturbations on such a system and proved that the existence of a negative energy mode in the continuous
spectrum caused the system to be structurally unstable. He also proved that such systems are structurally stable
otherwise.
This work is substantially different from previous efforts to characterize structural stability
in infinite dimensional Hamiltonian systems because the Vlasov-Poisson equation has a non-canonical
Hamiltonian sturcture. In this case it will be possible to have bifurcations into instability under much
weaker conditions on the perturbation operator as long as there is a signature change.
We are interested in further generalizing Krein's theorem to infinite-dimensional Hamiltonian systems.
In particular we are going to study the specific case of the Vlasov-Poisson equation. This case is
interesting because there exists a canonical transformation to action angle variables that makes it
possible to generalize some of the definitions of the finite-dimensional theory and apply them in
the infinite-dimensional case. The time evolution operator of the Vlasov-Poisson equation has the
entire real line as its continuous spectrum, but it also can have discrete eigenvalues embedded within
the continuous spectrum.
We will need the Fourier space formulation of the Vlasov-Poisson
equation.
In fourier space the Vlasov-Poisson equation takes the
form:
\begin{align}
\partial{f}{t}+{\rm i}kvf_k-\frac{4\pi e^2{\rm i}}{mk}f_0'\int_{\mathbb{R}}
f_kdv' &= 0
\end{align}
The Hamiltonian functional is
\begin{align}
H[f] &= \sum_k\left[-\frac{m}{2}\int_{\mathbb{R}}\frac{v|f_k|^2dv}{f_0'}+\frac{1}{8\pi}
|\phi_k|^2\right]
\end{align}
and the Poisson bracket is
\begin{align}
\{F,G\} &= \sum_{k=1}^{\infty}\frac{{\rm i}k}{m}\int_{\mathbb{R}}f_0'\left(\frac
{\delta F}{\delta f_k}\frac{\delta G}{\delta f_{-k}}-\frac{\delta F}{\delta f_{-k}}
\frac{\delta G}{\delta f_k}\right)
\end{align}
\subsection{$W_{1,1}$}
We are going to begin by defining our phase space $B$. We will require all of the Fourier coefficients $f_k$ to be
in the Sobolev space $W_{1,1}(\mathbb{R})$ and define the norm of a sequence $\{f_k\}$ to be:
\begin{align}
\|\{f_k\}\| &= \sum_{k} (1+k^2)\|f_k\|_{W_{1,1}}
\end{align}.
We are interested in the spectral properties of the time evolution operator $T$ on
this Banach space. We will prove that $T$ is unbounded, densely defined, and closable. We
will prove that the frequencies of the spectra are all confined to the real axis as long as
the Penrose criteria is satisfied and that
the Penrose criteria gives all of the eigenvalues. We will prove that the spectrum is composed
entirely of the continuous spectrum and discrete spectrum with no residual component.
We define
\begin{align}
D_1(T)= &\{\{f_k\}|vf_k\in W_{1,1}(\mathbb{R}),\\
&\sum_{k\neq 0}\|(1+k^2)(kvf_k-4\pi \frac{f_0'}{k}\int f_k(v)dv)\|_{W_{1,1}}<\infty\}.
\end{align}
\newtheorem{Dense}{Theorem}
\begin{Dense}
The operator $T$ with domain $D_1(T)$ is densely defined in the space $B$.
\end{Dense}
Define the set $G=\{\{f_k\}|f_k(\xi)\in C_c^{\infty}(\mathbb{R}), f_k)=0$ for all but a finite number of $k\}$.
Clearly $G\subset D_1(T)$. Let $\{f_k\}\in B)$. Because of the density of $C_c^{\infty}$ in $W_{1,1}$ for each $\epsilon>0$ we can construct a function $g_{k,\epsilon}\in C_c^{\infty}$
satisfying $\|f_k-g_{k,\epsilon}\|<\epsilon$. Then we can construct a sequence $\{h_k\}_n$ such
that the first $2n+1$ memebers of the sequence are $g_{k,\epsilon}$ and the
rest are zero. If we let $\epsilon$ go to zero like $1/n^2$ this sequence
will converge to $\{f_k\}$. Thus $D_1(T)$ is dense, and $T$ is densely defined.
\qed
\newtheorem{Closable}{Theorem}
\begin{Closable}
The operator $T$ with the domain $D_1(T)$ is closable in the space $B$.
\end{Closable}
Now we will prove that $T$ is closable. The operator $T$ will be closable if
the closure $\overline{G(T)}=\overline{(D_1(T),TD_1(T))}$ is the graph of an operator. Let
$(\{f_k\}_n,\{Tf_k\}_n)$ be a convergent sequence in $G(T)$. The operator $T$ will be closable if for every
convergent sequence that has $f_n\rightarrow f$ there is a single element of $H_s(T_1\times\mathbb{R})$ that we call
$Tf$ that satisfies $Tf_n\rightarrow Tf$ for all $f_n\rightarrow f$. This is equivalent
to the statement that $T$ is closable if and only if for every sequence
$f_n\rightarrow 0$ in $D_1(T)$ either $Tf_n\rightarrow 0$ or $Tf_n$ diverges.
Suppose $\{f_n\}\subset D_1(T)$ and $f_n\rightarrow 0$. Then we have $f_{k,n}\rightarrow 0$ for each $k$. Assume that $Tf_{k,n}$ is convergent for each $k$. We know that $\int f_{k,n}\rightarrow 0$.
Because each $f_{k,n}$ goes to zero $f_{k,n}(v)$ goes to zero for each $v$ and therefore $vf_{k,n}(v)$ goes to zero for each $v$. If $vf_{k,n}$ is convergent it must converge to $0$, so if
$\{Tf_{k,n}\}$
converges it must converge to $0$ and thus $T$ is closable. We denote the domain of the closure of
$T$ as $D(T)$.
\qed
At this point we will calculate the spectrum of the operator $T$. We split
the spectrum in discrete, continuous, and residual components.
\newtheorem{spectra}{Definition}
\begin{spectra}
Consider the operator $T$ and let $\lambda\in\sigma(T)$. Write the resolvent of $T$ as $R(T,\lambda)=(T-\lambda I)^{-1}$. We say $\lambda$
is in the discrete spectrum, or $\sigma_d(T)$ if $R(T,\lambda)$ fails to be injective. We say $\lambda$ is in the continuous spectrum, or $\sigma_c(T)$, if $R(T,\lambda)$ is injective but not bounded. We say $\lambda$ is in the
residual spectrum, or $\sigma_r(T)$, if $R(T,\lambda)$....
\end{spectra}
Using this definition we characterize the spectrum of the operator $T$.
\newtheorem{spectrumT}{Theorem}
\begin{spectrumT}
The discrete spectrum of $T$ consists of all the points ${\rm i}\omega$
in the complex plane where $\epsilon(\omega,k)=0$ for some $k\in\mathbb{Z}$. The
continuous spectrum consists of all ${\rm i}\omega$ with $\omega\in(\mathbb{R}/(-{\rm i}
\sigma_d(T)\cap\mathbb{R})$. The residual spectrum of $T$ is empty.
\end{spectrumT}
By the Penrose criterion from plasma physics we can identify all the points in the
discrete spectrum. If for any $k$ the image of $\mathbb{R}$ under the map
$v\rightarrow 1-\int\frac{4\pi e^2f_0'du}{kv-u}$ has a positive winding number there will exist discrete modes with frequency in the upper half plane. Because the system is Hamiltonian these modes will occur in quartets.
It is possible for there to be discrete modes with real frequencies. If
for real $v$ the previous map passes through the origin then there will be a mode.
We will prove that there is a continuous spectrum with frequencies spanning the entire real axis.
Let $f_{k,n}$ be a sequence of functions converge to zero and whose support also is contained
in an interval of length $2\epsilon(n)$ surrounding the point $v=u/k$, and that have zero average
value. Let $u\in\mathbb{R}$. We also choose sequences so that $\epsilon(n)\rightarrow 0$. Then
consider the resolvent of $T$, $R(T,{\rm i}u)=(T-{\rm i}uI)^{-1}$. For each $n$
\begin{align}
\|R(T,{\rm i}u)\| &\geq \frac{\|f_{k,n}\|_{W_{1,1}}}{\|(kv-u)f_{k,n}\|_{W_{1,1}}} \\
&\geq \frac{\|f_{k,n}\|_{W_{1,1}}}{\|kv-u\|_{W_{1,1}((u/k-\epsilon,u/k+\epsilon))}\|f_{k,n}\|_{W_{1,1}}} \\
&=\frac{1}{\|kv-u\|_{W_{1,1}((u/k-\epsilon,u/k+\epsilon))}}
\end{align}
In this expression the term $W_{1,1}(X\times(u/k-\epsilon,u/k+\epsilon))$ refers to the integral of $|f|+|f'|$ over the interval $(u/k-\epsilon,u/k+\epsilon)$.
Therefore the resolvent is an unbounded operator and ${\rm i}u$ is in the spectrum. If the frequency $u$ has an imaginary component ${\rm i}u_i$ then
$\|R(T,{\rm i}u)\|<1/u_i$ so unless $u$ is part of the discrete spectrum it is part of the resolvent set. Finally we prove that the residual spectrum of $T$
is empty.
The residual spectrum of $T$ is contained in the point spectrum of $T^*$. The dual of $B$ is the space $B^*$ defined by sequences of pairs $(g_k,h_k)\in W_{-1,1}$ with
$\sup_k (1+k^2)\|(g_k,h_k)\|_{W_{-1,1}}<\infty$. The operator $T^*(g_k,h_k)=(kvg_k-kh_k+\int\frac{g_kvf_0'-h_kf_0'-vh_kf_0''}{k},-vh_k)$ is the dual of $T$. If we search for
members $\lambda$ of the point spectrum we get two equations, one of which is $(\lambda+kv)h_k=0$. This forces $h_k=0$. The other equation is then $(kv-\lambda)g_k+\int\frac{g_kvf_0'}{k}=0$ which
forces $g_k=0$ and implies that the point spectrum of $T^*$ is empty. Thus the residual spectrum of $T$ is empty.
In the Banach space $B$ the perturbations of the time evolution operator take
the form of a bounded operator $\frac{4\pi e^2 f_p'}{mk}\int f_k$, the norm of which is
proportional to $\|f_p'\|_{W_{1,1}}$. We will use the norm of $f_p'$ to categorize the
size of the perturbation. Then we will write down a class of perturbations that
can have an arbitrarily small norm but an arbitrarily large effect on the
Hilbert transform of $f_0'$ at a chosen point. We use this perturbation
to produce an unstable system and we prove that the frequency of the resulting
unstable mode converges to the solutions of the equation $f_0'(v)=0$. The action of the
perturbation that we use will have a clear interpretation in terms of the Penrose criterion and
its topological effect on the zeros function $f_0'$ and the Penrose plot. All of the crossings of the x-axis
by the Penrose plot will have an orientation, and the action of the perturbation will be to translate those
crossings and to change the orientation to prevent the winding number from becoming negative. The winding number must
be positive because the plasma dispersion function is analytic in the upper half plane and thus has no poles. If an orientation
change occurs it will be accompanied by the creation of additional zeros of the distribution function.
We will then prove that by restricting to dynamically accessible perturbations we will eliminate all
perturbations that change the number of zeros of $f_0'$. Under this restriction, only zeros of the $f_0'(v)$ that
have a positive orientation can produce a nonzero winding number after a translation. Therefore only those
distribution functions with more than one real zero of $f_0'(v)$ will have structural instabilities, and those
instabilities will occur. The condition that $f_0'$ have multiple real zeros allows for the Hamiltonian system to
have a negative signature component, and the dynamical accessibility condition makes it possible for there to
be a Krein's theorem.
\newtheorem{Krein-VP}{Theorem}
\begin{Krein-VP}
Consider the formulation of the linearized Vlasov-Poisson equation in Fourier
space with stable equilibrium function $f_0$. We let phase space be the Banach space $B$.
Then the LVP system is structurally unstable under perturbations of
$f_0'$. The frequencies of the unstable modes are the solutions of the equation $f_0(v)'=0$.
\end{Krein-VP}
The time evolution
operator for a given mode $f_k$ is $T:f_k\mapsto \partial{f_k}{t}=-{\rm i}kvf_k+
\frac{4\pi e^2{\rm i}}{mk}f_0'\int_{\mathbb{R}}f_kdv'$. We will consider
perturbations of this operator that have the form $T_p=T+\frac{4\pi e^2{\rm i}}
{mk}f_p'\int_{\mathbb{R}}f_kdv'$.
We wish to construct an operator that is infinitesimal in our Banach space
but whose Hilbert transform is unbounded at a point. Consider the function
$\chi(v,h,d,\epsilon)$
\begin{align}
\chi &= \frac{hv}{\epsilon} & |v|<\epsilon \\
&= h sgn(v)& \epsilon<|v|<d+\epsilon \\
&= h+d+\epsilon-v & h+d+\epsilon>v>d+\epsilon \\
&= -h-d-\epsilon-v & h+d+\epsilon>-v>d+\epsilon \\
&= 0 & |v|>h+d+\epsilon
\end{align}
The figures show the graph of $\chi$ and its Hilbert transform.
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph3-1.pdf}
\label{sheardomain}
\caption{The perturbation $\chi$ for $\epsilon=e^{-10}$, $h=d=.1$.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph3-2.pdf}
\label{sheardomain}
\caption{The Hilbert transform of $\chi$.}
\end{center}
\end{figure}
\newtheorem{Chi}{Lemma}
\begin{Chi}
If we choose $d=h$ and $\epsilon=e^{-(1/h)}$ for any $\delta,\gamma>0$ we can choose
an $h$ such that $\|\chi\|_{H_s}|<\delta$ and ${\bf PV}\int \frac{\chi dv}{v}>1-O(h)$
and $|{\bf PV}\int\frac{\chi dv}{u-v}|<|\gamma/u|$ for $|u|>|d|$.
\end{Chi}
In the space $W_{1,1}$ the function $\chi$ has norm $h^2+2hd+h\epsilon$.
We can compute the value of the Hilbert transform of this
function at a given point $u$ by calculating the principal values:
\begin{align}
& {\bf PV}\int\chi\frac{1}{v-u}dv = \frac{hu}{\epsilon}\log\left(\frac{|u-\epsilon|}{|u+\epsilon|}\right)
+h\log\left(\frac{|d+\epsilon-u||u+d+\epsilon|}{|\epsilon+u||\epsilon-u|}\right)+\\
& (d+\epsilon+h-u)\log\left(\frac{|d+\epsilon+h-u|}{|d+\epsilon-u|}\right)+
(d+\epsilon+h+u)\log\left(\frac{|d+\epsilon+h+u|}{|d+\epsilon+u|}\right)
\end{align}
The next step is to analyze the asymptotics of this function as $h$, $d$, and
$\epsilon$ go to zero. The goal is to find a combination such that the norm of $\chi$
goes to zero, the maximum of the Hilbert transform of $\chi$ is $O(1)$ and there
is a band of vanishing width around the origin outside of which the Hilbert
transform can be made arbitrarily close to zero.
Note that the function can be written as a linear combination of translates of
the function $x\log x$:
\begin{align}
&{\bf PV}\int\chi\frac{1}{v-u}dv = \frac{h}{\epsilon}((u-\epsilon)\log(|u-\epsilon|)
-(u+\epsilon)\log(|u+\epsilon|))\\
& -(d+u+\epsilon)\log(|d+u+\epsilon|)-(d-u+\epsilon)
\log(|d-u+\epsilon|)\\
& +(d+u+\epsilon+h)\log(|d+u+\epsilon+h|)+(d-u+\epsilon+h)
\log(|d-u+\epsilon+h|)
\end{align}
The function $x\log x$ has a local minimum for positive $x$ at $x=1/e$.
This is the point at which the function is most negative. It has a zero at $x=0$
and $x=1$. For values of $u,d,\epsilon,h$ close to zero all of the arguments
of the $\log$ functions are less that $1/e$. Therefore for $|u|<d+\epsilon+h$
the $x\log x$ functions are all monotonically decreasing functions of the
argument $x$. Of the terms listed in the previous equation the $\frac{h}{\epsilon}((u-\epsilon)\log(|u-\epsilon|)
-(u+\epsilon)\log(|u+\epsilon|))$ has by far the largest coefficient as long as
$\epsilon$ is much smaller than $h$. We are going to choose $\epsilon=O(e^{-1/h})$
which guarantees the dominance of this term everywhere in the interval of interest.
The critical point of this function is $u=0$, and for small values of all the
parameters this is the maximum of the function. The value of the maximum is order
one as long as $\epsilon=0(e^{-1/h})$. Outside of the support of $\chi$ the Hilbert
transform decreases at least as fast as $O(1/u)$. At $u=d+\epsilon+h$ we calculate
the value of the function to be on the order of $(2d+2\epsilon+2h)\log(|2d+2\epsilon+2h|)$.
This can be made arbitrarily small by suitable choices of $d$ and $h$. If we arbitrarily take
$d=h$ we notice the norm is order $h^2$ which proves our theorem.
\qed
Thus we can
create a function $f_p'$ that has vanishing norm and whose Hilbert transform has
maximum of order one and is arbitrarily small and decreasing outside of an arbitrarily
small interval.
Now suppose that $f_0'=0$ has a solution somewhere on the real line, and further
suppose that for all values of $k$ the image of the real line under the plasma
dispersion function has zero winding number. Then it is possible to construct an
$f_p'$ to shift the Hilbert transform of $f_0'$ by an arbitrary amount, so as to make
it possible to increase the winding number to one.
We use the fact that one way to
compute the winding number is to draw a line segment from the origin to infinity and
to count the number of intersections with the contour with orientation.
\newtheorem{Winding}{Lemma}
\begin{Winding}
Consider an equilibrium distribution function $f_0'$. The winding number of the
Penrose plot around the origin is equal to $\sum_u sgn(f_0''(u))$ for all
$u\in\mathbb{R}_-$ satisfying $f_0'(u)=0$.
\end{Winding}
To calculate the winding number using this lemma one counts the number of roots of
$f_0'$ on the negative real line and adds them with a positive sign if $f_0''$ is positive
or a crossing from the lower half plane to the upper half plane, a negative sign if
$f_0''$ is negative or $u$ is a crossing from the lower half plane to the upper
half plane, and zero if $u$ is not a crossing of the x-axis.
To prove this lemma we note the following equivalent characterization of the
winding number from differential topology.
\newtheorem{diffwinding}{Definition}
\begin{diffwinding}
If $X$ is a compact, oriented, $l$-dimensional manifold and $f:X\mapsto\mathbb{R}^{l+1}$
is a smooth map, the winding number of $f$ around any point $z\in\mathbb{R}^{l+1}-f(X)$
is the degree of the direction map $u:X\mapsto S^l$ given by $u(x)=\frac{f(x)-z}{|f(x)-z|}$.
\end{diffwinding}
In our case the compact manifold is the real line plus the point at $\infty$ and $l=1$.
The degree of $u$ is the intersection number of $u$ with any point on the circle taken
with a plus sign if the differential preserves orientation and a minus sign if it reverses
it. Then our lemma is just a specialization of this definition to the negative x
direction on the circle. If more than one derivative of $f_0$ vanishes at a root of $f_0'$
we can still calculate the orientation by determining if there is a sign change in
$f_0'$.
Suppose the winding
number is zero, and consider the zeros on the real real axis. The point at $\infty$
always contributes a root on the positive real line with a positive $f_0''$.
Therefore there must exist a root on the positive real line with a negative $f_0''$.
Choose this root $u$
$f_0'=0$ and apply the perturbation $\chi$ to $f_0'$ in a neighborhood of the root.
Because the hilbert transform is arbitrarily small outisde of a region of width $d$
in a neighborhood of a root with negative $f_0''$ a $\chi$ with small $\epsilon$
will change the sign of $f_0''+\chi'$ to be positive at the root. This crossing
will not occur on the negative real axis because of the change in Hilbert transform.
Two additional roots will be created nearby the original root and these roots will have
negative $f_0''$ but they will not be moved because the Hilbert transform will be
arbitrarily small there. Thus there will be an additional positive crossing on the
negative real axis with no other change to the geometry of the Penrose plot. Thus
the winding number will increase to one.
Thus we can always construct a perturbation
that makes the LVP system unstable if there is a nonzero root of $f_0'$.
For the special case of the Maxwell distribution we have provided the Penrose
plot of the unstable perturbed system.
\qed
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph4-2.pdf}
\label{sheardomain}
\caption{$f_0'+\chi$ for a Maxwellian distribution.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph4-1.pdf}
\label{sheardomain}
\caption{Penrose plot for perturbed Maxwellian.}
\end{center}
\end{figure}
\subsection{Dynamical Accessibility}
Under the definitions of the previous section the concept of structural stability is
useless because every single distribution, even the Maxwell distribution, is structurally
unstable. The most obvious change would be to see what effect changing the norm has on
the structural stability. The spaces for which the resulting theorem is non-trivial are
those spaces where the norm bounds the $\sup$ of the perturbation function $f_p'$ but not
the $\sup$ of its Hilbert transform.
If the norm is strong enough to bound the Hilbert transform then it
is impossible to translate crossings of the x-axis in such a way as to increase the winding
number. Therefore only the point at infinity or actual embedded modes are structurally unstable, no
matter what the distribution function. On the other hand if the Hilbert transform is unbounded, all
solutions of $f_0'(v)=0$ become structurally unstable. The situation is even worse if the norm
does not bound the $sup$ when every single point becomes structurally unstable.
We need more structure in order to distinguish between distributions with only positive signature
and distributions with both negative and positive signature. To this end we introduce dynamically
accessible perturbations. Non-canonical Hamiltonian systems such as the Vlasov-Poisson equation
have a family of invariants called Casimir invariants. These invariants are functionals $C$ that
satisfy $\{C,F\}=0$ for all $F$. Casimir invariants depend on the structure of the phase space and Poisson
bracket but not the Hamiltonian. Dynamically accessible perturbations preserve the Casimir invariants, and
are generated by taking canonical transformations from the original Hamiltonian to the variables that
solve the perturbed Hamiltonian. In the Vlasov case one consequence is that to first order we can
write perturbations as perturbations of the equilibrium distribution function through rearrangements.
These rearrangements must be area preserving maps of the coordinates $(x,v)$. If we consider the simplest case
where we require the perturbed equilibrium to be a homogeneous function of $x$ then we have
the condtion $f_n(v)=f_0(V(v))$ where the map $V$ satisfies $[X(x,v),V(v)]=1$. This means that
$\partial_v V\partial_x X =1$. Along with the requirement that The maps $X$ and $V$ are invertible
we find that the derivative of $V$ cannot vanish, which means that the perturbed distribution
function $f_n$ has the same number of critical points as $f_0$. This rules out perturbations that
increase the number of zeros of $f_0'$.
\newtheorem{dynamicalaccessibility}{Definition}
\begin{dynamicalaccessibility}
Let $f$ be a solution of the nonlinear Vlasov-Poisson equation. A dynamically accessible perturbation
of $f$ is an area preserving rearrangement of the particle phase space $(x,v)$.
\end{dynamicalaccessibility}
If $f_0$ is an equilibrium a dynamically accessible perturbation cannot change the number of critical
points of $f_0$.
\newtheorem{criticalpoints}{Lemma}
\begin{criticalpoints}
Let $(X,V)$ be a dynamically accessible perturbation of $f_0$. Then $f_0(V)$ has the same number of
critical points as $f_0(v)$.
\end{criticalpoints}
To prove this we use the chain rule to calculate $\frac{d}{dv}f_0(V)=V'f_0'(V)$. Since $V'$ is never
0 this only has roots when $f_0'(V)=0$. Since $V$ is an invertible function of $v$ all the zeros of
this expression are the inverse image of the zeros of $f_0'$ under the map $V$. Therefore the two functions
have the same number of zeros.
We can consider the affect of a dynamically accessible perturbation in a neighborhood of an equilibrium.
If we write $V=v+h(v)$ and $X=x/(1+h'(v))$ then to first order $f_0(V)=f_0(v)+f_0'(v)h(v)$.
\subsection{More Detailed Analysis of Dynamical Accessibility to First Order}
Consider the following homogeneity preserving symplectic map:
\begin{align}
&v\mapsto v+h(v) \\
&x\mapsto x/(1+h'(v)) \\
&1+h'(v)>0
\end{align}
Consider the homogenous function $f(v)$ that satisfies $f(0)'=0$.
The function $f(v+h(v))$ has the series expansion:
\begin{align}
f(v+h(v))&=f(v)+h(v)f'(v)+\frac{h^2(v)}{2}f''(v)+...
\end{align}
The function $\frac{d}{dv}f(v+h(v))$ has the series expansion:
\begin{align}
\frac{d}{dv}f(v+h(v))&=(1+h'(v))(f'(v)+h(v)f''(v)+\frac{h^2(v)}{2}f'''(v)+...
\end{align}
Because (1+h'(v)) is non vanishing the critical points of this function will
be the zeros of $f'(v+h(v))$. Consider a neighborhood of $0$ and assume that $h(v)$
is small and has the form $h(v)=-av$ for some $a<1$ in a neighborhood of $0$. Suppose
that locally $f'(v)=bv+cv^2$. Then we can write $f'(v+h(v))=f'(v)+h(v)f''(v)$ to first
order and substitute to get:
\begin{align}
f'(v+h(v)) &= f'(v)+h(v)f''(v) \\
&= bv+cv^2-abv-2acv^2 \\
&= v(b(1-a)+c(1-2a)v)
\end{align}
This equation has two roots, $v=0$ and $v=\frac{b(1-a)}{c(1-2a)}$. For fixed $b$
and $c$ we can make the second root arbitrarily close to zero as we take the limit
$a\rightarrow 1$. This explicitly demonstrates that the second zero can occur in a region
where $h(v)$ is still equal to $-av$.
What is happening here? It is clear that if we expand around $v=0$ and compare our result
to that result we are neglecting a term that is non-negligible compared to terms that
we are keeping. For example, to the correct order in $v^2$ the function $f'(v+h(v))=
b(v-av)+c(v-av)^2$ is equal to $bv(1-a)+cv^2(1-a)^2$. The $a^2$ term that we dropped is
not negligible compared with $c$ or $2a$ if $a$ is not close to $0$. I have some ideas
for modifying the expansion, but I do not know if they make sense.
If we assume that the critical point is at $v=0$ like we have been doing consider
the series expansion of $f$ in a neighborhood of $v=0$.
\begin{align}
f'(v)&=vf''(0)+\frac{v^2}{2}f'''(0)+\frac{v^3}{6}f''''(0)+...
\end{align}
If we know that $f''(0)$ is non-vanishing we have a non degenerate critical
point. Now suppose we truncate the series somewhere. Is there an $\epsilon$
such that $f'(v)_{truncated}\neq0$ if $|v|<\epsilon$? The answer is yes, otherwise the
Taylor series would have problems converging in a neigborhood of $v=0$. If
there was no such $\epsilon$ then for each $\epsilon$ there would exist some
$n$ such that the truncation that leaves only $n$ terms in the series had a zero
for $|v|<\epsilon$. It is impossible for there to be an $n$ where $v=0$ is a double
root, it is always isolated. Then let $\epsilon=1/N$ and consider the subsequence
of the of Taylor expansion whose $N$th term has a zero for $v<1/N$. This subsequence
must converge to a function with a non-isolated critical point. However this is a
subsequence of a convergent sequence and must converge to the original sequence, which
produces a contradiction. Thus for each critical point of a given function there is
some $\epsilon$ such that if you truncate you cannot have a zero close to $v=0$.
Now let us consider the same function after the substitution $v\mapsto v+h(v)$.
Then we get a series expansion for $f'(v+h(v))$. This differs from $\frac{d}{dv}
f(v+h(v)$ by an initial $(1+h'(v))$ term.
\begin{align}
f'(v+h(v))&=(v+h(v))f''(0)+\frac{(v+h(v))^2}{2}f'''(0)+\frac{(v+h(v))^3}{6}f''''(0)+...
\end{align}
Now we consider a similar problem, but one dependent on the parameter $h(v)$. We
know that when we truncate there is some $\epsilon$ so that the second zero of $f'$
is never in that interval. If this is the case in this problem then as $h$ is sent to zero in regions where $h$ is zero we will have the original function and
once the support of $h$ is within $\epsilon$ we will no longer be creating any
new zeros.
Suppose that in a neighborhood of $v=0$ we have as in the previous examples $h=-cv$ with $c<1$. Then the full expansion of $f'$ becomes
\begin{align}
f'(v(1-c))&=v(1-c)f''(0)+\frac{v^2(1-c)^2}{2}f'''(0)+\frac{v^3(1-c)^3}{6}f''''(0)+...
\end{align}
If we just consider truncations of this we see that the same $\epsilon$ applies as did in the case of $f'(v)$, and in fact $\epsilon$ is forced to become even larger as $c$ becomes close to $1$. We can see that this expansion procedure is
correct, and we can see that it differs from the expansion procedure used in the
previous section. The question then becomes is it possible to reformulate the previous formula $f'(v+h(v))=f'(v)+h(v)f''(v)$ so that it agrees with a truncated power series expansion to some order? And if this is done, what happens to the Hilbert transform?
If we carry over the analysis of the previous section and combine it with the dynamical
accessibility constraint we see that not all crossings can be translated to produce
structural instability. All crossings with negative orientation can no longer be translated
because the perturbation would create two additional zeros. Furthermore, it is no longer possible
to create instabilities in the tail of the distribution because creating those instabilities would
involve the creation of additional zeros of $f_0'$. This result is much more interesting because
it captures the signature. Only when there is a negative signature in the continuous spectrum is
the distribution structurally unstable. In this case the points of structural instability are only points where there
is a crossing of positive orientation.
In the case where $f_0$ corresponds to the superposition of two Maxwell distributions that are close enough
to be stable the Penrose plot will cross the $x-axis$ with a positive orientation. We use $\chi$ to make
the system unstable without violating the dynamical accessibility constraint. We attach the Penrose plot before and after
the perturbation.
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph5-1.pdf}
\label{sheardomain}
\caption{Penrose plot of two separated Maxwellian distribution functions.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{graph5-2.pdf}
\label{sheardomain}
\caption{Penrose plot after the perturbation is applied.}
\end{center}
\end{figure}
\subsection{Frequency of Unstable Modes}
Now we must try to calculate the frequency of the solution that is produced by the
addition of the perturbation $f_p'$. The perturbation that we use satisfies $f_p'(0)=0$.Thus
if $f_0(u)'+f_p(0)'=0$ if $f_0(u)'=0$. If we make the Hilbert transform just the
correct magnitude we will produce a discrete mode embedded within the continuous
spectrum with the frequency $u$. This is true for any perturbation of the class $f_p$
that we introduced earlier. Therefore we are always an infinitesimal distance from
a spectrum with a discrete mode at frequency $u$ with respect to convergence of a bounded
perturbation.
Outside of the continuous spectrum the discrete eigenfrequency is a continuous function
with respect to the distance between the two operators. Thus for perturbations of the form
of $f_p$ we will always be able to restrict the eigenfrequency to a neighborhood of
$u$ whose size depends on the size of $f_p$. If we could prove that no perturbation
of the form of a function $f_p$ can be constructed that creates a discrete mode with
real frequency anywhere but at the frequency $u$ we could also prove that any bifurcation
into instability will produce an eigenfrequency in a neighborhood of $u$.
In order for there to be a discrete mode at $u$ the equation $f_0'(u)+f_p'(0)=0$.
If $f_0'(u)\neq 0$ then $f_p'(0)\neq 0$. However, the gap between $T$ and $T_p$ is
proportional to $\sup f_p'$ which must be at least as large as $f_0'(u)$. This means
that $T_p$ cannot converge to $T$ and that we cannot create an eigenfrequency $u$ with
an infinitesimal perturbation unless $f_0'(u)=0$. This means that all modes that
leave the axis will come from frequencies where $f_0'(u)=0$.
\section{Conclusion}
We have considered perturbations of the linearized Vlasov-Poisson equation through
changes in the equilibrium function. The effect of these perturbations on the spectral
stability of the equations is finely dependent on the class of perturbation functions
that are allowed. The two key constraints are dynamical accessibility and unboundedness
of the Hilbert transform. Without dynamical accessibility perturbations can change the
number of zeros of the $f_0'$ and this allows an instability along the tail of the distribution.
If the Hilbert transform is unbounded by the norm then all equilibrium distributions have
structural instabilities at crossings of the Penrose plot. If this quantity is bounded then
even those distributions with signature changes are sturcturally stable at the axis crossings.
The dynamical accessibility constraint prevents the creation or destruction of zeros of
$f_0'$. If the Hilbert transform is unbounded this means that there is a version of
Krein's theorem where axis crossings with positive orientation lead to structural instability.
\end{document}
|
\section{Introduction}
The advent of the AdS/CMP correspondence\footnote{CMP stands for Condensed Matter Physics. See \cite{Hartnoll:2009sz}\cite{McGreevy:2009xe} for recent reviews on the subject.} radically changes the status of the string theory, or quantum gravity. The holography is believed to be one of the fundamental principles of quantum gravity, but so far it has been formidable to acquire any experimental evidence. First of all, our observable universe is unique, so the holographic approach to our universe (if possible) is restricted to one and the only one particular example. On the other hand, we have investigated the string/gravity dual for the QCD, but again our QCD is unique, so experimental comparison of our holographic theory has been quite limited.
The AdS/CMP correspondence has completely changed the situation. We can engineer the condensed matter system as we like, and as a consequence we may have infinitely many experimentally testable holographic setups in principle. We believe that we will be able to compare various multiverse with condensed matter systems in the near future: we will soon realize that we are surrounded by quantum universes realized in condensed matter systems.
We are, however, still on the way to the above-mentioned paradise of holographic quantum gravities. Unlike the AdS/CFT correspondence and partially successful AdS/QCD correspondence, we do not have any concrete (experimentally testable) realizations of condensed matter systems in terms of quantum gravity. In this sense, we have not reached even the standard of AdS/QCD correspondence, where we can at least compare qualitative predictions with experiments, assuming $N=3$ is large enough.
In particular, less is known for the field theory dual of the non-relativistic AdS/CFT correspondence. The geometry that has the isometry corresponding to the Schr\"odinger group was first advocated in \cite{Son:2008ye}\cite{Balasubramanian:2008dm},\footnote{A geometric realization of the Schr\"odinger group was pioneered in the earlier work \cite{Duval:1990hj}, whose relation to \cite{Son:2008ye}\cite{Balasubramanian:2008dm} was discussed in \cite{Duval:2008jg}.} and its supergravity embedding has been discussed in the context of the string theory \cite{Herzog:2008wg}\cite{Maldacena:2008wh}\cite{Adams:2008wt}\cite{Hartnoll:2008rs}\cite{Bobev:2009mw}\cite{Donos:2009xc}\cite{Donos:2009zf} as well as in the M-theory \cite{Donos:2009xc}\cite{Ooguri:2009cv}\cite{Jeong:2009aa}. Unfortunately, we do not know the corresponding gauge theories except in some specific cases where the theory is supposed to be obtained from the discrete light cone quantization (DLCQ) of the $\mathcal{N}=4$ super Yang-Mills theory. The DLCQ is notoriously difficult to study, so in practice, we do not have any calculable Lagrangian description of the dual field theory for the non-relativistic AdS/CFT correspondence.
In this paper, as a first step to understand the nature of the non-relativistic AdS/CFT correspondence, we study the universal time-dependent deformations of the simplest Schr\"odinger invariant geometry. Our solutions are exact and universal in the sense that we can embed them in any supergravity constructions of Schr\"odinger invariant geometry. We study the field theory interpretations of our deformations, and we claim that the field theories dual to the Schr\"odinger invariant geometry should always admit such time-dependent exact deformations of the action, and they should also possess the states corresponding to our time-dependent solutions. The analysis furthermore reveals that the dual field theories always include certain operators that are not contained in the minimal operator contents of the Schr\"odinger invariant field theory. In other words, the Schr\"odinger invariant field theories that have a gravity dual predict an existence of these particular operators.
As a spin-off of our results, we argue that any time-dependent chemical potential can be treated exactly in our gravity dual approach. The time-dependent chemical potential turns out to be simply the time-dependent coordinate transformation of the bulk theory.
Our exact time-dependent deformations will show PP-wave singularities from the bulk gravitational theory viewpoint. It would be interesting to understand the nature of the singularity from the field theory perspective. Since we have not succeeded in finding the complete dual field theories, we cannot say much about the fate of the singularity from the field theory viewpoint. We hope that once the non-relativistic AdS/CFT correspondence is much better established, we would be able to attach the resolutions of PP-wave singularities from the dual non-relativistic field theory viewpoint. We leave this important issues for the future study.
\section{Universal Deformations of Schr\"odinger space-time}
We begin with the geometry with the Schr\"odinger invariance \cite{Son:2008ye}\cite{Balasubramanian:2008dm}:
\begin{align}
ds_{d+3}^2 = -2 \frac{dt^2}{z^4} + \frac{-2dt d\xi + dx_i^2 + dz^2}{z^2} \ , \label{schr}
\end{align}
where $i=1,\cdots d$. The light-like $\xi$ direction is compactified as $\xi \simeq \xi + 2\pi R$ so that the spectrum reproduces the quantization of the particle number in the dual field theory.
The metric \eqref{schr} is the solution of the Einstein equation with the massive (Proca) vector field
\begin{align}
S = \int d^d x_i d\xi dt dz \sqrt{-g} \left(\frac{1}{2}R - \Lambda - \frac{1}{4} F_{\mu\nu} F^{\mu\nu} - \frac{m^2}{2}A_\mu A^\mu \right) \ . \label{action}
\end{align}
where $F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu$. One can show that $A = A_\mu dx^\mu = -\frac{dt}{z^2}$ solves the equation of motion as well as the Einstein equation, provided
\begin{align}
\Lambda = -\frac{1}{2}(d+1)(d+2) \ , \ \ m^2 = 2(d+2) \ .
\end{align}
The geometry has an obvious invariance under the translation in $(t,x^i)$ as well as the Euclidean rotation in $x^i$. It is also invariant under the Galilean boost
\begin{align} x^i \to x^i - v^it \ , \ \ \xi \to \xi - v^i x^i + \frac{1}{2} v^2 t \ .
\end{align}
Furthermore, the geometry has the full non-relativistic conformal invariance \cite{Hagen:1972pd}\cite{Niederer:1972zz}.
The dilatation is generated by
\begin{align}
t \to \lambda^2 t \ , \ \ x^i \to \lambda x^i \ , \ \ z \to \lambda z\ , \ \ \xi \to \xi
. \label{dilatation}
\end{align}
The non-relativistic special conformal transformation is generated by
\begin{align}
t\to \frac{t}{1+at}\ , \ \ x^i \to \frac{x^i}{1+at}\ , \ \ z\to \frac{z}{1+at} \ , \ \ \xi \to \xi -\frac{a}{2}\frac{x^ix^i +z^2}{1+at} \ . \label{conformal}
\end{align}
\cite{SchaferNameki:2009xr} argues that the metric \eqref{schr} is the simplest geometrical realization of the Schr\"odinger invariance within the coset space construction.
The geometry is proposed to be dual to a non-relativistic Schr\"odinger invariant field theory. So far, there is no concrete proposal for what is the precise dual field theory corresponding to the geometry. It is suggested in \cite{Son:2008ye} that it would be obtained by a relevant deformation of the relativistic CFT dual to the $AdS_5$ space:\footnote{Note that the time $t$ here is a light-cone time $x^+$ in the original relativistic AdS/CFT coordinate.}
\begin{align}
\delta S_{\text {CFT}} = J\int d^{d+2}x O^t \label{dGKP}
\end{align}
after compactifying a light-cone direction because the GKPW prescription dictates that the deformation \eqref{dGKP} is induced by the Proca field $A_\mu$. The relativistic scaling dimension of $O^\mu$ is $d+2$, where $O^\mu$ is dual to $A_\mu$.
One of the main objectives of this paper is to understand the physics of this Schr\"odinger invariant geometry by introducing the exact time-dependent deformations. In section 2.1, we first study the scale invariant but non-conformal deformations, and in section 2.2, we study the conformal but non-scale invariant deformations. The both deformations sound peculiar from our experience in relativistic field theories: indeed, we would like to argue that they are special features of time-dependent non-relativistic field theories. In section 2.3, we extend our analysis to broader classes of exact time-dependent solutions.
\subsection{Scale invariant but non-conformal deformations}
We begin with the scale invariant but non-conformal deformations of the Schr\"odinger geometry. It was shown \cite{Nakayama:2009ww} that this is impossible without breaking further symmetries, so we investigate the solutions that are time-dependent explicitly.
The most general scale invariant deformations (up to a coordinate transformation) are given by
\begin{align}
ds_{d+3}^2 = -C\left(\frac{t}{z^2}\right) \frac{dt^2}{z^4} + D\left(\frac{t}{z^2}\right)\frac{-2dt d\xi + dx_i^2 + dz^2}{z^2}
\end{align}
for the metric and
\begin{align}
A = A_\mu dx^\mu = -E\left(\frac{t}{z^2}\right)\frac{dt}{z^2} \ ,
\end{align}
for the vector field. It is easy to see that $t/z^2$ is invariant under the scale transformation \eqref{dilatation} but not under the non-relativistic special conformal transformation \eqref{conformal}.
We have used the diffeomorphism invariance to remove the $dzdt$ component of the metric.
The $(zt)$ component of the Einstein equation tells us $D = \text{const} $, so we set $D=1$.
The other equations of motion can be solved exactly by
\begin{align}
C &= 2 + C_1 \frac{z^2}{t} + C_2 \left(\frac{z^2}{t}\right)^{d/2+2} + \epsilon^2\frac{d+2}{d+3} \frac{z^{2d+8}}{t^{d+4}} \cr
E &= 1 + \epsilon \frac{z^{d+4}}{t^{d/2+2}} \ ,
\end{align}
where $C_1$, $C_2$ and $\epsilon$ are integration constants.
The deformations by $C_1$ and $C_2$ do not depend on the background vector field, and they even exist without introducing the vector field ({\it e.g.} in locally AdS background).
Such PP-wave deformations in the AdS space was studied in \cite{Chamblin:1999cj}\cite{Brecher:2000pa}\cite{Kumar:2004jv}\cite{Sfetsos:2005bi}.\footnote{See also \cite{Chu:2006pa}\cite{Das:2006dz}\cite{Das:2006pw} for related null deformed backgrounds.} The metric is given by
\begin{align}
ds_{d+3}^2 = -\left( C_1 \frac{z^2}{t} + C_2 \left(\frac{z^2}{t}\right)^{d/2+2}\right) \frac{dt^2}{z^4} + \frac{-2dt d\xi + dx_i^2 + dz^2}{z^2} \ ,
\end{align}
which solves the vacuum Einstein equation with the negative cosmological constant.
In the AdS case, it was argued in \cite{Brecher:2000pa} that the deformation by $C_2$ is dual to the time-dependent vacuum expectation value (VEV) for a particular component of the energy momentum tensor:\footnote{Again, note that $t$ here corresponds to the light-cone direction $x^+$ in the original relativistic AdS/CFT coordinate.}
\begin{align}
\langle T_{tt} \rangle = \frac{C_2}{t^{d/2 + 2}} \ .
\end{align}
Later, we will discuss the similar operator has a VEV in the dual Schr\"odinger invariant field theory.
It is sometimes believed that the scale invariance implies the conformal invariance. This is not always true because there is no symmetric reason why this is so \cite{Coleman:1970je} a-priori, and indeed there are some known counterexamples \cite{Hull:1985rc}\cite{Riva:2005gd}\cite{Chu:2006pa}. However, it was shown that in $(1+1)$ dimension, the Poincar\'e invariance, unitarity, and the discreteness of the spectrum guarantees the equivalence between the conformal invariance and the scale invariance \cite{Zamolodchikov:1986gt}\cite{Polchinski:1987dy}. It is hoped that a similar statement should hold in higher dimensions with a suitable generalization of the assumptions \cite{Nakayama:2009qu}\cite{Nakayama:2009fe}\cite{Dorigoni:2009ra}.
In the unitary Sch\"odinger invariant field theories, it is conjectured that the scale invariance together with the Galilean invariance, rotation and translational invariance would imply the full non-relativistic conformal invariance \cite{Nakayama:2009ww}. It is easy to see that there is no such time-independent deformations in the above simple Sch\"odinger invariant geometry.\footnote{In higer dimensional supergravity embeddings, there could exist other terms that are scale invariant but not conformal invariant. We also note that the discussion here specializes in the case with the dynamical critical exponent $\mathcal{Z}=2$. When $\mathcal{Z} \neq 2$, the non-relativistic special conformal transformation cannot be constructed in the algebra. See \cite{Adams:2008wt}\cite{Hartnoll:2008rs}\cite{Singh:2009tq}\cite{Donos:2009xc} for examples of such geometries.} On the other hand, in the discussion above, the time-translational invariance is explicitly broken so that the conjecture does not apply.
\subsection{Conformal but not scale invariant deformations}
Now, we will present a more peculiar situation where the solution is invariant under the non-relativistic conformal transformation but not invariant under the dilatation. Such a geometry is impossible without breaking a further symmetry, and here again, we consider explicitly time-dependent solutions.
The most general non-relativistic conformal invariant deformations of the Schr\"odinger geometry are given by
\begin{align}
ds_{d+3}^2 = -C\left(\frac{t}{z}\right) \frac{dt^2}{z^4} + D\left(\frac{t}{z}\right)\frac{-2dt d\xi + dx_i^2 + dz^2}{z^2}
\end{align}
for the metric and
\begin{align}
A = -E\left(\frac{t}{z}\right)\frac{dt}{z^2} \ ,
\end{align}
for the vector field. It is easy to see that $t/z$ is invariant under the non-relativistic special conformal transformation \eqref{conformal} but not under the scale transformation \eqref{dilatation}. Again, we have used the diffeomorphism invariance to remove the $dzdt$ component of the metric.
The $(zt)$ component of the Einstein equation tells us $D = \text{const} $, so we set $D=1$.
The other equations of motion can be solved exactly by
\begin{align}
C &= 2 + C_1 \frac{z^2}{t^2} + C_2 \left(\frac{z}{t}\right)^{d+4} + \epsilon^2\frac{d+2}{d+3} \frac{z^{2d+8}}{t^{2d+8}} \cr
E &= 1 + \epsilon \frac{z^{d+4}}{t^{d+4}} \ ,
\end{align}
where $C_1$, $C_2$ and $\epsilon$ are integration constants. The deformations by $C_1$ and $C_2$ do not depend on the background vector field, and they even exist without introducing the vector field ({\it e.g.} in locally AdS background).
Such PP-wave deformations in the AdS space was studied in \cite{Chamblin:1999cj}\cite{Brecher:2000pa}\cite{Kumar:2004jv}\cite{Sfetsos:2005bi}. The metric is given by
\begin{align}
ds_{d+3}^2 = -\left( C_1 \frac{z^2}{t^2} + C_2 \left(\frac{z}{t}\right)^{d+4}\right) \frac{dt^2}{z^4} + \frac{-2dt d\xi + dx_i^2 + dz^2}{z^2} \ .
\end{align}
It was argued \cite{Brecher:2000pa} that the deformation by $C_2$ is due to the time-dependent VEV for a particular component of the energy momentum tensor:
\begin{align}
\langle T_{tt} \rangle = \frac{C_2}{t^{d + 4}} \ .
\end{align}
The closure of the non-relativistic conformal algebra ({\it i.e.} $i[H,K] = D$: see appendix A for details) demands that the non-relativistic special conformal invariance implies the scale invariance. Thus, it is impossible to break the scale invariance without spoiling the special conformal invariance. Here, the explicit time-dependence alleviates the situation, and without the conserved Hamiltonian, the conformal invariance without the dilation is possible.
\subsection{More general deformations}
Although the symmetry is more reduced, our deformations in the previous subsections are particular cases of more general time-dependent exact solutions of the simplest action \eqref{action} that allows the Schr\"odinger invariant geometry.
The more general time-dependent solutions are given by
\begin{align}
ds_{d+3}^2 = -C(t,z) \frac{dt^2}{z^4} + D(t,z)\frac{-2dt d\xi + dx_i^2 + dz^2}{z^2}
\end{align}
for the metric and
\begin{align}
A = -E(t,z)\frac{dt}{z^2} \ ,
\end{align}
where
\begin{align}
C(t,z) &= 2 + z^2 c_1(t) + c_2(t) z^{d+4} + \epsilon(t)^2 \frac{d+2}{d+3} z^{2(d+4)} \cr
D(t,z) &= 1 \cr
E(t,z) &= 1 + \epsilon(t) z^{d+4} \ .
\end{align}
$c_1(t)$, $c_2(t)$ and $\epsilon(t)$ are {\it arbitrary} functions of $t$.
There are several special choices of these functions:
\begin{itemize}
\item scale invariance: $c_1(t) = \frac{1}{t}$, $c_2(t) = \frac{1}{t^{\frac{d+4}{2}}}$ , $\epsilon(t) = \frac{1}{t^{d+4}}$ \ .
\item conformal invariance: $c_1(t) = \frac{1}{t^2}$, $c_2(t) = \frac{1}{t^{{d+4}}}$ , $\epsilon(t) = \frac{1}{t^{2(d+4)}}$
\item scale and conformal invariance: $c_1(t) = \delta(t)$.
\end{itemize}
We have discussed the first two choices in the previous subsections. The last one may need a comment. The solution contains a delta function and it looks singular, but in our approach, this is the only possible solution that is scale invariant and conformal invariant at the same time.
As we would like to discuss it later, however, the deformation by $c_1(t)$ can be gauged away by a coordinate transformation, so in practice, there is no non-trivial scale and conformal invariant time-dependent deformation of the Schr\"odinger invariant geometry.
The impossibility to obtain a scale and conformal invariant time-dependent solution may be regarded as a dual statement of the impossibility to construct a scale invariant but non-conformal field theory without breaking any translation invariance. The point is that the non-relativistic conformal algebra has a non-trivial involution anti-automorphism, which exchanges the Hamiltonain $H$ and the non-relativistic special conformal transformation $K$ (see appendix A for details). Therefore, from the representation theory viewpoint, it is very close to study the theory with no conserved Hamiltonian but invariant under all the other generators including the conformal transformation and to study the theory with no conformal invariance but invariant under all the other generators including the Hamiltonian.
Let us briefly discuss the singularity structure of the geometry. First of all, it is not difficult to see that all the curvature invariants are constant irrespective of the deformations. However, there is a PP-wave singularity at $z=0$ as well as $t=0$ (or $t = \pm \infty$) when $c_2(t)$ or $\epsilon(t)$ becomes infinite as $t\to 0$ (or $t = \pm \infty$).\footnote{The apparent singularity due to $c_1(t)$ can be removed by the coordinate transformation. We will give a field theory interpretation of this removal of singularity later in section 3.} In particular, the latter condition applies at $t=0$ to the scale/conformal invariant deformations discussed in section 2.2 and 2.3.
Since our deformations vanish as $z\to 0$, the singularity at $z=0$ is the same as that of the original Sch\"odinger space-time. See for example \cite{Hartnoll:2008rs} for the analysis of the singularity at $z=0$. On the other hand, the analysis of the PP-wave singularity structure has been done in the vacuum AdS PP-wave geometry in \cite{Chamblin:1999cj}. Our metric is within the analysis done in \cite{Chamblin:1999cj} except that our solution does not satisfy the vacuum Einstein equation but rather it has the source term from the Proca field. The PP-wave singularity appears whenever
\begin{align}
\mathcal{A} = z^5\partial_z\left(\frac{\partial_z(z^2C(t,z))}{z}\right)
\end{align}
is diverging, and this is precisely the above condition ({\it i.e.} diverging $c_2(t)$, $\epsilon(t)$ at $t=0$). Again note that the deformation by $c_1(t)$ does not introduce any PP-wave singularity at all.
\section{Field theory interpretation}
In this section, we will study the field theory interpretation of the exact time-dependent deformations introduced in the last section. We use the natural generalization of the GKPW prescription \cite{Gubser:1998bc}\cite{Witten:1998qj} of the AdS/CFT correspondence to our non-relativistic setup. See \cite{McGreevy:2009xe}\cite{Son:2008ye}\cite{Fuertes:2009ex}\cite{Volovich:2009yh}\cite{Leigh:2009ck} for its successful applications in non-relativistic AdS/CFT correspondence.
The deformation by $c_1(t)$ corresponds to the introduction of the time-dependent chemical potential in the action:
\begin{align}
M^2\int dt d^{d}x c_1(t) \rho(t,x) \ , \label{chemi}
\end{align}
where $\rho(t,x)$ is the particle number density. For instance, in the free Schr\"odinger theory with the action
\begin{align}
S = \int dt d^d x \left(i\Phi^* \partial_t \Phi - \frac{1}{2M}\partial_i \Phi^* \partial_i \Phi \right) \ ,
\end{align}
the particle number density is given by $\rho = \Phi^\dagger \Phi$. This identification was first proposed in \cite{Son:2008ye}, and we will confirm the identification by comparing the field theory expectation and the gravity prediction. In the local Schr\"odinger invariant field theories, such an operator always exists because it generates the particle number operator $N = \int d^d x \rho$ that lies in the non-relativistic conformal algebra (see Appendix A).
In the free Schr\"odinger theory, the introduction of this deformation modifies the two-point function as
\begin{align}
\left\langle \Phi(t_1,x_1) \Phi^\dagger(t_2,x_2) \right\rangle = \frac{1}{(t_1-t_2)^{d/4}} \exp\left(iM\frac{(x_1-x_2)^2}{t_1-t_2} + i M^2\int_{t_2}^{t_1} dt c_1(t) \right) \ \label{tptf}
\end{align}
up to a proportional factor.
It is important to notice that in the simple Schr\"odinger invariant field theories with Lagrangian description, the time-dependent chemical potential \eqref{chemi} can be exactly treated. The introduction of \eqref{chemi} is equivalent to the field redefinition $\Phi(t,x_i) \to e^{iM^2\int^t ds c_1(s)} \Phi(t,x_i)$. We will argue this is also true whenever the theories have a gravity dual.
To probe the implication of the gravity deformation by $c_1(t)$, let us consider a minimally coupled scalar field $\phi$ with the gravity:
\begin{align}
S_{\phi} = \int dt d\xi dzd^{d} x \sqrt{-g}\left( \partial^\mu \phi^* \partial_\mu \phi - m_0^2 |\phi|^2 \right) \ .
\end{align}
The equation of motion for $\phi$ is given by
\begin{align}
\partial_z^2 \phi - \frac{1}{z}(d+2) \partial_z \phi + \left(2iM\partial_t - k^2 - M^2c(t) - \frac{m^2}{z^2} \right) \phi = 0 ,
\end{align}
where $M$ is the momentum eigenvalue in $\xi$ direction, and $k_i$ is that for the $x^i$ directions: $\phi(z,t,x,\xi) = \phi(z,t) e^{ikx+iM \xi}$. $m^2 = m_0^2 + 4M^2$ is the effective mass of the KK mode in the light-cone direction $\xi$.
We can use the technique of separation of variables to solve the scalar equation of motion. It turns out that the superpositions of the wavefunction:
\begin{align}
\phi = z^{d/2 +1} K_\nu(pz) e^{-iEt+iM^2\int^t ds c_1(s)} \ ,
\end{align}
where $p = \sqrt{k^2-2ME}$, will give a complete basis of the solution.
In practice, this simply modifies the time-dependence of the scalar field by the phase factor $e^{iM^2\int^t ds c_1(s)}$ compared with the undeformed $c_1(t)=0$ case. Explicitly, the two-point function can be computed as
\begin{align}
\left\langle O(t_1,x_1) O^\dagger(t_2,x_2) \right \rangle &= \int dE dp e^{-iE(t_1-t_2) + ip(x_1-x_2)} (2ME-p^2)^\nu e^{iM^2\int_{t_1}^{t_2} dt c_1(t) } \cr
& = \frac{C_{\nu}}{({t_1-t_2})^{\frac{d+2}{2}+\nu}} \exp\left(iM\frac{(x_1-x_2)^2}{t_1-t_2} + i M^2\int_{t_2}^{t_1} dt c_1(t) \right) \ ,
\end{align}
where $C_\nu$ is a (cut-off dependent) constant.
We note the exact agreement with \eqref{tptf}.
This simplicity of the correlation functions actually suggests that the deformation by $c_1(t)$ is only apparent. To see it, we note that the coordinate transformation $\xi \to \xi - \frac{1}{2}\int^t ds c_1(s)$ will lead us to the original metric. It means that the time-dependent chemical potential in the Schr\"odinger invariant theory is always solved by the coordinate transformation in the dual gravity.
We emphasize that even though in the simplest field theory examples, the time-dependent chemical potential can be introduced/removed by the time-dependent field redefinition $\Phi(x,t)\to e^{iM^2\int^t ds c_1(s)} \Phi(x,t)$, it is not at all obvious this is always the case, when the system is strongly coupled and the Schr\"odinger symmetry is rather emergent. We have, on the other hand, showed that it is always possible to introduce the time-dependent chemical potential in the gravity approach. This will further constrain the candidates of the field theory duals.
The interpretation of $c_2(t)$ is given by assigning the VEV to a certain operator in the dual field theory:
\begin{align}
\langle T_{tt} \rangle = c_2(t) \ .
\end{align}
The scaling dimension of $T_{tt}$ is $d+4$. This operator does not always exist in non-relativistic Schr\"odinger invariant field theory, but the geometric construction requires the existence. The reason why we call it $T_{tt}$ is that it is exactly the $tt$ component of the energy-momentum tensor in the DLCQ construction of the Schr\"odinger invariant field theories. In the free Schr\"odinger theory, it is given by $\partial_t\Phi^\dagger \partial_t \Phi$.
Finally, $\epsilon(t)$ corresponds to the VEV of a certain (non-universal) operator of the non-relativistic CFT in the first oder approximation:\footnote{It looks peculiar that there is no corresponding (dual) operator insertion in the action. That would correspond to just the change of scale for $t$. Equivalently, the universal deformation by $\int dt d^dx \mathcal{H}$ does this job, where $\mathcal{H} = T_{00}$ is the Hamiltonian density. See the discussion in the following. Note that this $T_{00}$ is different from $T_{tt}$ introduced above. In the free Schr\"odinger theory, $T_{00} = \frac{1}{2m}\partial_i\Phi^\dagger\partial_i \Phi$. }
\begin{align}
\langle O_{t} \rangle = \epsilon(t) \ .
\end{align}
The scaling dimension of $O_t$ is $2(d+4)$.
In the DLCQ of the relativistic conformal field theory, we have discussed how $O^\mu$ was introduced to generate the Schr\"odinger invariant geometry. Here, the same operator has been given the VEV. Later, we will discuss the supergravity embedding of our deformations, and there, we will see the origin of the effective vector field corresponding to $O_t$. Various form fields in supergravity play the role of the operator.
So far, we have discussed the field theory interpretation of the exact time-dependent deformations of the simplest Schr\"odinger invariant field theory. Now, we would like to reverse the logic and ask the question: what would we expect for the universal deformations from the field theory viewpoint?
The minimal ingredient of the Schr\"odinger invariant field theory admits two universal deformations that preserve the Galilean boost and rotational invariance. They are given by the density operator $\rho$ and the trace of the energy momentum tensor $T^i_{\ i} = 2T_{00}$. The latter, however, only changes the normalization of the kinetic term with respect to the potential terms, so the universal deformations are simply given by adding the chemical potential in agreement with the gravity discussion above.
The other two deformations by $c_2(t)$ and $\epsilon(t)$ are not in the minimal list of the non-relativistic conformal field theory. Thus, the non-relativistic conformal field theory dual to the simplest Schr\"odinger invariant predicts the extra existence of the universal operators. In later section 4, we will discuss how these universal deformations appear in the general supergravity embedding of the Schr\"odinger invariant geometry.
\subsection{more on the correlation functions}
We have studied the exact modifications of correlation functions induced by $c_1(t)$.
The deformations given by $c_2(t)$ and $\epsilon(t)$ are more subtle than the deformation by $c_1(t)$ in terms of the correlation functions. These deformation will only affect the IR behavior of the scalar field in the geometry. To understand the situation, let us again consider the minimally coupled scalar field under the $c_2(t)$ deformation. The same analysis applies for $\epsilon(t)$ deformation. The equation of motion is
\begin{align}
\partial_z^2 \phi - \frac{1}{z}(d+1) \partial_z \phi + \left(2iM\partial_t - k^2 - M^2c_2(t)z^{d+2} - \frac{m^2}{z^2} \right) \phi = 0 . \label{seqq}
\end{align}
The UV behavior ({\it i.e.} $z\to 0$) of the scalar field is the same as in the undeformed case:
\begin{align}
\phi \sim e^{ipx-iwt}z^{d/2+1} \left(\frac{I_{-\nu}(|k|z) +A(k)I_\nu(|k|z)}{I_{-\nu}(|k|\epsilon) +A(k)I_\nu(|k|\epsilon)} \right) \ ,
\end{align}
where $\epsilon$ is the cutoff, $k^2 = p^2- 2M\omega$, and $\nu = \frac{1}{2}\sqrt{(d+2)^2 + 4m^2}$. In the undeformed case with $c_2(t) = 0$, the regularity of the Wick-rotated wavefunction at $z\to \infty$ determines $A(k)$ so that $\phi$ is given by $K_\nu(|k|z)$. Here the equation itself is modified toward $z\to \infty$ and in addition we do not have a good guiding principle to set the boundary condition there in our explicitly time-dependent setup.
To investigate the behavior $z\to \infty$ and study the boundary condition, we need to solve the equation toward $z\to \infty$.
Unfortunately, there is no analytic solution of the equation \eqref{seqq} except for $M=0$. When $M=0$, the solution is uniquely given by
\begin{align}
\phi \sim e^{ipx-iwt}z^{d/2+1} \frac{K_\nu(|p|z)}{K_\nu(|p|\epsilon)} \ .
\end{align}
The wavefunction with $M=0$ corresponds to the zero-norm state of the field theory, and the two-point functions between such states are not affected at all by the deformation $c_2(t)$ (and $\epsilon(t)$ by repeating the same argument). See \cite{Nakayama:2009ed} for the discussions on the zero-norm states in the Schr\"odinger invariant field theories and their gravity dual.
Instead of giving the analytic solutions of \eqref{seqq}, we can study the power series solution:
\begin{align}
\phi = z^{d+2-\Delta}\sum_{n=0}^{\infty} a_n(t)e^{ikx} z^n \ + z^{\Delta}\sum_{n=0}^{\infty} b_n(t) e^{ikx} z^n \ ,
\end{align}
where $\Delta$ is related to the conformal dimension of the operator and it is given by $\Delta = \frac{d+2}{2}+ \nu = \frac{d+2+\sqrt{(d+2)^2 + 4m^2}}{2}$.
The equations of motion will determine the higher powers $a_n(t)$ and $b_n(t)$ from the boundary data $a_0(t)$ and $b_0(t)$. For instance,
\begin{align}
a_2(t) = -\frac{2iM\partial_ta_0(t)-k^2 a_0(t)}{4(\nu-1)} \ , \ \ b_2(t) = -\frac{2iM\partial_t b_0(t)-k^2 b_0(t)}{4(\nu+1)} \ ,
\end{align}
and the higher terms are determined recursively.
The series coincide with the power expansion of the Bessel function, up to the $(d+2)$-th order. At the $(d+4)$-th order, they show a deviation:
\begin{align}
\delta a_{d+4}(t) &= \frac{M^2 c_2(t)}{2(\Delta-d-3)(d+4)} a_0(t) \cr
\delta b_{d+4}(t) &=-\frac{M^2 c_2(t)}{2(\Delta +1)(d+4)} b_0(t) \ .
\end{align}
Similarly, with the $\epsilon(t)^2$ deformation of the metric, the deviation appears at the $2d+8$-th order:
\begin{align}
\delta a_{2d+8}(t)&= -\frac{M^2 \epsilon(t)^2(d+2)}{2(d+3)(d+4)(3d+10-2\Delta)} a_0(t) \cr
\delta b_{2d+8}(t) &=-\frac{M^2 \epsilon(t)^2(d+2)}{2(d+3)(d+6)(d+4+\Delta)} b_0(t)
\end{align}
by assuming that the vector field does not affect the equation of motion for the scalar.
The special case of the Kaigorodov space \cite{Kaigorodov1}\cite{Kaigorodov2}, where $c_2(t) = c_2 = \text{const}$ was discussed in \cite{Brecher:2000pa}. It was argued that the two-point function does not show any {\it explicit} dependence on $c_2 $ up to contact terms, so we can {\it choose} the same boundary condition as in the case $c_2(t) = 0$, which will yield the independence of the two-point function on $c_2$. Similarly, when $\epsilon(t)$ is a constant, the deformation only gives {\it explicit} corrections to the contact terms, so we can {\it choose} the same boundary condition as in the case $\epsilon(t) = 0$, and we obtain no dependence on $\epsilon$ in the two-point functions.
Formally, one can study the perturbative corrections to the two-point functions by using the Witten diagram and computing the perturbative correction from the ``interaction" $\int dt dz d^{d}x c_2(t) z M^2|\phi|^2$. We do not know how to compute the same perturbative corrections from the field theory side because the expectation value of $T_{tt}$ alone does not seem to specify the ``state" we evaluate the correlation function. Even in the time-independent Kaigorodov case, we do not know the precise prescription to compute the correlation functions from the field theory side. Further studies are needed in this direction to reveal the nature of $c_2(t)$ and $\epsilon(t)$ deformations.
\section{Supergravity embedding}
\subsection{General argument}
We have discussed the three independent universal deformations ({\it i.e.} $c_1(t)$, $c_2(t)$ and $\epsilon(t)$) of the simplest Schr\"odinger geometry given by the metric \eqref{schr}, which is supported by a free Proca field with the definite mass. It would be interesting to see whether such deformations can be embedded in the string/M-theory solutions.
There are several supergravity solutions that possess the Schr\"odinger isometry with \cite{Hartnoll:2008rs}\cite{Bobev:2009mw}\cite{Donos:2009xc}\cite{Ooguri:2009cv}\cite{Donos:2009zf}\cite{Jeong:2009aa} or without supersymmetries \cite{Herzog:2008wg}\cite{Maldacena:2008wh}\cite{Adams:2008wt}. The simplest one is to consider the DLCQ of the AdS space. The compactification of the light-cone direction renders the isometry of the AdS space down to the Schr\"odinger group. In that case, as we have mentioned in the previous section, our deformations can also be applied in the DLCQ of the AdS space. We, however, prefer the geometry with $\frac{dt^2}{z^4}$ term as in \eqref{schr}.
We claim that our deformations are universal in the sense that they can be applied to any (known) solutions with the Schr\"odinger invariance embedded in the supergravity. In this sense, our exact time-deformations are universal and the dual field theory interpretations should be valid in all the field theory duals of such geometries. In other words, the existence of such exact time-dependent deformations are common features of the dual field theories.
We begin with the deformation given by $c_1(t)$. As we observed, the deformation by $c_1(t)$ is actually locally trivial (up to a possible change of the boundary condition). It is simply induced by the coordinate transformation of the geometry $\xi \to \xi + \frac{1}{2}\int^t ds c_1(s)$. Since this is just a coordinate transformation, there is no obstacle to do it in any supergravity embeddings. Obviously we can do the same thing in the DLCQ of the AdS compactification.
Now let us consider the non-trivial deformation given by $c_2(t)$. In the AdS compactifications of the supergravity, such a deformation, a generalization of the Kaigorodov space, was studied in \cite{Brecher:2000pa}\cite{Chu:2006pa}, where the deformation by $c_2(t)$ does not change the other equations of motion except for the $(tt)$ component of the Einstein equations which are solved by assigning a definite power of $z$. Similarly, in our case, it can be shown that the deformation does not change the equations of motion except for the $(tt)$ component of the Einstein equation in any supergravity compactification. Again, the $(tt)$ component of the Einstein equation is solved by assigning a definite power of $z$. Thus, the generalized time-dependent Kaigorodov deformation by $c_2(t)$ gives the exact deformations of the any known supergravity solutions with the Sch\"odinger invariance. We also note that the deformation by $c_2(t)$ does not change the equations motion for flux with the ansatz that is compatible with the Sch\"odinger invariance. We will explicitly see the structure below when we discuss a concrete example.
The deformations by $c_1(t)$ and $c_2(t)$ only deal with the metric. Now, we consider the deformation given by the vector field $\epsilon(t)$. In the typical supergravity embedding, the vector field in the effective $d+2$ dimensional compactifications comes from the flux in the supergravity action. The effective mass term comes from the non-trivial eigenvalue for the internal Laplacian for the flux. The details will vary with respect to the origin of the vector field, but quite generally we can find the solutions of the supergravity equations motion corresponding to the deformation $\epsilon(t)$ in the following manner.
We first investigate the equation of motion for the flux. Thanks to the Schr\"odinger invariance and our ansatz for the time-dependent flux, $g_{tt}$ component of the metric does not affect the flux equation motion at all. Furthermore, the time-dependence in the flux does not affect the equations of motion, so we regard the time-dependence of the flux as if it were constant. This agrees with our feasibility to introduce an arbitrary function $\epsilon(t)$ in the vector field. The flux equation motion, as a consequence, determines the $z$ dependence of the flux ansatz as in our effective field theory approach in section 2.
The introduction of the time-dependence in the flux (as $\epsilon(t)$) will back-react to the geometry in the Einstein equation. We note that the backreaction only affects the $(tt)$ component of the Einstein equation. In addition, the $(tt)$ component of the Einstein equation can be solved by introducing the $(tt)$ component of the metric deformations of order $\epsilon^2(t)$. A crucial point here is that this $(tt)$ component of the metric, which is PP-wave type, will not affect all the other components of Einstein equations as well as the flux equations of motion, so by a suitable choice of the $g_{tt}$ deformation, we are able to solve all the equations motion induced by the flux deformation induced by $\epsilon(t)$.
\subsection{Concrete example}
To make the above argument concrete, let us consider a particular M-theory compactification of the Schr\"odinger space $(d=1)$ based on the Sasaki-Einstein 7-fold. The original solution \cite{Donos:2009xc} is given by
\begin{align}
ds^2 &= \frac{1}{4}\left(q \frac{dt^2}{z^4} - \frac{2 dt d\xi}{z^2} + \frac{dx^2+dz^2}{z^2}\right) + ds^2(SE_7) \cr
G &= -\frac{3}{2^3z^4} dt \wedge d\xi \wedge dx \wedge dz + dt \wedge d\left(\frac{\tau}{2z^2}\right) \ , \label{gu}
\end{align}
where $\tau$ is a two-form on the Sasaki-Einstein 7-fold.
The flux equations of motion demand
\begin{align}
-*d* d\tau &= 24\tau \ , \ \ d*\tau = 0 \label{fleq} \ ,
\end{align}
where the Hodge star $*$ act on the internal space $SE_7$.
The $(tt)$ component of the Einstein equation demands
\begin{align}
D^i D_i q + 40q = -16|\tau|^2 - |d\tau|^2 \ ,
\end{align}
where $D_i$ is a covariant derivative on the internal space $SE_7$.
All the other equations motions are solved by \eqref{gu}.
It is clear that the first terms in $G$ gives the effective cosmological constant and the second term gives the effective massive vector field whose mass squared is determined by the flux equation of motion \eqref{fleq}.
In particular, for $SE_7=\mathbb{S}^7$, the explicit form of $\tau$ was constructed in \cite{Donos:2009xc} by splitting the parent $CY_4 = \mathbb{R}^8$ into $\mathbb{R}^8 = \mathbb{R}^4\times \mathbb{R}^4$ and considering a sum of terms which are $(1,1)$ with respect to the complex structure and primitive on one factor with a factor $dx^i$ on the other. In this simplest case, $q$ is a constant, which does not depend on the internal coordinate, but in general it may depend on the internal coordinate.
As discussed above, we can introduce the $c_1(t)$ and $c_2(t)$ deformation by simply replacing the metric with
\begin{align}
ds^2 &= \frac{1}{4}\left[\left(q - c_1(t) z^2 - c_2(t) z^5 \right)\frac{dt^2}{z^4} - \frac{2 dt d\xi}{z^2} + \frac{dx^2+dz^2}{z^2}\right] + ds^2(SE_7) \ . \label{eef}
\end{align}
The only equation of motion affected by the deformation is the $(tt)$ component of the Einstein equation, and it determines the power of $z$ appearing in the first term of \eqref{eef}.
As discussed in the last subsection, we can explicitly see that the introduction of $c_1(t)$ and $c_2(t)$ does not change the other equations motion.
We now studies the flux deformation.
\begin{align}
G &= -\frac{3}{2^3z^4} dt \wedge d\xi \wedge dx \wedge dz + dt \wedge d\left(E(t,z)\frac{\tau}{2 z^2}\right) \ .
\end{align}
Since we identified the second term in $G$ as the effective massive vector field in the Schr\"odinger geometry, the introduction of $E(t,z)$ is natural.
The 11-dimensional flux equations motion
\begin{align}
d*G + G\wedge G = 0
\end{align}
is insensitive to the $t$ dependence in $E(t,z)$. The $z$ dependence is uniquely fixed by
\begin{align}
E(t,z) = 1 + \epsilon(t)z^{5} \ .
\end{align}
together with \eqref{fleq}. Finally, the $(tt)$ component of the Einstein equation is now sourced by the new term coming from $\epsilon(t)$. This will be solved by introducing order $\epsilon^2(t)$ term in $g_{tt}$. The modification of $g_{tt}$ as well as the introduction of $\epsilon(t)$ in the flux do not affect all the other equations of motion so that we find the embedding of the solutions within the M-theory compactification.
It is clear that the above construction can be repeated in other string/M-theory realizations of the Schr\"odinger space-time. The only non-trivial part is to identify the effective massive vector field in the flux ansatz.
\section{Relativistic case}
In section 2.3, we have studied the time-dependent scale invariant and/or conformally invariant deformations of the Schr\"odinger invariant field theory.
A similar question can be addressed in the relativistic AdS/CFT correspondence. From the field theory viewpoint, it seems possible to deform the relativistic CFT so that only the translational invariance is broken while preserving dilation, special conformal transformation as well as Lorentz transformation. In the Lagrangian description, for instance, we may add the interaction
\begin{align}
\int d^d x (x^2) O(x) \ ,
\end{align}
where $O(x)$ is a primary scalar operator of conformal dimension $d+2$. At the first order, such deformations are scale and conformal invariant as well as Lorentz invariant, but they break the translational invariance due to the explicit $x^2$ dependence in the interaction.
Are there corresponding deformations in the CFT side? Surprisingly it seems very difficult to introduce such deformations contrary to the naive expectation from the field theory discussions. We first note that the special conformal transformation acts as ($a=0,1,\cdots ,d$)
\begin{align}
\delta x_a = 2(\epsilon^bx_b) x_a - (z^2 + x^bx_b) \epsilon_a \ , \ \ \delta z = 2(\epsilon^bx_b)z \ ,
\end{align}
so while $x^2/z^2$ is invariant under the dilation, it is not invariant under the special conformal transformation. In particular, any scalar field profile like $\phi = f(x^2/z^2)$ is invariant under the scale transformation, but not invariant under the special conformal transformation. Indeed, as a simple corollary, there would be no such scalar deformations possible from the bulk theory viewpoint.
While there is no invariant scalar perturbation, there are possible vector or metric perturbations. Let us consider the vector field profile
\begin{align}
A = A_{\mu}dx^\mu = K\left(\frac{x^2}{z^2}\right) \frac{dz}{z} + J\left(\frac{x^2}{z^2}\right) \frac{x_a dx^a}{z^2} \ . \label{vecd}
\end{align}
Again because of the non-invariance of $x^2/z^2$ under the special conformal transformation, each terms in \eqref{vecd} are not invariant under the special conformal transformation while it is obviously scale invariant.
In order for the deformation to be conformally invariant, $K(y)$ and $J(y)$ should satisfy
\begin{align}
K'(y) (1+y) + J(y) &= 0 \cr
2K(y) + J(y)(-1+y) &= 0 \cr
J'(y)(1+y) + J &= 0 \ .
\end{align}
This is an overdetermined system, but it has a solution:
\begin{align}
K(y) = \frac{c}{2} \frac{1-y}{1+y} \cr
J(y) = \frac{c}{1+y} \ ,
\end{align}
where $c$ is an integration constant.
Similarly, we can study the following metric perturbations:
\begin{align}
ds^2 = A\left(\frac{x^2}{z^2}\right) \frac{dz^2}{z^2} + B\left(\frac{x^2}{z^2}\right)\frac{dx^2}{z^2} + C\left(\frac{x^2}{z^2}\right) \frac{x_a dx^a dz}{z^3} + D\left(\frac{x^2}{z^2}\right) \frac{(x_a dx^a)^2}{z^4} \ .
\end{align}
The invariance under the special conformal transformation demands
\begin{align}
(1+y)A'(y) + C(y) &= 0 \cr
4A(y) - 4B(y) + (y-1)C(y) &= 0 \cr
(1+y)B'(y) &= 0 \cr
(1+y)C'(y) + C(y) + 2D(y) &= 0 \cr
C(y) + (-1+y) D(y) &= 0 \cr
(1+y)D'(y) + 2 D(y) & = 0 \ .
\end{align}
Again, the equations look overdetermined, but there is a nontrivial solution:
\begin{align}
A'(y) &= c'\frac{y-1}{(1+y)^3} \cr
B'(y) &= 0 \cr
C(y) &= c'\frac{1-y}{(1+y)^2} \cr
D(y) & = c'\frac{1}{(1+y)^2} \
\end{align}
with another integration constant $c'$.
We note that we have not imposed any equations of motion yet. In this sense, the solutions are quite restrictive. It is clear that arbitrary actions and equations of motion do not allow such a solution, and in general, there would be no scale and conformal invariant deformations of the relativistic conformal field theory that break the translational invariance from the gravity viewpoint. Again, this may be related to the fact that the representation theory of the relativistic conformal algebra has a non-trivial involution anti-automorphism replacing the momentum $P^\mu$ with the special conformal generator $K^\mu$, so it is likely that the difficulty to find a scale and conformal invariant but non-translational invariant deformations may be related to the difficulty to find a scale invariant but non-conformal deformations of the relativistic conformal field theories.
\section{Discussions and Summary}
What happens if we hit the singularity of the universe? Is it the end of the universe, or will the stringy effects remove it? How does the time begin or end? These are fundamental questions that should be addressed and hopefully answered in fundamental theories of quantum gravity. If the holography is one of the most fundamental nature of the gravity, the dual field theory approach would enable us to answer the question.
The AdS/CMP correspondence might give us a novel way to probe the singularities of the universe from our lab experiments. Our time-dependent deformations of the non-relativistic AdS/CFT correspondence contain the PP-wave singularity. Since they are universal, one may expect that they are realized in, for example, the cold atoms, or unitary fermion system.
In this paper, we have also clarified that the distinction between the scale invariance and the conformal invariance is more manifest in the time-dependent non-relativistic system. Although in the unitary Poincar\'e invariant field theories, or in the unitary Galilean invariant field theories, these two concepts might be equivalent, we have found otherwise in the time-dependent background from the gravity approach. It would be interesting to confirm our result from the field theory approach.
The subtle relation between the conformal invariance and the scale invariance is a fundamental question in theoretical physics that is yet to be solved. The probelm is one of the few good examples that can be shared by string theorists and the condensed matter physicists. We hope that further collaborations will give a theoretical as well as experimental clue to this elementary question that lies in the basic foundation of the world sheet formulation of the string theory, quantum gravity, as well as critical phenomena in condensed matter physics.
\section*{Acknowledgements}
The author would like to thank Gabriel Wong for the collaboration in the early stage of this work. He also thanks Ben Freivogel for the discussion of the PP-wave singularity.
The work was supported in part by the National Science Foundation under Grant No.\ PHY05-55662 and the UC Berkeley Center for Theoretical Physics and World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan.
|
\section{Introduction}
For over a century, Maxwell's demon has provided a setting in which to address the limitations that thermodynamics places on an observer free to perform measurements on a system and then act on their acquired information in some algorithmic way \cite{maxwell-1867, maxwell-1871, szilard-1929, vedral-review}. As is now well appreciated, information has an energetic value. A demon \footnote{Appropriately enough, Plato argued that the word `demon' comes from the Greek word \textit{da$\bar{e}$mones} meaning \textit{`one who has knowledge or wisdom'}.} in possession of information about a physical state may transform this information into mechanical work \cite{szilard-1929, zurek-quantum-szilard}. Specifically, given a system of dimension $D$ in a state $\rho$ and the presence of a reservoir at a temperature $T$, the demon can extract $W=kT(\ln D - S[\rho])$ amount of mechanical work from the reservoir, where $S[\rho]=- \mbox{Tr} [\rho \log \rho]$ is the von Neumann entropy of the state. The demon fails in his attempt to violate the second law of thermodynamics because the demon's memory, where he records the measurement results of the state, must be erased in order to operate in a cycle \cite{bennett82}. Such an erasure of memory can only occur with an unavoidable dissipation of heat \cite{landauer}.
As a concrete example, we may consider a gas of $N$ particles that resides in a piston on one side of a movable partition and in thermal equilibrium with a reservoir at a temperature $T$, Fig.~\ref{piston1}.
The demon, upon discovering what side of the partition the gas is on, may extract usable energy by deterministically acting on the piston system - if the gas is to the left of the partition, the demon puts the piston in contact with the reservoir, loads the partition with a mass $M$, tilts the piston as in Fig.~\ref{piston1} and allows the gas to expand slowly through the available empty volume. The expansion of the gas does work on the mass, elevating to a height $h$, which makes mechanical energy available. Indeed the expansion can provide a maximum mechanical energy of $NkT\ln 2$, which is extracted from the thermal reservoir.
\subsection{Confounding Maxwell's Demon}
In order to illustrate the new possibilities that quantum entanglement can bring to the story, it is amusing to consider cases in which a ``global demon'', with access to a large entangled quantum state, can confuse the traditional ``local demon'' who can only measure and act locally. The global demon can arrange that any thermodynamic process for the local demon can run ``backwards'' - heat can flow from a cold to hot and gases can contract instead of expand - driven by entanglement present in the state.
As a concrete example, consider a system that consists of a reservoir and piston that are both in local thermal states, but with the global state of the system being pure. This state is prepared by the global demon, who may manipulate the global state as he pleases. The global demon ensures that the entanglement is local, in the sense that any part of the total system is entangled with its surroundings. Furthermore, for simplicity, we constrain the reservoir subsystems to couple to each other much weaker than to the piston subsystem \footnote{Although we shall see later that this requirement is not needed.}. This situation is indicated in Fig.~\ref{piston1}.
Since thermality arises from entanglement between the different components of the system in a global pure state, when the local demon places the piston in contact with the quantum reservoir and releases the partition, instead of the gas expanding up the piston to fill the empty volume, the thermal state repeatedly expands forwards and then contracts backwards in the tube as energy flows into and out of the piston. The spontaneous contraction phase of the gas is identical to the usual spontaneous expansion of a gas but with time ``running backwards''.
For the local demon, the thermodynamic arrow of time has failed and he is unable to extract any net work despite having been given a piston that is on its own completely indistinguishable from a standard thermal state on one side of a partition, seemingly coupled to a standard thermal reservoir.
The mischievous global demon, however, may further confuse the local demon by thermalizing the entanglement in the (quantum) reservoir that he has access to, using a different (fully thermal, classical) reservoir. Since the full system of dimension $D$ is in a pure state, the global demon can extract $kT \ln D $ amount of work from the classical reservoir, which he may store as mechanical energy, leaving the quantum reservoir plus piston in a maximally mixed state. Then using only some of this mechanical energy he may convert the maximally mixed reservoir into a true thermal state at the original temperature $T$. At which point the thermodynamic arrow is returned, the global demon has converted the entanglement into mechanical energy and the local demon, unaware of all the entanglement that was present, will finally be able to extract $NkT \ln 2$ of work from the reservoir as he had originally hoped.
\begin{figure}[t]
\centering
\includegraphics[width=2cm]{piston1.eps}
\includegraphics[width=2cm]{piston2.eps}
\includegraphics[width=2cm]{piston3.eps}
\caption{\textbf{Uncorrelated case:} The piston is in contact with a thermal reservoir of particles, at temperature $T$. Mechanical energy is extracted from the reservoir by the demon choosing which way to tilt the piston, conditional on his information about what side the gas is located.}
\label{piston1}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3cm]{piston2v.eps}
\caption{\textbf{Entangled case:} The local thermodynamic states of the piston and reservoir are indistinguishable from the uncorrelated cases, however now entanglement is present in the system. In this case the piston both expands forwards and contracts backwards. Heat flows both into and out of the piston and the usual thermodynamic arrow no longer holds.}
\label{piston2}
\end{figure}
\subsection{The Arrow of Time}
There are many reasons to be interested in the arrow of time beyond simply fun stories about demons. The standard thermodynamic arrow applies to systems that are in isolation and amounts to a directionality for certain physical processes. The curious fact that certain processes are forbidden and certain states inaccessible is generally traced back to the initial conditions of the physical system. Furthermore, since all parts of the observed universe seem to obey a common ``arrow of time'', the problem of resolving the thermodynamic arrow's origin requires us to account for the particular initial conditions of our universe \cite{penrose-weyl-hyp}.
The notion of an ``arrow of time'' is quite broad - extending from the uniform large-scale expansion of the universe, to the radiative arrow, and through to the individual, psychological perception of time \cite{Zeh-book, Huw-price-book}. However, since the laws of physics are invariant under time reversal (strictly CPT invariant), it is believed that any form of temporal directionality will stem ultimately from the initial conditions of the universe.
Recent work by Partovi \cite{partovi} has considered a specific pure entangled state that produced heat flow from a cold body to a hotter one, contrary to the usual thermodynamic arrow. In \cite{partovi} Partovi identifies the large degree of entanglement present in the state as being responsible for this unusual behaviour. We will see, in fact, that the negative heat flow he considers does not necessarily depend on the presence of entanglement in the state and, by adopting an information theoretic framework, we explore these and similar possibilities as a way of elucidating the properties of entangled quantum states and the connection with varied notions of ``arrows of time''.
\subsection{The Thermodynamic Arrow and Correlations}
A paradigmatic setting that captures the essentials of the thermodynamic arrow of time is in the context of heat flow between two systems $A$ and $B$ that interact in isolation, so that their total energy is constant. For simplicity, we assume that they do no mechanical work on each other, but may exchange energy in the form of heat, for which conservation of energy implies that $Q_A+Q_B =0$, where $Q_A$ is the heat acquired by $A$ and $Q_B$ is the heat acquired by $B$.
Ultimately, the key ingredient that dictates which states are accessible via the set of all interactions on the total state $\rho_{AB}$ is the degree of correlations initially present between $A$ and $B$. These correlations between the two subsystems are quantified by the \textit{mutual information}, which is non-negative for all states and defined in terms of relative entropy as $I(A:B) = S[\rho_{AB} || \rho_A \otimes \rho_B]=S_A+S_B -S_{AB}$, where, for example, $S_A=- \mbox{Tr} [\rho_A \log \rho_A]$ is the von Neumann entropy for the state $\rho_A$. The mutual information is a convenient measure of how distinguishable the state $\rho_{AB}$ is from the completely uncorrelated product state $\rho_A \otimes \rho_B$.
Of particular interest is when the reduced states at the initial time $t_i$ of subsystems $A$ and $B$ are thermal states for their respective individual Hamiltonians, $H_A$ and $H_B$. Given temperatures $T_A$ and $T_B$, the reduced states then take the form $\rho_A (t_i)\propto \exp [-\beta_A H_A]$ and $\rho_B (t_i) \propto \exp [-\beta_B H_B]$, where $\beta_X =1/kT_X$ for subsystem $X\in\{A,B\}$.
We may consider the most general process of switching on time-dependent interactions $V_{AB}(t)$ between $A$ and $B$ that evolve the composite system into a new state $\rho_{AB} (t)$. For subsystem $X\in\{A,B\}$ at a later time $t$, the two thermodynamic variables of central interest are the average energy $U_X= \mbox{Tr} [H_X \rho_X(t)]$ and the entropy $S_X=- \mbox{Tr} [\rho_X(t) \log \rho_X(t)]$. Locally at $A$ the interaction results in the transformation $\rho_A(t_i) \rightarrow \rho_A(t_f) = \mbox{Tr} _B [\rho_{AB}(t_f)]$, and so $U_A \rightarrow U_A +\Delta U_A$ and $S_A \rightarrow S_A +\Delta S_A$, with similar expressions for $B$.
These changes in local thermodynamic variables are constrained by the fact that a thermal state $\exp [- \beta H]/ \mbox{Tr} [\exp[-\beta H]]$ minimizes the free energy function $F[\rho] = \mbox{Tr} [ \rho H] - S[\rho]/\beta$. Consequently we have that $ \mbox{Tr} [ \rho(t_i) H] - S[\rho(t_i)]/\beta< \mbox{Tr} [ \rho(t_f) H] - S[\rho(t_f)]/\beta$, which more explicitly yields
\begin{eqnarray}\label{freee}
\beta_A \Delta U_A - \Delta S_A &\ge& 0 \nonumber \\
\beta_B \Delta U_B - \Delta S_B &\ge& 0
\end{eqnarray}
for the change in the thermodynamic variables defined at $A$ and $B$. It must be emphasized that equations (\ref{freee}) are kinematical restrictions on \emph{any} transformation from initially thermal states, $\{\rho_A (t_i),\rho_B (t_i)$\}, to any other pair of local states $\{\rho_A(t_f),\rho_B(t_f)\}$.
We consider transformations on the composite system that satisfy $S[\rho_{AB}(t_i)] =S[\rho_{AB}(t_f)]$ (which includes unitary transformations) and so $\Delta I(A:B) = \Delta S_A +\Delta S_B$. The only other restriction that we place on the transformation is that $ \mbox{Tr} [ \rho_{AB} (t_i) (H_A+H_B)]= \mbox{Tr} [ \rho_{AB} (t_f) (H_A+H_B)]$. Since we don't consider mechanical work, this amounts to requiring that there is zero net heat flow into the composite system, and the total energy before the interaction process equals the total energy afterwards. Writing $Q_A= \Delta U_A$ and $Q_B= \Delta U_B$ for the heat gained by $A$ and $B$ respectively, we obtain for these energy preserving processes that
\begin{eqnarray}\label{flow}
\beta_A Q_A +\beta_B Q_B \ge \Delta I(A:B)
\end{eqnarray}
where $\Delta I(A:B)$ is the change in the mutual information between $A$ and $B$.
The special initial state $\rho_{AB}(t_i) = \rho_A(t_i) \otimes \rho_B(t_i)$ in which correlations vanish was historically called the condition of \textit{Stosszahlansatz}, or `molecular chaos'. If we consider $A$ and $B$ initially uncorrelated at some time $t_i$ then $I(A:B;t_i)=0$, however interactions allow correlations to form between the two subsystems, and so at later times $t_f$ we have $I(A:B;t_f) \ge 0$. We immediately deduce that $\beta_A Q_A +\beta_B Q_B \ge 0$, or more explicitly in terms of temperature
\begin{eqnarray}\label{uncorrelated}
Q_A \left ( \frac{1}{T_A} - \frac{1}{T_B}\right ) \ge 0.
\end{eqnarray}
In other words, heat can only flow from hot thermal states to cold thermal states in isolation. This is the standard thermodynamic arrow of time.
This situation can be generalized to $N$ initially uncorrelated thermal states and provides us with the constraint
\begin{eqnarray}\label{multiheat}
\sum _j \frac{Q_j}{T_j} \ge 0
\end{eqnarray}
which forbids certain types of heat flow; for example, if we partition the $N$ thermal states in two, such that all the temperatures in the first group are lower than those in the second, then it is impossible for heat to flow from the first to the second.
In general, however, initial correlations are expected to be present, $I(A:B;t_i) >0$, and so the change in mutual information can be negative. From (\ref{flow}) it is clear that there is no longer a constraint on the directionality of heat flow between $A$ and $B$ \cite{seth-loyd}. As we shall now see, the correlations that make up the total mutual information can arise from classical correlations, or from a combination of classical correlations and quantum entanglement.
\subsection{Entanglement correlations}\label{entanglementcorrelations}
Entangled quantum systems can possess far stronger correlations than are possible classically, and in fact the vast majority of quantum states are entangled \cite{Clifton-separablevol, peres-book}. Work has been done previously on how quantum correlations affect macroscopic thermodynamic observables such as the susceptibility or the heat capacity, and it has been shown that measurement of these macroscopic observables can act as entanglement witnesses \cite{witness-1, witness-2}. We shall show that the thermodynamic \emph{transformations} themselves also constitute entanglement witnesses. We should be careful to distinguish this research program from the one to do with with formal analogies \cite{HOH, vedral02} between irreversible entanglement transformations and the second law of thermodynamics, on which much progress has recently been made \cite{fernando-entanglement-theory}, or from attempts to connect the second law of thermodynamics to Bell inequalities \cite{durham08}. Here we are interested in the physical effects which entanglement between systems and reservoirs has on the thermodynamical transformations of the systems.
Recently in \cite{partovi} Partovi considered how the presence of entanglement in a system affects irreversible thermodynamic transformations.
The intriguing scenario he considered involves two subsystems $A$ and $B$ that are overall in a pure state $| \Psi _{AB} \>$ and possess strong quantum correlations. Furthermore, he arranged that the local reduced states $\rho_{A,B} = \mbox{Tr} _{B,A} [|\Psi_{AB} \> \< \Psi_{AB} |]$ are thermal states for the two subsystems and so locally indistinguishable from classical thermal mixtures. Despite being locally indistinguishable from classical thermal mixtures, the composite system behaves quite differently under any energy-conserving unitary. Since the total initial state is a pure state we have that $S[\rho_{AB}(t_i)]=0$, which remains true at any later time. From a Schmidt decomposition for $A$ and $B$ we find that $\rho_A(t)$ and $\rho_B(t)$ are isospectral, $\mbox{Spec}(\rho_A(t))=\mbox{Spec}(\rho_B(t))$, and so their entropies are always equal. Consequently, any unitary transformation occurring on the composite system will have $\Delta S_A = \Delta S_B$.
Once again, assuming (i) no mechanical work and (ii) no overall change in the total energy, the condition on the heat flow between $A$ and $B$ is
$\beta_A Q_A +\beta_B Q_B \ge \Delta S_A+\Delta S_B$.
However, since $\Delta S_A = \Delta S_B$ and $Q_A=-Q_B$ we have that
\begin{eqnarray}
-\Delta S_A( 1/\beta_A + 1/\beta_B)\ge 0
\end{eqnarray}
and so $\Delta S_A=\Delta S_B \le 0$. Energy conserving unitary interactions never \emph{increase} the local entropies of either subsystem, in stark contrast to the uncorrelated situation of classically mixed thermal states, where $\Delta S_A +\Delta S_B$ never decreases.
However, the model in \cite{partovi} does not actually require quantum entanglement. For example we may simply dephase $| \Psi_{AB} \>$ to turn it into a classically correlated separable state that would produce the exact same reversal of the thermodynamic arrow. In addition, while the pure state $| \Psi _{AB} \>$ allows local entropy decreases, and the reversal of the thermodynamic arrow, the isospectral constraint for the bipartite splitting of the pure state $| \Psi_{AB}\>$ is very restrictive - the only way in which two different temperatures for $A$ and $B$ is achieved in \cite{partovi} is to impose that the energy spectrum of $A$ be a scaled version of the energy spectrum of $B$. Another undesirable feature of the model is that the initial thermal states that are considered in \cite{partovi} (squeezed Gaussian states of two oscillators) are the \emph{only} thermal states that are attainable in this setting - all other thermal states are inaccessible.
In the next section we extend the work of \cite{partovi} to show that we may distinguish entangled systems from classically correlated system through violations of thermodynamic arrow and produce transformations between arbitrary thermal states without restrictions on the energy spectra. Furthermore, we analyze quantum systems in which the correlation structure is more complex and allows for a more subtle range of violations of the thermodynamic arrow.
\subsection{Classical correlations and entanglement witnesses}
Recently \cite{grois05} the mutual information $I(A:B)$ was given an operational meaning in terms of the minimal amount of local randomizing work that must be done on a bipartite system to reduce it to a product state and destroy all correlations. We will give here another thermodynamical interpretation - namely, the mutual information is a measure of the maximal amount that the isolated bipartite system may violate the thermodynamic arrow in the form of energy transfer from a colder subsystem to a hotter one.
The total correlations, as quantified by $I(A:B)$, also contain classical correlations, which may be defined as
\begin{eqnarray}
I_c(A:B) = \mbox{max}_{M_A\otimes M_B} H(A:B)
\end{eqnarray}
where $H(A:B)$ is the classical (Shannon) mutual information for the joint probability distributions generated by local POVM measurements $M_A \otimes M_B$ at $A$ and $B$ \cite{definition-of-Ic}. The classical mutual information $I_c(A:B)$ is always less than or equal to the total mutual information $I(A:B)$ and vanishes if and only if the bipartite state is a product state\cite{locking-classical}.
It is clear that for bipartite states $\rho_{AB}$ the classical mutual information always obeys the bound $ 0\le I_c(A:B) \le \log D$, where $D$ is the dimensionality of the smaller subsystem. On the other hand, while the quantum mutual information $I(A:B)$ is always positive, it can in general take on values larger than $\log D$. However, if we write the quantum mutual information as
\begin{eqnarray}
I(A:B) &=& S_A+S_B -S_{AB} \nonumber \\
&=& S_A - (S_{AB}-S_B) =S_B - (S_{AB}-S_A)\nonumber
\end{eqnarray}
and restrict ourselves to separable states $\rho_{AB}$ for which $S_{AB}$ is always greater than both $S_B$ and $S_A$ \cite{positive-conditional} we may deduce that $I(A:B) \le \mbox{min} \{S_A,S_B\}$ and so $I(A:B)|_{\mbox{\tiny sep}} \le \log D$ over the set of separable states. This upper bound is saturated for perfectly correlated, zero discord, separable states $\rho_{AB} = \frac{1}{D}\sum_i |i\>_A\<i| \otimes |i\>_B\<i|$, in which case $I_c(A:B) = I(A:B) = \log D$.
However, for entangled states we can have $I(A:B) > \log D$, and so from (\ref{flow}) we see that certain types of heat flow are possible that are forbidden classically. In particular a quantity of heat $I(A:B)/(|\beta_A -\beta_B|)$ can be transferred from the cold subsystem to the hot system, whereas classical correlations could only permit at most a quantity of heat $\log D/ (| \beta_A - \beta_B|)$ to flow from cold to hot. From this it follows that the reversal of the thermodynamic arrow of time is a new kind of entanglement witness where quantum correlations present in the state allow much larger violations of the arrow than are classically possible, and so the detection of such large reversals must imply an entangled state. Standard entanglement witnesses correspond to an observable $W$ that defines a plane $ \mbox{Tr} [W \rho] =0$ in state space separating entangled states from separable ones, while for us, it is the set of energetically accessible states that acts as the witness. Instead of local observables, it is the transformations of the state that signal entanglement.
The classical mutual information $I_c(A:B)$ has previously been considered in the context of the thermodynamic arrow as a candidate measure for the classical memory record that a system $A$ has of events that affect $B$ \cite{macc08}. The claim in \cite{macc08} was that any event that decreases the entropy of $B$ necessarily reduces the classical mutual information between $A$ and $B$, and so entropy-decreasing events at $B$ do not leave a memory trace in $A$. This was a proposed resolution to the empirical fact that we only observe entropy-decreasing events, namely the ``arrow of time dilemma''.
However, in \cite{jennings09} we demonstrate explicitly that this claim is false, and in reality quantum mechanics allows the exact opposite to occur. We show that large quantum correlations present in a state can be used up to reverse the thermodynamic arrow and at the same time increase the classical correlations between two subsystems. Quantum mechanics allows enhanced memory records of entropy-increasing events and so there can be no resolution of the arrow of time in terms of classical memory records.
The process of dividing correlations into quantum and classical components has many subtle aspects and counter-intuitive results exist, for example in the setting of multipartite correlations it is possible to have purely quantum multipartite correlations without any classical multipartite correlations being present \cite{ved03}. In the next section we consider multipartite quantum systems that are in a highly entangled pure state and analyse how the quantum correlations present allow arbitrarily large reversals of the thermodynamic arrow.
\section{Local Thermal States, Global States and a Hierarchy of Arrows}
We now turn to an analysis of entanglement-assisted violations of the thermodynamic arrow in a general multipartite setting. Once again, we consider an isolated system that may undergo a complex time-dependent interaction process between times $t_i$ and $t_f$, but with the total energy, defined via a total Hamiltonian $H_{\mbox{\tiny tot}}$, at time $t_i$ equal to the total energy at time $t_f$.
It is well known that most quantum states are highly entangled \cite{Clifton-separablevol}, however recently a deep theorem \cite{pop03} has shown that if the total system is large, then any randomly choosen pure state $|\Phi\>$ that satisfies the total energy constraint has the property that any sufficiently small subsystem is highly likely to be in a thermal state. For generic large systems in a pure state, thermality of its subsystems naturally arises from the high degree of entanglement present in the state. Consequently, instead of entangled thermal states such as in \cite{partovi} being artificial or exceptional, for large systems in a randomly chosen pure state they are actually quite typical.
\subsection{Local thermal states from an entangled state}
For simplicity we shall consider the multiple subsystems $S_1, \dots S_N$, to be qubits, but it is clear that similar arguments apply to any higher dimensional subsystems.
We take the reduced states, obtained by tracing out the other subsystems, to be initially thermal and the total system is assumed to be in a pure state $|\Psi \>$ with a fixed energy. We refer to the reduced states $\rho_1, \rho_2, \dots$ as the \emph{thermal marginals} for the subsystems.
However, not all sets of thermal marginals may be obtained from a global pure state $|\Psi \>$. For the case of qubits, if the marginal states for the individual qubits are given by $\{ \rho_i \}$, then the necessary and sufficient conditions for the existence of a pure state $| \Psi \>$ such that $ \mbox{Tr} _{j \ne i} | \Psi \> \< \Psi |= \rho_i$ are in terms of the smallest eigenvalue of each subsystem, $\lambda_i = \mbox{min(Spec(}\rho_i))$. The conditions are \cite{higuchi03}
\begin{eqnarray}\label{constraints}
\lambda_i &\le& \sum_{j \ne i} \lambda_j \nonumber\\
0\le \lambda_i &\le & 1/2 \hspace{0.5cm} i=1, \dots , N.
\end{eqnarray}
Without loss of generality, we take the Hamiltonians of each qubit to be equal, and choose the ground state to have zero energy while the excited state has energy 1. Thus, $H_i = \frac{1}{2} (I + Z_i)$, where $Z_i$ is the Pauli operator for the $Z$ direction, and for qubits with thermal marginals, the total expected energy of the system is then $E =\sum _i \lambda_i $. The parameter space for the set of states $|\Psi \>$ in terms of the smallest eigenvalues of its subsystems is defined by (\ref{constraints}) and is an $N$-dimensional polyhedron $ {\cal P} _N$, with the constant energy condition being a hyperplane that intersects $ {\cal P} _N$ on a subset $ {\cal T} _{N-1}$ of dimension at most $N-1$. Each point in $ {\cal T} _{N-1}$ corresponds to an accessible combination of thermal marginals, however there is in general more than one pure global state associated to such a point.
As an example, in the case $N=3$, the region of parameter space is a diamond formed from two tetrahedra, while the constant energy condition corresponds to a plane that slices the diamond in triangular cross-sections, as depicted in Fig.~\ref{colorpoly}.
Each point in the diamond defines three qubit thermal states, however, since the overall state is a pure state, any other pure state may be reached by a general unitary transformation. It is the restriction to energy-conserving unitaries which means that only thermal states with parameters in the triangular region $ {\cal T} _2$ are accessible. Such energy conserving unitaries could arise from time-dependent interactions $V_{ij}(t)$, or, for example, from time-independent interactions of the form $X_i Y_j -Y_iX_j$.
\begin{figure}[t]
\centering
\includegraphics[width=5.5cm]{diamond-region-smaller2.eps}
\caption{\textbf{The parameter space $ {\cal P} _3$:} The set of entangled states giving local thermal states possesses a diamond polygon parameter space. A triangular slice $ {\cal T} _2$ with constant total energy is shown. Dark regions have low energies and temperatures, while bright regions have high energies and temperatures.}
\label{colorpoly}
\end{figure}
The diamond polyhedron has some nice features. The origin is the bottom vertex and corresponds to $E=0$, where all the qubits are at $T=0$, while $E=3/2$ is the other extreme point, where each qubit is maximally mixed, and $T= \infty$.
The centroid of any given constant energy triangle corresponds to all three qubits having equal temperatures and so would correspond to the standard equilibrium one would expect if the three subsystems were initially uncorrelated. For standard thermally mixed states free to interact, the total system would tend to evolve towards this configuration. However, from our analysis we see that this is no longer imposed. The overall state is pure and the total system can undergo unitary evolution to any other set of local thermal states in $ {\cal T} _2$. The large degree of entanglement present has radically lifted the constraints of the thermodynamic arrow and allows the subsystems to transformation to otherwise inaccessible states.
The triangle formed by the intersection of $E=1$ and the diamond forms the widest part of the polyhedron and is special in the sense that there is enough entanglement for two of the qubits to form a singlet state, but not too much that the singlet becomes impossible. In terms of temperature, for points on this triangle it is possible to unitarily transform along $E=$constant to a situation where two qubits are maximally mixed while the third is in a pure state. For $E<1$ it is always possible to turn off the temperature of one qubit with the other two being at some finite temperatures, while for $E>1$ this is no longer possible and each qubit always has some non-zero temperature.
\begin{figure}[t]
\centering
\includegraphics[width=7cm]{triangle-arrows.eps}
\caption{\textbf{The parameter space $ {\cal T} _2$:} The standard thermodynamic arrow would constrain $C$ to evolve only into the cone of states depicted, and this drives it steadily towards equilibrium, while the entangled state $Q$ has no such restriction. Dotted lines are the isotherms for $T_a$, $T_b$ and $T_c$, while the solid internal lines represent $(T_a=T_b,T_c=0)$, $(T_b=T_c,T_a=0)$ and $(T_c=T_a,T_b=0)$. The center point corresponds to thermal equilibrium $T_a=T_b=T_c$. }
\label{triarrows}
\end{figure}
The diamond setting also provides a simple way to visualise the contrast between a system subject to the standard thermodynamic arrow and the highly entangled system discussed here. Fig.~\ref{triarrows} shows the constant energy region $ {\cal T} _2$ for a system of three qubits with thermal marginals $\rho_a$, $\rho_b$ and $\rho_c$. The centroid of the triangle corresponds to the equilibrium configuration $T_a=T_b=T_c$. The solid lines passing through the equilibrium point and intersecting the midpoint of the sides correspond to the situation where two of the subsystems have equal temperatures. These lines divide the triangle into six regions, labelled I-VI in Fig.~\ref{randomwalk}, corresponding to the six choices of temperature orderings, $T_a<T_b<T_c, \cdots$, $T_c < T_b<T_a$.
If the system is in a thermal configuration $C$ with no correlations then equation (\ref{multiheat}) forces the system to evolve only in the direction shown in Fig.~\ref{triarrows}. The thermodynamic arrow therefore imposes a cone of accessible configurations for the thermal system, and gradually drives the system inexorably to thermal equilibrium at the center of the region (a fun comparison is with black holes - the light cone of a person inside the event horizon always forces them to move radially inwards, eventually reaching the central singularity, while here we have in some sense a `thermal cone' that dictates how the system must evolve in time, ultimately reaching the central equilibrium point).
It makes physical sense in this context to only consider a small transfer of energy that perturbs the initial thermal configuration into a neighbouring thermal configuration. A state with $T_b=T_c$ has no formal restriction from (\ref{multiheat}) on the amount of heat that may be exchanged between $b$ and $c$, however the configuration $T_b=T_c$ is stable in the sense that a small transfer of heat will automatically induce a correcting directionality that returns the system to $T_b=T_c$.
This is illustrated in Fig.~\ref{randomwalk} where given an initial thermal configuration $C$, subject to random heat exchanges and the constraint (\ref{multiheat}), the system starts in region I where $T_a < T_b < T_c$. Heat is exchanged, cooling $c$ and heating $a$ and $b$ until the system crosses the $T_b=T_c$ line into region II. At which point only fluctuations around this line are permitted by (\ref{multiheat}). Subsequently, the system $a$ is heated up until the thermal configuration reaches the central equilibrium point, as shown in the diagram, at which point only small fluctuations from equilibrium are permitted.
However, no constraint is present for our entangled state, which is free to execute a random walk and roam over the accessible region of parameter space $ {\cal T} _2$, for example a given initial configuration $Q$ may start in region V, but is free to move towards and away from equilibrium and can visit every point in the accessible parameter space, see Fig.~\ref{randomwalk}.
\begin{figure}[t]
\centering
\includegraphics[width=6cm]{triangle-walk.eps}
\caption{Contrast in system evolution due to random transfers of energy between subsystems. For $C$, subject to the constraint (\ref{multiheat}), the thermodynamic arrow forces the subsystems to the equilibrium configuration. The system $Q$ is not restricted for the small fluctuations and performs a non-directed random walk through the accessible configuration space. }
\label{randomwalk}
\end{figure}
\subsection{A global pure state}
We now present a highly entangled pure state for which the thermodynamic arrow is removed for interactions between any two of its subsystems, and so allows the most stark deviation from standard thermodynamical behaviour.
In general for $N$ qubits in locally thermal states there exist a vast number of consistent pure states. However, it is straightforward to see that any set of thermal marginals $\{ \rho_i \}_{i=1}^N $ can be obtained from a global pure state with a fixed energy $E$, of the form
\begin{eqnarray}\label{A1states}
| \Psi \> = \sum_i x_i X_i | 0 \cdots 0 \> + \sqrt{1 -E/(N-1)} |1 \cdots 1\>
\end{eqnarray}
with $X_i$ the Pauli operator on qubit $i$, and each $x_i$ real and obeying the condition $\sum_i x_i^2 =E/(N-1)$.
Furthermore, these parameters are related to the parameters for $ {\cal P} _N$ through $\lambda_i =\sum _{j\ne i} x^2_j$.
The parameters $x_i$ specify a point on a hypersphere in $\mathbb{R}^N$ of radius $\sqrt{E/(N-1)}$ and any unitary transformation that satisfies the energy constaint $\sum_i Q_i=0$ then corresponds to $O(N)$ transformations on the subspace spanned by $\{ X_i|0 \cdots 0\> \}_{i=1}^N$. Heat transfer between any two qubits $a$ and $b$ can be mediated by an energy-conserving interaction Hamiltonian such as $V_{ab}=(Y_a X_b - X_a Y_b)/2$ and the evolution is then described by the unitary $U_{ab} = \exp [ -i \theta/2 (Y_a X_b -X_a Y_b)]$. This unitary allows heat flow in either direction between the thermal reduced states $\rho_a$ and $\rho_b$.
While the reduced states $\rho_a, \rho_b$ are thermal, it is not the case that $\rho_{ab}= \mbox{Tr} _{i \ne a,b} |\Psi \>\< \Psi|$ is a direct product of thermal states. The 2-qubit process we gave, detects the quantum correlations present between $a$ and $b$ that are not accounted for by merely probing the qubits individually. A state such as (\ref{A1states}) could clearly be used by a global demon to reverse standard thermodynamic behaviour for a local demon, as described in the introduction, with an interaction such as $V_{ab}$ causing heat to flow both into and out of a thermal state. In addition, we see that quantum correlations in a pure state allow far more freedom than would be allowed by classically correlated states.
\subsection{A candidate experimental state?}
\begin{figure}[t]
\centering
\includegraphics[width=4cm]{Ent-QA.eps}
\includegraphics[width=4cm]{Ent-QB.eps}
\includegraphics[width=4cm]{Ent-QC.eps}
\caption{\textbf{Extangled system:} The heat flow pattern into $A,B,C$ for the correlated state $\rho_{ABC}$ with $\lambda_A=0.15$, $\lambda_B=0.2 $, $\lambda_C=0.3$, $\gamma=0.4$ and parameters $t$ for the interaction $V_{AB}$ and $s$ for $V_{BC}$. Here a light-coloured plane is shown through $Q=0$
to highlight those regions where heat leaves the subsystem $Q<0$, while the dark regions visible above this plane correspond to heat entering the system $Q>0$.}
\label{correlatedABC}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=4cm]{Product-QA.eps}
\includegraphics[width=4cm]{Product-QB.eps}
\includegraphics[width=4cm]{Product-QC.eps}
\caption{\textbf{Uncorrelated system:} The heat flow pattern into $A,B,C$ for the ordinary product state $\rho_A \otimes \rho_B \otimes \rho_C$, with parameters $\lambda_A=0.15$, $\lambda_B=0.2 $, $\lambda_C=0.3$, $t$ for the interaction $H_{AB}$ and $s$ for $H_{BC}$. Here a light-coloured plane is shown through $Q=0$ to highlight those regions where heat leaves the subsystem $Q<0$, while the dark regions visible above this plane correspond to heat entering the system $Q>0$.}
\label{productABC}
\end{figure}
We now consider an interesting system consisting of 3 qubits $A$, $B$, and $C$ in a line, where $A$ is allowed to interact with $B$ and $B$ is allowed to interact with $C$. The state of the total system is given by a mixed state $\rho_{ABC}$, and has the property that $\rho_{AB}= \rho_A(T_A) \otimes \rho_B(T_B)$ and $\rho_{BC}= \rho_B(T_B) \otimes \rho_C(T_C)$ for thermal marginals $\rho_i(T_i)$, $i=A,B,C$ with $T_A <T_B<T_C$. Hence the pairs $AB$ and $BC$ obey the thermodynamic arrow individually and we would expect heat only to flow from $C$ to $B$ and from $B$ to $A$.
However, if $\rho_{AC}$ is given by
\begin{eqnarray}
\rho_{AC} &=&\frac{1}{2} ( ( \gamma + \lambda_C -\lambda_A) |10 \>\<10|+(\gamma -\lambda_C +\lambda_A) |01 \>\<01|\nonumber\\
&&\hspace{-1cm}+\sqrt{\gamma^2 - (\lambda_C -\lambda_A)^2}( |10 \>\<01|+|01 \>\<10|)\nonumber\\
&&\hspace{-1cm}+( \lambda_A+\lambda_C- \gamma) |00 \>\<00|+ (2-\lambda_A -\lambda_C -\gamma) |11\>\<11|\nonumber\\
&&\hspace{-1cm}+\sqrt{(\lambda_A +\lambda_C -\gamma)(2-\lambda_A -\lambda_C-\gamma)}(|00\>\<11|+|11\>\<00|) )\nonumber
\end{eqnarray}
then we are able to exploit correlations between $A$ and $C$ to for example transfer heat from $A$ to $B$. The state $\rho_{AC}$ at first glance looks complicated, but it may be obtained from a 3 qubit state $|\Psi \>_{ACD}$ of the form (\ref{A1states}) by tracing out $D$; the parameters $\lambda_A, \lambda_C, \gamma$ must then obey similar relations to (\ref{constraints}).
The total state now takes the form $\rho_{ABC} = \rho_{AC} \otimes \rho_B$ and we assume that the interaction Hamiltonians are once again given by $V_{AB}= \frac{1}{2}(X_AY_B -Y_AX_B)$ and $V_{BC}= \frac{1}{2}(X_BY_C -Y_BX_C)$. In the event that only one of these interactions is switched on we have that the thermodynamic arrow is obeyed and heat always flows in the direction $C\rightarrow B \rightarrow A$. However when both interactions are on, the heat flow becomes more complicated.
If we assume the system $\rho_{ABC}$ evolves under the unitary $U(t,s) = \exp[-itV_{AB} - isV_{BC}]$, then Fig.~\ref{correlatedABC} and Fig.~\ref{productABC} show the contrast between heat flows for the entangled state $\rho_{ABC}$ and for the uncorrelated product state $\rho_A\otimes \rho_B \otimes \rho_C$.
For $s=0$ heat can only flow from $B$ to $A$, however as can be seen from Fig.~\ref{correlatedABC} by switching on $s$ we are able to have system $A$, the coldest subsystem, emit heat into $B$. Fig.~\ref{productABC} shows the expected heat flow patterns given the standard uncorrelated product thermal states. Here the heat flow into $A$ is always positive (dark-coloured regions), while the heat flow into $C$ is always negative (light-coloured regions). It is clear that the presence of entanglement can dramatically alter this pattern.
Fig.~\ref{Delta-Q} shows the difference $ \Delta Q_A \equiv Q_A^{p}-Q_A^{e}$ between the flow for $A$ in the two cases of a product state ($Q^p_A$) and an entangled state $Q^e_A$. It is evident from this that the entanglement can increase the heat \emph{into} $A$ as well as out of $A$. Fig.~\ref{Delta-Q} also shows the region of parameter space for which heat leaves $A$ and both $B$ and $C$ gain heat, which corresponds to a clear violation of the expected thermodynamic behaviour.
\begin{figure}[t]
\centering
\includegraphics[width=4cm]{Delta-Q.eps}
\includegraphics[width=3cm]{Violation-Region.eps}
\caption{The first graph shows the difference in heat pattern for $A$ in the two cases of an entangled state and a separable state. Here $Q^p_A$ is the heat gained by $A$ in the product state case, while $Q^e_A$ is the heat gained by $A$ in the entangled state case. Once again, light regions correspond to negative values, while dark regions correspond to positive values. The second plot shows the regions of $s$-$t$ parameter space in which $A$ has lost heat and both $B$ and $C$ have gained heat.}
\label{Delta-Q}
\end{figure}
We can also ask if the system can act as an entanglement witness under heat flow. By analysing the mutual information for the system we can show that range exists in the parameters $(\lambda_A, \lambda_C,\gamma )$ for which the correlations present in the state $\rho_{ABC}$ exceed $\log 2$ and so allows heat flows that cannot be attributed to classical correlations between $A$ and $C$. Fig.~\ref{witness-region} shows this region as part of the diamond polygon parameter space discussed earlier. The set of states for which it may act as an entanglement witness is shown as the semi-transparent light-coloured region. However, it must be emphasized that for a generic, correlated mixed state it is not always possible to unitarily eliminate \emph{all} of the correlations present as the set of unitary orbits containing product states is a low dimensional sub-manifold of the manifold of all unitary orbits\cite{geometry-unitary-orbits}.
\begin{figure}[t]
\centering
\includegraphics[width=4.5cm]{Entanglement-witness-region.eps}
\caption{The region of parameter space for which reversal of the thermodynamic arrow in $\rho_{ABC}$ can act as an entanglement witness is shown in semi-transparent light region, while the remainder of the space only possesses correlations that could be attributed to a classically correlated state.}
\label{witness-region}
\end{figure}
\subsection{A quiver of thermodynamical arrows}
As already mentioned in general there exists a multitude of global states $\rho_{\mbox{\tiny tot}}$ associated with a given combination of thermal marginal states. For a given collection of marginals the full state can have a vast array of different consistent correlations. The set of such consistent states $ {\cal Q} $ contains states that differ in the correlations that exist between different combinations of subsystems. Consequently, given subsystems $S_1, \dots S_k$, the reduced state $\rho_{1 \cdots k}$ can vary for different global states $\rho_{\mbox{\tiny tot}}$ in $ {\cal Q} $.
It is easy to see that while we might demand the individual subsystems $S_1, \dots, S_N$ to be thermal states, this thermality may appear at various levels of coarse graining on the collection of subsystems. For example we might have for a pair of subsystems $S_a$ and $S_b$ that $\rho_{ab}= \mbox{Tr} [\rho_{\mbox{\tiny tot}}]= \rho_a \otimes \rho_b$, however for a triple of subsystems $S_a,S_b,S_c$ the state $\rho_{abc}$ is entangled. For such a situation, the thermodynamic arrow is present for any energy transfer process involving only the two subsystems $a$ and $b$, however for processes involving three or more subsystems the correlations present allow violations of the thermodynamic arrow.
Consequently, the different possible correlation structures that can occur in a state $\rho_{\mbox{\tiny tot}} \in {\cal Q} $ correspond to a hierarchy of thermodynamic arrows. The set of states $ {\cal Q} $ can be partitioned up as $ {\cal Q} = {\cal A} _1 \cup {\cal A} _2 \cup \cdots $, where a state $\rho$ is in $ {\cal A} _k$ if there exists subsystems $S_{a_1}, \dots , S_{a_k}$ such that
\begin{eqnarray}
\sum^k _{j=1} \frac{Q_j}{T_j} \ge 0
\end{eqnarray}
is observed for all transformations on $\rho_{a_1 \dots a_k}$ and no larger set of subsystems exist with this property. It is not clear, however, if a non-trivial partitioning can be defined in which, for a state in a given class, the thermodynamic arrow holds for \emph{all} marginals on $k$ subsystems.
This classification of states in terms of their largest thermal marginal depends on the particular correlation structure of the global state without reference to where the subsystem are situated, however in more practical situations it makes sense to include a notion of locality. For example, we might consider a collection of subsystems $S_1, S_2, \dots$ located at various points in space. For a given subsystem $S_a$ we may define the notion of an ``arrow range'', $R_a$, which is the largest $r_a$ such that the reduced state $\rho (r_a)$ on all subsystems within a distance $r_a$ of $S_a$ is a thermal marginal. For the entire state $\rho_{\mbox{\tiny tot}}$ we may then simply define a characteristic arrow range $\bar{R}$ as the average over the set $\{R_i \}$ of ranges for each subsystem $S_i$. States with large $\bar{R}$ do possess correlations, however these correlations are difficult to access in practice, requiring many coordinated, local pairwise interactions $V_{ij}$ in order to generate an effective multipartite interaction over the correlated state. On the other hand, states with small values for $\bar{R}$, for example states of the form (\ref{A1states}) or the state $\rho_{ABC}$ considered in the previous section, have correlations that are more easily accessible through local pairwise interactions.
This provides a simple generalization of the `local' Maxwell demon introduced in the first section. A demon should not only be finite in terms of its memory resources, but also in terms of the correlation range $R$ that it can probe.
\section{Conclusion}
Time, with its inexorable flow, is one of our oldest mysteries. In stark contrast, it is only in the last century that we have become aware of the quantum mechanical properties of Nature, and only in the past few decades that some of the subtleties and power of quantum correlations have revealed themselves.
In this article we have examined some aspects of entanglement correlations, how they can reverse the standard thermodynamic arrow of time, and have shown that in this regard thermodynamic transformations may act as entanglement witnesses.
We analysed how highly entangled multipartite states can make ordinarily forbidden thermodynamic states accessible. Different correlation structures can give rise to the same set of local thermal states, and so processes involving several subsystems are in general required to exploit the entanglement present. The set of transformations on a system then possess a hierarchy of thermodynamic arrows, whether considered globally or in terms of local correlation ranges. The key aspect is to allow interactions that activate the correlations present, which we demonstrated with a mixed state example $\rho_{ABC}$ in which the thermodynamic arrrow is in place for interactions involving only $A$ and $B$ or for $B$ and $C$, but by switching on an interaction with system $C$ this can be dramatically modified.
It would be of interest to further explore the multipartite correlation structures that can occur in a quantum state - for example we might speculate that in the early, dense universe that $\bar{R}$ is small enough to allow any random physical interactions to exploit the correlations present, producing a gradual disappearance of the thermodynamic arrow the closer we get to the initial state of the universe. Of course such notions of locality rely on a classical spacetime background, and so we could only push the issue of the special initial conditions into the sub-Planck scales, where a complete theory of quantum gravity would be required.
Another setting of interest is that of quantum field theory, where thermality arises naturally due to observers possessing causal horizons, such as in the case of black holes or for accelerated observers. Once again, thermality can be viewed as arising from entanglement across the horizon, however it is unclear if these correlations could actually be utilized in any thermodynamic transformations. At a deeper level, it may also be fruitful to explore connections with the thermal time hypothesis due to Connes and Rovelli \cite{thermal-time} or recent work by Padmanabhan \cite{paddy} that attempts to relate the cosmological arrow of time for our expanding universe with the thermodynamic arrow by exploiting the thermodynamics of local spacetime horizons.
Finally, in a more grounded setting, it would be of interest to see if a mixed state $\rho_{ABC}$, or one like it, could be realised experimentally to see how easily one might exploit quantum correlations present in a state to affect its thermodynamic behaviour.
|
\section{Spin Foams}
Spin foams are colored two-complexes and they naturally appeared in the context of Loop Quantum Gravity, see \cite{B}. They can be viewed as world-sheets of spin networks propagating in time, where the spin networks are colored graphs, originally invented by Penrose \cite{P}, representing the states of quantum geometry on a three-dimensional space. One can then imagine a spin foam interpolating between two fixed spin networks, so that by associating a complex number to the spin foam, which depends on the spin foam labels, and then summing over the labels, one can then obtain an invariant of the spin networks, provided that the sum over the labels is independent of the triangulation of the four manifold whose dual two-complex is the given spin foam. This sum over the labels is called the spin foam state sum, and the rules how to find triangulation independent spin foam amplitudes can be inferred from the path integrals for BF theories, see \cite{B}.
A $n$-dimensional BF theory can be defined by the action
\begin{equation} S = \int_M Tr\,( B \wedge F ) \,,\end{equation}
where $M$ is a $n$-dimensional compact manifold, $F = dA + A\wedge A$ is the curvature 2-form for a connection $A$ on the principal bundle $P_G (M)$ for a compact Lie group $G$, $B$ is a $\bf g$-valued $(n-2)$-form, where $\bf g$ is the Lie algebra of $G$. The path integral for this BF theory can be written as a functional integral
\begin{equation} Z(M) =\int DA \exp\left(i\int_M Tr\,( B \wedge F ) \right) \,,\end{equation}
which can be formally written as a functional integral over the flat connections on $M$
\begin{equation} Z(M) = \int DA \,\delta (F)\,,\label{fcpi}\end{equation}
where $\delta (F)$ denotes the product $\prod_{x\in M} \delta (F(x))$ and $\delta (x)$ is the Dirac delta-function.
All these expresions are formal and have to be defined, and the standard way is through a triangulation $\Delta(M)$ of $M$. The idea is that $Z(M)$ should be a manifold invariant, and since the BF theory is a topological theory, then a discretization of the path integral (\ref{fcpi}) will not depend on a chosen triangulation of $M$. Let us replace $\Delta(M)$ with its dual simplical complex $\Delta^* (M)$ and consider a discretization of (\ref{fcpi}) given by
\begin{equation} Z(M) = \int \prod_l dg_l \prod_f \,\delta (g_f)\,,\label{dpi}\end{equation}
where $l$ are the dual edges and $f$ are the dual faces, $g_l$ is the holonomy along an edge $l$ and $g_f = \prod_{l\in\partial f} g_l$ is the holonomy along the boundary loop $\partial f$ of $f$.
The group integrations in (\ref{dpi}) can be performed by using the Peter-Weyl theorem
\begin{equation} \delta (g) = \sum_{\Lambda} \,\textrm{dim}\Lambda \, Tr\left(D^{(\Lambda)}(g)\right)\,, \end{equation}
and
\begin{equation} \int_G dg \, D^{(\Lambda_1)}(g)\otimes\cdots \otimes D^{(\Lambda_m)}(g) = \sum_{\iota}\,C^{(\iota)}\otimes\overline{C^{(\iota)}}\,, \label{gii}\end{equation}
where $D^{(\Lambda)}(g)$ is the matrix of the group element $g$ in the irreducible representation $\Lambda$ and $C^{(\iota)}$ is the tensor for the intertwiner $\iota$ from the tensor product of representations $\Lambda_1 , \cdots, \Lambda_m$. Then
\begin{equation} Z(M) = \sum_{\Lambda,\iota}\prod_f \textrm{dim}\Lambda_f \, \prod_v A_v (\Lambda , \iota )\,,\label{sfss}\end{equation}
where the sum is over the colorings $\Lambda = (\Lambda_1 , \cdots, \Lambda_N )$ of the faces of the dual simplical complex by the irreps of $G$. The sumation has also to include the colorings $\iota = (\iota_1 , \cdots, \iota_M )$ of the dual edges by the corresponding intertwiners, and $A_v$ is the evaluation of the $n$-simplex spin network in terms of the products of $C^{(\iota)}$.
In the case when $G=SU(2)$, the irreps are called spins, and hence the name ``spin foam'' for the colored dual two-complex. For a three-dimensional manifold, the dual faces of a triangulation correspond to the edges, and the dual vertices correspond to the tetrahedrons, so that $A_v$ is the $6j$ symbol.
The spin-foam state sum (\ref{sfss}) can be shown to be independent of the triangulation, due to the invariance under the $n$-dimensional Pachner moves. However, the sum (\ref{sfss}) is infinite and therefore divergent in general case, so that one still does not have a well-defined invariant. Fortunately, the formal expression (\ref{sfss}) can be regularized by replacing the irreps of $G$ by the irreps of the quantum group $U_q ({\bf g})$, where $q$ is a root of unity. This can be done because there is a one-to-one correspondence between the irreps of a compact Lie group and the irreps of the corresponding quantum group when $q$ is not a root of unity. When $q$ becomes a root of unity, this correspondence still holds for the irreps of non-zero quantum dimension. Consequently, all the spin network evaluations, which are the simplex weights in (\ref{sfss}), can be replaced by quantum spin network evaluations. The invariance under the Pachner moves is maintained and (\ref{sfss}) becomes a finite sum, because when $q$ is a root of unity there are only finitely many irreps $\Lambda$ of non-zero quantum dimension.
Therefore
\begin{equation} Z(M) = \kappa^{N}\sum_{\Lambda,\iota}\prod_f \textrm{dim}_q \Lambda_f \,\prod_l A_l^{(q)}(\Lambda,\iota) \prod_v A_v^{(q)} (\Lambda , \iota )\,,\label{qsf}\end{equation}
is a manifold invariant, where the label $q$ denotes the quantum group evaluations of the relevant spin networks. The factor $\kappa$ is related to $q$, $N$ is the number of $0$-simplices in the triangulation and a dual-edge amplitude $A_l$ appears\footnote{One can set $A_l =1$ by changing the normalization of the spin network evaluations.}. In the three-manifold case and $G=SU(2)$ this is the Turaev-Viro invariant \cite{TV}, while in the 4-manifold case (\ref{qsf}) gives the Crane-Yetter invariant \cite{CY}.
\section{Chern-Simons and BF theory path integrals}
A BF theory is a theory of flat connections on a manifold, and in three dimensions one also has the Chern-Simons (CS) theory, which is also a topological theory of flat connections. Given that the spin-foam state sum invariants are derived from the BF theory, the question is what is the relation to the manifold invariants derived from the CS theory? Witten showed that the CS invariants can be derived from the CS path integral \cite{W}, and the path-integral approach can be used to find the relation between the spin foam and CS invariants.
A BF theory with a cubic interaction is given by the action
\begin{equation} S_{\lambda}(A,B) = \int_M Tr\,( B \wedge F + \lambda\, B\wedge B\wedge B ) \,,\end{equation}
where $\lambda$ is a parameter. In the $SU(2)$ group case, this action corresponds to the
three-dimensional (3d) General Relativity (GR) theory with a cosmological constant $\lambda$. It is easy to show that
\begin{equation} S_{\lambda}(A,B) = S_{CS}(A_+) - S_{CS}(A_-) \,,\label{spm}\end{equation}
where $A_\pm = A \pm \frac{B}{\sqrt\lambda}$, $S_{CS}(A)$ is the Chern-Simons action for a connection $A$
\begin{equation} S_{CS}(A) = \frac{k}{2\pi}\int_M \langle A \wedge dA + \frac{2}{3} A\wedge A\wedge A \rangle \,,\end{equation}
and $k=2\pi/\sqrt\lambda$. Witten has showed that the CS action is well-defined if $k$ is an integer \cite{W}, and this integer is related to the order of the root of unity taken for $q$ in the quantum group $U_q (\bf g )$. In the $SU(2)$ case, $q=\exp(\frac{\pi i}{k+2})$.
By using (\ref{spm}) one can formally write
\begin{equation} \int DA \,DB e^{ iS_{\lambda}(A,B)}=\int DA_+ DA_- e^{iS_{CS}(A_+) - iS_{CS}(A_-)}=|Z_{CS}(M)|^2 \,,\label{bfcs}\end{equation}
where $Z_{CS}(M)$ is the Chern-Simons path integral. Witten, Turaev and Reshetikhin have shown that this path integral can be rigorously defined \cite{W,RT}, and the corresponding manifold invariant is proportional to the quantum group evaluation of a surgery link of $M$. We denote this invariant as $Z_{WRT}(M)$.
On the other hand, the Turaev-Walker theorem \cite{T} implies
\begin{equation} Z_{TV} (M) = |Z_{WRT}(M)|^2 \,,\label{twt}\end{equation}
where $Z_{TV}(M)$ is the spin foam state-sum invariant (\ref{sfss}) for the 3-manifold $M$. From (\ref{twt}) and (\ref{bfcs}) it follows that $Z_{TV}$ is the path integral for the three-dimensional euclidean General Relativity theory with a cosmological constant.
Note that the parameter $\lambda$ can take only the restricted values $4\pi^2 /k^2$, $k\in \bf N$ in the case of the path integral $Z(\lambda,M)$ for 3d GR. A natural question is how to define this path integral for the case when $\lambda \ne 4\pi^2 /k^2$. This can be done by using the spin foam perturbation theory \cite{M2}, and the result can be again expressed in terms of the $Z_{WRT}(M)$ invariant. Namely
$$ Z(\lambda, M) = Z_{TV}(k,M)\big{|}_{k=\frac{2\pi}{\sqrt\lambda}}\,,$$
where the right-hand side of the equation represents the function $Z_{TV}(k,M)=|Z_{WRT}(k,M)|^2$ evaluated at a non-integer point $k=2\pi/\sqrt\lambda$.
\section{Spin network invariants}
Witten showed in \cite{W} that the Jones polynomial $J_K (q)$ of a knot $K$ embedded in a three-sphere $S^3$ can be obtained by analysing the following path-integral invariant of $K$
\begin{equation} \langle K \rangle = \int DA e^{iS_{CS}(A)}W_K (A) \,,\label{kwl}\end{equation}
where $W_K(A)$ is the trace of the holonomy of a connection $A$ around $K$. The invariant $\langle K \rangle$ depends on the representation of the $SU(2)$ group used to define the holonomy. When $K$ is an unknot, then
$$ \langle K \rangle = \textrm{dim}_q j =(-1)^{2j}\, {q^{2j+1} - q^{-2j-1}\over q - q^{-1}}\,,$$
where $j$ is the spin of the $SU(2)$ representation and the Jones polynomial corresponds to $j=\frac{1}{2}$.
Beside colored loops, one can consider more general objects, like spin networks $\Gamma$, which are oriented graphs whose edges are colored by the spins. In analogy to (\ref{kwl}), one can consider a spin-network invariant
\begin{equation} \langle \Gamma \rangle = \int DA \,e^{iS_{CS}(A)}W_\Gamma (A) \,,\label{wlcs}\end{equation}
where $W_\Gamma (A)$ is given by the product of the edge holonomies contracted by the intertwiner tensors $C^{(\iota)}$
$$ W_\Gamma (A) = \prod_{l\in E(\gamma)} \textrm{Hol}_{l} (A) \prod_{v\in V(\gamma)} C^{(\iota_{v})}\,,$$
where $E(\gamma)$ and $V(\gamma)$ are the sets of the edges and the vertices of the graph $\gamma$ corresponding to the spin network $\Gamma$.
One can be even more general, and consider that $\Gamma$ is embedded in an arbitrary closed 3-manifold $M$. In any case, it can be shown that the path-integral (\ref{wlcs}) can be rigorously defined as the Witten-Reshetikhin-Turaev invariant
$Z_{WRT} (M,\Gamma)$, which is proportional to the quantum group evaluation of the link $L_M \cup \gamma$ embedded in $S^3$, where $L_M$ is any surgery link of $M$ \cite{W,RT}.
Given that (\ref{wlcs}) can be rigorously defined in CS theory, a natural question is how to define the analogous invariant in the BF theory case. One way is to start from the path-integral expression
$$ \langle \Gamma \rangle_{BF} = \int DA DB \,e^{i\int_M \langle B\wedge F \rangle} W_\Gamma (A)\,,$$
and to use the spin-foam discretization to define it. By repeating the procedure from section 1 for $\Gamma$ embedded in the handle-body $H$ obtained from thickening the dual one-complex of the triangulation of $M$, where each edge of $\Gamma$ goes along some one-handle of $H$ and each vertex of $\Gamma$ is placed in some zero-handle of $H$, then a Turaev-Viro type state sum is obtained. In the $SU(2)$ case this sum takes the form \cite{M,M1}
\begin{equation} Z_{TV}(\Gamma,M) = \sum_{J,\iota}\prod_f \textrm{dim}_q j_f \, \prod_{v\notin V_{\gamma}} \{6j\}_v^{(q)} \prod_{v\in V_{\gamma}} A_v^{(q)} (j ,j_\Gamma, \iota,\iota_\Gamma )\,,\label{sfmi}\end{equation}
where $V_\gamma$ are the vertices of the dual 1-complex whose thickenings contain the vertices of the graph $\gamma$ of $\Gamma$. $A_v^{(q)} (j ,j_\Gamma, \iota,\iota_\Gamma)$ is the quantum group evaluation of a tetrahedron spin network with an insertion corresponding to the spin network vertex residing in the 3-handle of $v$, where $\{j_\Gamma , \iota_\Gamma \}$ are the spins and intertwiners of $\Gamma$.
By using the Chain-Mail formalism of J. Roberts \cite{R}, it can be shown that $Z_{TV}(\Gamma,M)$ can be represented as the quantum group evaluation of a link embedded in $S^3$ \cite{M1}, and hence a relationship to the invariant $Z_{WRT}(\Gamma,M)$ can be obtained. The reason is that a Chain-Mail link $CHL_M$ for $M$ can be constructed by using the handle-body $H$, which was a thickening of the dual one-complex $\Delta_1^* (M)$ of a triangulation of $M$. Let $N_p$ be the number of $p$-handles of $H$, then the link $CHL_M$ consists of $N_2$ loops runing along the one-handles of H, which correspond to the boundaries of the faces of $\Delta_1^* (M)$ and $N_1$ loops running around the middle sections of the one-handles of $H$.
Then it can be shown for $G=SU(2)$ and a trivalent spin network\footnote{In the $SU(2)$ group case, a spin network vertex of valence $m > 3$ requires an intertwiner. However, such a vertex can be represented as a tri-valent spin network with $m-2$ vertices.} that
\begin{equation} Z_{TV}(\Gamma,M) = \eta^{-N_0 - N_2}\langle CHL_M \cup \gamma, \Omega^{N_1 + N_2} , j_\Gamma \rangle \,,\label{chm}\end{equation}
where
$$ \eta^2 = \sum_{j=0}^{k/2}(\dim_q j )^2 = {k+2\over \sin^2\left( \frac{\pi}{k+2} \right)}$$
and $\Omega$ is the coloring of each loop of $CHL_M$. The $\Omega$-element is a linear combination of all possible colorings of a loop, and it is given by the sum
$$\Omega = \sum_{j=0}^{k/2} (\dim_q j )\, C(j) \,,$$
where $C(j)$ denotes the coloring of a loop by spin $j$.
The derivation of (\ref{chm}) is the same as the derivation of the Lie group state sum (\ref{sfss}) from the group integral (\ref{dpi}), since both derivations are based on the same graphical calculus, which is associated to the equation (\ref{gii}), see \cite{B,M1}.
Given that the pair $(CHL_M,\Gamma)$ is a surgery link for the spin-network $\Gamma$ embedded in the connected sum $M\# \overline{M}$, where $\overline{M}$ is the manifold $M$ with the opposite orientation and $\Gamma$ is embedded in the obvious way in the $M$ part of $M\# \overline{M}$, it follows from (\ref{chm}) that
\begin{equation} Z_{TV}(\Gamma,M) = Z_{WRT}(\Gamma,M)\overline{Z_{WRT}(M)} \,.\label{tvw}\end{equation}
The formula (\ref{tvw}) has several interesting consequences:
\begin{itemize}
\item It gives a state-sum representation of $Z_{WRT}(\Gamma,M)$, since
$$Z_{WRT}(\Gamma,M) = \left(\overline{Z_{WRT}(M)}\right)^{-1} Z_{TV}(\Gamma,M)\,.$$
\item It allows for a path-integral interpretation, which is
$$Z_{TV}(\Gamma,M) = \int DA \,DB e^{ iS_{\lambda}(A,B)}W_\Gamma (A_+)\,, $$
where $\lambda = 4\pi^2 /k^2$.
\item It gives the relation to Turaev's shaddow-world invariants, see \cite{T}. Namely, Turaev has defined a state-sum invariant of a spin network $\Gamma$ embedded in $M$, which is based on the $Z_{TV} (M)$ sum in the following way. Let $M'$ be a 3-manifold obtained by excising a tubular neighborhood of $\gamma$ from $M$, where $\gamma$ is the graph of $\Gamma$. Triangulate $M'$ and construct the corresponding $Z_{TV} (M')$ such that the $6j$ symbols for the terahedrons which have a triangle on the triangulation of the boundary surface $\Sigma$ of $M'$ are excluded. Instead of the boundary tetrahedrons weights, one includes in the state sum the weight associated with a graph $\gamma'\cup \delta$ where $\gamma'$ is a natural projection of $\gamma$ onto $\Sigma$ and $\delta$ is the dual one-simplex of the triangulation of $\Sigma$ induced by the triangulation of $M'$.
The graph $\gamma'\cup\delta$ divides the surface $\Sigma$ into disjoint discs, which can be colored by spins $s$. One also has the spins associated to the edges of $\gamma'$, which are the spins of the spin network $\Gamma$, and the spins associated to the edges of $\delta$, which are the spins of a coloring $J$ of $\Delta (M')$. There is a natural way to associate the weights to such a coloring of $\Sigma$ in terms of the quantum $6j$ symbols, see \cite{T,M1}. By making the product of the surface weights and summing over the surface spins $s$ one obtains the weight $w_\Sigma (\Gamma,j)$ so that
$$ \tilde{Z}_{TV} (\Gamma, M) = \sum_{J}\prod_{j\in J} \dim_q j \prod_{v\in M'} \{6j\}_v \,w_\Sigma (\Gamma , j )\,.$$
Turaev has shown that $\tilde{Z}_{TV} (\Gamma, M)$ is related to $Z_{WRT}$ by the same formula as $Z_{TV}$ to $Z_{WRT}$, which means that $\tilde{Z}_{TV}$ and $Z_{TV}$ coincide. One can also show this directly, by using the Chain-Mail techniques, and one obtains that $Z_{TV} = \tilde{Z}_{TV}$ \cite{M1}.
\item In the case $M= \Sigma \times I$, where $I=[0,1]$, $\tilde{Z}_{TV}$ coincides with the spin foam state-sum invariants constructed in \cite{M3} for spin networks embedded in $\Sigma$. In particular, when $\Sigma = S^2$ then $\tilde Z_{TV} (\Gamma,S^2 \times I)$ is proportional to the quantum group evaluation of the spin network $\Gamma$, which means that the evaluation of a spin network can be expressed as a sum of products of $6j$ symbols. This can be used for obtaining the large-spin asymptotics of the quantum group evaluation of a spin network, since the large-spin asymptotics of the quantum $6j$ symbols is known \cite{MT}.
\end{itemize}
\section{Conclusions}
The spin foam state sum invariants of 3-manifolds and embedded graphs can be constructed by using the essentially finite category of representations of the quantum group $U_q (\bf g)$ for $q$ a root of unity. In the $SU(2)$ group case these state sums can be identified with the path integrals for the 3d Euclidean GR with a non-zero cosmological constant $\lambda$. An interesting question is how to define these state sums when $\lambda =0$, since then the finite category of the quantum group representations is replaced by the infinite category of the Lie group representations. It turns out that it is possible to define the path integral in this case, by regularising the group integral (\ref{dpi}) \cite{BG}. A finite invariant can be obtained only in special cases, when the second twisted cohomology group vanishes. This is in agreement with the results of Witten \cite{W2}, who showed that the $\lambda =0$ path integral can be represented as an integral over the space of flat connections of the Reidemeister torsion, and the integral exists if the $B$ field does not have certain zero-modes.
Note that we have discussed the spin-foam and related CS theory invariants for the case of a compact Lie group $G$. The case when $G$ is noncompact is an outstanding problem, and sveral strategies have been suggested so far:
\begin{itemize}
\item Repeat the steps of sect 1. for the category of unitary irreps of $G$. Because the unitary irreps of non-compact Lie groups are infinite-dimensional, just defining the classical spin network evaluations is a non-trivial task. Then one has to regularize the resulting formal sums by finding a finite category of quantum group representations. However, nobody has succeded so far in finding an appropriate category of the quantum group irreps, see \cite{FM,BC}.
\item Try to regularize the group integral (\ref{dpi}), by using the approach of \cite{BG}.
\item Try to find an analytic continuation of the compact group results, see \cite{FM,W3}.
\end{itemize}
\bibliographystyle{amsalpha}
|
\section{Introduction}
\label{intro}
LISA (Laser Interferometer Space Antenna) is a joint space mission by ESA and NASA which is planned to be launched
at the end of the next decade. It consists of three identical free-falling satellites, orbiting around the Sun and
marking the vertices of a nearly equilateral triangle with $\simeq 5 \cdot 10^{6}$ km (1/30 AU) sides,
located $20^{\circ}$ behind the Earth and lying in a plane that makes an angle of $60^{\circ}$ with the
ecliptic \cite{LISA}. LISA target is the detection of gravitational waves (GWs) through the measure
of the relative and differential motions between the spacecrafts. Among the astrophysical goals of LISA is to detect GWs
originated by events like black holes coalescence or capture of
compact objects by black holes \cite{Bender}. This requires a strain sensitivity of
$10^{-20} \ \mbox{Hz}^{-1/2}< S_h^{1/2} < 10^{-17} \ \mbox{Hz}^{-1/2}$ in the $10^{-4}$ to $10^{-1}$ Hz frequency band range.
However, LISA reference masses interact also with time-dependent (e.g. planets and their satellites,
quadrupolar pulsation of the Sun, etc.) and static (e.g. ID and LDM) local gravitational fields,
and therefore they depart from ideal unperturbed orbits around the Sun.
In this paper, we focused on the perturbing effects caused by the static components ID and LDM.
As the discriminating feature between ID and LDM concerns only their coupling with the electromagnetic field,
we expect that ID and LDM will produce similar gravitational forces on reference masses \footnote{This is correct only in Newtonian physics: in fact, DM and ID are expected to be gravitationally bound to the Galaxy and to the Solar System respectively; therefore, while gravitoelectric effects are identical for the two, that would not be the case for the gravitomagnetic ones, i.e., effects depending on the speed of the considered objects.}.
In fact, only ID reflects the solar light and can be studied through observations of zodiacal light \cite{Giese},
while LDM is ``dark" and its presence can be investigated by means of gravitational perturbations
induced on orbiting bodies. It is worth mentioning that other possibilities for
detection of DM are the searches for particular electro-weak processes beyond the Standard Model at particle
colliders. Thus we assume that ID and LDM induce identical gravitational effects on LISA orbits,
with the irrelevant difference that the first is luminous and the second is dark.
As we will show, the linear perturbation theory holds, being other gravitational forces acting on LISA reference masses much smaller than the Sun pull; under this approximation, small perturbations to ideal keplerian motions induced by different sources in the Solar System can be studied separately.
\section{Interplanetary dust}
\label{id}
The ID cloud is a dust composed by grains with typical sizes of $10-100\ \mu$m pervading the Solar System.
The distribution of ID particles is studied by the observations of solar light scattering (i.e. the zodiacal light confined to the ecliptic plane) and by their thermal emission, which is the dominant component of night-sky light in the $5-50\ \mu$m wavelength \cite{Levasseur}.
Physical models of ID density distribution $\rho$ should account for its symmetries, that can be easily
described in the Solar System Baricentric (SSB) reference frame $(x, y, z)$: i) invariance under rotation
about the $z$ axis, and ii) invariance under reflections in the $(x, y)$ ecliptic plane \footnote{The distribution is not symmetric with respect to the ecliptic plane, but to one defined by the total angular momentum of the Solar System. The differences between the two planes are negligible for our scopes.}.
Usually one also assumes a static distribution and a power-law radial profile.
As a consequence, all the proposed models factorize $\rho$ into two functions \cite{Giese}
\begin{equation}
\rho(r, \beta_{\odot}) = \rho_0 \left(\frac{r_0}{r} \right)^{\alpha} f \left(\beta_{\odot} \right),
\end{equation}
where $r$ is the distance from the Sun, $\beta_{\odot}$ is the helioecliptic latitude,
$\rho_0\simeq 9.6 \times 10^{-20}$ kg/$\mbox{m}^3$ is the density value
at $r_0=1$ AU (Earth's orbit), as estimated by the integration of the meteoroid mass distribution far away from the Earth, in order to avoid its gravitational attraction on ID particles, and $f$ is a given function.
The typical value of the parameter $\alpha$, determined from zodiacal light photometer measurements on
Helios 1 and 2 \cite{Grun}, is $1.3$.
The expression of $f \left(\beta_{\odot} \right)$ is still uncertain; however, the analytical formula which best
reproduces the observations of zodiacal light is the so-called \textsl{ellipsoid model}
\begin{equation} \label{eqn: model}
\rho(r, \beta_{\odot}) = \rho_0 \left(\frac{r_0}{r} \right)^{\alpha} \frac{1}{\left[1 + \left(\gamma_E \sin \beta_{\odot} \right)^2 \right]^{\alpha/2}},
\end{equation}
where $\gamma_E=\sqrt{a^2-b^2}/b$, $a$ and $b$ are the semi-major and semi-minor axes of an oblate ellipsoid
respectively. Beyond $\sim 3$ AU in the ecliptic plane and $\sim 1.5$ AU off the ecliptic plane, no reliable density values can be obtained from the zodiacal light. Therefore we assume $a=3$ AU and $b=1.5$ AU for the semi-axes and $\gamma_E=\sqrt{3}$ \cite{Giese}. It is worth noticing that the total ID mass amount inside the considered region is of the order of $10^{17}$ kg, i.e., $\approx 10^{-8} \ M_{\oplus}$ .
The ellipsoid model of ID density depends on two parameters, $\alpha$ and $\gamma_E$, and, in order to
numerically study gravitational effects of ID on LISA orbits, we will consider four
cases: i) \textsl{spherical homogeneous distribution} ($\alpha=0,\gamma_E=0$), ii) \textsl{spherical distribution with power-law
density profile} ($\alpha=1.3,\gamma_E=0$), iii) \textsl{ellipsoidal homogeneous distribution}
($\alpha=0,\gamma_E=\sqrt{3}$), and iv) \textsl{ellipsoidal distribution with power-law density
profile} ($\alpha=1.3,\gamma_E=\sqrt{3})$.
\section{Method}
\label{method}
As the first step we define the unperturbed LISA orbits\footnote{In this paper we neglect post-Newtonian and relativistic effects.}. The cartesian coordinates of each LISA reference mass ($k=1,2,3$) are related to the keplerian orbital elements defined in the SSB through the following equations:
\begin{equation} \label{eqn: orbita}
\left\{
\begin{array}{l}
x_k(t) = a_k \left[\left(\cos \Omega_k \cos \omega_k - \sin \Omega_k \sin \omega_k \cos i_k \right) \left(\cos \psi_k(t) - e_k \right) \right] \\
+ a_k \left[\left(- \cos \Omega_k \sin \omega_k - \sin \Omega_k \cos \omega_k \cos i_k \right) \sqrt{1 - e_k^2} \sin \psi_k(t) \right] \\
\\
y_k(t) = a_k \left[\left(\cos \Omega_k \cos \omega_k + \cos \Omega_k \sin \omega_k \cos i_k \right) \left(\cos \psi_k(t) - e_k \right) \right] \\
+ a_k \left[\left(- \sin \Omega_k \sin \omega_k + \cos \Omega_k \cos \omega_k \cos i_k \right) \sqrt{1 - e_k^2} \sin \psi_k(t) \right]\\
\\
z_k(t) = a_k \left[\sin \omega_k \sin i_k \left(\cos \psi_k(t) - e_k \right) + \cos \omega_k \sin i_k \sqrt{1 - e_k^2} \sin \psi_k(t) \right] \ ,
\end{array}
\right.
\end{equation}
where $a_k$ is the semi-major axis, $e_k$ the eccentricity, $i_k$ the inclination of the orbit with respect to the ecliptic plane, $\omega_k$ the argument of periapsis, $\Omega_k$ the longitude of ascending node and $\psi_k(t)$ the eccentric anomaly \cite{Boccaletti}.
We chose the initial conditions providing rigid triangular configuration of
side $l\simeq 5 \times 10^6\ \mbox{km} \simeq 1/30$ AU
\begin{equation}
\left\{
\begin{array}{l}
a_k = 1 \ \mbox{AU}\\
\\
e_k = \left(1 + \frac{2}{\sqrt{3}} \xi + \frac{4}{3} \xi^2 \right)^{1/2} - 1 \\
\\
\tan i_k = \frac{\xi}{1 + \xi/\sqrt{3}} \\
\\
\omega_k = \frac{\pi}{2} \\
\\
\Omega_k = \frac{2}{3} \pi (k - 1) - \frac{\pi}{2} \\
\\
M_k(t) = 2 \pi t - \frac{2}{3} \pi (k - 1) - \pi ,
\end{array}
\right.
\label{Initial}
\end{equation}
where $\xi = l/(2 a_k) \simeq 1.67 \times 10^{-2}$ and $M_k(t)$ is the mean anomaly.
Such orbits minimize the variation of the inter-spacecraft distance between the $i$-th and $j$-th
satellite $\Delta L_{i,j} $ $\equiv |\vec{r}_i(t)-\vec{r}_j(t)|$ \cite{Jiang}, which turns out to be independent of time at the first order in $\xi$ \cite{Dhurandhar}.
This requirement must be fulfilled in order to operate LISA successfully.
In fact, the measured quantities for the GW searches are the differential relative motion between the couples $(i,j)$ and $(j,l)$ of the three reference masses, $\delta L_{i,j,l} = \Delta L_{i,j} - \Delta L_{j,l}$ (see Fig. (\ref{variazione_distanze_imp})). Each couple can be regarded as an unequal arm interferometer and the
doppler shift induced by the relative differential motions can be measured by a suitable Time Delay Interferometry
(TDI) \cite{TINTO} with a strain sensitivity in Fig. (\ref{sensitivity}) \cite{Bender2}.
\begin{figure}[!hbt]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width = 5.9 cm, height = 5.0 cm]{dist12.pdf}&
\includegraphics[width = 5.9 cm, height = 5.0 cm]{diff1223.pdf}
\end{tabular}
\caption{Left, relative motion of 1 and 2 reference masses $\Delta L_{1,2}$ as a function of time. Right, differential relative motion $\delta L_{1,2,3}$ between couples 1-2 and 2-3 of reference masses as a function of time. For different couples, plots are identical but shifted in time.}
\label{variazione_distanze_imp}
\end{center}
\end{figure}
\begin{figure}[!hbt]
\begin{center}
\includegraphics[width = 8.9 cm, height = 7.2 cm]{sensitivity.pdf}
\caption{LISA sensitivity curve as a function of frequency. The curve in the dashed region
is calculated from ref. \cite{Bender2}, where it is discussed the sensitivity curve up to $10^{-6}$ Hz.}
\label{sensitivity}
\end{center}
\end{figure}
The keplerian motions of the LISA reference masses are periodic. We can distinguish their harmonics
by looking at the modulus of the Fourier transform of $\delta L_{i,j,l}(t)$ \footnote{The Fourier transform of $F(t)$ is as usual $\widetilde{F}(\omega)=
\int_{-\infty}^{+\infty} F(t) \exp(- i \omega t) dt$.}, as in
Fig. (\ref{spettriimp}), where we have considered a finite observation time $T$ and, to overcome the
spectral leakage, we have applied a Blackman tapering function \cite{Proakis}.
\begin{figure}[!hbt]
\begin{center}
\includegraphics[width = 8.6 cm, height = 7.2 cm]{fourierimpert.pdf}
\caption{$|\widetilde{\delta L}_{1,2,3}(\omega)|$ windowed by a Blackman function with $T=30$ y
(1 $\mbox{y}^{-1} \approx 3.1709792 \times 10^{-8}$ Hz).
}
\label{spettriimp}
\end{center}
\end{figure}
It is worth noticing the absence of the harmonics at 3, 6, 9 \ldots \ $\mbox{y}^{-1}$ due to the
presence of a discrete symmetry in the equations of motion of the constellation. In fact,
as a consequence of the initial conditions in Eq. (\ref{Initial}), the LISA triangle rotates as a ``quasi-rigid body"
with period of one year and, after integer multiples of one third of year, the dynamical configuration
is identical to the initial one under any cyclic permutation of the reference mass indices.
However, such a feature is only of theoretical interest, because slightly different initial conditions,
due for instance to the unavoidable injection errors of LISA spacecrafts in their orbits (position $\simeq 2$ Km and velocity $\simeq 2$ mm/s, \cite{Sweetser}), will produce harmonics of the 3 $\mbox{y}^{-1}$ frequency.
To study the effects of ID on LISA orbits, we apply the perturbation theory in the six-dimensional
space of parameters and use the Gauss planetary equations.
Such equations provide time evolution of the orbital parameters under a generic perturbing acceleration field
$\vec{\gamma}(\vec{r})$ and read
\begin{equation} \label{eqn: gausseq}
\left\{
\begin{array}{l}
\frac{d a_k}{dt} = \frac{2}{n_k \sqrt{1 - e_k^2}} \left[e_k A_r \sin \theta_k + A_t \left(\frac{p_k}{r} \right) \right]\\
\\
\frac{de_k}{dt} = \frac{\sqrt{1 - e_k^2}}{n_k a_k} \left\{ A_r \sin \theta_k + A_t \left[\cos \theta_k + \frac{1}{e_k} \left(1 - \frac{r}{a_k} \right) \right] \right\} \\
\\
\frac{d i_k}{d t} = \frac{1}{n_k a_k \sqrt{1 - e_k^2}} A_n \left(\frac{r}{a_k} \right) \cos (\omega_k + \theta_k)\\
\\
\frac{d \Omega_k}{dt} = \frac{1}{n_k a_k \sin i_k \sqrt{1 - e_k^2}} A_n \left(\frac{r}{a_k}\right) \sin (\omega_k + \theta_k)\\
\\
\frac{d \omega_k}{d t} = - \cos i_k \frac{d \Omega_k}{d t} + \frac{\sqrt{1 - e_k^2}}{n_k a_k e_k} \left[-A_r \cos \theta_k + A_t \left(1 + \frac{r}{p_k} \right) \sin \theta_k \right]\\
\\
\frac{d M_k}{d t} = n_k - \frac{2}{n_k a_k} A_r \left(\frac{r}{a_k} \right) - \sqrt{1 - e_k^2}\frac{d \omega_k}{dt} \ \ ,
\end{array}
\right.
\end{equation}
where $p_k = a_k(1 - e_k^2)$ is the semi-latus rectum, $\theta_k$ is the true anomaly, $n_k = \sqrt{4 \pi^2/a_k^3}$ is the keplerian mean motion and $A_r, A_t$ and $A_n$ are the components of $\vec{\gamma}$ along the versors $\hat{r} = \vec{r}/ \left\| \vec{r} \right\|$ in the radial direction, $\hat{t}$, orthogonal to $\hat{r}$ in the osculating plane and in the direction of $\dot{\vec{r}}$, and $\hat{n} = \hat{r} \times \hat{t}$ \cite{Boccaletti}.
For the generic point $\vec{r} = \left(x, y, z \right)$ close to the ecliptic plane ($i_k \ \approx 10^{-2}$ rad $\forall k$), the accelerating field $\vec{\gamma}(\vec{r})$ has been calculated from the internal gravitational potential of the distribution in Eq. (\ref{eqn: model}),
$$\Phi(R, z) = \varphi \sum_{n = 0}^{\infty}
\left(
\begin{array}{c}
1 - \frac{\alpha}{2} \\
n \\
\end{array}
\right) \times$$
\begin{equation}
\times \int_0^{\infty} \left( \frac{R^2}{\tau + a^2}\right)^{1 - \frac{\alpha}{2}} \left[\frac{\tau + a^2}{R^2} \frac{1}{\tau + b^2} \right]^n \frac{z^{2n}}{\left(\tau + a^2 \right) \sqrt{\tau + b^2}} d \tau,
\end{equation}
where $R = \sqrt{x^2 + y^2}$, $\varphi = 2 \pi \rho_0 G r_0^{\alpha} a^{3 - \alpha} \sqrt{1 - \gamma_E^2/\left(1 + \gamma_E^2 \right)} / \left(2 - \alpha\right)$ and $G$ is the Newton gravitational constant. The gravitational potentials of the considered distributions can be easily obtained for the appropriate values of $\alpha$ and $\gamma_E$.
We then numerically integrated the Gauss equations for the three LISA satellites adopting different ID distributions\footnote{We made use of the \emph{Mathematica 7} numerical integrator \emph{NDSolve}; see the documentation webpage http://reference.wolfram.com/mathematica/ref/NDSolve.html}. In order to check the effectiveness of the linear perturbation theory of the studied perturbations, we have compared our results with the numerical solutions obtained by keeping the orbital elements on the RHS unperturbed, and they matched exactly within the integration time (30 years). In addition, the accuracy of the numerical integrations ($\approx 10^{-30}$) have been verified by comparison with analytical solutions of Gauss equations (approximated to the fourth order in $e$ and with unperturbed orbital parameters on the RHS) for a spherical homogeneous distribution of matter \cite{Cerdonio} and the agreement has been satisfactory. To get the perturbed orbits, and, therefore, the perturbative effects of ID on LISA constellation, we substituted the unperturbed orbital parameters appearing in Eq. (\ref{eqn: orbita}) with the solutions of Eq. (\ref{eqn: gausseq}).
\section{Results}
\label{results}
In Fig. (\ref{diffmotellalpha}), we report the plot representing
\begin{equation}
\delta l_{1,2,3} = \delta L_{1,2,3}^{pert} - \delta L_{1,2,3}^{unpert},
\end{equation}
i.e., the difference between perturbed and unperturbed differential motions of the constellation for the four considered distributions.
\begin{figure}[!hbt]
\begin{center}
\includegraphics[width = 8.5 cm, height = 7.3 cm]{dust.pdf
\caption{(color on line) Time evolution of $\delta l_{1,2,3}$, for the spherical ($\gamma_E = 0$) distribution (thin dashed line, $\alpha = 0$, thin solid line, $\alpha = 1.3$) and ellipsoidal ($\gamma_E = \sqrt{3}$) (thick dashed line, $\alpha = 0$, thin solid line, $\alpha = 1.3$) distribution. It is worth noticing that ID induces effects of the order of $10^2 \ \mu\mbox{m}$, $10^{12}$ times smaller than those due to the keplerian differential motions.}
\label{diffmotellalpha}
\end{center}
\end{figure}
We note that the perturbative effects of ID are very small, in fact the LISA constellation opens with an amplitude of the order of $10^2 \ \mu$m after 5 complete orbits. Fig. (\ref{diffspectraellalpha}) shows the modulus of the Fourier transform of the difference between perturbed and unperturbed differential motions $\delta l_{1,2,3}$, for the ellipsoidal ID distribution with radial profile, and we see that ID enhances resonance peaks, in particular the 2 $\mbox{y}^{-1}$ harmonic. In addition, new harmonics corresponding to integer multiples of 3 $\mbox{y}^{-1}$ appear, as a consequence the ID gravitational field that breaks the permutation symmetry of the equations of motion.
\begin{figure}[!hbt]
\begin{center}
\includegraphics[width = 8.8 cm, height = 7.2 cm]{findiff1223pert.pdf}
\caption{Difference between $\delta L_{1,2,3}^{pert}$ and $\delta L_{1,2,3}^{unpert}$ spectra of frequencies after 30 orbits, windowed with Blackman function, obtained with $\alpha = 1.3$ and $\gamma_E = \sqrt{3}$; we note the presence of the harmonics
corresponding to the integer multiples of 3 $\mbox{y}^{-1}$, contrary to Fig (\ref{spettriimp}).}
\label{diffspectraellalpha}
\end{center}
\end{figure}
Therefore, our analysis for the different ID distributions listed at the end of Sec. (\ref{id}) shows that curves in Fig. (\ref{diffmotellalpha}) and Fig. (\ref{diffspectraellalpha}) change their amplitudes of a factor $\approx 2.5$ by varying $\gamma_E$ from $\sqrt{3}$ to 0, while they do not depend significantly on $\alpha$. Additionally, it is easy to show that LISA sensitivity curve is not significantly affected by the ID perturbing effects in the $10^{-4} - 10^{-1}$ Hz frequency band.
\subsection{Optimal ID signal resolution}
\label{resolution}
It is of some interest to investigate the problem of resolving the contributions to the
differential relative motions due to Sun and ID by means of optimal filtering.
From the theory of signals resolution we known that if a noisy signal $v(t)$
can be a linear combination of two given signals $f(t)$ and $g(t)$ of unknown amplitudes $A$ and $B$
\begin{equation}
v(t) = A f(t) + B g(t) + n(t),
\end{equation}
where $n(t)$ is a zero mean gaussian stochastic process with correlation $\langle n(t)n(t')\rangle = S_0\delta(t-t')$,
the four following hypothesis are possible:
\begin{enumerate}
\item $H_0$: neither signal $A$ nor signal $B$ is present;
\item $H_1$: signal $A$ alone is present;
\item $H_2$: signal $B$ alone is present;
\item $H_3$: both signal $A$ and $B$ are present,
\end{enumerate}
which are to be verified by comparing the expectation values of $A$ and $B$, with zero (test of hypothesis).
The best estimate of $A$ and $B$ turns out to be \cite{Helstrom}
\begin{eqnarray}
\hat{A} &=& \frac{1}{1 - \lambda^2} \int_0^T \left[f(t) - \lambda g(t) \right] v(t) dt\nonumber \\
\hat{B} &=& \frac{1}{1 - \lambda^2} \int_0^T \left[g(t) - \lambda f(t) \right] v(t) dt \ ,
\end{eqnarray}
with standard deviations
\begin{equation}
\sigma_{\hat{A}} = \sigma_{\hat{B}} = \frac{S_0^{1/2}}{\sqrt{2 \left(1 - \lambda^2 \right)}} \ ,
\end{equation}
where $\lambda = \int_0^T f(t)g(t)dt$, $S_0$ is the noise power density and $f(t)$ and $g(t)$ are normalized to unit energy, $\int_0^T f^2(t) dt = \int_0^T g^2(t) dt = 1.$
In our case, $v(t)$ represents the perturbed differential motions, $Af(t)$ and $Bg(t)$ are the contributes due to Sun and ID respectively. We stress that $f(t)$, $g(t)$, $\hat{A}$ and $\hat{B}$ are almost independent of the initial conditions of LISA reference masses.
It is worth noticing that the two contributions to the perturbed motion are almost orthogonal,
($\lambda \simeq 10^{-2}$)
as can be seen in Fig. (\ref{orthogonal}), where the normalized functions $f(t)$ and $g(t)$ are plotted.
\begin{figure}[!hbt]
\begin{center}
\includegraphics[width = 8.3 cm, height = 6.5 cm]{segnali.pdf}
\caption{(color on line) Time evolution of the two normalized functions $f(t)$ (blue solid line) and $g(t)$ (red dashed line) for the case of an ellipsoidal distribution with radial profile. }
\label{orthogonal}
\end{center}
\end{figure}
As a crude estimation, a strain sensitivity $S_h \approx 10^{-11} \ 1/ \sqrt{\mbox{Hz}}$ at the 2 $\mbox{y}^{-1}$ $ (\approx 6.4 \cdot10^{-8}$ Hz) frequency (the most powerful ID harmonic), is required to achieve a unitary signal-to-noise ratio (SNR = $\hat{A}/\sigma_{\hat{A}} = \hat{B}/\sigma_{\hat{B}}$.) in a 5 years observation time. However, in the literature there exist extrapolations of the strain sensitivity $S_h$ only up to $10^{-6}$ Hz \cite{Bender2} and the measure of such a small signal amplitude depends on the noise level of LISA at very low frequencies ($ 10^{-7}$ Hz), which is still largely unknown.
\section{Discussion}
\label{disc}
We have calculated the ID perturbing effects on the LISA differential motions.
Using the estimated ID density, we found a continuous opening of the LISA orbits of the order of $10^2 \ \mu$m after 5 years at $2\ \mbox{y}^{-1}$ frequency. Moreover, due to the particular LISA orbits, which are nearly circular and very close to the ecliptic plane, we found similar effects for the distributions we have taken into account. In fact, the ID tidal forces vary significantly over a scale much larger than the LISA triangle sides.
All the results we presented for ID also hold in the presence of LDM. In fact, we have shown that the numerical solutions of Eq. (\ref{eqn: gausseq}), for the LISA orbits, are almost independent on $\alpha$ and $\gamma_E$ parameters within the investigated range and of the injection errors. In addition, the perturbations on LISA relative differential motions $\delta L_{i,j,k}$ scale linearly with $\rho_0$.
Thus we can account for LDM just by a rescale of $\rho_0$: for instance, if we assume that the LDM density value is close to the average galactic dark matter (GDM) density value, $\rho_{0,GDM} = 5 \cdot 10^{-22} \ \mbox{kg/m}^3$ as obtained from the galaxy rotation curve, the effect of DM is expected to be 0.5\% that of ID.
As a consequence, LISA could provide interesting upper limits on both $\rho_{0,ID}$ and $\rho_{0,LDM}$, depending on low frequencies noise and $\rho_{0}$ value, by means of direct gravitational field measurements instead of observations of zodiacal light in the case of ID. At present time, the best upper limits $\rho_{0,LDM} < 3 \cdot 10^{-16}$ kg/$\mbox{m}^3$ are based on the study of the precession of the perihelions of Mercury, Earth and Mars \cite{Khirplovich}.
However, we stress that a thorough study of LISA displacement noise below $1 \ \mu$Hz, including local gravitational field fluctuations, thruster noise, orbit determination and injection errors \cite{Bik} etc., is required to establish the relevance of the ID effects in LISA physics.
\section{Conclusions}
The study of the Solar System gravitational field acting on LISA reference masses is of some relevance also for the detection of GWs. In fact, the non-radiative gravitational perturbations on LISA keplerian motions must be subtracted at the highest accuracy from the relative differential motion in order to measure the GW contributions. In this paper we have calculated the time evolutions of the orbital parameters due to some ID distributions and estimated the opening induced on the LISA constellation, $\approx 10^2 \ \mu$m after 5 years. The Fourier transform of the perturbation of the differential relative motion has shown an enhancement of the resonances characterizing the unperturbed spectra, in particular the peak at $2\ \mbox{y}^{-1}$ frequency. As a consequence of the LISA orbits, such effects are similar for the
studied distributions of ID and do not affect the LISA sensitivity band for GW detection.
On the other hand, the discrimination of very small perturbations from the keplerian differential relative motion depends crucially on the low frequency displacement noise, that in the ideal case should be $\simeq 10^{-11}\ 1/\sqrt{\mbox{Hz}}$ at $2\ \mbox{y}^{-1}$ frequency to detect ID at the density $\sim 10^{-20}\ \mbox{kg/m}^3$ measured by the zodiacal light.
Unfortunately, a reliable estimate of this noise level at very low frequency is still lacking and requires
futher investigations.
Finally, we investigate the possibility of constraining the LDM density with LISA. Even though gravitational
perturbations
due to ID and LDM are undistinguishable, the study of the deviations from the keplerian orbits of LISA reference masses could provide interesting upper limits on LDM density.
\section{Acknowledgments}
MS was supported by the Swiss National Science Foundation and the Dr. Tomalla Foundation. GM was supported by the Italian INFN (Istituto Nazionale di Fisica Nucleare). MC and GM are grateful to Professor G. Benettin for useful discussions.
|
\section{Introduction}
\begin{comment}
On the coding front, Reed-Solomon (RS) codes have been the choice
for a wide range of applications in digital communication and data
storage systems. In most of the cases, the classical hard-decision
(HD) decoding algorithm of Berlekamp-Massey (BM) with low complexity
has been employed. However, HD decoding algorithms do not take advantage
of the soft-information available at the receiver. The potential gain
from making use of such information is very significant as exemplified
by the improvement of the soft-decision Koetter-Vardy (KV) algorithm
\cite{Koetter-it03} over the hard-decision Guruswami-Sudan (GS) counterpart
\cite{Guruswami-it99}. One of the current interests of researchers
is to find soft-decision decoding methods for RS codes with improved
performance at a reasonable expense of complexity.
\end{comment}
{}The design of a computationally efficient soft-decision decoding algorithm
for Reed-Solomon (RS) codes has been the topic of significant research
interest for the past several years. Currently, there are several
soft-decision decoding algorithms for RS codes which exhibit a wide
range of trade-offs between computational complexity and error performance.
Among such decoding methods is a class of algorithms called multiple
errors-and-erasures decoding. The algorithms belonging to this class
first construct a set of erasure patterns based on the available soft
information and then run an errors-and-erasures decoding algorithm,
such as the Berlekamp-Massey (BM) algorithm, multiple times. Each
time one erasure pattern in the set is used for decoding. By doing
this, the algorithm outputs a list of candidate codewords and then
chooses the best codeword from the list. Several algorithms of this
type, including the popular generalized minimum distance (GMD) decoding
algorithm, are discussed in \cite{Forney-it66,Lee-globecom08,Bellorado-isit06,Nguyen-aller09}.
In \cite{Nguyen-aller09}, the authors proposed a rate-distortion
(RD) approach for constructing the set of erasure patterns. The main
idea is to choose an appropriate distortion measure so that the decoding
is successful if and only if the distortion between the error pattern
and erasure pattern is smaller than a fixed threshold. After that,
a set of erasure patterns is generated randomly (similar to a random
codebook generation) in order to minimize the expected minimum distortion.
The approach was also extended to analyze multiple-decoding for decoding
schemes beyond conventional errors-and-erasures decoding.
\begin{singlespace}
One of the drawbacks in the RD approach is that the mathematical framework
is only valid as the block-length goes to infinity. Therefore, we
also consider the natural extension to a rate-distortion \emph{exponent}
(RDE) approach that studies the behavior of the probability, $p_{e}$,
that the transmitted codeword is not on the list as a function of
the block-length $N$. The overall error probability can be approximated
by $p_{e}$ because the probability that the transmitted codeword
is on the list but not chosen is very small compared to $p_{e}$.
Hence, our new approach essentially focuses on investigating the exponent
at which the error probability decays as $N$ goes to infinity.
\end{singlespace}
The proposed RDE approach can also be considered as the generalization
of the RD approach since the latter is a special case of the former
when the RDE function tends to zero. Using the RDE analysis, our proposed
approach also helps answer the following two questions: (i) What is
the maximum rate-distortion exponent achievable at or below a given
number of decoding attempts (or a given size of the set of erasure
patterns)? (ii) What is the minimum number of decoding attempts required
to achieve a rate-distortion exponent at or above a given level?
The paper is organized as follows. In Section \ref{sec:MultipleBMA},
we review multiple errors-and-erasures decoding algorithms and highlight
the connection between multiple errors-and-erasures decoding and rate-distortion.
Then, in Section \ref{sec:EXP-approach}, we propose a RDE approach
to construct a good set of erasure patterns for a finite length codewords.
Next, we discuss how to compute the RDE function which is required
in the proposed approach. Finally, simulation results are presented
in Section \ref{sec:Simulation} and conclusion is provided in Section
\ref{sec:Conclusion}.
\section{Multiple errors-and-erasures Decoding\label{sec:MultipleBMA}}
In this section, we discuss several multiple errors-and-erasures decoding
algorithms. While each algorithm uses a different set of erasure patterns,
the common trend is that one either erases or tries several different
candidates for each symbol in the least reliable positions (LRPs).
One focuses on the LRPs because the hard-decision made at these positions
are more likely to be incorrect.
Let $\mathbb{F}_{m}$ be the Galois field with $m$ elements denoted
as $\alpha_{1},\alpha_{2},\ldots,\alpha_{m}$. We consider an $(N,K$)
RS code of length $N$ and dimension $K$ over $\mathbb{F}_{m}$.
Assume that we send a codeword $\mathbf{c}=(c_{1},c_{2},\ldots,c_{N})$
over some channel and $\mathbf{r}=(r_{1},r_{2},\ldots,r_{N})$ is
the received vector. A well-known decoding threshold states that a
single attempt of errors-and-erasures decoding succeeds if and only
if\vspace{-2mm} \begin{equation}
2v+e<d_{min}=N-K+1\label{eq:decthreshold}\end{equation}
\vspace{-6mm}~
\hspace{-3.5mm}where $e$ is number of erased symbols and $v$ is
the number of errors in unerased positions. A multiple errors-and-erasures
decoding is considered to succeed if the decoding threshold (\ref{eq:decthreshold})
is satisfied for at least one attempt of decoding. Intuitively, the
best case is when one erases an error and the worst case is when ones
wastes an erasure on a hard-decision symbol that turns out be correct.
The first algorithm of this type is called Generalized Minimum Distance
(GMD) decoding \cite{Forney-it66} where the set of erasure patterns
is obtained by successively erasing the $0,2,4,\ldots,d_{min}-1$
LRPs (with the assumption that the minimum distance $d_{min}$ is
odd). Recent work by Lee and Kumar \cite{Lee-globecom08} proposes
a soft-information successive (multiple) error-and erasure decoding
(SED) which constructs the set of erasure patterns in a more exhaustive
way. Specifically, SED$(l,f)$ tries to erase all possible combinations
of an even number less than or equal to $f$ of positions within the
$l$ LRPs. The SED algorithm hence yields better performance but at
increased complexity.
In an attempt to answer the question how to build a good set of erasure
patterns in terms of performance-versus-complexity, in \cite{Nguyen-aller09},
we proposed a probabilistic method based on rate-distortion theory
and random coding arguments instead of the deterministic methods which
had been used in previously proposed algorithms. Basically, after
defining $x^{N}$ and $\hat{x}^{N}$ as an error pattern and an erasure
pattern whose letters $x_{i}$'s and $\hat{x}_{i}$'s are in the alphabets
$\mathcal{X}$ and $\hat{\mathcal{X}}$ respectively, a letter-by-letter
distortion measure $\delta:\mathcal{X}\times\hat{\mathcal{X}}\to\mathbb{R}_{\geq0}$
is chosen properly so that the condition (\ref{eq:decthreshold})
can be reduced to the form\vspace{-1mm}\begin{equation}
d\left(x^{N},\hat{x}^{N}\right)<N-K+1\label{eq:distorthreshold}\end{equation}
\vspace{-6mm}~
\hspace{-3.5mm}where the distortion between an error pattern and
an erasure pattern $d\left(x^{N},\hat{x}^{N}\right)=\sum_{i=1}^{N}\delta(x,\hat{x}_{i})$
is smaller than a fixed threshold. In general, an appropriate distortion
measure $\delta(j,k)$ for every $j\in\mathcal{X}$ and $k\in\hat{\mathcal{X}}$
should be specified.
\begin{example}
Consider a specific class of multiple errors-and-erasures (Berlekamp-Massey)
top-$\ell$ decoding (mBM-$\ell$) for an positive integer $\ell$
smaller than the field size $m$ where at each codeword index, up
to the $\ell$-th most likely symbols are taken care of. In this case,
$\mathcal{X}=\hat{\mathcal{X}}=\mathbb{Z}_{l+1}$ and $x^{N}\in\mathcal{X}^{N}$
where at each index $i$, $x_{i}=0$ implies that using up to the
$\ell$-th most likely symbols as the hard-decision all gives an error,
$x_{i}=j$ implies that the $j$-th most likely symbol is correct
for $j=1,2,\ldots,\ell$; $\hat{x}^{N}\in\hat{\mathcal{X}}^{N}$ where
at each index $i$, $\hat{x}_{i}=0$ implies that an erasure is applied
and $\hat{x}_{i}=k$ implies that the $k$-th most likely symbol is
used as the hard-decision for $k=1,2,\ldots,\ell$. For example, mBM-1
is the case of multiple conventional errors-and-erasures decoding.
The letter-by-letter distortion measure for mBM-1 is chosen in the
following way\vspace{-1mm}\begin{equation}
\begin{array}{cc}
\delta(0,0)=1 & \delta(0,1)=2\\
\delta(1,0)=1 & \delta(1,1)=0.\end{array}\label{eq:dstmb1}\end{equation}
\vspace{-4mm}
It is also possible to choose appropriate distortion measures that
work for $\ell>1$ and other decoding schemes such as algebraic soft-decision
(ASD) decoding. Still, the main idea is to convert the decoding threshold
of the corresponding decoding scheme into the form of (\ref{eq:distorthreshold}).
Thus, by viewing $x^{N}$ (resp. $\hat{x}^{N}$) as a source sequence
(resp. reproduction sequence) and choosing a suitable distortion measure,
the task of designing a good set of erasure patterns turns out to
be how to best approximate the source sequence with a minimum number
of reproduction sequences. In other words, it becomes a covering problem
where one wants to cover the most-likely error patterns with the fewest
number of balls whose centers are erasure patterns. The main steps
in the RD based algorithm are given here briefly, but more detail
can be found in \cite{Nguyen-aller09}.
\emph{Step 1}: Empirically compute the reliability matrix whose entries
are $\Pr(c_{i}=\alpha_{j}|r_{i})$ for $i=1,2,\ldots N$ and $j=1,2,\ldots m$.
From this, determine probability matrix $\mathbf{P}$ where $p_{i,j}=\Pr(x_{i}=j$)
for $i=1,2,\ldots,N$ and $j\in\mathcal{X}$.
\emph{Step 2}: Compute the RD function using $\mathbf{P}$. Determine
the test-channel input-distribution matrix $\mathbf{Q}$ where $q_{i,k}=\mbox{\ensuremath{\Pr}}(\hat{x}_{i}=k)$
for $i=1,2\ldots,N$ and $k\in\mathcal{\hat{X}}$ that achieves a
point on the RD curve corresponding to a chosen rate $R$.
\emph{Step~3: }Randomly generate a set $\mathcal{B}$ of~$2^{R}$~erasure
patterns using the distribution matrix $\mathbf{Q}$ in the correct
reliability order of the codeword positions.
\emph{Step~4:} Run multiple attempts of the corresponding decoding
scheme using the set $\mathcal{B}$ to produce a list of candidate
codewords.
\emph{Step~5:} Use Maximum-Likelihood (ML) decoding to pick the best
codeword on the list.\vspace{-2mm}
\end{example}
\section{RATE-DISTORTION EXPONENT APPROACH\label{sec:EXP-approach}}
In the RD approach, the set $\mathcal{B}$ of $2^{N\bar{R}}$ (or
$2^{R}$) erasure patterns can be generated randomly so that%
\footnote{We denote the rate and distortion by $R$ and $D$, respectively,
using unnormalized quantities, i.e., $R=N\bar{R}$ and $D=N\bar{D}$.%
}\vspace{-2mm} \[
\lim_{N\rightarrow\infty}\frac{1}{N}E_{x^{N},\mathcal{B}}[\min_{\hat{x}^{N}\in\mathcal{B}}d(x^{N},\hat{x}^{N})]<\bar{D}.\]
\vspace{-4mm}~
Thus, for large enough $N$, with high probability we have $\min_{\hat{x}^{N}\in\mathcal{B}}d(x^{N},\hat{x}^{N})<N\bar{D}=D$.
Basically, \cite{Nguyen-aller09} focuses on minimizing the average
minimum distortion with little knowledge of how the tail of the distribution
behaves. In this paper, we instead focus on directly minimizing the
probability that the minimum distortion is not less than the pre-determined
threshold $D=N-K+1$ (due to the condition (\ref{eq:distorthreshold}))
with the help of an error-exponent analysis. The exact probability
of interest is $p_{e}=\Pr(x^{N}:\min_{\hat{x}^{N}\in\mathcal{B}}d(x^{N},\hat{x}^{N})>D)$
that reflects how likely the decoding threshold (\ref{eq:decthreshold})
is going to fail.
In other words, every error pattern $x^{N}$ can be covered by a sphere
centered at an erasure pattern $\hat{x}^{N}$ except for a set of
error patterns of probability $p_{e}$. The RDE analysis shows that
$p_{e}$ decays exponentially as $N\to\infty$ and the maximum exponent
attainable is the RDE function. In our context, we have a source sequence
$x^{N}$ of $N$ independent non-identical source components. We denote
the rate-distortion exponent by $F(R,D)$ using unnormalized quantities
(i.e., without dividing by $N$) and note that exponent used by other
authors in \cite{Blahut-it74,Marton-it74,Csiszar-1981} is often the
normalized version $\bar{F}(R,D)\triangleq\frac{F(R,D)}{N}$.
The original RDE function $F(R,D)$, defined in \cite{Blahut-it74}
for a single source $x$, is given by%
\footnote{All logarithms are taken to base 2.%
}\vspace{-2mm}\[
F(R,D)=\max_{\mathbf{w}}\min_{\mathbf{\tilde{p}}\in\mathcal{P}_{R,D}}\sum_{j}\tilde{p}_{j}\log\frac{\tilde{p}_{j}}{p_{j}}\]
\vspace{-3mm}~
\hspace{-3.5mm}where $p_{j}\triangleq\Pr(x=j)$, $w_{k|j}\triangleq\Pr(\hat{x}=k|x=j)$,
and\vspace{-2mm}
\[
\mathcal{P}_{R,D}=\left\{ \mathbf{\tilde{p}}\bigg|\begin{array}{c}
\sum_{j}\sum_{k}\tilde{p}_{j}w_{k|j}\log\frac{w_{k|j}}{\sum_{j}\tilde{p}_{j}Q_{k|j}}\geq R\\
\sum_{j}\sum_{k}\tilde{p}_{j}w_{k|j}\delta_{jk}\geq D\end{array}\right\} .\]
\vspace{-2mm}%
\begin{comment}
\[ \begin{array}{ccc} \mathcal{P}_{R,D} & = & \Big\lbrace\tilde{p}: \sum_{j}\sum_{k}\tilde{p}_{j}w_{k|j}\log\frac{w_{k|j}}{\sum_{j}\tilde{p}_{j}Q_{k|j}}\geq R\\ & & \sum_{j}\sum_{k}\tilde{p}_{j}w_{k|j}\delta_{jk}\geq D\Big\rbrace.\end{array}\]
\end{comment}
{}
The RDE was first extensively discussed in \cite{Blahut-it74,Marton-it74}
and their results show that there exists a set $\mathcal{B}$ of roughly
$2^{N\bar{R}}$ codewords, generated randomly using the test-channel
input distribution matrix $\mathbf{Q}$, that achieves $\bar{F}(R,D)$.
This gives the upper bound that for every $\epsilon>0$, we have\vspace{-1mm}
\begin{equation}
p_{e}\leq2^{-N[\bar{F}(R,D)-\epsilon]}.\label{eq:BlahutUB}\end{equation}
\vspace{-6.5mm}~
\hspace{-3.5mm}for $N$ large enough (see \cite[p. 229]{Blahut-1987}).
An exponentially tight lower bound for $p_{e}$ can also be obtained
for $N$ large enough (see \cite[p. 236]{Blahut-1987}) and this gives\vspace{-2mm}\[
\lim_{N\to\infty}-\frac{1}{N}\log p_{e}=\bar{F}(R,D).\]
\vspace{-5mm}~
\hspace{-3.5mm}\noindent\hspace{1em}{\itshape Proposed algorithm:}
In the RDE approach proposed here, instead of computing the RD function,
we need to compute the RDE function $F(R,D)$ along with the optimal
test-channel input distribution matrix $\mathbf{Q}$ (see Section
\ref{sec:EXP-func}). This distribution serves as a replacement for
the distribution used in Step 2 of the RD based algorithm given in
the previous section. Apart from this, the other steps of that algorithm
are unchanged for the proposed RDE-based algorithm.
\begin{remrk}
\label{rem:The-RDE-approach}The RDE approach possesses several advantages.
First, it can help one estimate the smallest number of decoding attempts
to get to a RDE of $F$ (or get to an error probability of roughly
$2^{-N\bar{F}}$) or, similarly, allow one to estimate the RDE (and
error probability) for a fixed number of decoding attempts. Second,
it provides a converse based on the sphere-packing bound lower bound
for $p_{e}$. This implies that, given an arbitrary set $\mathcal{B}$
of roughly $2^{N\bar{R}}$ erasure patterns and any $\epsilon>0$,
the probability $p_{e}$ cannot be made lower than $2^{-N[\bar{F}(R,D)+\epsilon]}$
for $N$ large enough. Thus, no matter how one chooses the set $\mathcal{B}$
of erasure patterns, the difference between the induced probability
of error and the $p_{e}$ for the RDE approach becomes negligible
for $N$ large enough.\vspace{-4.3mm}
\end{remrk}
~
\begin{remrk}
It is interesting to note that the RDE approach for ASD decoding schemes
contains the special case where the codebook has only one entry (i.e.,
ASD decoding is run one time). In this case, the distribution of the
codebook that maximizes the exponent implicitly generates the optimal
multiplicity matrix. This is similar to the line of work \cite{El-Khamy-dimacs05,Ratnakar-it05,Das-isit09}
where various researchers tried to find the multiplicity matrix that
optimizes the error-exponent obtained by either applying a Chernoff
bound \cite{El-Khamy-dimacs05,Ratnakar-it05} or using Sanov's theorem
\cite{Das-isit09}.\vspace{-2mm}
\end{remrk}
\section{COMPUTING THE RDE FUNCTION\label{sec:EXP-func}}
In this section, we first present an extension of Arimoto's numerical
method for computing the RDE function \cite{Arimoto-it76} that works
for any chosen discrete distortion measure. Next, we consider some
special case where we can give an analytical treatment of the function.\vspace{-1mm}
\subsection{Numerical computation of RDE function}
For each discrete source component $x_{i}$, given two parameters
$s\geq0$ and $t\leq0$, the Arimoto algorithm given in \cite{Arimoto-it76}
computes the RDE function numerically as follows.
$\bullet$ Step 1: Choose an arbitrary all-positive distribution vector~$\underline{q}^{(0)}=\left(q_{1}^{(0)},q_{2}^{(0)},\ldots,q_{|\hat{\mathcal{X}|}}^{(0)}\right)$.
$\bullet$ Step 2: Iterate the following steps at $\tau=0,1,\ldots$
\vspace{-4mm}
\begin{eqnarray*}
w_{k|j}^{(\tau)} & = & \frac{q_{k}^{(\tau)}2^{t\delta_{jk}}}{\sum_{k}q_{k}^{(\tau)}2^{t\delta_{jk}}}\\
q_{k}^{(\tau+1)} & = & \frac{\left\{ \sum_{j}p_{j}2^{-st\delta_{jk}}(w_{k|j}^{(\tau)})^{(1+s)}\right\} ^{\frac{1}{1+s}}}{\sum_{k}\left\{ \sum_{j}p_{j}2^{-st\delta_{jk}}(w_{k|j}^{(\tau)})^{(1+s)}\right\} ^{\frac{1}{1+s}}}.\end{eqnarray*}
\vspace{-3mm}~
\hspace{-3.5mm}for $j\in\mathcal{X}$ and $k\in\hat{\mathcal{X}}$.
It is shown by Arimoto that $w_{k|j}^{(\tau)}\rightarrow w_{k|j}^{\star}$
and $q_{k}^{(\tau)}\rightarrow q_{k}^{\star}$ as $\tau\rightarrow\infty$.
Using the resulting $w_{k|j}^{\star}$ and $q_{k}^{\star}$, we can
compute\vspace{-2mm}\begin{eqnarray}
F & = & \sum_{j}\tilde{p}_{j}^{\star}\log\frac{\tilde{p}_{j}^{\star}}{p_{j}}\vspace{-4mm}\label{eq:Ffromq}\\
R & = & \sum_{j}\sum_{k}\tilde{p}_{j}^{\star}w_{k|j}^{\star}\log\frac{w_{k|j}^{\star}}{\sum_{j}\tilde{p}_{j}^{\star}w_{k|j}^{\star}}\vspace{-2mm}\label{eq:Rfromq}\\
D & = & \sum_{j}\sum_{k}\tilde{p}_{j}^{\star}w_{k|j}^{\star}\delta_{jk}\vspace{-2mm}\label{eq:Dfromq}\end{eqnarray}
\vspace{-2mm}~
\hspace{-3.5mm}where $\tilde{p}_{j}^{\star}=\frac{p_{j}(\sum_{k}q_{k}^{\star}2^{t\delta_{jk}})^{-s}}{\sum_{j}p_{j}(\sum_{k}q_{k}^{\star}2^{t\delta_{jk}})^{-s}}$.
\vspace{1mm}
However, in the context we consider, the source (error pattern) $x^{N}$
comprises independent but not necessarily identical source components
$x_{i}$'s. The complexity is a problem if we consider a group of
source letters $(j_{1},j_{2},\ldots,j_{N})$ as a supper-source letter
$\mathcal{J}$, a group of reproduction letters $(k_{1},k_{2},\ldots,k_{N})$
as a super-reproduction letter $\mathcal{K}$ and apply the Arimoto
algorithm straightforwardly . Instead, we can avoid this computational
obstacle by choosing the initial distribution still arbitrarily but
following a factorization rule $q_{\mathcal{K}}^{(0)}=\prod_{i=1}^{N}q_{k_{i}}^{(0)}$.
Then, we can verify that this factorization rule still holds for $w_{\mathcal{K}|\mathcal{J}}^{(\tau)}$
and $q_{\mathcal{K}}^{(\tau)}$ after every step of the Arimoto algorithm.
This leads to\vspace{-1mm} \[
\begin{array}{ccc}
w_{\mathcal{K}|\mathcal{J}}^{\star}=\prod_{i=1}^{N}w_{k_{i}|j_{i}}^{\star} & \mbox{and} & q_{\mathcal{J}}^{\star}=\prod_{i=1}^{N}q_{k_{i}}^{\star}.\end{array}\]
Combining with $\delta_{\mathcal{JK}}=\sum_{i=1}^{N}\delta_{j_{i}k_{i}}$
and $p_{\mathcal{J}}=\prod_{i=1}^{N}p_{j_{i}},$ we have\vspace{-2mm}
\[
\tilde{p}_{\mathcal{J}}^{\star}=\prod_{i=1}^{N}\tilde{p}_{j_{i}}^{\star}.\]
\vspace{-3mm}
This gives the following proposition.
\begin{prop}
(Factored Arimoto algorithm for RDE function) Consider a discrete
source $x^{N}$ of independent but non-identical source components
$x_{i}$'s. Given parameters $s\geq0$ and $t\leq0$, the exponent,
rate and distortion are given by\vspace{-2mm}\[ \left.F\right|_{s,t}=\sum_{i=1}^{N}\left.F_i\right|_{s,t},\,\,\, \left.R\right|_{s,t}=\sum_{i=1}^{N}\left.R_i\right|_{s,t},\,\,\, \left.D\right|_{s,t}=\sum_{i=1}^{N}\left.D_i\right|_{s,t}\]where
the components $\left.F_i\right|_{s,t},\left.R_i\right|_{s,t},\left.D_i\right|_{s,t}$
are computed parametrically by the Arimoto algorithm.\vspace{-1mm}
\end{prop}
\subsection{Analytical computation of RDE function}
In this subsection, we consider the case m-BM1 whose distortion measure
is given in (\ref{eq:dstmb1}). We study the setup that RS codewords
defined over Galois field $\mathbb{F}_{m}$ are transmitted over the
$m$-ary symmetric channel ($m$-SC) which for each parameter $p$
can be modeled as\[
\Pr(r|c)=\begin{cases}
p & \mbox{if}\,\, r=c\\
(1-p)/(m-1) & \mbox{if}\,\, r\neq c.\end{cases}\]
Here, $c$ (resp. $r$) is the transmitted (resp. received) symbol
and $r,c\in\mathbb{F}_{m}$. With this channel model, we consider
$p$ not too small so that $p>(1-p)/(m-1)$. Therefore, at each index
$i$ of the codeword, the hard-decision is also the received symbol
and then it is correct with probability $p$. Thus, we have $p_{i,1}\triangleq\Pr(x_{i}=1)=p$
for every index $i$ of the error pattern $x^{N}$. That means, in
this context we have a source $x^{N}$ with i.i.d. binary components
$x_{i}$. Since the components $x_{i}$ are i.i.d we can treat each
$x_{i}$ as a binary source $X$ with $\Pr(X=1)\triangleq p$ and
$\Pr(X=0)=1-p\triangleq\bar{p}$ and first compute the RDE function
for this source $X$.
According to \cite{Blahut-it74}, for any admissible $(R,D)$ pair
we can find two parameters $s\geq0$ and $t\leq0$ so that $F(R,D)$
can be parametrically evaluated as\begin{eqnarray*}
F(R,D) & = & sR-stD+\max_{q_{1}}\left(-\log f(q_{1})\right)\\
& = & sR-stD-\log\min_{q_{1}}f(q_{1})\end{eqnarray*}
\vspace{-4mm}where\[
f(q_{1})=\bar{p}\left(\sum_{k}q_{k}2^{t\delta_{0k}}\right)^{-s}+p\left(\sum_{k}q_{k}2^{t\delta_{1k}}\right)^{-s}\]
and $R,D$ are given in terms of optimizing $q^{\star}$ which we
will discuss later.
For the distortion measure in (\ref{eq:dstmb1}) and note that $q_{0}=1-q_{1}$,
we have\vspace{-1mm} \[
f(q_{1})=\bar{p}\left((1-q_{1})2^{t}+q_{1}2^{2t}\right)^{-s}+p\left((1-q_{1})2^{t}+q_{1}\right)^{-s}\]
which is a convex function in $q_{1}$. We then see that\vspace{-2mm}\[
\frac{\partial f}{\partial q_{1}}=0\Leftrightarrow q_{1}^{\star}=\frac{1+2^{t}}{1-2^{t}}\left(\frac{1}{1+2^{t}}-\frac{\bar{p}^{\frac{1}{s+1}}}{2^{\frac{st}{s+1}}p^{\frac{1}{s+1}}+\bar{p}^{\frac{1}{s+1}}}\right)\triangleq\beta.\]
In order to minimize $f(q_{1})$ over $q_{1}\in[0,1]$, we consider
three following cases where the optimal $q_{1}^{\star}$ is either
on the boundary or at a point with zero gradient.
$\bullet$ Case 1: $0\leq p\leq\frac{2^{t}}{1+2^{t}}$ then $\beta\leq0$.
Since $f$ convex, it is non-decreasing in the interval $[\beta,\infty)$
and therefore in the interval $[0,1]$. Thus, the optimal $q_{1}^{\star}=0$
and we can also compute from (\ref{eq:Ffromq}), (\ref{eq:Rfromq}),
(\ref{eq:Dfromq}) that\vspace{-2mm}\[
\begin{array}{ccc}
D=1; & R=0; & F=0=D_{KL}(u||p)\end{array}\]
\vspace{-7mm}~
\hspace{-3.5mm}where in this case $u=p$.
$\bullet$ Case 2: $1\geq p\geq\frac{1}{1+2^{t(2s+1)}}$ then $\beta\geq1$.
Since $f$ convex, it is non-increasing in the interval $(-\infty,\beta]$
and therefore in the interval $[0,1]$. Thus, the optimal $q_{1}^{\star}=1$
and similarly we get\vspace{-2mm}\[
\begin{array}{ccc}
D=\frac{2\bar{p}}{p2^{2ts}+\bar{p}}; & R=0; & F=D_{KL}(u||p)\end{array}\]
\begin{comment}
\begin{eqnarray*}
D & = & \frac{2\bar{p}}{p2^{2ts}+\bar{p}};R=0;F\\
R & = & 0\\
F & = & I(u||p)\end{eqnarray*}
\end{comment}
{}where in this case $u=1-\frac{D}{2}$. We can further see that $D\in[2(1-p),1]$
and $u\in[1-D,p]$.
$\bullet$ Case 3: $\frac{2^{t}}{1+2^{t}}<p<\frac{1}{1+2^{t(2s+1)}}$
then $\beta\in(0,1)$. In this case, the optimal $q_{1}^{\star}=\beta$.
We then can find $w_{k|j}^{\star}=\frac{q_{k}^{\star}2^{t\delta_{jk}}}{\sum_{k}q_{k}^{\star}2^{t\delta_{jk}}}$
according to \cite{Blahut-it74} and plug in (\ref{eq:Ffromq}), (\ref{eq:Rfromq}),
(\ref{eq:Dfromq}) to get%
\footnote{The binary entropy function is $H(u)\triangleq-u\log u-(1-u)\log(1-u)$
and the Kullback-Leibler divergence is $D_{KL}(u||p)\triangleq u\log\frac{u}{p}+(1-u)\log\frac{1-u}{1-p}$.%
}\vspace{-2mm}\begin{eqnarray*}
D & = & \frac{2^{t}}{1+2^{t}}+1-u\\
R & = & H(u)-H(u+D-1)\\
F & = & D_{KL}(u||p)\end{eqnarray*}
where $u=\frac{2^{\frac{st}{s+1}}p^{\frac{1}{s+1}}}{2^{\frac{st}{s+1}}p^{\frac{1}{s+1}}+\bar{p}^{\frac{1}{s+1}}}$
. With this notation of $u$, we can express $q_{1}^{\star}=\frac{1-D}{3-2(u+D)}$
and $q_{0}^{\star}=\frac{2(1-u)-D}{3-2(u+D)}.$ We can see that $D\in(1-p,1)$.
It can also be verified that, in this case, by varying $s$ and $t,$
$u$ spans $(1-D,1-D/2)$ and $R$ spans $(0,H(1-D))$.
Based on the above analysis, we obtain the following lemmas and theorems.
\begin{lemma}
\label{lem:Lemmah}Let $h(u)=H(u)-H(u+D-1)$ map $u\in\left[1-D,1-D/2\right)$
to $R$. Then, the inverse mapping of $h$,\vspace{-1mm} \[
h^{-1}:(0,H(1-D)]\to\left[1-D,1-D/2\right),\]
is well-defined and maps $R$ to $u$.\end{lemma}
\begin{proof}
We first notice that $h(u)$ is strictly decreasing since the derivative
is negative over $\left[1-D,1-D/2\right)$, hence the mapping $h:\left[1-D,1-D/2\right)\rightarrow(0,H(1-D)]$
is one-to-one. From the analysis above, one can also see that $h$
is onto.\end{proof}
\begin{thm}
Using mBM-1 with $2^{R}$ decoding attempts where $R\in(0,NH(1-\frac{D}{N})]$,
the maximum rate-distortion exponent that can be achieved is\vspace{-1mm}\begin{equation}
F=N\, D_{KL}\left(h^{-1}\left(R/N\right)||\, p\right).\label{eq:ComputeF}\end{equation}
\vspace{-6mm}\end{thm}
\begin{proof}
First, note that in our context where we have a source sequence $x^{N}$
of $N$ i.i.d. source components, the rate and exponent for each source
component is now $\frac{R}{N}$ and $\frac{F}{N}$. From Case 3 in
the analysis above and from Lemma \ref{lem:Lemmah}, we have\vspace{-1mm}
\[
F/N=D_{KL}(u||p)=D_{KL}\left(h^{-1}\left(R/N\right)||\, p\right)\]
and the theorem follows.\end{proof}
\begin{lemma}
Let $g(u)=D_{KL}(u||p)$ map $u\in[1-D,p]$ to $F$. Then, the inverse
mapping of $g$,\vspace{-1mm} \[
g^{-1}:[0,D_{KL}(1-D\,||\, p)]\rightarrow[1-D,p]\]
is well-defined and maps $F$ to $u$.\end{lemma}
\begin{proof}
We first see that $g(u)$ is a strictly convex function and achieved
minimum value at $u=p$ and therefore $g(u)$ is strictly decreasing
over $[1-D,p]$. Thus, the mapping $g:[1-D,p]\to[0,D_{KL}(1-D\,||\, p)]$
is one-to-one. From the analysis above, one can also see that $g$
is onto.\end{proof}
\begin{thm}
In order to achieve a rate-distortion exponent of $F\in\left[0,N\, D_{KL}\left(1-D\,||\, p\right)\right]$,
the minimum number of decoding attempts required for mBM-1 is $2^{R}$
where\vspace{-1mm}\[
R=N\left[H\left(g^{-1}\left(F/N\right)\right)-H\left(g^{-1}\left(F/N\right)+D/N-1\right)\right]^{+}\]
\end{thm}
\begin{proof}
We also note that the rate, distortion and exponent for each source
component is $\frac{R}{N},\frac{D}{N}$ and $\frac{F}{N}$ respectively.
Combining all the cases in the above analysis, we have \[
R/N=\left[H\left(g^{-1}\left(F/N\right)\right)-H\left(g^{-1}\left(F/N\right)+D/N-1\right)\right]^{+}\]
and the theorem follows.
\end{proof}
\section{SIMULATION\label{sec:Simulation}}
\begin{figure}[!t]
\begin{minipage}[c][1\totalheight][t]{0.485\textwidth}%
\includegraphics[width=0.4\paperwidth]{R11isitbw}\vspace{-3mm}\caption{\label{fig:simulation} Performance of various decoding algorithms
for the (255,239) RS code over an AWGN channel.}
\end{minipage}
\end{figure}
Simulations of the proposed algorithm were conducted for the (255,239)
RS code over an AWGN channel with BPSK as the modulation format. In
Fig. \ref{fig:simulation}, the mBM-2(RD,11) curve belongs to the
algorithm mBM-2 using RD approach proposed in \cite{Nguyen-aller09}
while the mBM-2(RDE,11) one corresponds to the algorithm mBM-2 using
RDE approach proposed in this paper. The label SED(12,12) denotes
the algorithm presented in \cite{Lee-globecom08}. While all these
three algorithms use the same number of $2^{11}$ erasure patterns,
at a FER of $10^{-4}$, the mBM2(RDE,11) provides a performance gain
of roughly 0.4 dB over the SED(12,12) and outperforms the mBM2(RD,11)
by about 0.1 dB. The conventional HDD and the GMD algorithms have
modest performance since they use only one or a few decoding attempts.
Compared to the conventional HDD, the proposed algorithm mBM-2(RDE,11)
gives approximately a 0.9 dB gain. It also outperforms the Koetter-Vardy
(KV) algorithm \cite{Koetter-it03} with infinite multiplicity $(\mu=\infty)$.
The performance of mBM-2(RDE,11) is roughly the same as the performance
of mASD-3(RDE,11). This implies that, for this setup, algorithms based
on multiple trials of BM decoding perform as good as algorithms based
on running the more complicated ASD decoding multiple times. In Fig.
\ref{fig:simqSC}, we simulate the performance mBM-1(RDE,11) for the
same RS code over an $m$-SC channel. One curve reflects the simulated
frame-error rate (FER) and the other is the approximation derived
from $2^{-F}$ where $F$ is given in (\ref{eq:ComputeF}) with $R=11$.
\begin{figure}[t]
\centering{}\includegraphics[scale=0.4]{berR11qSC_new}\caption{\label{fig:simqSC} Performance of mBM-1(RDE,11) and its approximation
$2^{-F}$ where $F$ is given in (\ref{eq:ComputeF}) for the (255,239)
RS code over an $m$-SC($p$) channel.}
\end{figure}
\vspace{-1mm}
\section{CONCLUSION\label{sec:Conclusion}}
A RDE-based algorithm has been proposed for multiple decoding attempts
of RS codes. The RDE analysis shows that this approach has several
advantages. Firstly, the RDE approach achieves a near optimal performance-versus-complexity
trade-off among algorithms that consider running a decoding scheme
multiple times (see Remark \ref{rem:The-RDE-approach}). Secondly,
it can help one estimate the error probability using exponentially
tight bounds for $N$ large enough. Simulations are used to confirm
that algorithms using this approach achieve a better trade-off than
previously known algorithms. Along with this, a numerical method is
given to compute the required RDE function.
Our future work focuses on extending this approach to analyze multiple
decoding attempts for ISI channels. In this case, it makes sense for
the decoder to consider multiple candidate error-events during decoding.
Extending the RD approach directly gives a RD problem for Markov sources
in the large distortion regime. Some work is required though because
this is a well-known open problem.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
One of the most interesting systems currently under study
is that of fermions with
interactions fine-tuned to unitarity~\cite{Leggett}.
This system is interesting due
to the absence of any scale parameter other than the density. The
system is strongly interacting and thus presents a challenge for many-body
methods.
Recently, these systems have been probed with high-frequency or
short-wavelength external probes. By bombarding the fermions with
radio-frequency (RF) radiation, which causes a transition from one
atomic state to another, one can probe the structure of the unitary
Fermi gas. Previous investigations~\cite{Strinati,Schneider,Zwerger}
have shown that the high frequency tail of the RF line shape is related to the
contact parameter, first introduced by Shina
Tan~\cite{Shina1,Shina2}.
In this paper, we address two problems. The first one concerns the
dynamic structure factor $S(\q,\omega)$. This factor, in principle,
can be measured by Bragg scatterings~\cite{Drummond}.
At the Fermi momentum and energy scales, the dynamic structure function
depends on the complicated many-body physics of the unitary Fermi gas.
For much larger $\q$ and $\omega$, the picture becomes simpler.
There is a dominant coherent peak in the
response function at $\omega=\epsilon_\q$.
Far away from this peak, the strength of the
dynamic structure factor should be proportional to Tan's parameter,
the same parameter that characterizes the RF tail. It is easy to
understand why this is true:
in order to absorb an external ``virtual photon''
carrying large energy and momentum and far away from the coherent peak,
the photon has to hit a particle when it is near a second particle,
so that it can give its momentum and energy to both. The absorption
rate, therefore, depends on the probability of finding two particles
at short distance from each other---which is characterized by Tan's
contact parameter.
The problem of the dynamic structure factor is more complicated than that
of the RF response. In particular, care should be taken to not
violate conservation laws. In this paper we use the operator product
expansion (OPE) to facilitate the calculation. The operator product expansion,
introduced by Wilson, is a standard method of quantum field theory.
The use of the OPE for cold atoms was pioneered by Braaten and
Platter~\cite{Braaten:2008uh}. The end result for the dynamic
structure factor is Eq.~(\ref{DSF}).
The second problem considered in this paper is the
calculation of the energy deposition into a unitary
Fermi gas by small, rapid variations of the inverse scattering length.
It can be solved using the same OPE methods.
The usefulness of the operator product expansion can be illustrated as
follows. Suppose we need to compute the following Green's function
\begin{equation}
G_{AB}(\omega,\q) =
\int\! dt\, d\x\, e^{i\omega t - i\q\cdot\x}\, \< A(t,\x) B(0,{\bf 0})\>
\end{equation}
for large $\omega$ and $\q$. Here ``large'' means energy and momentum
much larger than the typical energy and
momentum scales of the state with respect to which the average
$\<\ldots\>$ is taken. For the ground state of a unitary Fermi gas,
these scales are the Fermi energy and Fermi momentum. Let us also
recall that one can associate a local operator $O$ with a scaling
dimension $\Delta_O$. In our counting scheme, the dimension of momentum is
1 and of energy is 2 (the particle mass $m$ is set to 1).
Assuming that the operator product expansion
exists, one can expand the product $A(t,\x)B(0,{\bf 0})$ in terms of
local operators,
\begin{equation}
A(t,\x) B(0,{\bf 0}) = \sum_i |\x|^{\Delta_i-\Delta_A-\Delta_B}
f_i \left( \frac{|\x|^2}t \right) O_i(0).
\end{equation}
Here $f_i$ are functions of one variable $|\x|^2/t$,
and $\Delta_i$ are the dimensions of $O_i$. In contrast to the
OPE in relativistically invariant theories, in nonrelativistic
theories the OPE coefficients are not constant, but are in general
functions of this variable. This identity is to be interpreted as an
operator identity; in particular, we can take the expectation value of
both sides with respect to any state, including thermodynamic states.
Taking the average and performing a Fourier transform, one finds
\begin{equation}
G_{AB}(\omega, \q) = \sum_i
\frac1{\omega^{(5+\Delta_i-\Delta_A-\Delta_B)/2}}
c_i \left( \frac{q^2}\omega\right) \<O_i\>,
\qquad q \equiv |\q|.
\end{equation}
On the right hand side, the higher the dimension of $O_i$, the more
rapidly its contribution decays in the limit of large momentum/energy
(to be precise, the limit considered in this paper will be $\omega\to\infty$, $\q\to\infty$, $q^2/\omega=\textrm{fixed}$.) Thus,
the leading behavior of the Green function is dominated by those few
operators in the OPE with smallest scaling dimension.
The expectation values of the operators $O_i$, are, in general, not
computable theoretically because they depend on many-body physics.
Thus, they should be considered as numbers parameterizing the many-body
state. The OPE coefficients $c_i$, however, depend only on few-body
physics (although, the number of bodies increases with increasing
complexity of the operator $O_i$), and hence can be computed reliably,
at least for simple operators $A$, $B$ and $O_i$. In this way, the
functional dependence of $G_{AB}$ on frequency and wavenumber can be
expressed in terms of a few numbers which have to be determined
experimentally or numerically.
We now discuss the operators of lowest dimension. The unitary fermions are
described by the Lagrangian
\begin{equation}
{\cal L} = i\psi^\+ \d_t \psi - \frac{|\nabla\psi|^2}{2}
+ \psi_2 \psi_1\phi^*
+ \psi_1^\+ \psi_2^\+ \phi
- c_0^{-1} \phi^*\phi.
\end{equation}
We set the fermion mass to one. The indices $1$, $2$ refer to the two
spin polarizations. If one integrates out $\phi$, the saddle point
for this field is at $\phi=c_0\psi_2\psi_1$. This Lagrangian is
therefore equivalent to that with a four-Fermi interaction
$c_0\psi_1^\+\psi_2^\+\psi_2\psi_1$. We will use dimensional
regularization, where setting $c_0^{-1}=0$ corresponds to fine-tuning
the interaction to infinite scattering length. The propagator of the
$\phi$ field is completely determined by its self-energy, and is equal
to
\begin{equation}\label{scalar-prop}
D(\omega,\q) = -\frac{4\pi}{\sqrt{q^2/4-\omega-i0}}\,.
\end{equation}
Let us discuss the operators which may have nonzero expectation value
in a unitary Fermi gas. For our applications, we will need to
consider only operators which do not carry particle numbers.
Moreover, we assume the ground state to be isotropic, so we need to
look only at operators with vanishing orbital angular momentum. The following
two operators have the lowest dimensions: $n_i=\psi_i^\+\psi_i$ ($i=1,2$) and
$\phi^*\phi$. The operators $n_i$ are the particle number densities and
have dimension 3. The operator $\phi^*\phi$ has dimension equal to 4. This
can be seen by computing the dimension of $\phi$ from
Eq.~(\ref{scalar-prop}). An alternative way to find the dimension of
$\phi$ is to use the operator-state correspondence, according to which
the dimension of $\phi$ is the ground state energy of a system of one
spin-up and one spin-down particle in a harmonic trap with unit
oscillator frequency. This ground state energy is 2.
As we shall see, the expectation value $\<\phi^*\phi\>$ can be
identified with Tan's contact parameter. Together, these two
operators dominate the high-momentum behavior of correlation
functions, including the dynamic structure factor.
\section{Single-particle Green's function}
To establish the relationship between $\<\phi^*\phi\>$ and Tan's
contact parameter, let us apply the method of the
OPE to the one-particle Green's function
\begin{equation}
iG(\omega,\q) = \int\!dt\,d\x\, e^{i\omega t- i\q\cdot\x}\,
\< T \psi_1(t,{\bf x}) \psi^\dagger_1(0,{\bf 0}) \>,
\end{equation}
where $T$ denotes time ordering.\footnote{Y.~Nishida~\cite{Nishida}
has essentially arrived at a similar derivation.}
At the end we will be interested only
in the case of $t<0$, where the Green's function is
$-\<\psi^\+(0)\psi(t,\x)\>$,
but the Feynman diagrams are readily available for the time-ordered
Green's function.
According to the previous discussion, we can write
\begin{equation}\label{fermionprop}
G(\omega,\q) = C_n(\omega,\q)\<\psi_2^\+\psi_2\>
+ C_{\phi^*\phi}(\omega,\q) \<\phi^*\phi\> +\cdots
\end{equation}
(We shall see why there is no $\<\psi_1^\+\psi_1\>$ term below.)
To compute $C_n$, we use the Feynman diagram in
Fig.~\ref{fig:Cn_fermionprop}.
\begin{figure}[ht]
\includegraphics[width=0.25\textwidth]{Cn-fermionprop.pdf}
\caption{The diagram that gives $C_n$ in Eq.~(\ref{fermionprop}).
The large momentum and energy flow along lines carrying label ``q.''}
\label{fig:Cn_fermionprop}
\end{figure}
Physically, this diagram describes the interaction of a particle
carrying large momentum and frequency $(\omega,\q)$ with particles already
in the medium. The energy and momentum of the particles in the medium
can be neglected compared to $(\omega,\q)$. The ``hard'' particle
interacts with particles of type 2 in the medium with the result,
\begin{equation}
C_n(\omega,\q) = -\frac{4\pi}{(\omega-q^2/2+i0)^2
\sqrt{q^2/4-\omega-i0}}\,.
\end{equation}
One can see immediately that $C_n(\omega,\q)$ has singularities only
in the lower-half plane, and thus does not contribute to the Green's
function for $t<0$. This can already be seen from the Feynman diagram in
Fig.~\ref{fig:Cn_fermionprop}: one can go directly from the
initial point to the
final point, following the direction of the propagators (fermions and
scalars). Recall that the propagators are all retarded.
Consider now the OPE coefficient of $C_{\phi^*\phi}$. It
can be found by
computing the fermion propagator in the background field of $\phi$.
One diagram that contributes to $C_{\phi^*\phi}$
is as in Fig.~\ref{fig:ferm-prop}.
\begin{figure}[ht]
\includegraphics[width=0.4\textwidth]{ope1b}
\caption{A contribution of $\<\phi^*\phi\>$ to the fermion propagator}
\label{fig:ferm-prop}
\end{figure}
The dashed lines carry zero energy and momentum, and go to the
``condensate'' $\<\phi^*\phi\>$. Fig.~\ref{fig:ferm-prop} is not
the sole diagram contributing to $C_{\phi^*\phi}$;
there is an infinite number of diagrams contributing to this coefficient,
like those depicted in
Fig.~\ref{fig:ferm-prop-others} and obvious subsequent iterations.
\begin{figure}[ht]
\includegraphics[width=0.75\textwidth]{ope1c}
\caption{The other diagrams contributing to the fermion propagator}
\label{fig:ferm-prop-others}
\end{figure}
It can be seen that, in all diagrams of the type of
Fig.~\ref{fig:ferm-prop-others}, one can go from one end of the diagram to the
other by
following the arrows, which is not true in the case of
Fig.~\ref{fig:ferm-prop}. Thus, if our ultimate goal is to compute the
Green's function for $t<0$, it is sufficient to just evaluate the diagram in Fig.~\ref{fig:ferm-prop}.
From the Feynman rules, the contribution of Fig.~\ref{fig:ferm-prop} to
$C_{\phi^*\phi}$ is
\begin{equation}
-\frac 1{(\omega -\epsilon_\q+i0)^2(-\omega-\epsilon_\q+i0)} ,
\end{equation}
where $\epsilon_\q=q^2/2$.
With this information we can find the leading nontrivial contribution
to the Green's function for $t<0$:
\begin{multline}
\int\!{d\x}\, e^{-i\q\cdot\x}\<\psi_1^\dagger(0,{\bf 0})\psi_1 (t,\x)\> =
-i\int\!\frac{d\omega}{2\pi}\, e^{-i\omega t}
C_{\phi^*\phi}(\omega,\q) \<\phi^*\phi\>\\
= \frac1{q^4}\exp \left(i\frac{q^2}2 t\right) \<\phi^*\phi \>
+ \cdots , \qquad t<0.
\end{multline}
In the limit $t\to-0$, we find the tail of the distribution function to be
\begin{equation}
n_\q = \frac{\<\phi^*\phi\>}{q^4},
\end{equation}
which allows us to establish $\<\phi^*\phi\> = C$, where
$C$ is Tan's contact parameter~\cite{Shina1,Shina2}.
One can also show that the diagrams in
Fig.~\ref{fig:ferm-prop-others} do not contribute to the imaginary
part of $G(\omega,\q)$ for $\omega<0$. The latter receives only a
contribution from Fig.~\ref{fig:ferm-prop}, which is
\begin{equation}
\omega < 0: \qquad
\Im G(\omega,\q) = \frac{\<\phi^*\phi\>}{q^4}\pi\delta(\omega+\epsilon_\q).
\end{equation}
The peak at $\omega=-\epsilon_\q$ is
already discussed in Refs.~\cite{Combescot,Randeria-bendback}. At
this level, we are not able to resolve the structure of the peak.
\section{RF spectroscopy}
Consider now a system where, in addition to the `up' and `down' fermions
included in the earlier Lagrangian, a third species of fermion is added
which does not interact with the other two fermions. Suppose we turn on
a photon field which converts atoms of type one into
atoms of type three. The absorption rate is proportional to the imaginary
part of the Green's function of
\begin{equation}
O_{13} = \psi_3^\+\psi_1.
\end{equation}
We thus need to compute the OPE expansion of
\begin{equation}
\Pi(\omega, \q) = -i \int\!dt\,d\x\,
e^{i\omega t - i\q\cdot\x}\, \<T O_{13}(t,\x) O_{13}^\+(0,{\bf 0})\>
= C_n \<n_1\> + C_{\phi^*\phi}\<\phi^*\phi\> + \cdots
\end{equation}
The coefficient $C_n$ is obtained from the diagram of Fig.~\ref{fig:n-O13},
\begin{equation}\label{n-coeff}
C_n = \frac 1{\omega - \epsilon_\q + i0} .
\end{equation}
The external wavy lines in this diagram represent the operator $O_{13}$. The interpretation of this formula is obvious. The dominant part of
the response function is the same as a collection of noninteracting
particles with zero momentum, and has a peak at $\omega=\epsilon_\q$.
\begin{figure}[h]
\includegraphics[width=0.45\textwidth]{n-O13.pdf}
\caption{Diagram contributing to $C_n$}
\label{fig:n-O13}
\end{figure}
Let us now compute the diagram which determines the coefficient of
$\phi^*\phi$ in the OPE (Fig.~\ref{fig:phiphi-O13}).
\begin{figure}[t]
\includegraphics[width=0.50\textwidth]{phiphi-O13.pdf}
\caption{Diagram contributing to $C_{\phi^*\phi}$.
The external dashed lines, which
carry zero energy and momentum,
go to the ``condensate'' $\<\phi^*\phi\>$ which is not depicted for
simplicity.}
\label{fig:phiphi-O13}
\end{figure}
We find
\begin{equation}
\begin{split}
\Pi(\omega,\q) &= \cdots -i \int\! \frac{dp_0\,d\p}{(2\pi)^4}\,
\frac1{(p_0-\epsilon_\p+i0)^2}\,\frac1{p_0 + \omega - \epsilon_{\p+\q}}\,
\frac1{p_0+\epsilon_\p-i0} \< \phi^*\phi\>\\
& \qquad\qquad\qquad + (\omega\to-\omega,\q\to-\q)
+ \cdots\\
&= \cdots+\int\!\frac{d\p}{(2\pi)^3}\, \frac1{4\epsilon_\p^2}\,
\frac1{\omega-\epsilon_p-\epsilon_{\p+\q}+i0} \<\phi^*\phi\>
+ (\omega\to-\omega,\q\to-\q)
+\cdots
\end{split}
\end{equation}
This integral is divergent in the infrared. We can isolate the divergent
part of the integral as
\begin{equation}
\frac{\<\phi^*\phi\>}{\omega - \epsilon_\q+i0}
\int\!\frac{d\p}{(2\pi)^3}\, \frac1{4\epsilon_\p^2}\,
+ \int\!\frac{d\p}{(2\pi)^3}\, \frac1{4\epsilon_\p^2}\left(
\frac1{\omega-\epsilon_\p-\epsilon_{\p+\q}+i0} -
\frac1{\omega-\epsilon_\q+i0} \right) \<\phi^*\phi\>.
\end{equation}
It is clear that the first (divergent) part is actually included in the
first term in the OPE (\ref{n-coeff}), i.e., it represents the contribution of the
condensate $\<\phi^*\phi\>$ to the number density $n$. Subtracting
this contribution, the second part gives $C_{\phi^*\phi}$,
\begin{equation}
C_{\phi^*\phi} = \int\!\frac{d\p}{(2\pi)^3}\, \frac1{4\epsilon_p^2}\left(
\frac1{\omega-\epsilon_p-\epsilon_{\p+\q}+i0} -
\frac1{\omega-\epsilon_\q+i0} \right)
+ (\omega\to-\omega, \q\to-\q).
\end{equation}
We now restrict ourselves to the imaginary part at $\q=0$. We have
\begin{equation}
I(\omega) = -\frac1\pi \Im \Pi(\omega, {\bf 0}) =
\frac{\<\phi^*\phi\>}{4\pi^2} \frac1{\omega^{3/2}}\,,
\end{equation}
which coincides with the result obtained previously by other
authors~\cite{Schneider,Zwerger}.
\section{Dynamic structure factor}
Having tested the OPE method on the simple examples above, we will now
discuss the dynamic structure factor. The dynamic structure factor can be
defined as the imaginary part of the Green function of the density,
\begin{equation}
G_{nn} (\omega,\q) = -i \int\!dt\,d\q\, e^{i\omega t- i\k\cdot\x}\,
\< T n(t,\x) n(0,{\bf 0})\> = C_n \<n\> + C_{\phi^*\phi} \< \phi^*\phi\>
+ \cdots
\end{equation}
Namely,
\begin{equation}
S(\q,\omega) = -\frac1\pi \Im G_{nn}(\omega,\q), \qquad
\omega > 0.
\end{equation}
The coefficient $C_n$ is
\begin{equation}
C_n = G(\omega,\q) + G(-\omega,-\q) =
\frac{2\epsilon_\q}{\omega^2-\epsilon_\q^2+i0}\,,
\end{equation}
and corresponds to the one-particle peak $\sim\delta(\omega-\epsilon_\q)$
in the dynamic structure factor. This peak dominates all sum rules. However,
we are interested in the structure factor far away from this peak, hence
we need to compute the OPE coefficient of the next operator, $\phi^*\phi$.
The diagrams contributing to $C_{\phi^*\phi}$ are sketched
schematically in Fig.~\ref{fig:phiphi-nn}.
\begin{figure}[ht]
\includegraphics[width=0.75\textwidth]{phiphi-nn.pdf}
\caption{Schematic diagrams contributing to $C_{\phi\phi}$.
Each diagram represents a number of Feynman graphs where the photon lines
are attached to different fermion lines}
\label{fig:phiphi-nn}
\end{figure}
The contribution from the first diagram is exactly twice the diagram
in the RF case,
\begin{equation}
C^{(1)}_{\phi^*\phi} = 2 \int\!\frac{d\p}{(2\pi)^3}\,
\frac1{4\epsilon_\p^2}\left(
\frac1{\omega-\epsilon_\p-\epsilon_{\p+\q}+i0} -
\frac1{\omega-\epsilon_\q+i0} \right)
+ (\omega\to-\omega,\q\to-\q).
\end{equation}
The second diagram gives
\begin{equation}
C^{(2)}_{\phi^*\phi} = 2 \int\!\frac{d\p}{(2\pi)^3}\,
\frac1{4\epsilon_\p\epsilon_{\p+\q}}\left(
\frac1{\omega-\epsilon_\p-\epsilon_{\p+\q}+i0}-
\frac1{\omega+\epsilon_\p+\epsilon_{\p+\q}-i0}\right).
\end{equation}
The third diagram is the product of three pieces, each of which can
be computed separately. At the end one finds
\begin{multline}
C^{(3)}_{\phi^*\phi} = \left[ -2 \int\!\frac{d\p}{(2\pi)^3}\,
\frac1{2\epsilon_\p} \frac1{\omega-\epsilon_\p - \epsilon_{\p+\q}+i0}
\right]^2 \frac{-4\pi}{\sqrt{\frac12\epsilon_\q-\omega-i0}}
+ (\omega\to-\omega, \q\to-\q).
\end{multline}
Evaluating the imaginary part of the correlation function
for positive $\omega$, we find
that it is zero when $\omega<q^2/4$. This should be the case since
$\omega=q^2/4$ is the threshold for knocking out two particles from the
medium (recall that all the energy scales associated with the medium,
like the chemical potential, are negligible).
For $\omega>q^2/4$, the dynamic structure factor is
\begin{multline}\label{DSF}
\frac{S(\q,\omega)}{\<\phi^*\phi\>}
= -\frac1\pi\Im C_{\phi^*\phi} = \frac1{2\pi^2}
\frac{\sqrt{\omega-q^2/4}}{(\omega-q^2/2)^2}
+ \frac1{2\pi^2\omega q} \ln \frac{\omega+q\sqrt{\omega-q^2/4}}
{|\omega-q^2/2|}\\
- \frac1{\pi^2 q^2\sqrt{\omega-q^2/4}} \left(
\ln^2 \frac{\omega+q\sqrt{\omega-q^2/4}}{|\omega-q^2/2|}
-\pi^2\theta(q^2/2-\omega)\right).
\end{multline}
In Fig.~\ref{fig:fixom} we plot the function
$\omega^{3/2}S(\omega,\q)/\<\phi^*\phi\>$ as a
function of the ``Bjorken $x$'' variable, $x=q^2/2\omega$. This
can be thought of as a plot of $S$ at fixed $\omega$ as a function of $q^2$.
\begin{figure}[ht]
\includegraphics[width=0.5\textwidth]{fixom.pdf}
\caption{A plot of $\omega^{3/2}S(\q,\omega)/\<\phi^*\phi\>$ as a
function of $x=q^2/2\omega$.}
\label{fig:fixom}
\end{figure}
In Fig.~\ref{fig:fixk},
we plot the same quantity as a function
of $x^{-1}=2\omega/q^2$.
This plot is basically a plot of $S$ as a function of $\omega$ at fixed $q$.
\begin{figure}[ht]
\includegraphics[width=0.5\textwidth]{fixk.pdf}
\caption{A plot of $q^{3/2}S(\q,\omega)/\<\phi^*\phi\>$ as a
function of $x^{-1}=2\omega/q^2$.}
\label{fig:fixk}
\end{figure}
As we see from the plots, the dynamic structure factor is positive for
all $\omega$ and $\q$, as it should be. It vanishes in the zero
momentum, finite frequency limit $\q=0$, as it should due to the
conservation of particle number. At small $q^2/\omega$, the behavior
of the dynamic structure factor is
\begin{equation}\label{S-smallq}
S(\q,\omega) = \frac4{45\pi^2} \frac{q^4}{\omega^{7/2}}\<\phi^*\phi\>.
\end{equation}
The dynamic structure function has a singularity near $\omega=q^2/2$,
which is the single-particle peak. It would be incorrect to take
the limit $x= q^2/2\omega\to1$ in our formula at fixed $\q$ or fixed
$\omega$. Our result is strictly valid in
the regime of $\omega\to\infty$, $\q\to\infty$, fixed $x=q^2/2\omega\neq1$.
The weaker singularity at $\omega=q^2/4$ ($x=2$) is associated with the fact
that near this point, the two particles that are kicked out have small
relative momentum, and thus interact strongly with each other.
\section{Response of unitary gas to rapid changes of the inverse
scattering length}
We consider in this section the following problem. Assume we make the
inverse scattering length $\alpha\equiv a^{-1}$ exhibit periodic
behavior in time,
\begin{equation}
\alpha(t) = a^{-1}(t) = \alpha_0 \cos\omega t,
\end{equation}
and assume the amplitude $\alpha_0$ to be small, and the frequency
of the oscillations to be large compared to the Fermi energy,
$\omega\gg\epsilon_{\rm F}$. The
question is: at what rate is energy deposited into the system?
We know that the inverse scattering length is coupled to the operator
$\phi^*\phi$, i.e., introducing a finite inverse scattering length
corresponds to adding into the Lagrangian a term
\begin{equation}
{\cal L} \to {\cal L} + \frac{\alpha}{4\pi} \phi^*\phi.
\end{equation}
The energy deposited into the system, in unit volume and in unit time,
can be computed from the formula
\begin{equation}
\frac{d\epsilon}{dt} = \frac{\alpha^2_0}{2(4\pi)^2} \omega \,
|\Im G_{OO}(\omega,{\bf 0})|,
\end{equation}
where $G_{OO}$ is the Greens function of the the operator
$O=\phi^*\phi$. The OPE of two $O$ operators is easy to compute,
\begin{equation}
-i \<T\, \phi^*\phi(t,\x)\, \phi^*\phi(0)\> =
\<\phi^*\phi\> [D(t,\x) + D(-t,-\x)] + \cdots,
\end{equation}
where $D(t,\x)$ is the scalar propagator [Eq.~(\ref{scalar-prop})].
From this we find $\Im G_{OO}(\omega,{\bf 0}) =
-4\pi\omega^{-1/2}\<\phi^*\phi\>$. Thus we find the rate of energy
deposition,
\begin{equation}
\frac{d\epsilon}{dt} = \frac{\alpha_0^2}{8\pi}\<\phi^*\phi\> \sqrt\omega\,.
\end{equation}
If $\alpha(t)$ has the form of a sudden pulse, i.e., is nonzero only
in a finite time interval, the total energy deposited into the system,
per unit volume, will be related to the Fourier transform
$\tilde\alpha(\omega)=\int\!dt\,e^{i\omega t}\alpha(t)$ by
\begin{equation}
\epsilon = \frac{\<\phi^*\phi\>}{2\pi}\int_0^\infty\!\frac{d\omega}{2\pi}\,
\sqrt\omega\, |\tilde\alpha(\omega)|^2.
\end{equation}
\section{Conclusion}
In this paper we have shown that the structure of a unitary Fermi gas,
at short distance and length scales, can be obtained from the operator
product expansion. We computed the OPE of the density operator at two
points and derived an expression for the dynamic structure factor.
The same technique is applied to the problem of finding the energy
deposited into a unitary Fermi gas by a rapid, small oscillation of the
scattering length.
As mentioned above, the two-body contribution to $S(\q,\omega)$ vanishes
for $\omega<q^2/4$: this is the two-body threshold.
If $\omega>q^2/6$, it is possible to transfer the energy and
momentum to three particles instead. In general, when $n-1<x<n$, where $x$ is
the ``Bjorken $x$'' variable,
the dominant
contribution to the dynamic structure factor $S(\omega,\q)$ in the large
$\omega$, fixed $x$ regime is due to the lowest dimensional operator in
the OPE that contains $n$ creation and $n$ annihilation operators.
The dimension of this operator is $2\Delta_n$, where $\Delta_n$ is the
ground state energy of a system of $n$ particles in a harmonic
potential. Thus we find
\begin{equation}
S(\q,\omega) \sim \<O_n^\+ O_n\> \frac{f_n(x)}{\omega^{\Delta_n-1/2}}\,,
\end{equation}
in the limit $\omega\to\infty$, $x=\textrm{fixed}$, and $n-1<x<n$ ($n\ge3$).
For example, the $n=3$
body contribution to the dynamic structure
factor is down by $\omega^{-3.77272}$
since $\Delta_3\approx4.27272$. The larger $n$ is, the faster the
contribution of $n$-body physics decreases with increasing momentum.
The operator product expansion cleanly separates ``hard''
(large-momentum and/or large-energy) physics from soft physics. The
computation of the OPE coefficients involves computing diagrams in
vacuum, with a finite number of particles in the intermediate state.
Thus the OPE is a natural way of separating few-body from
many-body physics. It application to cold atom physics should be
further investigated.
\acknowledgments
The authors thank G.~Baym, E.~Braaten, A.~L.~Fitzpatrick, E.~Katz,
Y.~Nishida, L.~Platter, S.~Shenker, and S.~Tan for discussions. This
work is supported, in part, by DOE grant No.\ DE-FG02-00ER41132
and by University of Washington Royalty Research Fund grant No.\ 65-8195.
After this work was finished, we learned about
Ref.~\cite{Braaten:2010dv} which has some overlap with our calculation
of the RF response. We also learned from M.~Randeria and E.~Taylor that
they considered
$S(\q,\omega)$ in the large-$\omega$, small-$q$ regime
and obtained the same $\<\phi^*\phi\>q^4/\omega^{7/2}$ behavior as
in Eq.~(\ref{S-smallq}), but with
a different numerical coefficient~\cite{Randeria}.
|
\chapter{Implications of noncanonical neutrinos on cosmology}
Violating Lorentz invariance in QFT through a deformation of the anticommutation rules of quantum fields yields the possibility of the neutrinos having a IR modified dispersion relation such that mimics the role of the cosmological constant in the present cosmic acceleration. This possibility is further studied.
\section{Motivation}
Our last work on LIV in neutrino physics caused by a deformation of the anticommutation rules of the fermion fields opens the possibility of the neutrinos having a dispersion relation very different from the relativistic one. Experimental data from high energy physics have shown that the neutrino dispersion relation differs from that of a relativistic massless particles at low energies. The scale of energy of that deviation is $\lambda < 1 eV$ and the differences between the dispersion relation of the different neutrino species can be parameterized up to first order corrections as
\begin{equation}
E_i(p) = p + \frac{m_i^2}{2 p} + \mathcal{O}(\frac{\lambda^3}{p^2})\, ,
\end{equation}
where we suppose that $m_i \sim \lambda$, with $\vert\Delta m^2_{12}\vert = 7.9^{+0.27}_{-0.28} \times 10^{-5} (eV)^2$ and $\vert\Delta m^2_{13}\vert = 2.6 \pm 0.2 \times 10^{-3} (eV)^2$ as shown by neutrino oscillation experiments.
However, a modification in the dispersion relation of the neutrinos should have a visible impact on cosmology, due to the effective temperature of the cosmic neutrino background being at least one or two orders of magnitude below the scale $\lambda$ ($T_\nu = 1.95\, K = 1.68 \times 10^{-4}\, eV$). Moreover cosmological observations seem to imply the existence of a cosmological constant with energy density
\begin{equation}
\rho_\Lambda = \rho_c \Omega_\Lambda = \frac {3 H_0 \Omega_\Lambda}{8 \pi G} = \frac {3 \Lambda}{8 \pi G} = 3.0 \times 10^{-11}\, (eV)^4 \, .
\end{equation}
The similarity between the energy scale of the cosmological constant and the energy scale of neutrino masses seem to suggest that both phenomena are connected. Might neutrinos be the Dark Energy?
Let us check the order of magnitude of a cosmological constant-like contribution to the energy density of the universe. Neutrinos would mimic a cosmological constant if their dispersion relation satisfies $E(p_d \frac{a_d}a )n_\nu = const$, where $n_\nu$ is the number density of neutrinos ($n_{\nu} \propto a^{-3}$, $n_{\nu, now} \sim n_{\gamma, now} \sim 10^{-12}\, (eV)^3 $), $p_d$ is the momentum of a neutrino at neutrino decoupling ($T_d \sim 1.5\, MeV$) and $a_d$ is the scale factor at neutrino decoupling. Therefore, the contribution of neutrinos to the energy density will be constant if they have a dispersion relation $E(p) \sim \frac {\lambda^4}{p^3}$ for $p \ll \lambda$. If we assume that the average momentum of the neutrinos at neutrino decoupling is similar to the decoupling temperature and that the number of photons and neutrinos is more or less conserved during the expansion of the universe after they decouple, we get, for $\lambda^2\sim\Delta m_{12}^2$:
\begin{equation}
\Omega_{\nu, now} \approx 0.68 \Omega_\gamma \frac {E_\nu (T_\nu)}{T_\nu} \sim 0.68\Omega_\gamma \frac {\lambda^4}{T^4_\nu} \sim 0.68 \times 10^{-5} \frac {(10^{-5}\, (eV)^2)^2}{(10^{-4}\, eV)^4} \sim 1\, .
\end{equation}
It is known that $\Omega_\Lambda = 0.73$ in $\Lambda$CDM and the contribution of this neutrinos seems to be of the same order, so maybe we can explain what is the Dark Energy by considering some specific neutrino dispersion relation.
We will now see how a polynomial dispersion relation in the ratio $\lambda/p$ is able to reproduce the effect of a cosmological constant. However, a detailed computation will show that the resulting effective cosmological constant is too high.
\section{A polynomial dispersion relation}
We will consider a universe composed by the following:
\begin{itemize}
\item \textbf{Photons}: Their contribution to the total energy density today is $\Omega_\gamma= 4.92\times10^{-5}$. They form a Plankian spectrum with temperature $T_\gamma= 2.73\, K= 2.35\times10^{-4}\, eV$, which implies an energy denisity of $\rho_\gamma = 4.64 \times 10^{-34}\, g/cm^3 = 2.01\times 10^{-15} \, (eV)^4$.
\item \textbf{Baryons}: By baryons it is meant ordinary mater: nuclei and electrons (the contribution of electrons to the energy density is negligible compared to that of nulei). $\Omega_b = 0.0415 \pm 0.0016$.
\item \textbf{Dark Matter}: It doesn't interact with photons, it is neither baryonic. $\Omega_{DM} = 0.236^{+0.016}_{-0.024}$.
\item \textbf{Primordial Neutrinos}: Consider 3 families of neutrinos whose abundance would be $\rho_{\nu_i} \sim \frac {0.68}3 \rho_\gamma$ if they were massless relativistic particles. They form a Plankian distribution with temperature $T_\nu = 1.95 \, K = 1.68 \times 10^{-4}\, eV$. In this section we will consider that their energy dispersion relation is
\begin{equation}
E_{\nu_i}(p)=p\left( 1+\frac{c_i}{p^2}\right) ^2\, .
\label{disp1}
\end{equation}
We can obtain more information about the coeficient $c_i$ from neutrino oscillation experiments: $ \vert c_2-c_1\vert\equiv\frac{\Delta m^2_{12}}4 = (1.98 \pm 0.07) \times 10^{-5} (eV)^2$, $ \vert c_3-c_1\vert\equiv\frac{\Delta m^2_{13}}4 = (6.5 \pm 0.5)\times10^{-4} (eV)^2$.
\end{itemize}
\section{Equal momenta approximation}
As a first approximation we will consider that today the momentum of every neutrino is equal to the temperature $T_\nu$. Taking into account the dispersion relation (\ref{disp1}), the contribution from each of the neutrino species to the total energy density will be:
\begin{equation}
\Omega_{\nu_i} = \frac{\rho_{\nu_i}}{\rho_c} \sim \frac{0.68}3\Omega_\gamma \left( 1+\frac{c_i}{T_\nu^2} \right)^2 \sim
1.12 \times10^{-5} +793 \frac{c_i}{(1 eV)^2}+1.40\times10^{10} \frac{c_i^2}{(1 eV)^4}\, .
\label{approx1}
\end{equation}
The first term equals the contribution of massless relativistic neutrinos to the energy density today. The second term plays the role of an effective spatial curvature contribution to the energy density $\equiv -\frac k{(aH_0)^2}$. The coefficients $c_i$ must be chosen to make this contribution compatible with the upper bounds on a spatial curvature term from WMAP. The last term mimics the contribution of a cosmological constant to the total energy density.
It is easy to see that the condition for $\vert c_2 -c_1 \vert$ is easy to satisfy consistently with all the experimental data. However, the condition for $\vert c_3-c_1\vert$ forces at least one of the $c_i$ to be of order $10^{-4} (eV)^2$, and this causes the last term in (\ref{approx1}) to be too large compared to the experimental data for $\Omega_\Lambda$.
A particular example is
$$
\begin{array}{l}
c_1=-2.23\times 10^{-4} (eV)^2 \, ,\\
c_2=-2.03\times 10^{-4} (eV)^2 \, ,\\
c_3=+4.27\times 10^{-4} (eV)^2 \, ,\\
\Rightarrow \Omega_\Lambda \sim 3800 \gg 0.73\, .
\end{array}
$$
\section{Taking into account the Planckian spectrum}
It might be possible that this disease is cured by taking into account that the momenta of the neutrinos are distributed forming a Planckian spectrum. We must have into account that there are two degrees of freedom ($\nu$ and $\bar{\nu}$) per neutrino species. The distribution in thermal equilibrium at temperature T is then
\begin{equation}
\frac{d n_i (p)}{d p} = \frac 1{\pi^2} \frac {p^2}{e^{(E_i(p)-\mu_i)/T}+1}\, .
\end{equation}
The neutrinos were in thermal equilibrium before they decoupled. Then they keep their distribution with $p=p_d \frac {a_d}a$. The chemical potential can be neglected since
\begin{equation}
\frac {\mu_i} T = \frac {n_{\nu_i}-n_{\bar{\nu_i}}}{n_\gamma}\equiv L_i<10^{-9}\, .
\end{equation}
At energies close to decoupling ($T_d \sim 1.5 \, MeV$) we can use the approximation $E_d \approx p_d$. Therefore, at neutrino decoupling we have:
\begin{equation}
n_i \approx \frac 1{\pi^2} \int dp p^2 (e^{p/T} +1)^{-1} = \frac {3 T^3 \zeta(3)}{2 \pi^2} \approx 0.1827 \times T^3\, ,
\end{equation}
\begin{equation}
\left\langle p\right\rangle_i \approx \frac 1{n_i}\frac 1{\pi^2} \int dp p^3 (e^{p/T} +1)^{-1} = \frac {7 \pi^2 T^4}{120 n_i} \approx 3.1514 \times T\, ,
\end{equation}
\begin{equation}
\left\langle \frac 1p\right\rangle_i \approx \frac 1{n_i}\frac 1{\pi^2} \int dp p (e^{p/T} +1)^{-1} = \frac {T^2}{12 n_i} \approx \frac{0.4561}T\, .
\end{equation}
$\left\langle p^{-3} \right\rangle_i$ diverges due to the contribution of small momenta to the integral, unless we take into account that $E(p) = p\left( 1+\frac {c_i}{p^2}\right) ^2$. This integral cannot be solved analytically, so we will solve it numerically using the values for the $c_i$ in the previous example:
\begin{equation}
\begin{array}{l}
\left\langle p^{-3} \right\rangle _i = \frac 1{n_i}\frac 1{\pi^2} \int \frac {d p}p (e^{p\left( 1+\frac {c_i}{p^2}\right) ^2 /T_d} +1)^{-1}\\
\approx \left\lbrace
\begin{array}{r}
\frac {1.23609}{n_i} \approx 6.76 T_d^{-3} \;\text{for} \; c_i = -2.23 \times 10^{-4}\, (eV)^2\, ,\\
\frac {1.23926}{n_i} \approx 6.78 T_d^{-3} \;\text{for} \; c_i = -2.03 \times 10^{-4}\, (eV)^2\, ,\\
\frac {1.21415}{n_i} \approx 6.64 T_d^{-3} \;\text{for} \; c_i = 4.27 \times 10^{-4}\, (eV)^2\, .
\end{array}
\right.
\end{array}
\end{equation}
The coefficients coming out of this strict derivation are of order one, which means the dispersion relation (\ref{disp1}) gives a too large constant contribution to the total energy density. In fact,
\begin{equation}
\Omega_\nu\equiv\sum_{i}\Omega_{\nu_i}^{(\gamma)}+\Omega_{\nu_i}^{(k)}+\Omega_{\nu_i}^{(\Lambda)}\, ,
\end{equation}
where $\sum_{i}\Omega_{\nu_i}^{(\gamma)}=3.37 \times 10^{-5}$, $\sum_{i}\Omega_{\nu_i}^{(k)}<0.02$, $\sum_{i}\Omega_{\nu_i}^{(\Lambda)}\sim 8100$. This means that the simplest energy-momentum dispersion relation which we have thought is ruled out by the cosmological data.
\section{Discussion}
Although it is possible to find {\it ad hoc} dispersion relations for the neutrinos that mimic the effect of the Dark Energy and are consistent with the cosmological data, the simplest examples are ruled out. This indicates that maybe our assumption that neutrinos were driving the present accelerated expansion was too ambitious.
Thus cosmology imposes great constraints in the low energy behavior of the dispersion relation of neutrinos. Any nontrivial behavior should be non-polynomial in $\lambda/p$ in order not to be ruled out by cosmological observations. Otherwise it would enter in conflict with the cosmological bounds on Hot Dark Matter (HDM)\footnote{If the neutrino dispersion relation mimics that of a massive particle in the IR, then one should obviously choose the simplest explanation and work with massive neutrinos.}, spatial curvature, or a cosmological constant, and we have proven that this bounds are tight.
The effect of neutrinos with a modified dispersion relation in the IR will be further studied in future work. In particular, the effect of a modified dispersion relation in the distance to neutrino decoupling surface or the matter power spectrum will be addressed elsewhere.
\chapter{Conventions and Abbreviations}
Throughout this thesis and unless it is otherwise stated, we will use units in which $\hbar = c = 1$. Moreover we will use the convention in which the signature of the metric is $+\, -\, -\, -$.
\bigskip
We will use the following abbreviations:
\bigskip
\textbf{ACM :} Asymptotic Cosmological Model \bigskip
\textbf{ADM :} Arnowitt-Deser-Misner formalism \bigskip
\textbf{BAO :} Baryon Acoustic Oscillations \bigskip
\textbf{$\bm{\beta\beta 0\nu}$ :} Neutrinoless Double Beta Decay \bigskip
\textbf{CMB :} Cosmic Microwave Background \bigskip
\textbf{C$\bm \nu$B :} Cosmic Neutrino Background \bigskip
\textbf{CSM :} Cosmological Standard Model \bigskip
\textbf{DE :} Dark Energy \bigskip
\textbf{DM :} Dark matter \bigskip
\textbf{DSR :} Doubly (or Deformed) Special Relativity \bigskip
\textbf{EFT :} Effective Field Theory \bigskip
\textbf{FRW :} Friedman-Robertson-Walker \bigskip
\textbf{GR :} General Theory of Relativity \bigskip
\textbf{GUT :} Grand Unified Theories \bigskip
\textbf{HDM :} Hot Dark Matter, for instance massive neutrinos \bigskip
\textbf{HST :} Hubble Space Telescope \bigskip
\textbf{ISW :} Integrated Sachs-Wolfe effect \bigskip
\textbf{IR :} Infrared \bigskip
\textbf{$\bm \Lambda$CDM :} the Concordance Model, in which the universe is dominated today by a mixture of Cold Dark Matter (CDM) and a Cosmological Constant ($\Lambda$) \bigskip
\textbf{LIV :} Lorentz Invariance Violation \bigskip
\textbf{LQG :} Loop Quantum Gravity \bigskip
\textbf{LSND :} Liquid Scintillator Neutrino Detector \bigskip
\textbf{LSS :} Large Scale Structure \bigskip
\textbf{MSW :} Mikheev-Smirnov-Wolfenstein \bigskip
\textbf{NCQM :} Noncommutative Quantum Mechanics \bigskip
\textbf{PMNS :} Pontecorvo-Maki-Nakagawa-Sakata \bigskip
\textbf{QED :} Quantum Electrodynamics \bigskip
\textbf{QFT :} Quantum Field Theory \bigskip
\textbf{QNCFT :} Quantum Theory of Noncanonical Fields \bigskip
\textbf{SM :} Standard Model of particle physics \bigskip
\textbf{SME :} Lorentz Violating Standard Model Extension \bigskip
\textbf{SNe :} Type Ia Supernovae \bigskip
\textbf{SNO :} Sudbury Neutrino Observatory \bigskip
\textbf{SR :} Special Relativity \bigskip
\textbf{SSB :} Spontaneous Symmetry Breaking \bigskip
\textbf{ST :} String Theory (or Theories) \bigskip
\textbf{UV :} Ultraviolet \bigskip
\textbf{vev :} Vacuum expectation value \bigskip
\textbf{VSL :} Varying Speed of Light \bigskip
\textbf{WMAP :} Wilkinson Microwave Anisotropy Probe \bigskip
\textbf{WMAP5 :} WMAP five year dataset \bigskip
\chapter{Introduction}
By the beginning of the XXI century, the General Theory of Relativity (GR) and the Standard Model of particle physics (SM) seem to offer a very accurate description of all known physical phenomena... All of them? Not really, yet neutrino oscillations and cosmology still resist a full understanding.
\section{The Accelerating Universe}
Since the discovery of GR by Albert Einstein in 1915, there has been great interest in applying its equations to determine the behavior of the universe as a whole. Einstein was the first to open the Pandora's box in 1917, presenting a model of a static universe which required the presence of a `cosmological constant' in the field equations. Friedman found in 1922 that the equations admitted expanding universe solutions without the presence of this term. The expansion of the universe was also studied by Lema\^itre, Robertson and Walker, and confirmed seven years later by Hubble. Thus Einstein rejected the cosmological constant, calling it his `greatest blunder'.
Consider the Einstein's equations of GR
\begin{equation}
R_{\mu\nu} -\frac 12 g_{\mu\nu}R = 8\pi G T_{\mu\nu}\, ,
\end{equation}
where $R_{\mu\nu}$ is the Ricci curvature tensor, $g_{\mu\nu}$ is the metric of spacetime, $R = R_{\mu\nu}g^{\mu\nu}$ is the curvature scalar and $T_{\mu\nu}$ is the stress-energy tensor of the fluid filling spacetime. An isotropic and homogeneous universe is described by the Friedman-Robertson-Walker (FRW) metric
\begin{equation}
ds^2 = dt^2 - a(t)^2 \left[ \frac{dr^2}{1-kr^2} + r^2 d\Omega_3^2 \right] \, ,
\end{equation}
where $\{ t, r, \theta, \phi \}$ are the comoving coordinates, $a(t)$ is the scale factor, $k$ is the spatial curvature, and $\Omega_3$ is the solid angle of three-dimensional space $d\Omega_3^2 = d\theta^2 + sin^2 \theta d\phi^2$. The present value of the scale factor $a_0 = a(t_0)$ is usually normalized to one.
According to GR, this metric evolves following the relation (known as Friedman Equation)
\begin{equation}
\frac{8\pi G}{3}\rho = H^2 + \frac{k}{a^2}\, ,
\end{equation}
where $\rho$ is the total energy density and $H$ is the Hubble parameter, the rate of expansion of the universe ($H = \frac{da}{adt}$). Moreover the energy density of massive (matter) or massless (radiation) particles decreases with the scale factor as $\rho_m \sim a^{-3}$ or $\rho_r \sim a^{-4}$, respectively. This scenario, together with the conditions of an initially radiation-dominated universe which undergoes an adiabatic expansion, makes up the `Cosmological Standard Model' (CSM, as named by Guth \cite{Guth:1980zm}, by analogy with the Standard Model of particle physics). The solutions of the resulting differential equation for the scale factor are always decelerating ($\ddot a < 0$).
The CSM has achieved a number of successes, including the correct prediction of the relative densities of nuclei heavier than hydrogen (Nucleosynthesis) or the prediction of the Cosmic Microwave Background (CMB), the most perfect black body radiation ever measured. Since the first observations by Hubble it has however been very difficult to directly measure the precise effect of the cosmic expansion on the motion of galaxies.
This has been connected to the difficulty of finding 'standard candles', astronomical objects with known luminosity $L$ (energy released per unit time). Given the luminosity of an object and its spectrum, it is possible to directly measure its redshift ($z = \frac{a_0}a -1$) and bolometric flux $F$ (the energy detected per unit of area and time), where the relation between luminosity and flux is encoded in the luminosity distance $d_l$ to the astronomical object, defined as $d_l\equiv \left(\frac{L}{4\pi F } \right)^{1/2}$. The luminosity distance vs. redshift (which is in one to one correspondence with time) relation enables us to determine the evolution of the scale factor and the Hubble parameter with time\footnote{This relation is computed in the framework of metric spacetimes with a FRW metric. For objects of sufficient size, it is possible to compare this quantity to their angular size distance $d_A\equiv l/\Delta\theta$, where $l$ is the transverse size of the object and $\Delta\theta$ is its angular size (the angle that it subtends in the sky). In a metric spacetime, $d_l = d_A (1+z)^2$. This relation can be used to test the validity of metric theories as a good description of spacetime \cite{Bassett:2003vu,Sahni:2006pa}. However experimental data seem to support this relation \cite{DeBernardis:2006ii}. Thus we will not leave the framework of generally covariant metric spacetimes. },
\begin{equation}
d_l (z) = c(1+z) \int^z_0 \frac {dz'}{H(z')} \, ,
\end{equation}
and therefore gives us information about the relative importance of radiation, matter or spatial curvature in the evolution of the universe. As we told above, for any universe driven by a combination of those (the massless and massive particles contained in the Standard Model of particle physics (SM) or not, and the spatial curvature allowed in the FRW metric) will undergo a decelerated expansion, as gravity acts as an attractive force.
However, measurements of distant Type Ia Supernovae (SNe) (widely accepted standard candles) distance vs. redshift relation reveal that the universe is presently undergoing an accelerated expansion \cite{Riess:1998cb,Perlmutter:1998np}, in contradiction with the decelerated expansion predicted by GR!
\subsection{A conventional explanation: bring back the cosmological constant!}
Now it is not a secret that GR admits a cosmological constant $\Lambda$, a constant term in the Einstein-Hilbert lagrangian
\begin{equation}
S = \int d^4 x \sqrt{-g} [\frac{R -2 \Lambda}{16 \pi G} + \mathcal{L}_m ]\, ,
\end{equation}
where $g$ is the determinant of the metric and $\mathcal{L}_m$ is the Lagrangian of the matter content of spacetime from which the stress-energy tensor is derived. The cosmological constant term $\Lambda$ is able to produce the desired accelerated expansion, as we will explain below. It modifies the Friedman equation giving
\begin{equation}
\frac{8 \pi G}{3}\rho = H^2 + \frac{k}{a^2} - \frac \Lambda 3 \, . \label{11fried}
\end{equation}
Given that the universe is expanding, at late times the effect of the energy density $\rho$ and curvature $k$ will be negligible and the universe will expand almost exponentially with
\begin{equation}
H = \sqrt{\frac \Lambda 3} \, .
\end{equation}
We arrive then to the Concordance Model ($\Lambda$CDM), in which the universe is homogeneous and isotropic in cosmological scales, with a spatial curvature $k$ and a cosmological constant $\Lambda$, and is filled by a mixture of radiation (mainly photons and neutrinos), massive particles of the standard model (mainly baryons, as electrons have a much smaller mass) and new massive particles which are non-baryonic and electrically neutral which we call Dark Matter (DM). If we divide \eqref{11fried} by $H^2$ and regroup terms we arrive to the cosmic balance equation,
\begin{equation}
\Omega_r + \Omega_b + \Omega_{DM} + \Omega_k + \Omega_\Lambda = 1 \, , \label{11balan}
\end{equation}
where $\Omega_i = \rho_i/\rho_c$ for $i=r,b,DM$, $\rho_c = \frac{3H^2}{8\pi G}$ is the critical density, $\Omega_k = -\frac k{a^2 H^2}$ and $\Omega_\Lambda = \frac \Lambda{3 H^2}$. The $\Omega$'s represent the relative importance of each of the components of the universe at a given time, and this relative importance vary.
This simple explanation fits comfortably the data collected by the SNe searchers, from the first to the last datasets \cite{Riess:1998cb,Perlmutter:1998np,Astier:2005qq,Riess:2006fw,Kowalski:2008ez}, which comprehends a few hundreds of SNe. It also provides a good explanation of the measured spectrum of CMB anisotropy \cite{Jaffe:2000tx,Spergel:2003cb,Spergel:2006hy,Dunkley:2008ie} and of the spectrum of matter anisotropy \cite{Peacock:2001gs,Tegmark:2003ud,Eisenstein:2005su,Cole:2005sx} if the universe is taken to be spatially flat, $k=0$ \cite{de Bernardis:2000gy}. In particular, the joint fit of the model to the whole set of experimental data gives, in the present time:
$$
\begin{array}{cc}
H_0 = 72 \pm 3 \, km/sMpc\, , &
\Omega_r = (4.8 \pm 0.5)\times 10^{-5} \, , \\
\Omega_b = 0.044 \pm 0.004 \, , &
\Omega_{DM} = 0.22 \pm 0.2 \, , \\
-0.018 < \Omega_k < 0.008 \, , &
\Omega_\Lambda = 0.726 \pm 0.015 \, .
\end{array}
$$
Most of the errors in the data are dominated by the error in the determination of $H_0$. If this is the end of the story, what is the point of the cosmic acceleration that we are not understanding?
\subsection{The Problems of $\Lambda$CDM}
\paragraph{The Cosmological Constant Problem}
Although a low value of the cosmological constant is compatible with the observations, theory predicts that it has a high value. It would not be any problem if the theory were classical, but the theory describing the source of the gravitational interaction $\langle T_{\mu\nu}\rangle$ is QFT in curved spacetime. Thus, the semiclassical Einstein equations valid for cosmology at curvatures well below the Planck scale ($E_P = 1.2209 \,\times\, 10^{19}\, GeV$ / $l_P = 1.6165 \,\times\, 10^{-35}\, m$)\footnote{At this scales gravity cannot be treated as a classical interaction any more, as we will explain later.} are
\begin{equation}
R_{\mu\nu} -\frac 12 g_{\mu\nu}(R + \Lambda) = 8\pi G \langle\hat T_{\mu\nu}\rangle\, ,
\end{equation}
where the curvature terms are treated classically and $\langle\hat T_{\mu\nu}\rangle$ is the expectation value of the stress-energy tensor computed for a QFT in curved spacetime \cite{BD}. In order to renormalize the contributions of the matter Lagrangian to the total action of matter and gravitation, it is necessary to add counterterms also in the gravitational Lagrangian that cancel some of the infinities arising in the matter Lagrangian. Some of these counterterms are of the same form of the ones appearing in the classical Lagrangian ($R$, $\Lambda$) and some present more than two derivatives of the metric.
The term of interest is the divergent term which has the same form as the cosmological constant $\Lambda$. If we compute the value of this contribution in the frame of Effective Field Theory (EFT) \cite{Burgess:2007pt}, we assume that the SM is valid up to a high energy scale (which we will call ultraviolet (UV) scale), which we assume to be of the order of the Planck scale. Then this contribution is $\sim 10^{71} GeV^4 $, that is, $120$ orders of magnitude greater than the observed value, $\sim 10^{-47} GeV^4 $ \cite{Weinberg:1988cp}. In order to explain the data, the quantum corrections $ 8\pi G \langle\hat T^{vac}_{\mu\nu}\rangle$ should cancel the first $120$ significant figures of the classical contribution $\Lambda$, but not the next one! Such an unnatural cancelation is called the Cosmological Constant Problem.
The solutions to the Cosmological Constant Problem among the quantum gravity community are not very convincing. In String Theory (ST) the natural value of the cosmological constant is negative and of order $E_p^4$. This has led the ST community to resort to anthropic arguments in order to save the face (see Ref.~\cite{Carroll:2000fy} for arguments and references). In Loop Quantum Gravity (LQG) it has been claimed that a positive small cosmological constant is possible in a solution called Kodama state \cite{Smolin:2002sz}. However the dispersion relation of particles living in this spacetime have at energies well below the Planck scale a modified dispersion relation of the form
\begin{equation}
E^2 = \mathbf{p}^2 + m^2 + \alpha \frac{E^3}{E_p} \, ,
\end{equation}
which has been already almost ruled out by observation \cite{Jacobson:2002ye,Mattingly:2005re,Collaborations:2009zq}.
The problem is that we have quantum corrections coming from the UV behavior of a QFT to an otherwise infrared (IR) scale, i.e. of lower energy than the typical energies in particle physics, of GR. Is there any possibility that the quantum corrections coming from QFT do not stop the cosmological constant being IR? EFT relies in some basic principles, and the one responsible of the quantum corrections to the vacuum energy being of the order of the UV scale is the decoupling theorem \cite{Appelquist:1974tg}. Any mechanism able to avoid UV quantum corrections to the stress-energy tensor must violate this property. It would then be possible to have a QFT which is valid in a window of energies bounded by an IR and an UV scale, in which the main quantum corrections to the Einstein's equations come from the IR.
\paragraph{The Horizon Problem} The present homogeneous, isotropic domain of the universe is at least as large as the present horizon scale $c t_0 \sim c/H_0$, where $H_0$ is the present value of the Hubble parameter. This domain has grown up with the scale factor. As inhomogeneity cannot be erased with the decelerated expansion, we can conclude that the size of the patch of the universe from which this domain originated was at $t = t_i$ at least as big as $l_i \sim c t_0 a_i/ a_0$. If we compare this with the causal horizon at the same time, $l_c = c t_i$, then the ratio between the two scales is,
\begin{equation}
\frac{l_i}{l_c}\sim \frac {t_0}{t_i}\frac{a_i}{a_0}\, .
\end{equation}
Thus, if we assume some random initial conditions at $t_i$, which we will assume to be of Planckian size because we consider that GR is valid up to that scale, and $a\sim 1/T$ under radiation domination, we get
\begin{equation}
\frac{l_i}{l_c} \sim \frac{10^{17}s}{10^{-43}s}\frac{2.73 K}{10^{32} K} \sim 10^{28}\, ,
\end{equation}
and therefore that patch was not causally connected, so it did not need to be homogeneous at all! This is a big problem, as the initial conditions should be very fine-tuned to reproduce our current universe. This is the Horizon Problem.
Unless there is something we are missing. Let us assume that the scale factor grows with a power of time, and therefore $\dot a \sim a/t$. Thus we can rewrite
\begin{equation}
\frac{l_i}{l_c} \sim \frac {\dot a_i}{\dot a_0} \, .
\end{equation}
From the last expression we can read that the size of our universe was initially larger than that of a causal patch by the ratio of the corresponding expansion speeds. Only if these are of the same order can the initial conditions at $t_i$ thermalize and give rise to an homogeneous universe. And as gravity has been decelerating this rate for a wide period of time, there must have been a fast accelerated expansion before the decelerated one that has correctly described so much physics. This points out to new particle or gravitational physics in the UV, but below the Planck scale.
\paragraph{The Flatness Problem} Current observations tell us that the curvature term in the Friedman equation is very close to zero. In particular, if we write the cosmic balance equation \eqref{11balan} and neglect $\Omega_\Lambda$, the cosmological parameter $\Omega(t)$ follows
\begin{equation}
\Omega(t) \equiv \frac{\rho(t)}{\rho_c(t)} = 1 + \frac k{(Ha)^2} \, .
\end{equation}
Therefore, at early times,
\begin{equation}
\Omega(t_i) - 1 = (\Omega_0 - 1)\frac{(Ha)^2_0}{(Ha)^2_i} = (\Omega_0 - 1)\frac{(\dot a)^2_0}{(\dot a)^2_i} < 10^{-56} \, .
\end{equation}
Taking into account that this difference can be interpreted as the relative difference between kinetic and potential energy in the cosmic fluid \cite{Mukh}, this means that the initial velocities must be fine-tuned to a degree of $10^{-28}$ in order to explain the current flatness of the universe! This is the flatness problem.
This fine-tuning would not be so high if the expansion rate at Planckian time would be smaller than or at least as small as the present expansion rate, which would in turn require an early accelerated expansion, as we have seen before.
\subsection{Ingredients for an alternative to $\Lambda$CDM}
The problems of $\Lambda$CDM have their origin in our attempt of extending the range of validity of our current theories of gravity and particle physics outside the region in which those theories have been experimentally tested and validated. We have no idea of what the correct theory of gravity will be at extremely high (or low) curvatures compared to the ones that have been tested; and we have arguments that point out that a quantum theory of gravity may violate some of the principles underlying GR and QFT \cite{Garay:1994en}.
At this point it becomes evident that one must find a way of extending either GR, QFT or both in order to explain the problems of cosmology at very early or recent times. This extensions can be performed either following a top-down or a bottom-up approach. In a top-down approach the fundamental principles of the theories are changed, a new theory is derived, and the phenomenological consequences are derived. Examples of top-down approaches to the problems of cosmology are ST \cite{Polchinski}(and its applied branch of String Cosmology) or LQG \cite{Rovelli:1997yv,Thiemann:2001yy} (from which Loop Quantum Cosmology is derived). The aforementioned approaches have departed from a solid theoretical and mathematical base but have grown up on the sidelines of the experimental results, and some of their implications for cosmology are in conflict with observational data. We will show a couple of examples below.
In contrast, a bottom-up approach is based on experimental and observational data, which are used to build up laws and models which can later on be incorporated to a well founded theory. In this PhD thesis we will try to show that it is possible to build a phenomenological extension to $\Lambda$CDM at the level of the cosmological evolution equations. This extension, which we have called the Asymptotic Cosmological Model (ACM), incorporates a modification of the laws of cosmic expansion such that an epoch of (asymptotically exponential) accelerated expansion arises both in the IR (corresponding to nowadays Hubble rate) and in the UV (corresponding to the very early universe). The model fits successfully the current observational data. The study of the origin of the modifications is precluded, although some possibilities ar pointed out in the following.
\paragraph{New physics in the IR}
We have mentioned that in order to solve the Cosmological Constant Problem, we need to render the UV corrections to the vacuum energy density IR. This is not possible in the frame of local EFT, due to the decoupling theorem. In fact, this is a problem not only of $\Lambda$CDM, but of GR coupled to an EFT in general. Even if there is no classical contribution to the cosmological constant/vacuum energy in the gravitational lagrangian, the renormalization of the matter lagrangian will in general induce a nonzero term. \footnote{With the remarkable exception of supersymmetric theories with unbroken supersymmetry. However, the absence of exact supersymmetry in nature discards this explanation.} The problem is that the power counting of a local EFT implies that the corrections to this term will be much greater than the observed value. Therefore a solution to the Cosmological Constant Problem will presumably require a departure from locality in QFT, which is parameterized by some new IR scale.
It has already been noticed that local EFT should fail to be a good description of nature in the presence of gravitational effects \cite{Cohen:1998zx}. In any EFT with a UV cutoff $E_{NP}$, the entropy of a system is an extensive quantity and therefore grows with its volume, $S \sim L^3 E_{NP}^3$. However blackhole physics seems to reveal that the maximum entropy contained in a certain box of size $L$ must be the Bekenstein entropy, $S < \pi E^2_p L^2$. This means that the entropy predicted by EFT can exceed the Beckenstein bound for sufficiently large sizes. Therefore it has been argued that EFT overcounts the degrees of freedom in some way \cite{'tHooft:1993gx,Susskind:1994vu}, and that the number of degrees of freedom of a given system must saturate from an extensive behavior (proper of EFT) to a proportionality with area, when the size of the region investigated is of order $L \sim E_p^2 / E_{NP}^3$. The authors of Ref.~\cite{Cohen:1998zx} propose an even more restrictive domain of validity, based on the premise that the size of the region under consideration should not exceed the Schwarzschild radius associated with their temperature $T\sim E_{NP}$ for EFT to be valid. If EFT is used to describe particles up to temperatures $T$, then the associated thermal energy of the system will be of order $U\sim L^3 E_{NP}^4$ and the associated entropy will be of order $S \sim L^3 E_{NP}^3$. Then the Schwarzschild radius will be $r_s \sim U/E_p^2$. If we want the system under description not to form a blackhole, then its size needs to be greater than this quantity, $L > r_s$, and therefore the momenta of the particles we are describing will present an IR cutoff of order $\lambda \sim 1/L$, where
\begin{equation}
\lambda \sim E_{NP}^2 / E_p \, .
\end{equation}
This would generate, aside from corrections $\propto \frac 1{E_{NP}^n}$ to the tree level effective action, also IR corrections $\propto \lambda^2$. As a consequence, in a given experiment at energies $\sim p$ the predictions of EFT for a given quantity $Q$ will have corrections from both irrelevant terms in the action of dimension $D$ and gravitational physics, of the form
\begin{equation}
\frac{\delta Q}{Q} \sim \left[ \left(\frac{p^{D-4}}{E_{NP}^{D-4}}\right) + \left( \frac{E_{NP}^4}{p^2 E_p^2}\right) \right]\, .
\end{equation}
The maximum possible accuracy will be attained by some value $E_{NP}$ of the maximum energy scale that can be probed. As an example given in Ref~\cite{Cohen:1998zx}, if the momenta of the particles are of the order of the electroweak scale (as in the new collider experiments, like LHC), and the EFT is the standard model, new physics appears at dimension $6$ operators ($5$ if we allow for lepton number violation), the minimum uncertainty of $10^{-13}$ is achieved when the maximum scale that can be probed with EFT techniques is $10^8 GeV$. The minimum scale that could be probed would then be of the order of the $MeV$. However if new physics appears at $10 Tev$, then the SM would be valid up to that energies and down to $10^{-2} eV$, around the energy scale which is behind neutrino physics or the cosmological constant.
In the same context, if one tries to compute the quantum corrections of the matter content of the universe to the cosmological constant in the context of EFT, one should notice that for this approach to be valid at a patch of the universe of the size of the causal horizon $\sim 1/H_0$, one should introduce an IR cutoff of the order of $H_0 \sim 10^{-26} eV$ and therefore the associated UV cutoff of the EFT would be $10^{-2} eV$. As a result, quantum corrections would be of the same order of the observed value of he vacuum energy.
It has been claimed that any EFT approach to a theory including gravity will be naturally attached to a range of validity bounded by an UV and an IR scale \cite{Carmona:2000gd}. However, if the UV scale is taken to be the Planck scale, then the IR will also equal the Planck scale and the range of validity of the EFT will be constrained to a point. It may be that the effective fundamental scale of gravity in the EFT would be decreased by the existence of large extra dimensions \cite{ArkaniHamed:1998rs,Antoniadis:1998ig,Sundrum:1998ns,ArkaniHamed:1998kx}. The effective gravity scale has been proposed to lie in the order of magnitude of the few $TeV$ scale, and the IR scale (corresponding to extra dimensions of the size of a millimeter) would be of the order of the scale of the cosmological constant or neutrino physics.
Anyway, if local EFT is to be modified in the IR in order to solve the Cosmological Constant problem, it is also possible that the quantum corrections induced in the gravitational Lagrangian add not only terms with a higher number of derivatives of the metric suppressed by an UV scale, as in the standard case, but also other nonlinear or even nonlocal terms suppressed by the IR scale.
In fact, there is no reason to believe that the action of the classical theory of gravity is just given by a term proportional to the curvature scalar and a constant term, as in $\Lambda$CDM. With the premises of a metric theory with general covariance and the equivalence principle, it is possible to build an infinite plethora of theories. Einstein decided to choose the simplest one for the shake of aesthetics: he required his equations to be of second order, and this is only possible with a Lagrangian of the form $\sqrt{-g}(R-2\Lambda)$. However in modern theoretical physics we have learned that everything that is not forbidden is compulsory \cite{Weinberg:2005kr}. Additional terms in the action are not forbidden by the symmetries of the theory as long as they behave as scalars under general coordinate transformations.
Terms with a higher number of spatial derivatives can be neglected by dimensional analysis, as they must be supplied by a certain scale. If this scale is a UV scale such as the Planck scale, then the effect of these terms can be neglected at most tests, with the remarkable exception of the very early universe. Terms of these type arise naturally when one tries to renormalize a QFT in curved space \cite{BD}. In this context even if they are not originally in the classical gravitational action, terms of the following form appear in the action:
\begin{equation}
S_g^{UV} = \frac{1}{16\pi G} \int d^4 x \sqrt{-g} \left[ -2\Lambda + R + \frac 1{E_{UV}^2}\left( \alpha R^2 + \beta R_{\mu\nu}R^{\mu\nu} + \gamma R_{\mu\nu\rho\sigma}R^{\mu\nu\rho\sigma} + \delta \square R \right)\right]\, , \label{GRUV}
\end{equation}
and there is no reason to believe that terms with even higher derivatives of the metric appear in the gravitational action, suppressed by greater powers of the inverse of the UV scale $E_{UV}$.
However, if these terms are supplied by an IR scale, like the cosmological constant, their effect should be observable at cosmological distances. Aside from a cosmological constant, terms involving an infrared scale should be at least non-polynomial or nonlocal,
\begin{equation}
S_g = \frac{1}{16\pi G} \int d^4 x \sqrt{-g} \left[ R -2\Lambda + \alpha E_{IR}^2 \square^{-1}R + \beta\frac{E_{IR}^4}R + ... \right]\, , \label{GRIR}
\end{equation}
where the effect of these terms can na\"{\i}vely be neglected in some local tests of gravity but heavily influence the cosmic evolution at the present time (provided that $\Lambda \sim E_{IR}^2$).
\paragraph{New physics in the UV}
We have also pointed out that, in order to solve the Horizon Problem and the Flatness Problem of the universe, some sort of new physics in the UV is required. As the universe expands it also cools down from an ``initial'' state at Planckian temperature $\sim T_p = k_B^{-1} E_p \sim 10^{32} K$, whose description requires some knowledge of the correct theory of quantum gravity. Lacking any knowledge of such a theory, the best we can do is to assume that quantum gravity provides us with some random initial conditions that can, since some initial time $t_i \gtrsim t_P$, be treated with classical gravity together with some extra corrections suppressed by an UV scale. This new physics could maybe induce an accelerated expansion just after the regime in which gravity must be described by a quantum theory.
Many models have been built in order to explain this accelerated expansion. First models resorted to the coupling of the expansion to a phase transition in the context of GUT \cite{Sato:1980yn,Guth:1980zm,Linde:1981mu,Albrecht:1982wi}. Later it was discovered that the details of the particle physics behind the accelerated expansion were not important \cite{Linde:1983gd}. The addition of terms with higher derivatives suppressed by a UV scale \cite{Starobinsky:1980te} has also been used to explain this accelerated expansion (called inflation \cite{Guth:1980zm}). As an additional consequence the quantum fluctuation of fields in that regime is enough to cause the fluctuations of the CMB spectrum, and therefore clusters, galaxies, stars, planets and life. As said, the details of the mechanism are unimportant, as several {\it a priori} very different models lead to very similar conclusions.
As an example of new physics in the UV leading to an early accelerated expansion, let us introduce the very simple model of Ref.~\cite{Starobinsky:1980te}. Assume that the Einstein-Hilbert action is just a low energy approximation to a more general theory,
\begin{equation}
S = \int d^4 x \sqrt{-g} \left[ R + \Lambda - \alpha \frac{R^2}{E_p^2} + ... \right] \, .
\end{equation}
It has been shown \cite{BD} that such terms appear in the renormalized action of classical GR coupled to a QFT in curved spacetime. If we consider the theory modified just by the $R^2$ term, it turns out to be conformally equivalent (with a conformal factor $F = 1 + 2\alpha R/E_p^2$) to a theory of GR coupled to a scalar field $\varphi$,
\begin{equation}
S = \int d^4 x \sqrt{-g} \left[ R + \Lambda + 16 \pi G(g^{\mu\nu}\partial_\mu\varphi\partial_\nu\varphi - \frac{E_p^2}{12\alpha}\varphi^2) \right] \, ,
\end{equation}
with $\alpha$ a few orders of magnitude above unity. The potential of the scalar field can be seen as an effective mass term with $m^2 = \frac 1{6\alpha} E_p^2$ A detailed study of the dynamical system will prove that this scenario is an example of chaotic inflation \cite{Linde:1983gd}, in which almost random initial conditions lead to an exponential expansion and to an almost flat power spectrum for the primordial fluctuations in the metric of spacetime.
Let us assume that the initial condition of the scalar field is at least locally an homogeneous condensation. The equations of motion in the homogeneous approximation (FRW metric and $\varphi \approx \varphi(t)$) can be written as:
\begin{eqnarray}
\ddot \varphi + 3 H \dot \varphi + m^2 \varphi & = & 0 \, , \\
\frac{8\pi G}3 \left( \frac 12 \dot\varphi^2 + \frac 12 m^2 \varphi^2 \right) & = & H^2 \, ,
\end{eqnarray}
where we are neglecting the effect of the spatial curvature ($H^2 \gg k/a^2$) for the sake of simplicity. The effect of spatial curvature can be studied numerically. The system can be then reduced to a first order dynamical system for $\dot\varphi (\varphi)$,
\begin{equation}
\frac{d \dot \varphi}{d \varphi} = -\sqrt{-12\pi G(\dot \varphi^2 + m^2\varphi^2)} - m^2\frac{\varphi}{\dot\varphi}
\end{equation}
It should be noticed that this dynamical system possesses an attractor to which all other solutions converge with time. The attractor behaves as $\dot \varphi^2 = \frac{m^2}{12\pi G}$ and $sign(\dot\varphi)\neq sign(\varphi)$
for $\dot \varphi^2 \ll m^2 \varphi^2$ and then forms a double decaying orbit around the origin. If a solution meets the attractor at one of its flat branches, in this regime the field varies slowly with time and the universe expands almost exponentially, with
\begin{equation}
H^2 \approx \frac{8\pi G}3 \left( \frac 12 m^2 \varphi^2\right) \sim const \, .
\end{equation}
until $\varphi \sim \frac 1{\sqrt{12\pi G}}$. When the value of the field reaches the Planck scale its decrease rate is high and the field rapidly decays to $\varphi = 0$, providing the exit to inflation and transferring the energy of the scalar field to the matter content of the model (reheating). As the original theory is related to the scalar-metric theory by a conformal factor which depends only on $R$, which is approximately constant during the inflationary regime, then in the original theory the same inflationary solution exists.
One can arbitrarily complicate this model, but the general prediction of an early accelerated expansion (in most models close to an exponential expansion) leading to a flat, thermalized universe with small perturbations obeying a scale free power spectrum is rather general. Most of the models include modifications of gravity or a set of scalar fields which mimics the previous mechanism. Noticeably, the universe is devoid of matter (or emptied by the mechanism) until the accelerated expansion stops, and then refilled by the energy released.
\paragraph{Agreement with observation}
In the past decade a vast amount of observational data has been collected, that severely constrains most of the models that have been built for cosmology, especially those derived from top-down approaches like ST or LQG. At this point, it is no longer sufficient to build models which qualitatively explain the features in the universe, but also which provide us with competitive fits of the currently gathered data and desirably with predictions to be tested by future missions.
There is a list of cosmological observations that should be accommodated in an extension of $\Lambda$CDM:
\begin{itemize}
\item
In any sensible model of inflation quantum (or thermal) fluctuations are produced, which seed the perturbations that will later produce the large scale structure of the universe, galaxies, and so on. Most inflationary models predict that the primordial scalar fluctuations will have an almost scale invariant power spectrum $P(k)$,
\begin{equation}
P(k) \equiv \frac{\vert \Phi_k\vert^2 k^3}{2\pi^2} \propto k^{n_s -1} \, ,
\end{equation}
where $\Phi_k$ is the Fourier transform of the gauge-independent scalar perturbation of the FRW metric in GR without a shear, and $n_s$ is called 'scalar spectral index' (scale invariant spectrum means $n_s \approx 1$). The Starobinsky $R^2$ model \cite{Starobinsky:1980te} and in general chaotic inflation \cite{Linde:1983gd} predict $n_s \lesssim 1$. In fact, the scalar spectral index has managed to be measured in the CMB spectrum by WMAP \cite{Dunkley:2008ie}, and it turns to be $n_s = 0.963^{+0.014}_{-0.015}$.
\item
Models for the present accelerated expansion must be confronted with a much wealthier sample of data. Given that Type I a Supernovae (SNe) can be taken as standard candles, it has been possible to determine the luminosity distance $d_L$ vs. redshift $z = a_0/a -1$ relation,
\begin{equation}
d_l (z) = c(1+z) \int^z_0 \frac {dz'}{H(z')} \, ,
\end{equation}
for a few hundred SNe \cite{Riess:2006fw,Kowalski:2008ez} ranging for $z<2$. Besides supernovae, the reduced distance to the surface of last scattering $R_{CMB}$,
\begin{equation}
R_{CMB} = \Omega_m^{1/2} H_0 \int_0^{1089} dz/H(z) \, ,
\end{equation}
can be also read from the WMAP data \cite{Komatsu:2008hk}. Both measurements involve the integral of the inverse of the Hubble parameter with respect to the redshift, but SNe observations sample the present era whereas the reduced distance $R_{CMB}$ provides an integrated information of the evolution of the universe since its youth.
A slightly different dependence is found in the distance parameter of Baryon Acoustic Oscillations (BAO) $A(z)$,
\begin{equation}
A(z)= \Omega_m^{1/2} H_0 H(z)^{-1/3} z^{-2/3} \left[\int_0^z dz'/H(z')\right]^{2/3} \, ,
\end{equation}
measured at redshift $z=0.35$ by SDSS \cite{Eisenstein:2005su}. Although some other data can be confronted with cosmological models, these are the most widely used probes in cosmology.
\end{itemize}
In conclusion, GR can be extended to include new IR and UV phenomena, and it should be extended in order to satisfactorily explain both the initial conditions of the universe and the current accelerated expansion. It should also be extended in order to comprehend the new corrections arising from the renormalization of the matter lagrangian coupled to gravity, especially if this Lagrangian is also modified in the IR or the UV.
The effect of this extension in cosmological scales should be to provide us with an accelerated expansion in the very early universe and another accelerated expansion in the present age. These accelerated expansions are in most cases asymptotically exponential expansion. In this PhD thesis a cosmological model will be built, in which these to epochs of accelerated expansion will be phenomenologically incorporated. The resulting model will be used to fit the available observational data.
\subsection{Review on alternative models to explain the accelerated expansion}
This subsection is a brief review on the main alternatives proposed in order to explain the problems concerning the present accelerated expansion and the very early universe.
The main paradigm explaining the Horizon and Curvature problems in the very early universe is inflation; either occurring due to a modification of GR in the UV or the addition of new fields at high energies in extensions of QFT such as ST. The only real alternative to the inflationary paradigm is that of Varying Speed of Light (VSL) \cite{Albrecht:1998ir}. We will explain both alternatives further below.
With the increasing interest on the present accelerated expansion, a wide plethora of models has risen. Aside from modifications of gravity (such as $f(R)$-theories, higher dimensional models, alternatives to Dark Matter explaining the smallness of the cosmological constant), these models include new forms of matter (new fields of any type, perfect fluids with exotic equations of state,...) or rejection of spatial homogeneity as a good approximation in the present universe (effective cosmological constant arising from homogenization procedure \cite{Kolb:2005da,Wiltshire:2007zj}, local void as an explanation to the apparent accelerated expansion \cite{Celerier:1999hp,Tomita:2000jj,GarciaBellido:2008nz,Celerier:2007jc},...). Some modifications of gravity try even to unify the explanation of inflation, the present accelerated expansion and/or Dark Matter. We will briefly introduce also some models of modified gravity and of exotic sources of the gravitational field.
\paragraph{Scalar fields for inflation}
We have already mentioned that it is possible to drive the universe into an accelerated expansion era with the help of a set of scalar fields with an appropriate potential. These scalar fields might arise as fundamental fields in a QFT beyond the standard model (like in Ref.~\cite{Guth:1980zm}) or as new modes in a modification of general relativity (as in Ref.~\cite{Starobinsky:1980te}). As noticed in Ref.~\cite{Linde:1983gd}, the particular potential behind the accelerated expansion is unimportant, provided it satisfies some very general conditions, the 'slow-roll conditions'.
Since then, there has been much effort directed to find some viable inflation mechanisms in theories of quantum gravity. In particular, the string theory community has viewed the inflationary paradigm as a playground in which their ideas can be tested. Many possible realizations of inflationary scenarios have been found.
As ST lives in $10$ dimensions, it needs the compactification of $6$ extra dimensions in order to describe our world. The moduli, which are the parameters describing this compactification, are stabilized dynamically in the quantum theory, producing different vacua. There are numerous such solutions, with different values of the vacuum energy. The set of these vacua is called the moduli space of supersymmetric vacua. If we expand this set to include non-supersymmetric vacua\footnote{We are talking about $10^{500}$ possible vacua!}, we arrive to the `landscape' \cite{Susskind:2003kw}. It was first proposed as a way of solving the cosmological constant problem: if there are multiple vacua and the cosmological constant differs from one vacua to the other, and considering that human life is only possible if the value of this constant lies in a certain range, then we must be living in one of the 'anthropically permitted' vacua.
The landscape offers multiple options for inflationary model-building, such as hybrid inflation \cite{Linde:1993cn}. In this model, inflation is driven by two scalar fields, with a potential
\begin{equation}
V(\phi,\chi) = \frac 12 m^2 \phi^2 +\frac 12 \chi^2 \phi^2 + \frac \lambda 4 (\chi^2 - \frac{M^2}4)^2 \, .
\end{equation}
The main new ingredient in the hybrid inflation as compared to standard inflation is that the corresponding quantum fluctuations are non-gaussian due to the interaction. In contrast to single field inflation, hybrid inflation naturally gives $n_s \geq 1$ \cite{Bassett:2005xm}, and thus is ruled out as the mechanism realized in nature.
Other inflationary scenarios arise in the context of $Dp$-branes moving through extra dimensions. In the brane-world scenario, our spacetime corresponds to one of these $D3$-branes. The SM particles live stuck to the $D$-brane, for they are modes of open strings whose ends lie on it. Graviton and other open string modes are in contrast not confined. In 'brane-inflation' the inflaton corresponds simply to the position of this brane \cite{Dvali:1998pa}. A simple model is $D3$-brane/$\bar D3$-brane, in which brane and antibrane experiment an attractive force and finally annihilate each other, causing the end of inflation \cite{Burgess:2001fx}. Here the inflaton $\varphi$ is the relative position of brane and anti-brane and the potential $V(\varphi)$ is due to the tensions and interaction. Energy is released with the annihilation and cosmic superstrings are left as a remnant. More realistic but much more complicated is the KKLMMT scenario \cite{Kachru:2003sx}. In general, most of these models give $n_s \geq 0.98$ and predict the creation of a cosmic string network which has not been observed \cite{HenryTye:2006uv}, and are therefore ruled out.
However, despite the interest of string theorist, stringy scenarios are seen by phenomenologists and cosmologists as very degenerate and very little predictive. The idea of the landscape and the 'anthropic selection of vacua' is often received with skepticism and polemic.
\paragraph{Alternatives to inflation: VSL}
For more than a decade, inflationary cosmologies have not had any real competitor. However, in Ref.~\cite{Albrecht:1998ir} an alternative mechanism was first proposed under the name of Varying Speed of Light (VSL). The proposal is that the local speed of light, as measured by comoving observers, might have varied with time in the very early universe from a very large value to its current value. This implies obviously a departure from SR at very high energies, and therefore mediated by some UV scale.
Assume that in the preferred frame in which the universe is approximately homogeneous, the speed of light varies as $c(t)$. Therefore, $ds^2 = c(t)^2 dt^2 - a(t)^2 d\mathbf{x}^2$ (although $g_{00}=1$) but the Friedman equations are still valid,
\begin{eqnarray}
H^2 & = & \frac{8\pi G}3 \rho - \frac{k c(t)^2}{a^2} \, , \label{vslf1} \\
\frac{\ddot a}a & = & -\frac{4 \pi G}3 \left(\rho + 3\frac p{c(t)^2}\right) \, . \label{vslf2}
\end{eqnarray}
If the universe is radiation dominated, $p = \rho c(t)^2/3$ and thus $a \propto t^{1/2}$ as usual. As a consequence of the violation of invariance of time translations even at the local level, the energy conservation condition is modified,
\begin{equation}
\dot\rho + 3H\left( \rho + c(t)^{-2}p\right) = \frac{3 k c(t) \dot c (t)}{4 \pi G a^2} \label{vslcont}
\end{equation}
The horizon problem is straightforwardly explained within this model. The ratio of the event horizon to the causal horizon at $t = t_i$ is now given by
\begin{equation}
\frac{l_i}{l_c}\sim \frac {c_0 t_0}{\int^{t_i}_0 c(t) dt }\frac{a_i}{a_0} \, ,
\end{equation}
and thus can be made arbitrarily small with an appropriate choice of the function $c(t)$. A ``superluminal'' propagation of light at early times suffices. In the original paper \cite{Albrecht:1998ir}, a phase transition between an early value of the speed of light $c_i \sim 10^{30} c_0$ and the present one $c_0$ at $t \lesssim t_p$ is proposed.
The flatness problem can also be understood. Taking Eqs.~\eqref{vslf1},\eqref{vslf2},\eqref{vslcont} into account, we get that in a universe dominated by a perfect fluid with $p = \omega \rho c(t)^2$ the cosmological parameter $\Omega(t)$ evolves as
\begin{equation}
\dot \Omega = \Omega (\Omega - 1) H (1+3\omega) +2 \frac{\dot c}c(\Omega -1) \, .
\end{equation}
The inflationary cosmology solves the flatness problem because the inflaton field satisfies $1 + 3\omega < 0$ and thus $\Omega \approx 1$ is stable. The varying speed of light is able to solve the problem because the term proportional to $(1 + 3\omega)$ can be compensated by the term proportional to $\dot c$ as long as $\vert \frac{\dot c}c \vert > \Omega H$.
Some models providing concrete realizations of VSL can be found in Refs.~\cite{Moffat:1992ud,Albrecht:1998ir,Barrow:1999is} (see Ref.~\cite{Magueijo:2003gj} for a review). A covariant theory was presented in Ref.~\cite{Magueijo:2000zt}, as a gauged Fock-Lorentz symmetry theory in which $c$ is allowed to vary as a scalar field does in the manifold, $c(x)$. Another similar covariant realization is Rainbow Geometry \cite{Magueijo:2002xx}, in which the metric of the spacetime as seen by a test particle of a given energy depends on this energy. However, the interpretation of this scenario seems obscure. A more clear interpretation is found if the speed of light is wavelength dependent, $c(E)$, due to a modified dispersion relation. In the very early universe the wavelength of the photons becomes comparable to the Planck length or other UV scale and thus the global effect can be seen as VSL \cite{Alexander:2001ck}. This approach allows a deeper description of VSL cosmology \cite{Alexander:2001dr} and structure formation \cite{Magueijo:2002pg}. Note however that this idea is also not absent of criticism \cite{Ellis:2007ah}.
\paragraph{$f(R)$-theories}
There has been an interest in modifying GR since it was formulated \cite{Weyl:1919fi}, as a consequence of the success of GR in the description of the precession of the perihelion of Mercury. As a consequence of the difficulties in the renormalization of GR, extensions of the Einstein-Hilbert action with quadratic terms in the action (like in Eq.~\eqref{GRUV}) were studied; later it was proved that these corrections provided a mechanism of inflation \cite{Starobinsky:1980te}, as we have explained. Nonlinear corrections in the scalar curvature to the Einstein-Hilbert action have also found motivation in ST \cite{Candelas:1985en,Nojiri:2003rz}.
However, the motivation for extending the Einstein-Hilbert action can be much more mundane: there is no fundamental principle fixing the form of the gravitational action. Thus if we allow for new scales in a gravitational theory, new terms in the action will generally appear, which must respect general covariance. In this sense the gravitational action can be arbitrarily extended as in Eqs.~\eqref{GRUV} or \eqref{GRIR}. As a toy model one may consider actions which depend only on the curvature scalar,\footnote{More general models have also been considered. For instance, models which depend on the squared curvature scalars \cite{Carroll:2004de,Mena:2005ta} or which are nonlocal \cite{Wetterich:1997bz,Deser:2007jk,Capozziello:2008gu}.}
\begin{equation}
S=\int d^4 x\sqrt{-g}[\frac 1{16 \pi G}f(R) + L_m] \, .
\end{equation}
These are called $f(R)$-theories \cite{Faraoni:2008mf}. In fact, one can distinguish between three types of $f(R)$-theories depending on how the variational principle is implemented on the action. In metric $f(R)$, the only gravitational field appearing in the action is the metric $g_{\mu\nu}$ and variation with respect to the metric provides the equations of motion. In Palatini $f(R)$ the scalar curvature is the result of the contraction of the inverse of the metric of the Ricci tensor, which in turns depends only on the connnection (treated as an independent field). The matter Lagrangian is considered to depend only on the metric, but not on the conection. In metric-affine $f(R)$, not only the Ricci tensor is made of the connection, but also the matter Lagrangian contains this field. Both of the three formalisms are equivalent if $f(R) = R + \Lambda$, but differ if $f''(R) \neq 0$. In this thesis mainly metric $f(R)$-theories will be treated.
In a metric $f(R)$ theory, the equations of motion are,
\begin{equation}
f'(R)R^\mu_\nu-\frac12
f(R)\delta^\mu_\nu+(\delta^\mu_\nu\square-\nabla^\mu\nabla_\nu)f'(R)=8\pi G
T^\mu_\nu \, ,
\end{equation}
which, particularized to a FRW metric, become
\begin{eqnarray}
8 \pi G \rho & = & -3(\dot{H}+H^2)f'(R_{(0)})-\frac 12 f(R_{(0)}) -3 H
f''(R_{(0)})\dot{R}_{(0)} \, ,\\
-8 \pi G p & = & -(\dot{H} + 3 H^2)f'(R_{(0)})-\frac 12 f(R_{(0)})
+f^{(3)}(R_{(0)})\dot{R}_{(0)}^2 + f''(R_{(0)})\ddot{R}_{(0)} \, ,
\end{eqnarray}
where
\begin{equation}
R_{(0)}(t) = -6(\dot{H} +2H^2)
\end{equation}
is the curvature scalar of the Friedman-Robertson-Walker (FRW) metric. For a $f(R)$-type modification in the IR it is then possible to find mechanisms for the present acceleration of the universe \cite{Carroll:2003wy}; in the same way in a $f(R)$-type modification in the UV it is possible to find mechanisms for inflation \cite{Starobinsky:1980te}. It has also been tried to unify inflation and the present cosmic acceleration in the context of these theories \cite{Nojiri:2003ft,Sotiriou:2005hu}.
Metric $f(R)$ theories are formally equivalent to scalar-tensor theories (as we explained in the context of the Starobinski model), with a potential for the scalar field
\begin{equation}
V(\varphi) = \frac 1{16\pi G}\frac{Rf'(R)-f(R)}{f'(R)^2}\, ,
\end{equation}
where $R(\varphi)$ can be obtained from
\begin{equation}
\varphi \equiv \sqrt{\frac 3{16\pi G}} log f'(R)\, .
\end{equation}
The scalar presents an effective mass
\begin{equation}
m^2 = \frac 12 \frac {d^2 V}{d \varphi^2} = \frac 16\left(\frac{R f'(R)-4 f(R)}{f'(R)^2} + \frac 1{f''(R)}\right)\, .
\end{equation}
Thus, $f(R)$ can be seen as theories of a propagating massless spin-2 field (the graviton) together with a (generally massive) scalar field.\footnote{In more general modifications, more effective propagating degrees of freedom may arise.} This scalar field may act as an inflaton in the very early universe but becomes a problem when building an $f(R)$ theory consistent with solar system tests of gravity.
In order to present a viable alternative to $\Lambda$CDM, a modified gravity theory must pass a number of tests. In particular, an $f(R)$-theory must:
\begin{itemize}
\item
possess the correct cosmological dynamics. Many models present non-smooth transitions, for instance between the radiation dominated and matter dominated era, which were naively unexpected (for instance, see Ref.~\cite{Amendola:2006kh}). It has been also pointed out that some $f(R)$ theories may possess different future time singularities \cite{Abdalla:2004sw}, classified in Ref.~\cite{Nojiri:2005sx}. Future time singularities can be cured by adding $R^2$ terms to the action, as shown in \cite{Abdalla:2004sw}. Both of these problems may also be surmounted by starting from a given evolution $a(t)$ and integrating the ordinary differential equation that relates it to the family of $f(R)$ which present that solution \cite{Capozziello:2005ku}.
\item
not suffer from instabilities. The simple model in Ref.~\cite{Carroll:2003wy} was shortly found to have an instability \cite{Dolgov:2003px}. However this instability was later cured by the addition of new terms to the action relevant at high curvatures \cite{Nojiri:2003ft}, much like future time singularities. Metric $f(R)$-theories suffer from this instability if $f''(R)<0$ \cite{Faraoni:2006sy}. If this is the case the energy of the corresponding scalar field may become negative. This can be understood in terms of the effective gravitational coupling appearing in metric $f(R)$ theories $G_{eff} = G/f'(R)$. If $f''(R)<0$ then $d G_{eff}/d R >0$. Large curvature causes gravity to become stronger, so curvature further increases, causing a positive feedback mechanism. Also an unprotected curvature singularity arises in the nonlinear regime of IR modified gravity, which makes it difficult to describe the strong gravity regime of relativistic stars \cite{Frolov:2008uf}. In fact this singularities that appear in the spherically symmetric case are connected to the future singularities of the FRW spacetime mentioned above \cite{Nojiri:2009uu}, and are as well cured with the addition of new terms in the action which are relevant at high curvature \cite{Abdalla:2004sw}.
\item
be free from ghosts. Although $f(R)$ is ghost-free, more general modifications of gravity may contain ghost fields. Extensions containing also the Gauss-Bonnet scalar $\mathcal{G}= R^2 - 4 R_{\mu\nu}R^{\mu\nu} + R_{\mu\nu\rho\sigma}R^{\mu\nu\rho\sigma}$ present no ghosts \cite{Nunez:2004ji}.
\item
have the correct Newtonian and post Newtonian limit and be compatible with the data of CMB and LSS. These conditions have in common the fact that they depend on the behavior of inhomogeneous perturbations of a homogeneous background metric. In the conformally equivalent scalar-tensor theory, the problem is that the propagating scalar degree of freedom may spoil the behavior of this perturbations; it gives rise to a modification of the Newton's law and modifies cosmological perturbation dynamics \cite{Chiba:2003ir}. In most models, the mass of the scalar is naturally light ($ m < 10^{-3} \, eV$) and thus the effect on Solar System tests rules them out as viable gravity theories. However it must be taken into account that the effective mass of the scalar depends on the local curvature, and therefore differs in cosmological, Solar System and on Earth tests. This is known as the `chameleon effect' \cite{Khoury:2003aq,Khoury:2003rn}. Thus it is possible to have scalars modifying gravity at sub-millimeter scales in the Solar System (and therefore remaining undetected) while offering the long-ranged modified cosmological dynamics. Examples can be found in Refs.~\cite{Faulkner:2006ub,Hu:2007nk,Nojiri:2007as,Cognola:2007zu}. However it is not clear wether the physics of the conformally equivalent scalar-tensor theory always provides a good picture of the physics in the $f(R)$ theory \cite{Carloni:2009gp}. The behavior of scalar perturbations may change due to the redefiniton of fields done after the conformal transformation. The behavior of cosmological perturbations in metric $f(R)$-theories is still work in progress: it has been defined in the covariant approach \cite{Carloni:2007yv} and in the Newtonian gauge \cite{delaCruzDombriz:2008cp}. It is then clear that the evolution of the energy density involves a fourth order differential equation. A similar situation holds for Palatini $f(R)$-theories.
\item
have well-posed Cauchy problem. It has been shown that metric $f(R)$-theories present a well-posed Cauchy problem for several sources (perfect fluids, scalar fields) while Palatini $f(R)$-theories do not \cite{LanahanTremblay:2007sg}.
\end{itemize}
Remarkably, these problems do not only arise in modified gravity theories, but also have their equivalent in other approaches to the problem of acceleration like quintessence models or dark fluids. This can be trivially seen as one can write the effect of a mofification of gravity as a scalar-tensor theory (as shown above) or as a dark fluid model. In the same way as in singular $f(R)$ theories, the problem of singular quintessence or dark fluid models can be cured with the addition of terms with higher powers of curvature to the action, such as the ones that appear in the renormalization of field theories in curved space \cite{Nojiri:2009pf}.
To sum up, $f(R)$-theories are a simple toy model which shows the difficulties of departing from GR in a consistent way and in agreement with the experimental and observational data. This is possible under certain conditions that impose a great fine-tuning of the function $f(R)$. The origin of the constraints is the presence of new effective propagating degrees of freedom, i.e. the presence of differential equations of order greater than two.
This is a general problem of theories beyond GR. We will try to surmount this problem (at least in an effective way) in the formulation of the Asymptotic Cosmological Model (ACM) in one of the works presented in this PhD thesis, and in the treatment of cosmological perturbations in the ACM.
\paragraph{Higher Dimensional Models}
Inspired by ST, it has been proposed that our universe might be a four-dimensional hypersurface (a 3-brane) embedded in a higher dimensional bulk spacetime. The SM fields are considered to live in the brane while the gravitational field (or fields) are considered to live in the whole bulk spacetime. The question is how to recover the standard Newton's law instead of its higher dimensional version in the brane. This can be achieved through compactification of the extra dimension \cite{ArkaniHamed:1998rs,Antoniadis:1998ig,ArkaniHamed:1998kx}, whose size can be much larger than in the case of Kaluza-Klein models; thus this approach is called `Large Extra Dimensions'. This can also be achieved by embedding the brane with positive tension into an Anti-deSitter bulk \cite{Randall:1999ee,Randall:1999vf}; this approach is called `Warped Extra Dimensions'. In both cases, the extra dimension is finite and has a size of the order of the millimeter.
Dvali {\it et al.} proposed another solution: it suffices to add to the action of the brane an Einstein-Hilbert term computed with the intrinsic curvature of the brane. This term is necessary for the renormalizability of the quantum field theory in the brane interacting with the bulk gravity. Thus the standard Newton's law is recovered even for infinite size extra dimensions \cite{Dvali:2000hr}. The case of a 3-brane embedded in a 4+1 Minkowski spacetime is of special interest. We will explain this model in the lines below.
Consider the action
\begin{equation}
S_{(5)} = \frac 1{2\kappa^2}\int d^5 X \sqrt{-\bar{g}}\bar{R} + \int d^4 x \sqrt{-g}(\lambda_b + l_m) +\frac 1{2\mu^2}\int d^4 x \sqrt{-g}R \, ,
\end{equation}
where the first term corresponds to the Einstein-Hilbert action in 4+1 dimensions, with a five-dimensional bulk metric $\bar{g}_{AB}$ and curvature scalar $\bar{R}$; the third term corresponds to the Einstein-Hilbert action of the induced metric
$g_{ab}$ on the brane, with its corresponding intrinsic curvature scalar $R$; the second term corresponds to the matter action, which might include a brane tension (an effective cosmological constant) $\lambda_b$. The following quantities can be defined,
\begin{eqnarray}
\kappa^2 & = & 8\pi G_{(5)} = M^{-3}_{(5)} \, \\
\mu^2 & = & 8\pi G_{(4)} = M^{-2}_{(4)}\, .
\end{eqnarray}
In Ref.~\cite{Dvali:2000hr}, it was shown that the Newtonian potential is recovered in the brane for distances smaller than $r_0 = M_{(4)}^2/2 M_{(5)}^3$, with an effective Newton's constant $G_N = \frac 43 G_{(4)}$.
The cosmology of the model was derived by Deffayet in Ref.~\cite{Deffayet:2000uy}. Consider the brane-isotropic metric
\begin{equation}
ds^2 = n^2(\tau,y)d\tau^2 -a^2(\tau,y)\gamma_{ij}dx^i dx^j - b^2(\tau,y)dy^2 \, ,
\end{equation}
where $\{\tau,\mathbf{x}\}$ are coordinates on the brane ($y=0$) and $\gamma_{ij}$ is the maximally symmetric 3-metric (with spatial curvature $k$). The stress-energy tensor of the bulk is reduced to a possible bulk cosmological constant $\rho_B$, while the stress-energy tensor of the brane can be expressed as
\begin{equation}
T^A_B = \frac{\delta(y)}b diag(\rho_b,-p_b,-p_b,-p_b,0) \, .
\end{equation}
Assuming that the metric $\bar{g}_{AB}$ is symmetric under the change $y \rightarrow -y$, the corresponding Friedman equation reads
\begin{equation}
\pm \sqrt{H^2 - \frac{\kappa^2}6\rho_B - \frac C{a_0^4} + \frac k{a_0^2}} = r_0 \left( H^2 + \frac k{a_0^2} \right) - \frac {\kappa^2}6 \rho_b \, ,
\end{equation}
where $C$ is a constant of integration, the subscript $0$ means evaluation at $y = 0$, and $H = \frac{\dot a_0}{n_0 a_0}$, with dot denoting derivation with respect to the time variable $\tau$. The restriction of the fields to the brane also leads to the continuity equation for the brane, $\dot{\rho}_b + 3(\rho_b + p_b)\frac{\dot a_0}{a_0} = 0$ .
If we choose the $+$ sign and neglect $\rho_B$, $C$ and $k$ for simplicity, we arrive to the modified Friedman equation
\begin{equation}
\frac{8\pi G_{(4)}}{3} \rho_b = H^2 - H/r_0 \, ,
\end{equation}
which produces a late accelerated expansion tending asymptotically to an exponential expansion with $H = 1/r_0$, as in the case of $\Lambda$CDM. However, this modified Friedman equation is distinguishable from the one appearing in $\Lambda$CDM,
\begin{equation}
\frac{8\pi G_N} {3} \rho = H^2 - \Lambda/3 \, . \label{DGPFried}
\end{equation}
Current constraints on the expansion history from supernova luminosity distances, the CMB, and the Hubble constant exclude the simplest flat DGP model at about $3\sigma$. Even including spatial curvature, best-fit open DGP model is a marginally poorer fit to the data than flat $\Lambda$CDM. Moreover, its substantially different expansion history raises serious challenges for the model from structure formation \cite{Song:2006jk}.
Moreover, the DGP model is also plagued with anomalies \cite{Charmousis:2006pn,Gregory:2007xy}. First, it was discovered that ghosts are present in perturbation theory \cite{Nicolis:2004qq}. Second, if general bulk metrics are allowed, then the energy spectrum of the solutions is unbounded from below (negative mass blackholes are allowed) and therefore the theory is unstable \cite{Gregory:2007xy}. Furthermore there are solutions (positive mass black holes) in which `pressure singularities' are present: pressure diverges for finite energy density. Third, the DGP vacuum is unstable as it may tunnel to alternative vacua \cite{Gregory:2007xy}.
Thus, we will not consider DGP as a viable alternative to explain the current accelerated expansion. However, it is interesting how a simple modified Friedman equation such as \eqref{DGPFried} is able to produce an accelerated expansion in the absence of strange fluids.
\paragraph{Quintessence}
Aside from modifications of gravity such as the ones described above, it is also possible to explain the present acceleration of the universe as a consequence of the existence of a new form of matter, different form massless or massive particles, which presents at least at later cosmic times an effective negative pressure.
There are several ways of implementing this idea, but one of the most popular ones is to introduce phenomenologically a self-interacting scalar field. Extensions of GR including scalar fields have a long history \cite{Brans:1961sx}. This mechanism has also been used in QFT in order to explain some of the unnaturally small parameters of the SM \cite{Peccei:1977hh,Peccei:1977ur}. In analogy, it is possible to build dynamical scalar field actions which mimic the effect of a cosmological constant, and which could render its value small \cite{Wetterich:1987fk,Wetterich:1987fm}. However, this would imply a new force mediated by the scalar field, despite new interactions mediated by light scalar fields are highly constrained \cite{Turyshev:2008dr}. Thus there must be a fine-tuning in the action at least in the effective mass of the scalar or its coupling to observed matter.
Typically the action of one of such `quintessence' theories has the form
\begin{equation}
S = \int d^4 x \sqrt{-g} \left[ \frac 1{16 \pi G} R + \frac 12 g^{\mu\nu}\partial_\mu \varphi \partial_\nu \varphi - V(\varphi) +\mathcal{L}_m [g^{\mu\nu},\varphi] \right] \, ,
\end{equation}
where $\varphi$ is the quintessence scalar field\footnote{This scalar field is given a variety of names, depending on its role in the particular model in which it is introduced. It may be the dilaton (the Goldstone mode associated to the spontaneous breaking of dilatation symmetry), the cosmon (a particular case of the dilaton), the inflaton (if we consider a model of inflation),...}, $V$ is the effective self-interaction potential, which may be induced in a more fundamental theory, and $\mathcal{L}_m$ is the matter Lagrangian, which may include couplings between ordinary matter and the quintessence.
In the homogeneous approximation, in which the metric is FRW and the quintessence and matter fields are decoupled and depend only on the time variable, the equations of motion become
\begin{eqnarray}
\ddot{\varphi}+3 H \dot{\varphi}+V'(\varphi)&=0 \\
\frac{8\pi G_N}{3}(\rho+\frac{\dot{\varphi}^2}2 +V(\varphi))&=H^2\\
-3(1+\omega)H\rho&=\dot{\rho} \, .
\end{eqnarray}
Note that if the quintessence currently varies sufficiently slow with time ($\dot \varphi^2 \ll V(\varphi) $ ), then the quintessence is approximately frozen at its present value $\varphi_0$ and acts as an effective cosmological constant $\Lambda = 8 \pi G V(\varphi_0)$; for a sufficiently slow variation, the negative effective pressure $p_\varphi = \frac{\dot{\varphi}^2}2 - V(\varphi)$ required by the accelerated expansion is guarantied \cite{Wetterich:1987fm,Ratra:1987rm,Caldwell:1997ii}.
Some specific quintessence models deserve more attention. One of this classes is `tracker models' \cite{Zlatev:1998tr,Steinhardt:1999nw}, which try to explain the coincidence problem\footnote{Why is the value of the vacuum energy density of the same order that the matter energy density today?}. The idea underlying tracker models is the following: the energy density of quintessence component somehow tracks below the background density for most of the history of the universe, falling from a Planckian value in the very early universe with the expansion history. Only recently it grows and becomes the dominant energy density, driving the universe to its current accelerated expansion. The tracker field $\varphi$ rolls down the potential $V(\varphi)$ following an attractor-like solution. Examples of tracker potentials are exponentials $V(\varphi) = M^4 (e^{M_p/\varphi}-1)$ and inverse power laws $V(\varphi) = M^{4+n}\varphi^{-n} $.
For instance for inverse power law potentials $V(\varphi) = M^{4+n}\varphi^{-n} $ it is possible to arrive to a quintessence-dominated period when $\varphi\sim M_p$. The value of the scale $M$ must be fit by observations (typically $M \gtrsim 1 \, GeV$). For a given background component with equation of state $\omega_B \equiv \rho_B/ p_B$, the subdominant quintessence then adapts itself and adopts an equation of state
\begin{equation}
\omega_\varphi \approx \frac{n \omega_B /2 -1}{n/2 +1} \, .
\end{equation}
Thus for high $n$ if quintessence is subdominant it mimics the equation of state of radiation and its energy density decreases at the same rate. When matter takes over radiation, the pressure of quintessence becomes negative and later on begins to dominate. However for these potentials the condition $\dot \varphi^2 \ll V(\varphi) $ never holds and thus they are not able to reproduce equation of state of a cosmological constant. Thus the model is disfavored by observations. The coincidence problem is also just moved from the choice of the value of $\Lambda$ in $\Lambda$CDM to the choice of the scale $M$.
Other interesting type of quintessence models are k-essence models, in which the kinetic term of the scalar field is non-standard \cite{ArmendarizPicon:2000dh}. Such kinetic terms may arise in ST \cite{Gross:1986iv}. Typically the action of a k-essence model can be written as
\begin{equation}
S = \int d^4 x \sqrt{-g} \left[ \frac 1{16 \pi G} R + \frac 1{\varphi^2} f(X) +\mathcal{L}_m [g^{\mu\nu},\varphi] \right] \, ,
\end{equation}
where $f(X)$ is a certain function of $X\equiv g^{\mu\nu}\partial_\mu \varphi \partial_\nu \varphi$. The energy density and pressure of the k-essence read
\begin{eqnarray}
\rho_k & = & (2 X f'(X) - f(X) )/\varphi^2 \, ,\\
p_k & = & f(X)/\varphi^2 \, .
\end{eqnarray}
As shown in Ref.~\cite{ArmendarizPicon:2000dh}, k-essence models are able to reproduce the behavior of tracker models in the early universe, while showing a behavior more similar to a cosmological constant in the present universe. However, these models are quite {\it ad hoc} and do not fit the data as well as an effective cosmological constant does. Moreover they do not really solve the coincidence problem, as the transition from a matter dominated to a k-essence dominated universe occurs just as in $\Lambda$CDM.
Similar tracker behaviors arise much more naturally in vector-tensor models such as the one proposed in Ref.~\cite{Jimenez:2008au}. However, such a model presents an undesirable tension when compared to SNe, CMB and BAO observations \cite{Jimenez:2009py}.
\paragraph{Dark Energy}
When phenomenologically describing the exotic component of the universe either with a self-interacting scalar field, a k-essence or other types of fields, we are making an {\it a priori} assumption (the field content of our model) that is going to condition our following analysis. For this reason, it may be better from a phenomenological point of view to consider just the essential features of a new matter component, independently of a field theory representation. The essential ingredient of the matter content that must be characterized in GR is just the stress-energy tensor.
In the homogeneous approximation, the only two nonzero elements of the stress-energy tensor are the energy density $\rho_X$ and the pressure $p_X$. Thus, the effective description of a new fluid component amounts to the relation between this quantities, the equation of state $p_X = p_X (\rho_X)$. The most simple parametrization of this relation is a linear equation of state $p = \omega \rho$ \cite{Turner:1998ex}; \footnote{Despite it is a flagrant abuse of language, $\omega$ is often referred to as the equation of state. } for instance radiation has $\omega = 1/3$, massive matter has $\omega = 0$ (strictly speaking $\rho \gg p$) and a vacuum energy has $\omega = -1$. Naturally, this parametrization is also the most used in data analyses, such as \cite{Spergel:2003cb,Astier:2005qq,Spergel:2006hy,Dunkley:2008ie}. In contrast, an example of a nonlinear equation of state could be a gas of strings, the Chaplygin gas $p_X=-K/\rho_X$ \cite{Kamenshchik:2001cp,Alvarenga:2001nm}, which is able to mimic the dark matter at high densities but leads to an accelerated expansion at low densities.
The relation between the pressure and the energy density does not contain all the information about the fluid if one considers the evolution of inhomogeneous perturbations \cite{Hu:1998kj}. Then, it is necessary to consider the scalar part of the non-diagonal terms of the stress-energy tensor, i.e.: the momentum density $(\rho_X + p_X)\bm u_X = (\rho_X + p_X)\bm \nabla \theta_X $ and anisotropic stress $\Pi_{Xij} = \partial_i\partial_j \Pi_X$. A complete description of the fluid should involve an expression relating the pressure and anisotropic stress as a function (or a functional) of the energy density and the momentum density of the fluid at each point.
This description is of course too general to be predictive, and one usually restricts oneself to the use of perfect fluids ($\Pi = 0$) with a certain $p(\rho)$ equation of state. However, the effect of some specific anisotropic stress has been investigated in the literature (see for instance Refs.~\cite{Hu:1998kj,Capozziello:2005pa,Ichiki:2007vn}).
In light of the vast amount of concrete models, a phenomenological description of the equation of state has been proposed as a way of parameterizing the present acceleration of the universe, such as
\begin{equation}
p_X = \left[\omega_0 + \omega_a(1-a) + \mathcal{O}\left((1-a)^2\right)\right] \rho_X \, ,
\end{equation}
where the normalization $a_0 = 1$ is taken. Although we agree with the bottom-up spirit of such parameterizations, this choice is very anthropocentric and sheds little light on the physics of the dark component or gravitational physics behind the expansion. Is it possible to make a better choice?
\subsection{The Asymptotic Cosmological Model}
The first thing one should note when thinking about an extension of the cosmological standard model is that it is difficult (impossible?) to distinguish by means of gravitational physics between modified gravitational physics and new forms of matter. This is due to the Einstein equation. Any modification of the Einstein tensor can be moved to the rhs of the equation and become a modification to the stress-energy tensor and viceversa. This term will have a null covariant divergence as a consequence of being derived from a scalar action, either from the gravitational or the matter part of the Lagrangian. Thus, we will consider the departure to be encoded in the gravitational part of the Lagrangian without loss of (phenomenological) generality.
Secondly, in order to build a concrete model we have to make a number of assumptions based on phenomenological grounds. We have mentioned that it is experimentally well established that gravity should be a metric theory \cite{Bassett:2003vu,Sahni:2006pa,DeBernardis:2006ii} and thus we will assume our model to be a metric theory. We will also consider that the cosmological principle is a good approximation at sufficiently large scales, so that the FRW metric will be a good approximation to the metric of spacetime. The FRW metric depends just on the spatial curvature of spacetime, $k$. In $\Lambda$CDM, $k$ is severely constrained by the observations to be zero, so in our model we will assume that $k = 0$.
In GR with no spatial curvature, the evolution of the universe in the homogeneous approximation is obtained by an algebraic relation between the Hubble rate and the energy density. Furthermore if we assume that $\dot H < 0$, then there is a one-to-one correspondence between $t$ and $H$, and thus we can always find an algebraic relation between the energy density (of a fluid composed of SM particles and DM) and the Hubble parameter. Therefore we are going to define our model through the relation between $H$ and $\rho$.
Third, we will consider that the standard relation between $H$ and $\rho$, the first Friedman equation \eqref{11fried}, will be valid just in a window limited by the presence of an UV and IR scale. From the point of view of the phenomenologist, every theory must be valid in the regime in which it has been tested, but might fail in the regime in which it has not. GR has been well tested in a range from the millimeter scale to the galactic scale and QFT has been tested from the $eV$ to the $TeV$ scale, and thus they might fail beyond this scales. In our model, the effect of the new UV and IR scales will be responsible of the accelerated expansion in the very early universe and the present epoch, respectively.
We will take as a basic principle that the Hubble rate is constrained to a bounded interval between an upper scale $H_+$ and a lower scale $H_-$. Therefore in the infinite past and infinite future the universe will asymptotically tend to the deSitter spacetime. We have called this model the Asymptotic Cosmological Model.
\section{Neutrino Oscillations}
The development of QFT has been the most active sector of theoretical physics in the past century. Quantum Electrodynamics (QED) has provided the most precise predictions for experiments which have later confirmed those predictions figure by figure. The Standard Model of particle physics (SM), built up through a continuous interaction with experiment, has managed to provide a framework in which to understand the electromagnetic, weak and strong interactions, and to propose a mechanism to explain the apparition of mass (yet to be confirmed).
The model consists on a $SU(3)\times SU(2) \times U(1)$ gauge theory, in which the $SU(2)\times U(1)$ symmetry is ``broken'' through a Higgs mechanism by the vacuum expectation value (vev) of a scalar field. The spinors of the theory are divided in three families (or flavors) and can be either quarks ($SU(3)$-interacting) or leptons ($SU(3)$-noninteracting), and their masses are attained by their Yukawa interactions with the scalar field.
Let us denote the fields in the standard model by
\begin{equation}
\begin{array}{ccc}
\Psi_{L \alpha}=\left(\begin{array}{c}\nu_{\alpha} \\
l_{\alpha}\end{array}\right)_L, & L_\alpha = (l_\alpha)_R, & \\
Q_{L \alpha}=\left(\begin{array}{c}u_{\alpha} \\
d_{\alpha}\end{array}\right)_L, & U_\alpha = (u_\alpha)_R, & D_\alpha = (d_\alpha)_R, \\
\Phi=\left(\begin{array}{c}\varphi^+ \\
\varphi^0 \end{array}\right), &
\tilde{\Phi}=\left(\begin{array}{c}\varphi^{0*} \\
-\varphi^- \end{array}\right), & \\
G_\mu, & W_\mu, & B_\mu,
\end{array}
\end{equation}
where in the first line we have denoted the leptons, in the second line the quarks, in the third line the Higgs boson and its conjugate, and in the last line the gauge fields carrying the $SU(3)$, $SU(2)$ and $U(1)$ interactions respectively, whose respective field strengths will be denoted $G_{\mu\nu}$, $W_{\mu\nu}$ and $B_{\mu\nu}$. The index $\alpha=1,2,3$ labels the flavor, the coupling of the gauge fields are respectively $g_3$, $g$ and $g'$. The covariant derivatives will be denoted as $\nabla_\mu$, and $A \overleftrightarrow{\nabla_\mu} B \equiv A \nabla_\mu B - (\nabla_\mu A) B$. The Yukawa couplings will be denoted $Y_L$, $Y_U$, $Y_D$. Then, the Lagrangian of the SM is just the sum of the following terms:
\begin{eqnarray}
\mathcal{L}_{lepton} & = & \frac i2 \bar{\Psi}_{L \alpha}\gamma^\mu \overleftrightarrow{\nabla_\mu} \Psi_{L \alpha} + \frac i2 \bar{L}_\alpha\gamma^\mu \overleftrightarrow{\nabla_\mu} L_\alpha\, , \\
\mathcal{L}_{quark} & = & \frac i2 \bar{Q}_\alpha\gamma^\mu \overleftrightarrow{\nabla_\mu} Q_\alpha + \frac i2 \bar{U}_\alpha\gamma^\mu \overleftrightarrow{\nabla_\mu} U_\alpha + \frac i2 \bar{D}_\alpha\gamma^\mu \overleftrightarrow{\nabla_\mu} D_\alpha\, , \\
\mathcal{L}_{Yukawa} & = & -[(Y_L)_{\alpha\beta}\bar{\Psi}_{L \alpha}\Phi L_\beta + (Y_U)_{\alpha\beta}\bar{Q}_\alpha\tilde{\Phi}U_\beta +(Y_D)_{\alpha\beta}\bar{Q}_\alpha\Phi D_\beta ] + h.c. \, ,\\
\mathcal{L}_{Higgs} & = & (\nabla_\mu \Phi)^\dagger \nabla^\mu \Phi + \mu^2 \Phi^\dagger \Phi - \frac \lambda{3!}(\Phi^\dagger \Phi)^2 \, , \\
\mathcal{L}_{gauge} & = & -\frac 12 Tr(G_{\mu\nu}G^{\mu\nu}) -\frac 12 Tr(W_{\mu\nu}W^{\mu\nu}) -\frac 14 B_{\mu\nu}B^{\mu\nu}\, .
\end{eqnarray}
In the SM, the scalar field acquires a vev through the Higgs mechanism, and thus the Yukawa terms turn into mass terms for the quarks $u$, $d$ and the electrically charged leptons $l$, but not the neutrinos $\nu$. Therefore, the neutrinos are massless electrically neutral particles which only interact weakly (exchanging $W^\pm$ or $Z$ bosons, which are combinations of the $W$ and $B$ bosons). In a given experiment involving a charged lepton $l_\alpha$, the presence of a neutrino $\nu_\alpha$ as a final state can be detected as a violation of energy-momentum conservation.
These particles are generally detected in deep underground laboratories, in which the presence of more intensely interacting particles has been damped by the Earth crust\footnote{This is due to the fact that at energies below the electroweak scale neutrinos interact very weakly and otherwise the signal to noise ratio is small. However it is possible to detect neutrinos in accelerator experiments in which the source is very luminous, of high energy and controllable. However, detection of accelerator neutrinos at installations far away the source also need to be underground in order to improve the signal to noise ratio}. Their detection usually involves the exchange of a $W^\pm$ boson and the detection of the charged lepton $l_\alpha$ associated to the incoming neutrino $\nu_\alpha$.
Therefore, if a $\nu_\alpha$ is produced somewhere in a process involving $l_\alpha$, the massless and weakly interacting neutrino can propagate huge distances until being detected in another process involving the same $l_\alpha$. This was the idea underlying
the Homestake experiment headed by Raymond Davis Jr. and John N. Bahcall \cite{Bahcall:1976zz}. The experiment designed by Davis managed to measure the flux of $\nu_e$ coming from the sun as a result of the nuclear reactions happening at its core. The aim of the experiment was to confirm the theoretical predictions, derived by Bahcall, of the Standard Solar Model. However, the experiment measured around one third of the expected neutrinos! This was named the `solar neutrino problem'. At that time it was impossible to decide whether the problem was in neutrino physics or in the current understanding of the sun.
\subsection{A conventional explanation: neutrinos do have mass!}
It was first proposed by Pontecorvo \cite{Pontecorvo:1957qd} that neutrinos may change their nature (flavor) along their propagation through long distances in the same sense the neutral kaons do. He named this process `neutrino oscillations'. This explanation fitted the results of the Homestake experiment well but has profound implications \cite{Pontecorvo:1967fh}.
This phenomenon is shocking because in the context of the SM it cannot take place. When a neutrino is produced in a certain interaction within the SM of particle physics, it is produced by definition in a state which is a flavor eigenstate of definite momentum ($\nu_e$ if the interaction involves an electron, $\nu_\mu$ if the interaction involves a muon, or $\nu_\tau$ if the interaction involves a $\tau$). In the SM, the energy of these three states $E_\alpha$ is degenerate and depends only on their momentum, $E_\alpha = p$. Thus if a neutrino is produced in some flavor eigenstate, this is also a propagation eigenstate and it will be detected in the same flavor eigenstate, i.e.: its interaction with matter will produce the same charged lepton that was involved in its creation.
For neutrino oscillations to occur, two circumstances must concur. First, it is necessary that there exist different eigenstates of the effective Hamiltonian of the neutrinos with different energies $E_i$ for the same linear momentum $\mathbf{p}$. Let us call the neutrino propagation eigenstates $\{\nu_1,\nu_2,\nu_3\}$ with energies $\{ E_1(p),E_2(p),E_3(p)\}$. In a Lorentz invariant theory the only possibility of the neutrinos to have different energy eigenstates is that they have different masses $m_i$: $E_i = \sqrt{\mathbf{p}^2 + m_i^2}$. Therefore it is often stated that neutrino oscillations imply that the neutrinos are massive.
Second, it is necessary that these eigenstates do not coincide with the flavor eigenstates. The reason is that the eigenstates of the effective Hamiltonian (states of definite energy for a definite momentum) coincide with the eigenstates of the propagation, but those must not coincide with the eigenstates of definite flavor for a mixing to happen. This is quite natural in the context of QFT, and we already have an example in the flavor physics of quarks, parameterized by the Cabibbo-Kobayashi-Maskawa matrix connecting mass eigenstates of negatively charged quarks with their ``flavor'' eigenstates, defined as the eigenstates resulting of their charged current weak interaction with a positively charged quark. In analogy, a unitary matrix (called the Pontecorvo-Maki-Nakagawa-Sakata or PMNS matrix \cite{Maki:1962mu}) connects the mass eigenstates of neutrinos $\nu_i$ with the flavor eigenstates $\nu_\alpha$,
\begin{equation}
\nu_\alpha = U_{\alpha i} \nu_i \, ,
\end{equation}
which are defined as the states resulting of the charged current weak interaction of a charged lepton $l_\alpha$. This state is a linear combination of states with definite and different energies, which propagate in a definite way.
However, if there exists different states with different energies, it is no longer true that the neutrino is produced in a flavor eigenstate, due to energy-momentum conservation (unless the laws of energy-momentum conservation are only an approximation for sufficiently high energy-momentum transfers). If we reject this possibility, what happens is that in the production, a $\nu_1,\nu_2$ or $\nu_3$ can appear with different probabilities of sum one, which depends of the nature of the charged lepton involved in the production. Similarly, any of this neutrinos will produce, in the interaction with a detector, one of any of the three charged leptons, with different probabilities depending on the nature of the neutrino which has been detected.
It is assumed that this mass is an infrared (IR) scale, i.e.: it is much lower than the energy of the neutrino, and therefore the difference of energy between two mass eigenstates will be much lower than the total energy of any of these states, and the oscillations are assumed to take a significant place in neutrinos propagating through very long distances. Also, the effect of the mass in the total energy of the neutrino is often neglected, and it is commonly stated that neutrinos are produced and detected in states of definite flavor.
When we define a particle in a theory with a symmetry group, it is characterized by the eigenvalue of the invariant (under the symmetry group) operators that can be defined combining elements of the symmetry group. In the Lorentz invariant case, these operators are the mass and the spin of a particle. The particles also have definite charges under the gauge group. In the case of neutrinos, these states are not the flavor eigenstates, as they do not have a definite mass, but the propagation eigenstates $\{\nu_1,\nu_2,\nu_3\}$. Therefore, although flavor eigenstates can formally be defined in the context of a theory exhibiting neutrino oscillations it is an abuse of notation to call them particles.
Therefore it has been necessary to extend the SM of particle physics to include at least mass effects. The Standard Model of Neutrino oscillations assumes that there exist a new field, the sterile neutrino $\nu^R_i$, which is a singlet of the entire gauge group and which is necessary to include perturbatively renormalizable mass terms for the neutrinos in the Lagrangian (see Ref.~\cite{GonzalezGarcia:2007ib} for a review). The new terms that are added to the SM Lagrangian are:
\begin{equation}
-\mathcal{L}_{\nu_R} = \bar\nu^R_i \gamma^\mu \partial_\mu \nu^R_i
+ Y^D_{i\alpha} \tilde \Phi^\dagger \bar \nu^R_i \gamma^\mu \partial_\mu \Psi^L_\alpha
+ m^M_{ij} \bar \nu^R_i \nu^{R,c}_j + h.c.\, , \label{righthand}
\end{equation}
where $\nu^R$ is the sterile neutrino singlet, with flavor index $i$ running from $1$ to $n$; $Y^D$ is an adimensional matrix of coefficients in flavor space, which will give rise to Dirac mass terms after the Higgs acquires its expectation value; and $m^M$ is a Majorana mass matrix allowed by all the symmetries (the only of such terms in the whole lagrangian).
This terms lead to the so-called `see-saw mechanism'. Let us assume that sterile neutrinos are a signal of new physics at some UV scale $E_{NP}$ a few orders of magnitude below the Planck scale. Then, the $m^M$ matrix has components naturally of the same order of this scale, while the Dirac mass terms associated to neutrinos will be naturally of the order of the vev of the scalar field. Then, the diagonalization of the resulting total mass matrix lead to $3+n$ eigenstates with different masses. $n$ of them are mostly sterile, have a mass comparable to the UV scale, and are very weakly coupled to the charged leptons. The other three are mostly lefthanded, have a mass of order $v^2/E_{NP}$. \footnote{This idea fits well in the frame of GUT theories with symmetry group $SU(5)$, in which the GUT scale is $M_{GUT}\sim 10^{15}\, GeV$; however, the absence of proton decay \cite{Babu:1998wi} within the predicted window has ruled out the simplest of these theories.} All of these particles are Majorana fermions, i.e.: they are their own antiparticles, in contrast with Dirac fermions like the charged leptons or quarks.
In this case the SM with Majorana neutrinos is a good effective field theory (EFT) \cite{Burgess:2007pt} at energies below the almost sterile neutrino mass. In this case operators of dimension greater than $4$ cannot be omitted, but they are corrected with an appropriate negative power of the UV scale. Then the first corrections to the SM should come from terms of dimension $5$. If we compute all the $5$-order terms compatible with the Poincar\'e and gauge symmetry, the only new term which is added is
\begin{equation}
-\mathcal{L}_{NP} = \frac {Z_{ij}} {E_{NP}} \bar \Psi^L_i \tilde \Phi \tilde \Phi^T \Psi^{L,c}_j +h.c. \, .
\end{equation}
In fact, this result tells us that if we are expecting any new physics at a scale $E_{NP}$ we should expect that this new physics (not necessarily sterile neutrinos) provides the neutrinos with an effective mass term.
The proposal that massive neutrinos should suffice to explain the solar neutrino problem in terms of neutrino oscillations in the propagation of neutrinos from the Sun to the earth needed only to be contrasted with more experimental data. The so-called solar neutrino problem was further studied with similar detectors as Kamiokande, SAGE, GALLEX, Superkamiokande and the Sudbury Neutrino Observatory (SNO). SNO was the first experiment that proved that neutrino oscillations were behind the solar neutrino problem \cite{Ahmad:2002jz}, although neutrino oscillations were first detected by Superkamiokande \cite{Fukuda:1998mi}. In the latter case, the oscillation was produced in the neutrinos produced by muon decay in the atmosphere (atmospheric neutrinos). Neutrino oscillations produced in the propagation of neutrinos generated in artificial nuclear reactors \cite{Eguchi:2002dm} and in nuclear decays in the Earth's crust and mantle (geo-neutrinos) \cite{Araki:2005qa} have been also measured by KamLAND.
The aforementioned experiments have produced a vast amount of data about neutrino oscillations. Solar neutrino oscillation are affected also by the so-called MSW effect (neutrino oscillations due to its interaction with matter), so they do not provide a clean picture of vacuum neutrino oscillations. Atmospheric neutrino experiments have observed $\nu_\mu$ disappearance with $\Delta m^2 \sim 3 \times 10^{-3} eV^2$ \cite{Fukuda:1998mi}, while reactor antineutrino experiments have measured $\bar\nu_e$ disappearance at $\Delta m^2 \sim 8 \times 10^{-5} eV^2$ \cite{Eguchi:2002dm,Araki:2005qa}. The minimum joint description of solar, atmospheric and reactor neutrinos requires Dirac neutrinos with different masses
\begin{eqnarray}
m^2_1 -m^2_2 & = & 7.9 \pm 1.1 \times 10^{-5} eV^2 \\
\vert m^2_1 -m^2_3 \vert & = & 2.6 \pm 0.6 \times 10^{-3} eV^2
\end{eqnarray}
and the four independent degrees of freedom in the PMNS matrix\footnote{This is true for Dirac neutrinos. In the case of Majorana neutrinos there are six independent parameters in the PMNS matrix.}, which have been constrained \cite{GonzalezGarcia:2007ib}. In fact, the phenomenology involving massive Dirac neutrinos depends only in the squared absolute value of the elements of this matrix, which we will call populations and denote by
\begin{equation}
r_{i\alpha} = \vert U_{\alpha i} \vert^2 = \left(
\begin{array}{ccc}
0.69 \pm 0.03 & 0.31\pm 0.03 & < 0.01 \\ 0.16\pm 0.07 & 0.36 \pm 0.11 & 0.46 \pm 0.11 \\ 0.14 \pm 0.08 & 0.30 \pm 0.10 & 0.52 \pm 0.12
\end{array}
\right) \, .
\end{equation}
The population $r_{i\alpha}$ represent the portion of a state $\nu_\alpha$ in a state $\nu_i$. Although there exist nine populations in a three-mixing problem, the sum-one condition makes only four of them linearly independent. In the Standard Model of Neutrino Oscillations these parameters are constants, but in a more general oscillation problem they could be functions of the momentum $\mathbf{p}$.
It seems that massive neutrinos fit the current experimental data well. What is then the problem of this explanation?
\subsection{Doubts arising with massive neutrinos}
\paragraph{The LSND anomaly and other suggestive experimental results}
However, not all the ongoing experiments have their results well described by this simple picture. The Liquid Scintillator Neutrino Detector (LSND) reported evidence for $\bar \nu_\mu \rightarrow \bar \nu_e$ at $\Delta m^2 \sim 1 eV^2$ \cite{Athanassopoulos:1995iw,Athanassopoulos:1996jb}. This results have later been confirmed by the same group \cite{Athanassopoulos:1997pv,Aguilar:2001ty}, which has spent a huge effort in finding sources of experimental error. The final sample tests both $\bar \nu_\mu \rightarrow \bar \nu_e$ and $\nu_\mu \rightarrow \nu_e$ oscillations in a window of $20\, MeV < p < 200 \, MeV$.
In order to confirm or deny the results of the LSND collaboration, the KARMEN and MiniBOone collaborations have been running similar experiments. The MiniBOone collaboration has reported no evidence of $\nu_\mu \rightarrow \nu_e$ oscillations for neutrino beams with $p > 475 \, MeV$ \cite{AguilarArevalo:2007it,AguilarArevalo:2008rc} but has found a tiny excess of $\nu_e$ in beams with $200 \, MeV > p > 475 \, MeV$ \cite{AguilarArevalo:2008rc} (although this excess is incompatible with the oscillation mechanism proposed by LSND). The same group has also performed a $\bar \nu_\mu \rightarrow \bar \nu_e$ search at $200 \, Mev < p < 3000 \, MeV$, but no significant signal was found \cite{AguilarArevalo:2009xn}. These results have been qualified as `inconclusive' \cite{AguilarArevalo:2009xn}.
Furthermore, if we consider that the neutrino mass arises as a consequence of EFT with new physics in the UV (maybe close to $10^{15} \, GeV$), then the corresponding dimension $5$ operator giving mass to the neutrinos is a Majorana mass term which violates lepton number by two units. This should lead to new currently unobserved phenomena related to lepton-nonconservation such as neutrinoless double-beta decay ($0\nu\beta\beta$). These experiments are also interesting because, if $0\nu\beta\beta$ is observed, it will provide a measurement of the absolute value of the neutrino mass scale (in contrast to the relative values provided by neutrino oscillations). The search for these phenomena has been performed in the last years and has become quite controversial.
A subgroup of the Heidelberg-Moscow collaboration led by H. V. Klapdor-Kleingrothaus published a series of controversial papers claiming to have observed neutrinoless double-beta decay in $^{76}Ge$ with $6\sigma$ CL \cite{KlapdorKleingrothaus:2001ke,KlapdorKleingrothaus:2004wj,KlapdorKleingrothaus:2006ff}. This ``evidence'' has been hotly argued \cite{Aalseth:2002dt}, and is not accepted by most of the scientific community. Their claim, in which the expectation value of the mass matrix in the $\nu_e$ state is found to be $m_{\beta} \sim 0.2-0.5 \, eV$ has yet to be confirmed by other experiments.
Even if the `Klapdor's evidence' turns out to be wrong, $0\nu\beta\beta$ experiments would give us a hint on the absolute value of neutrino masses in the (near?) future if neutrinos are Majorana particles: it would be just a matter of being sensitive to lower masses.
However if these experiments persist without any positive signal of $0\nu\beta\beta$, then a solution to the absence of lepton number violating phenomena is to impose lepton number conservation by hand on the model (lepton number conservation is just an accidental symmetry of the standard model, but it stops to be a symmetry if we allow for dimension 5 operators or other neutrino mass terms), so that $m^M = 0$. If $n=3$, the sterile neutrinos can be seen as the right handed components of a Dirac spinor (therefore sterile Dirac neutrinos are called `righthanded neutrinos'). Then, processes violating lepton number would be forbidden. However, the reason of the smallness of the neutrino mass would be precluded, worsening the hierarchy problem of the SM.
Apart from $\beta\beta 0\nu$, beta decay experiments in general are also able to determine the absolute scale of neutrino masses. However, the expected scale of neutrino masses are well below the sensitivities of most experiments. Recently, a handful of experiments such as Troitsk Neutrino Mass Experiment \cite{Lobashev:1999tp} and the Mainz Neutrino Mass Experiment \cite{Weinheimer:1999tn} have tried to measure the endpoint region of the tritium beta decay with enormous precision. These experiments rely on the low energy exchange in the beta decay of light nuclei ($18.6 \, keV$ in the case of tritium). Both experiments have set $m_\beta < 2.3 \, eV$, but, surprisingly, the midpoint of the confidence regions lie in the negative mass squared sector. In particular, the Troitsk Experiment gives $m_\beta = -2.3 \pm 2.5(stat.) \pm 2.0(syst.) \, eV$, while the Mainz experiment gives $m_\beta = -0.6 \pm 2.2(stat.) \pm 2.1(syst.) \, eV$. In order to work out a more precise result, the Karlsruhe Tritium Neutrino Experiment (KATRIN) \cite{Osipowicz:2001sq}, which expects to achieve a sensitivity of $m_{\beta} > 0.2 \, eV$, has already begun to gather data.
\paragraph{Are massive neutrinos a too mild assumption?}
There have been attempts of conciling the data of LSND and MiniBOone with other neutrino experiments simply by adding more sterile neutrinos to the model. At this point one might wonder, if this is how one should proceed. Taking into account that sterile neutrinos are almost undetectable particles, for which we have little evidence, is it not just a new problem of adding ``epicycles'' to the SM? Or is it just that we are driving the SM beyond its range of validity and that we should look for radically new particle physics?
In the previous section about the Cosmological Constant Problem we have argued that QFT should be modified in the IR in order to be consistent with results such as the Bekenstein entropy or the smallness of the cosmological constant. In agreement with this point of view, there is a striking coincidence that should not be overlooked. If we consider the cosmological constant as the vacuum energy of some field theory, the value of this vacuum energy is roughly $\rho_v \sim (10^{-2.5} \, eV)^4$. In comparison, the scale of neutrino oscillations (in this case, the scale driving solar and reactor neutrino oscillations in the SM with massive neutrinos) is of the order of $\Delta m_\odot^2 \sim (10^{-2.5} \, eV)^2$.
We wonder if these numbers are unrelated or if they are derived from the same extension of local relativistic QFT. It might be that the same IR physics which causes the quantum corrections to the vacuum energy to be of the order of an IR scale $\lambda \sim 10^{-2.5} eV$ also causes a modification in the neutrino propagator of the same order. If this is true, we have argued that the corresponding departure from QFT should be nonlocal, and therefore it might even be that the description of neutrino oscillations requires some kind of non-locality.
\subsection{Ingredients for an alternative to mass-driven oscillations}
Alternative mechanisms to mass-driven neutrino oscillations must share a basic characteristic: they must include some new ingredient, which must have different eigenvalues for different neutrino species, and the corresponding eigenstates must not coincide with the flavor eigenstates, i.e. the coupling of these neutrino species to the charged leptons via $W$ exchange must be non-diagonal.
We will assume that the mechanism responsible of the phenomena we have listed above is some oscillation mechanism during the propagation of neutrinos. Other possible mechanisms would be nonstandard couplings to either the gravitational potential (violation of the equivalence principle), a torsion field or even matter.
\paragraph{Lorentz Invariance Violation}
Let us think about how one could go beyond the mechanism of mass terms driving the oscillation mechanism. We have stated that an oscillation mechanism can appear only if different neutrino families have different energies $E_i(p)$ for the same momentum $p$. Lorentz invariance restricts these different energies to be a consequence of different masses. Therefore if we want to go beyond the mechanism of mass-driven oscillations, this has to be done at the cost of abandoning Lorentz invariance. Given that the CPT theorem assures that CPT is a symmetry of any local field theory with Lorentz invariance, then CPT could also be violated. Although problematic from the point of view of self-consistency, CPT violation could be very handy in order to explain the discrepancies between LSND and MiniBOone.
It has been argued that Lorentz Invariance Violation (LIV) is an expectable phenomenon in a theory of quantum gravity. It arises naturally in LQG at scales at which the discreteness of spacetime cannot be neglected; Lorentz Invariance is restored as an approximated symmetry at low energies. In this case the departure from Lorentz symmetry is characterized by the Planck scale. The dispersion relation of particles reflects the scale of the spacetime fabric, and takes the form \cite{AmelinoCamelia:2008qg}
\begin{equation}
E^2 = \mathbf{p}^2 + m^2 + \eta \mathbf{p}^2 \left( \frac{E}{E_p} \right)^n \, , \label{dispdef}
\end{equation}
where $m$ is the mass of the tested particle, $\eta$ is a phenomenological parameter, presumably of order $1$, and $n>0$ is assumed to be an integer. This is the kind of models that are being contrasted with experimental tests of Lorentz invariance. They are explicitly rotation invariant, as it is boost invariance what is usually assumed to be broken by quantum gravity. The reason is the presence of a new very high energy scale (typically the Planck scale $E_p$, although this might vary in particular models) which is either an invariant of the theory or a quantity in whose value every observer should agree.
Violations of Lorentz Invariance may also arise in ST as a result of Spontaneous Symmetry Breaking (SSB). In order for Lorentz SSB to occur, it would be required that some Lorentz tensor would acquire an effective negative mass term. In relativistic perturbatively renormalizable quantum field theories, it is impossible to arrive to an effective negative mass term for any vector field.
However it has been proven that this is not the case in string theory \cite{Kostelecky:1988zi}. For instance, in the open bosonic string, it is possible that a vector field $B_\mu$ acquires a nonzero vacuum expectation value due to an effective potential $V$ in the Lagrangian $\mathcal{L}_B$ of the type
\begin{equation}
\mathcal{L}_B = - \frac 14 B_{\mu\nu} B^{\mu\nu} - V(B_\mu B^\nu \pm b^2) - B_\mu J^\mu \, , \label{bblag1}
\end{equation}
where $B_{\mu\nu} = \partial_\mu B_\nu -\partial_\nu B_\mu$ is the field strength, $V(x) = \frac 12 \lambda x^2 $ is the potential and $J^\mu$ is the matter current. Depending on the sign appearing in the argument of the potential, the vacuum expectation value of the Bumblebee will be timelike or spacelike (the case of a lightlike expectation value is ill-defined). It is usual to take the vacuum expectation value of the vector field, $b^\mu$, as independent of spacetime coordinates so that the broken symmetries are global Lorentz symmetries, although solitonic solutions are also plausible.
In the same sense in which Lorentz symmetry can be spontaneously broken in string theories, it is also possible to find models in which CPT symmetry is also spontaneously broken \cite{Kostelecky:1991ak,Kostelecky:1995qk}. Taking to account this spontaneous symmetry breaking, it is reasonable to consider that the low energy limit of such a quantum theory of gravity admits a description as an ordinary QFT including departures from Lorentz invariance and can be studied with the techniques of EFT.
In the Standard Model Extension (SME) \cite{Colladay:1996iz,Colladay:1998fq}, the starting point is the minimal SM of particle physics and its associated scale, the electroweak scale $m_W$. New Lorentz- and possibly CPT-violating terms\footnote{The CPT theorem ensures that a local Lorentz invariant theory preserves CPT symmetry \cite{Streater:1989vi}. Thus, a local theory violating CPT must violate Lorentz invariance. The converse is not true; Lorentz violating but CPT conserving theories exist.} arising from a spontaneous breaking of Lorentz invariance in a more fundamental theory are assumed to be suppressed at least by the ratio $\frac{m_W}{E_p}\approx 10^{-17}$. The new terms also respect the gauge invariance of the SM. A minimal standard model extension can be defined to include just power-counting renormalizable terms (as in Ref.~\cite{Colladay:1998fq}), while terms of dimensions greater than four can be included using the standard EFT technics \cite{Bluhm:2005uj}.
Both in the case of an explicit or a spontaneous Lorentz Symmetry Breaking, the symmetry group is reduced and thus a preferred frame arises. These frame is often identified with the frame in which the CMB is almost isotropic \cite{Coleman:1997xq}. However our laboratory frame does not coincide with the CMB rest frame; an obvious consequence is the need of removing the dipole in the observed CMB spectrum, as it only expresses the boost that relates both frames \cite{Jaffe:2000tx}. This boost between the preferred frame (in which the laws of physics are assumed to be isotropic) and the laboratory frame introduces a new directional effect, and therefore an anisotropy in the laws of physics in the laboratory frame. Thus one needs to consider the anisotropic couplings of the SME in order to build a realistic model.
Fortunately, the rotational and translational motions of the Earth are essentially non relativistic in the approximately inertial Sun-centered frame relevant for neutrino experiments, so in practice any direction-dependent effects are suppressed to parts in a thousand or more \cite{Bluhm:2003un}. Anisotropic effects would also be excluded if instead of a Lorentz violation a new symmetry principle (DSR) would replace Lorentz invariance \cite{AmelinoCamelia:2000mn,AmelinoCamelia:2002fw,Magueijo:2001cr,Magueijo:2002am}.
The possibility of neutrino oscillations driven or corrected by LIV has been considered in several works, which we will review below.
\paragraph{Agreement with observational and experimental data}
As it should always be in any field of science, it will be our priority to find alternatives to mass driven oscillations which are fully consistent with the experimental data. Many models of LIV driving neutrino oscillations have been proposed, although few of them have managed to explain the energy behavior that has been experimentally observed.
In order to fit the existing experimental data (aside from the controversial results of LSND and $0\nu\beta\beta$), it is sufficient for these models to meet one property: that the new effect be a correction proportional to the square of an infrared scale $\lambda \sim 1 - 10^{-3} eV$ at energies $E\gg \lambda$.\footnote{Surprisingly, the tandem model presented in Ref.~\cite{Katori:2006mz} manages to fit the experimental data without resorting to this mechanism.} This is due to the fact that all the neutrino experiments that have been carried out involve high energy neutrinos (compared to the scale $\lambda$).
It has not been easy for researchers of LIV in neutrinos to find this energy dependence in their models. In one of the works presented in this PhD thesis, we will show that it is possible to find such a behavior if we invoke nonlocal physics in the IR, as we suggested as a solution to the Cosmological Constant Problem.
Also cosmology will impose important bounds on any alternative to neutrino masses, as well as it imposes a competitive bound on the sum of the neutrino masses in the case of massive neutrinos \cite{Lesgourgues:2006nd,Seljak:2006bg}.
In order to have some information about the low energy behavior of neutrinos, it would be necessary to get information directly or indirectly about neutrinos of energies close to this scales, such as tritium decay close to the endpoint of the spectrum or detection of the Cosmic Neutrino Background (C$\nu$B), which has not yet been observed. These two experiments are completely unfeasible in the near future. However the effect of the C$\nu$B in the cosmic matter power spectrum of large scale structure might be observable.
In conclusion, although neutrino masses are a simple extension of the SM perfectly able to describe the current experimental results on neutrino oscillations (aside of the poorly understood LSND and MiniBOone), it is not the only possibility. An alternative mechanism to neutrino masses requires LIV, but it must also be consistent with the experimental and observational data. A sufficient condition is that the deformation of the dispersion relation of neutrinos involves the square of some IR energy scale.
\subsection{Review on alternatives to neutrino masses}
In this subsection I want to review some of the attempts to offer an alternative description of neutrino oscillations with massless neutrinos. As mentioned this requires to go beyond Special Relativity (SR). Neutrino oscillations have been suggested to be a good probe of physics beyond SR in Ref.~\cite{Coleman:1997xq}, maybe arising in the low energy limit of a theory of quantum gravity. A folkloric example is a quantum black hole swallowing a neutrino of a given flavor and evaporating into a neutrino of a different one \cite{Hawking:1982dj}.
In particular, it has been stated that the addition of Lorentz violating but CPT conserving terms to the SM Lagrangian would induce differences in the maximum attainable velocities of the neutrinos which may depend on their nature and helicity. It was also first stated that massless neutrinos could oscillate in vacua in the presence of LIV. More complicated models have been built in order to fit the vast amount of experimental data that has been collected in the past decade.
\paragraph{Velocity-driven oscillations}
Coleman and Glashow introduced a simple model of oscillation between electron and muon neutrino \cite{Coleman:1997xq},
\begin{eqnarray}
\nu_\mu & = & \cos \theta \nu_1 + \sin \theta \nu_2\, , \\
\nu_e & = & \cos \theta \nu_2 - \sin \theta \nu_1\, ,
\end{eqnarray}
where $\nu_1$ and $\nu_2$ are propagation eigenstates with different velocities and dispersion relations
\begin{equation}
E_1 = p (1 + v_1) \, , \, E_2 = p (1 + v_2) \, ,
\end{equation}
and $\theta$ is a phase characterizing the mixing. The differences between the speed of the neutrino and the speed of light $v_i$, are assumed to be $\vert v_i \vert \ll 1$. The populations of the flavor eigenstates in the propagation eigenstates are
\begin{equation}
r_{1\mu} = r_{2 e} = \cos^2\theta\, , \, r_{1 e} = r_{2 \mu} = \sin^2 \theta \, ,
\end{equation}
and the probability of an electron neutrino of traversing a distance $R$ without changing its flavor is
\begin{equation}
P_{e\rightarrow e}(R) = r_{1 e} \sin^2 \left( (v_1-v_2) p R /2 \right) \, , \label{velosc}
\end{equation}
while the same probability, if the oscillation mechanism is driven by mass differences, is
\begin{equation}
P_{e\rightarrow e}(R) = r^m_{1 e} \sin^2 \left( (m^2_1-m^2_2) R /4 p \right) \, . \label{masosc}
\end{equation}
It is also possible that both masses and difference between velocities are behind the neutrino oscillations. In this case, the diagonalization problem would be more complicated and momentum dependent. The effective Hamiltonian would be of the type
\begin{equation}
\mathcal{H}_{\alpha\beta} = p \left(\delta_{\alpha\beta} + V_{\alpha\beta} \right) + \frac 1{2p} M^2_{\alpha\beta}\,
\end{equation}
in the base of flavor eigenstates. The diagonalization of the effective Hamiltonian should be performed for any given momentum $p$ of the neutrino, resulting in a mixing of the type
\begin{equation}
P_{e\rightarrow e}(R) = r_{1 e}(p) \sin^2 \left( (E_1(p)-E_2(p)) R /2 \right) \, .
\end{equation}
The limits of low and high momentum of this expression should coincide with the expressions for the mass driven \eqref{masosc} and velocity driven oscillations \eqref{velosc}, respectively.
The clear distinction between the behavior with momentum $p$ between \eqref{velosc} and \eqref{masosc} has enabled the experimentalists to clearly favor mass driven oscillations over velocity driven oscillations \cite{GonzalezGarcia:2007ib}, at least if the velocity of the neutrinos is independent of their frequency. In general, a deviation from the energy behavior of \eqref{masosc} would be a smoking-gun signal of LIV effects \cite{Brustein:2001ik}.
This has therefore led most of the studies of Lorentz violation in neutrino physics to consider Lorentz violation as a subleading mechanism that can be tested within the framework of mass driven oscillations. In some of these studies \cite{Coleman:1998ti,Brustein:2001ik,GonzalezGarcia:2004wg,Morgan:2008zza,Diaz:2009qk} the LIV has been studied at the level of a phenomenological deformation of the dispersion relation of the type of \eqref{dispdef}.
\paragraph{The Standard Model Extension and Neutrino Oscillations}
As we mentioned before, the Minimal SME includes all the Lorentz non-invariant terms with or without CPT violation which are perturbatively renormalizable and respect the gauge symmetries of the SM. The new terms can be divided into CPT preserving (CPT-even) or CPT violating (CPT-odd). The new terms are supposed to be suppressed with respect to the Lorentz conserving terms of the same kind by powers of the electroweak scale over the Planck scale. In this sense, some of the dimensionful terms can be interpreted as new IR scales.
As an example, we will give the explicit form of the terms in the fermion sector. In the lepton sector, the new terms are:
\begin{eqnarray}
\mathcal{L}_{lepton}^{CPT-even} & = & \frac i2 (c_L)_{\mu\nu\alpha\beta} \bar{\Psi}_{L \alpha}\gamma^\nu \overleftrightarrow{\nabla_\mu} \Psi_{L \beta} + \frac i2 (c_R)_{\mu\nu\alpha\beta} \bar{L}_\alpha\gamma^\nu \overleftrightarrow{\nabla_\mu} L_\beta\, , \\
\mathcal{L}_{lepton}^{CPT-odd} & = & -(a_L)_{\mu\alpha\beta} \bar{\Psi}_{L \alpha}\gamma^\mu \Psi_{L \beta} -(a_R)_{\mu\alpha\beta} \bar{L}_\alpha\gamma^\mu L_\beta\, .
\end{eqnarray}
All the coefficients $c$ and $a$ are understood to be hermitian matrices in flavor space as well as Lorentz tensors. These new terms are of dimension $4$ and $3$, respectively. Therefore the coefficients $c_L$ and $c_R$ are dimensionless and may have symmetric and antisymmetric parts but can be assumed to be traceless. The coefficients $a_L$ and $a_R$ have dimensions of energy. If we consider that these vectors are timelike, in the preferred frame of reference their only nonzero component will be $(a_L)_0$ and $(a_R)_0$, and will play the role of an IR scale.
In the quark sector the new terms are analogous to those appearing in the lepton sector, with coefficient matrices $c_Q$, $c_U$, $c_D$, ... Other Lorentz violating terms appear in the Yukawa, Higgs and gauge sectors of the model.
The Minimal SME has been constrained by a huge amount of experiments, among which we could stress high precision tests of QED, neutrino and neutral meson oscillations (for a living review on tests of Lorentz Invariance see Ref.~\cite{Mattingly:2005re}, for constraints on the SME parameters see Ref.~\cite{Kostelecky:2008ts}). Despite the long list of experiments, a vast sector of the Minimal SME remains yet unexplored. Meson oscillation experiments have set the CPT-violating coefficients of the quark sector $a_\mu < 10^{-20} GeV$. The most stringent bounds come from precision tests in QED. Bounds on a spacelike vector vev $b_\mu$ are of order $10^{-25} GeV$ to $10^{-31} GeV$ depending of the experiment and the kind of particle.
The lepton sector has also been constrained with the data provided by neutrino oscillations. The possibility of LIV-induced neutrino oscillations arises in the model without sterile neutrinos, and the SME has also been extended to include the righthanded Dirac neutrinos by Kostelleck\`{y} and Mewes \cite{Kostelecky:2003cr}. The complete SME with righthanded neutrinos has many parameters and thus several simpler models have been built within this framework in order to gain predictivity. These models include neutrinos with \cite{Kostelecky:2003cr,Mewes:2004ge,Katori:2006mz,Diaz:2009qk} or without \cite{Kostelecky:2003cr,Kostelecky:2003xn,Kostelecky:2004hg,Mewes:2004ge} mass terms, in which the Lorentz-violating coefficients of the SME are treated as subleading terms or as driving the oscillation phenomenon, respectively.
Aside from the aforementioned anomalous energy dependence of the oscillation phenomena, more new effects are expected if the SME with righthanded neutrinos is taken into account \cite{Kostelecky:2003cr}. Anisotropic terms induce periodic variations in the oscillations: those induced by the Earth's rotation and sidereal translation. These effects can be averaged away by integrating the data over time, but anisotropic effects would not be erased, as asymmetries between the east-west and north-south directions or between the ecliptic plane and the ecliptic north-south would appear. Other possible effects arising from the SME would be discrete symmetry violations, like effects violating lepton number (as neutrino-antineutrino oscillations), CPT (deviations from the relationship $P_{\nu_\alpha\rightarrow \nu_\beta} = P_{\bar\nu_\beta\rightarrow \bar\nu_\alpha}$ or both. CPT violation is extremely interesting as it could provide an explanation to the discrepancy between the LSND \cite{Aguilar:2001ty} and MiniBoONE \cite{AguilarArevalo:2008rc} experiments.
The most general modified Dirac equation of the neutrino fields within the SME with righthanded neutrinos would be \cite{Kostelecky:2003cr}
\begin{equation}
\left(i\Gamma^\nu_{\bar\alpha\bar\beta}\partial_\nu - M_{\bar\alpha\bar\beta}\right)\nu_{\bar\beta} = 0 \label{SMEDiraceq}\, ,
\end{equation}
where $\bar\alpha$ and $\bar\beta$ run for the three possible neutrino flavors and their charge conjugates and
\begin{eqnarray}
\Gamma^\nu_{\bar\alpha\bar\beta} & : = & \gamma^\nu \delta_{\bar\alpha\bar\beta} + c^{\mu\nu}_{\bar\alpha\bar\beta}\gamma_\mu + d^{\mu\nu}_{\bar\alpha\bar\beta} \gamma_5\gamma_\mu + e^\nu_{\bar\alpha\bar\beta} + i f^\nu_{\bar\alpha\bar\beta}\gamma_5 +\frac 1 2 g^{\lambda\mu\nu}_{\bar\alpha\bar\beta}\sigma_{\lambda\mu} \, , \\
M_{\bar\alpha\bar\beta} & := & m_{\bar\alpha\bar\beta}+ i m_{5\bar\alpha\bar\beta} \gamma_5 + a^\mu_{\bar\alpha\bar\beta}\gamma_\mu + b^\mu_{\bar\alpha\bar\beta}\gamma_5\gamma_\mu + \frac 12 H^{\mu\nu}_{\bar\alpha\bar\beta}\sigma_{\mu\nu} \, .
\end{eqnarray}
The coefficients $m$ and $m_5$ are a linear combination of the righthanded and lefthanded neutrino mass matrices $m_R$ and $m_L$, and thus Lorentz invariant. The coefficients $c$, $d$, $H$ are CPT conserving but Lorentz violating, while $a$, $b$, $e$, $f$, $g$ are CPT violating. Hermiticity implies that all the coefficients are Hermitian in flavor space.
The effective Hamiltonian for the neutrinos in the Minimal SME reads
\begin{equation}
(\mathcal{H}_{eff})_{\bar\alpha\bar\beta} = \mathcal{H}_0 -\frac 12 \left\{\gamma^0 \delta\Gamma^0,\mathcal{H}_0\right\} -\gamma^0(i\delta\Gamma^j\partial_j-\delta M) \, ,
\end{equation}
where $\mathcal{H}_0 = -\gamma^0(i\gamma^j\partial_j - M_0)$ is the general Lorentz conserving Hamiltonian, $M_0$ is the Lorentz conserving part of $M$ and $\delta\Gamma$, $\delta M$ are the Lorentz violating parts of $\Gamma$, $M$. The general treatment is possible but rather horrible, due to the vast amount of parameters and the fact that the coefficient matrices do not necessarily commute.
Therefore the model has to be simplified in order to gain some predictive power. It is standard to get rid off righthanded neutrinos and present a model for oscillations among active (lefthanded) neutrinos and their antiparticles. Further simplification can be achieved if oscillations between neutrinos and antineutrinos $\nu \leftrightarrow \bar \nu$ are highly suppressed. Then, the effective Hamiltonian for the lefthanded ultrarelativistic neutrinos reads
\begin{equation}
\left(h_{eff} \right)_{\alpha\beta} = p \delta_{\alpha\beta} + \frac 1{2p}\left(m_L^2\right)_{\alpha\beta} + \frac 1p \left[ \left(a_L\right)^\mu p_\mu - \left(c_L\right)^{\mu\nu}p_\mu p_\nu\right]_{\alpha\beta}
\end{equation}
and its complex conjugate for antineutrinos, with the opposite sign for the CPT-odd coefficient $a_L$. In the previous formula $m_L$ is the combination of $m$ and $m_5$ relevant for lefthanded neutrinos (see \cite{Kostelecky:2003cr} for the details), $c_L = c + d$ and $a_L = a + b$. Even simpler models can be built in order to explain the oscillations data.
`Fried-chicken' models \cite{Kostelecky:2003cr} are those in which spatial isotropy is imposed,
\begin{equation}
\left(h_{eff}^{FC} \right)_{\alpha\beta} = p \delta_{\alpha\beta} + \left[ \frac 1{2p}\left(m_L^2\right) + \left(a_L\right)^0 - \frac 43 \left(c_L\right)^{00} p \right]_{\alpha\beta}
\end{equation}
so that the study of modified dispersion relations is not complicated by direction dependent effects. Most models of Lorentz-violating neutrino oscillations fall in this class. However, isotropy in the laboratory frame is difficult to motivate in a Lorentz violating theory.
Anisotropic models, in contrast, need a well defined spacetime reference frame, which is usually taken to be the Sun-centered celestial equatorial frame of coordinates $\left\{X,Y,Z,T\right\}$, in order to report the values of its coefficients. An interesting model of this class is the so called `Bicycle Model' presented in Ref.~\cite{Kostelecky:2003xn}, in which most of the oscillations data can be explained only with two free parameters and no mass term.
It must be remembered that the $a_L$ coefficient has dimensions of mass and represents some infrared scale, presumably of order $\frac {m_W^2}{E_p}$. If the dimensionless coefficient $c_L$ is of order $\frac {m_W}{E_p}$, then for momenta well below the electroweak scale (and this happens in every single neutrino oscillations experiment), the dominant source of the oscillation in a massless model ($m_L = 0$) is the $a_L$ term. The argument of the oscillation's sine function will be $\sim \Delta a_L L$ instead of the standard $\Delta m^2 L / 2E$. This behavior is disfavored by the experimental data.
However, this na\"{\i}ve energy behavior was surmounted by the Bicycle Model. In this model, the coefficients $c_{Lee}^{TT} = \frac 32 \check{c}$ and $a_{Le\mu}^Z = a_{Le\tau}^Z = \frac{\check{a}}{\sqrt 2}$ are the only nonzero coefficients. The resulting neutrino oscillation probabilities are
\begin{eqnarray}
P_{\nu_e \rightarrow \nu_e} & = & 1 - 4 \sin^2\theta \cos^2 \theta \sin^2\left(\Delta_{31}L/2\right) \, , \\
P_{\nu_e \rightarrow \nu_\mu} & = & P_{\nu_e \rightarrow \nu_\tau} = 2 \sin^2\theta \cos^2 \theta \sin^2\left(\Delta_{31}L/2\right) \, , \\
P_{\nu_\mu \rightarrow \nu_\mu} & = & P_{\nu_\tau \rightarrow \nu_\tau} = 1 - \sin^2\theta \sin^2\left(\Delta_{21}L/2\right) \nonumber \\
& & -\sin^2\theta \cos^2 \theta \sin^2\left(\Delta_{31}L/2\right)- \cos^2 \theta \sin^2\left(\Delta_{32}L/2\right)\, , \\
P_{\nu_\mu \rightarrow \nu_\tau} & = & \sin^2\theta \sin^2\left(\Delta_{21}L/2\right) \nonumber \\
& & -\sin^2\theta \cos^2 \theta \sin^2\left(\Delta_{31}L/2\right) + \cos^2 \theta \sin^2\left(\Delta_{32}L/2\right)\, ,
\end{eqnarray}
where
\begin{eqnarray}
\Delta_{21} & = & \sqrt{(\check{c} p)^2 + (\check{a} \cos\Theta)^2} + \check{c} p \, ,\\
\Delta_{31} & = & \sqrt{(\check{c} p)^2 + (\check{a} \cos\Theta)^2} \, , \\
\Delta_{32} & = & \sqrt{(\check{c} p)^2 + (\check{a} \cos\Theta)^2} - \check{c} p \, ,\\
sin^2 \theta & = & \frac 12 \left[ 1 - \check{c}p/\sqrt{(\check{c} p)^2 + (\check{a} \cos\Theta)^2}\right] \, ,
\end{eqnarray}
and $\Theta$ is the angle between the celestial north pole and the direction of propagation. The same probabilities are valid for antineutrinos (which are obtained by changing $\check{a}\rightarrow - \check{a}$).
In this model, electron neutrino propagating at low energies compared with the scale $\check{a}/\check{c}$ oscillate into muon and tau neutrinos. However, at greater energies electron neutrinos do not oscillate, but there is a maximal mixing between muon and tau neutrinos, which oscillate with a wavelength $\sim \frac{\check{c}E}{\check{a}^2\cos^2 \Theta}$. Thus, high energy oscillations mimic mass driven oscillations with a direction-dependent effective mass squared difference $\Delta m^2_\Theta = \check{a}^2 \cos^2\theta/\check{c}$. The model failed to explain the LSND anomaly, but did quite well in fitting the then-current neutrino data with $\check{c} = 10^{-19}$ and $\check{a} = 10^{-20} GeV$ . More recent tests using crossed data have ruled out the bicycle model and some simple extensions \cite{Barger:2007dc}.
The big difference between the predictions of this model and the standard mass driven oscillations mechanism is the directional dependence which induce small semiannual variations in the neutrino flux that could be observed and `compass' asymmetries. Directional variations have been tightly constrained by MINOS \cite{:2008ij}. Nevertheless, the model can be modified to have no directional dependence by changing $a_L^Z$ to $a_L^T$ with a small impact on the fits \cite{Kostelecky:2003xn}.
An extension of the Bicycle Model, the `Tandem Model' \cite{Katori:2006mz} was proposed to fit also the anomalous data of LSND. The model includes similar parameters to the ones in the Bicycle Model and a single mass term $m_{\tau\tau}$. The effective Hamiltonian can be written in the flavor basis as
\begin{equation}
h^\nu_{TM} = \left(
\begin{array}{ccc}
(1 +\check{c}) p & \check{a} & \check{a} \\ \check{a} & p & \check{a} \\ \check{a} & \check{a} & p + \check{m}^2/2p
\end{array}
\right) \, ,
\end{equation}
where $\check{c} = (c_L)^{TT}_{ee}$ and $\check{a}=(a_L)^T_{e\mu} = (a_L)^T_{e\tau} = (a_L)^T_{\mu\tau}$. The corresponding effective Hamiltonian for antineutrinos is obtained by changing $\check{a} \rightarrow -\check{a}$.
The Tandem Model provides a good description of existing neutrino-oscillations data with
\begin{eqnarray}
\frac 12 \check{m}^2 & = & 5.2 \times 10^{-3} eV^2 \, , \nonumber \\
\check{a} & = & -2.4 \times 10^{-19} GeV \, , \nonumber \\
\check{c} & = & 3.4 \times 10^{-17} \, .
\end{eqnarray}
Solar neutrino oscillations can be fit without the need of any MSW effect \cite{Wolfenstein:1977ue,Mikheev:1986gs}. Atmospheric neutrino oscillations can be fit as in the Bicycle Model, as it reduces to it at high energies. Long-baseline experiments like KamLAND \cite{Araki:2004mb} lie in the low energy regime in which the $\check{c}$ term can be neglected. In this regime, the CPT violation is manifest as a difference in the $\nu_e$ and $\bar\nu_e$ survival probabilities which can be tested. Short-baseline experiments like LSND \cite{Aguilar:2001ty} and MiniBOone \cite{AguilarArevalo:2008rc} can be fit due to this CPT violation. The effect lies on the edge of experimental sensitivity due to the short paths that are being tested. In fact, the tandem model predicted a low energy excess in the MiniBOone experiment prior to its discovery, although the observed excess is qualitatively greater. This discovery remarks the interest of the search for hybrid mass and LIV-driven oscillation mechanisms. Nevertheless, the observations of LSND and MiniBOone must be confirmed by other ongoing or planned short-baseline experiments.
Finally, the case of mass-driven neutrino oscillation modified by subleading LIV-effects has also been studied in the framework of the SME. The constraints of the parameters of the model have been gathered in Ref.~\cite{Diaz:2009qk}.
\paragraph{Neutrino Oscillations arising from a noncanonical quantization procedure?}
Motivated by the apparition of noncommutative spacetime \cite{connes} in ST \cite{Connes:1997cr,Seiberg:1999vs,Yoneya:2000bt}, it has been proposed that the most general extension of Noncommutative Quantum Mechanics (NCQM) \cite{Mezincescu:2000zq,Duval:2000xr,Gamboa:2000yq} to field theory is to make the fields, which are the degrees of freedom in a field theory, noncommutative. The field noncommutativity is introduced in a scalar quantum field theory at the level of the commutation relations that define the way from the classical theory to the QFT.
In this scalar quantum 'noncommutative' or noncanonical field theory (QNCFT) \cite{Carmona:2002iv,Carmona:2003kh}, the (equal-time) commutation relations read
\begin{eqnarray}
\left[\Phi(\mathbf{x}),\Phi^\dagger(\mathbf{x}')\right] & = & \theta \delta^{(3)}(\mathbf{x}-\mathbf{x}') \nonumber\, , \\
\left[\Phi(\mathbf{x}),\Pi^\dagger(\mathbf{x}')\right] & = & i \delta^{(3)}(\mathbf{x}-\mathbf{x}') \, ,\\
\left[\Pi(\mathbf{x}),\Pi^\dagger(\mathbf{x}')\right] & = & B \delta^{(3)}(\mathbf{x}-\mathbf{x}') \nonumber \, ,
\end{eqnarray}
while the free Hamiltonian remains unmodified.\footnote{This commutation relations have also been extended in Ref.~\cite{Balachandran:2007ua}, in which the Dirac deltas have been turned into Gaussian distributions. Note that this change implies that the theory becomes nonlocal.} The resulting free theory can be fully solved and proves it to be a theory of particles $a$ and antiparticles $b$ with Lorentz and CPT violating dispersion relations
\begin{eqnarray}
\frac{E_a(\mathbf{p})}{\omega_{\bm{p}}}& = & \sqrt{1+\left(\frac{B/\omega_{\bm{p}} - \theta \omega_{\bm{p}}}2\right)^2}+\frac{B/\omega_{\bm{p}} + \theta \omega_{\bm{p}}}2 \, , \\
\frac{E_b(\mathbf{p})}{\omega_{\bm{p}}} & = & \sqrt{1+\left(\frac{B/\omega_{\bm{p}} - \theta \omega_{\bm{p}}}2\right)^2}-\frac{B/\omega_{\bm{p}} + \theta \omega_{\bm{p}}}2 \, ,
\end{eqnarray}
where $\omega_{\bm{p}} = \sqrt{\mathbf{p}^2 + m^2}$. Note that the parameters introduced in the commutation relations act as an UV and an IR scale. Thus, relativistic QFT is recovered in the limit in which the energies of the particles are between but far from these values.
QNCFT was extended in Ref.~\cite{Arias:2007zz} to spinor fields. In this case, Hamiltonian of the spinor fields is the standard massless one, but the anticommutation relations are changed. In their model a massless spinor field with just two flavors is considered. The theory is quantized by imposing the (equal-time) anticommutation relations
\begin{equation}
\left\{ \Psi^i_\alpha (\mathbf{x}), \Psi^{j\dagger}_\beta (\mathbf{x}') \right\} = \mathcal{A}_{\alpha\beta}\delta^{ij}\delta^{(3)}(\mathbf{x}-\mathbf{x}') \, , \label{paola}
\end{equation}
where $i,j$ are the spinor indices and $\mathcal{A}$ is an hermitian matrix. The new parameters in $\mathcal{A}$ are in this case adimensional, and therefore the theory amounts to be a field theory description of the model introduced in Ref.~\cite{Coleman:1997xq} of velocity driven oscillations or of the SME with just a $c_L^{00}$ nonzero coefficient, and therefore is ruled out by experiment.
The authors of \cite{Arias:2007zz} try to surmount this problem by making the eigenvalues of $\mathcal{A}$ momentum-dependent, but this has not much sense as the anticommutation is just a function of the spacetime coordinates. We will try to solve this problem in one of the papers presented in this thesis.
\subsection{Neutrino Physics and new physics in the IR}
In this thesis we want to present an alternative to neutrino oscillation driven by neutrino masses which does not modify the field content of the SM, respects the gauge invariance principle and does not imply lepton number violating phenomenon. We also want this alternative to be consistent with all the present experimental and observational data and physically distinguishable from the standard mechanism with future data.
We will seek a realistic model in the framework of QNCFT, for it is a simple extension of QFT. We will try to find a modification of the dispersion relation of neutrinos which involves the square of an IR scale at momenta well above this scale.
The solution appears to invoke a nonlocal extension of the anticommutation relations \eqref{paola}, similar to the one in Ref.~\cite{Balachandran:2007ua} and to the nonlocal extension of QFT in the IR that was suggested in the previous section. We will refer to this modified neutrino physics as noncanonical neutrinos.
\subsection{Probing new neutrino physics in cosmology}
The aforementioned alternative mechanism for neutrino oscillations will be difficult to be distinguished from the mass mechanism by means of particle physics experiments. The reason is that particle physics experiments always probe neutrinos of energies well above the IR scale, and in this regime the behavior of both mechanism is the almost the same.
In order to distinguish between massive neutrinos and noncanonical neutrinos, it will be necessary to probe neutrinos of momenta comparable to or be below the IR scale, and such neutrinos can be found in the C$\nu$B. However in order to consistently explore the cosmological effect of such a Lorentz violating, nonlocal neutrinos, it will be necessary to extend the concept of QFT in curved spacetimes \cite{BD} to more general field theories.
Doing this consistently is a difficult task and is beyond the scope of this PhD thesis. However, as an first step, we will explore how the scalar QNCFT presented in Refs.~\cite{Carmona:2002iv,Carmona:2003kh} may couple to gravity. The symmetry properties of this theory have been studied in Ref.~\cite{Carmona:2009ra}. A more detailed introduction to QNCFT and its symmetry properties have been included in this work. The requirement of keeping this symmetry properties in the curved spacetime case will be essential for determining the precise form of the coupling.
\section{Objectives}
We have tried to argue that problems in cosmology might be solved by an extension of QFT in the IR and the UV, and that the effect of gravity might be behind a modification of QFT responsible of ill-understood phenomena such as neutrino oscillations. Although others have considered top-down approaches to these problems, we think that tracking the experimental data from a bottom-up perspective is the key of scientific progress. Thus we will try to explore departures from GR and QFT in the UV and the IR that might be behind the present problems in cosmology and the recently explored physics of neutrinos.
In the first of the papers presented in this thesis, we will build a cosmological model based on a phenomenological departure from the cosmological standard model both in the UV and the IR, the Asymptotic Cosmological Model. The model will be defined in the homogeneous approximation. The kinematics of the universe will be determined. The horizon problem will be automatically solved. Through a simulation using Monte Carlo Markov Chains, the model will be contrasted with the available observational data which do not depend on the behavior of inhomogeneous perturbations. Also particular realizations of the model in the homogeneous approximation will be considered among the most popular theories and models in cosmology.
We will also try to depart from the homogeneous approximation in the Asymptotic Cosmological Model. In the second paper presented in this thesis, the linearized treatment of scalar perturbations in the model will be derived with just one assumption based on phenomenological consistence. The resulting model will turn out to be distinguishable from the most popular theories and models in cosmology. The effect of the departure from the standard cosmological model in the CMB spectrum will be computed using a modification of the CAMB code \cite{CAMB}.
An alternative to neutrino masses as a mechanism for neutrino oscillations which is compatible with all the experimental data within the framework of QNCFT will be presented in the third of the papers presented in this thesis. The extension will not add new fields to the SM and respect all its gauge and discrete symmetries, but will violate Lorentz invariance and locality. This violation will be mediated by a new IR scale controlling the modification of the dispersion relation and the presence of new physics in the IR.
In order to build a theory of quantum fields in curved spacetime for particles with modified dispersion relations, it will also be important to distinguish if Lorentz symmetry is violated or deformed, as we will see in the fourth of the works presented in this PhD thesis. This distinction could be crucial in order to derive the quantum corrections to the cosmological constant, for instance. In the fouth of the papers presented in this compendium we will explain how the symmetry properties determine the formulation of an extension of QFT to curved spacetime. For the sake of simplicity we will choose the case of a QNCFT of a scalar field in the $B=0$ limit.
A comment on the implications of this mechanism of neutrino oscillations in cosmology will be included in Appendix A. Here a na\"{\i}ve calculation of the effect of a modified dispersion relation of neutrinos in the expansion of the universe is presented.
\chapter{Interpretation of neutrino oscillations based on new physics in the infrared}
An interpretation of neutrino oscillations based on a modification of
relativistic quantum field theory at low energies, without the need to
introduce a neutrino mass, is seen to be compatible with all observations.
\section{Introduction}
The now well-established observation of deficit of solar neutrinos,
atmospheric neutrinos, and neutrinos from reactors and accelerators finds
a coherent interpretation in terms of neutrino
oscillations between three neutrino flavors of different masses~\cite{GonzalezGarcia:2007ib}.
In the minimal standard model (SM), and in contrast with the rest
of the matter particles, the neutrino is assumed to be a zero mass,
left-handed fermion. Therefore neutrino oscillations is our first
glimpse of physics beyond the SM.
Massive neutrinos are introduced in extensions of the SM which
normally invoke new physics at high energies. In particular, one can
consider a Majorana mass term for the neutrino, generated by a
five-dimensional operator in the SM Lagrangian which would be
suppressed by the inverse of a certain high-energy scale.
Another possibility is to enlarge the field content of the SM with
a right-handed neutrino, which allows mass to be generated by
the usual Higgs mechanism. One has to account however for the smallness
of the neutrino mass, which is achieved by the see-saw mechanism~\cite{Minkowski:1977sc,Ramond:1979py,Gell-Mann,Yanagida:1979as,Mohapatra:1979ia},
again invoking a grand-unification scale.
The presence of new physics at high energies has been explored in
several attempts to find alternatives to the standard neutrino
oscillation mechanism.
This new physics might include Lorentz and/or CPT
violations. These two low-energy symmetries are being questioned at
very high energies in the framework of quantum gravity and string theory
developments~\cite{AmelinoCamelia:1997gz,Gambini:1998it,Alfaro:1999wd,Alfaro:2001rb,Alfaro:2001gk,Jacobson:2001tu,Kostelecky:2000mm}, and in fact simple models with Lorentz and/or
CPT violations are able to generate neutrino oscillations,
even for massless neutrinos~\cite{Coleman:1997xq,Kostelecky:2003cr,Kostelecky:2003xn}. Some of them are considered in
the context of the Standard Model Extension (SME)~\cite{Colladay:1996iz,Colladay:1998fq}, which
is the most general framework for studying Lorentz and CPT violations
in effective field theories.
However, all these alternative mechanisms involve new energy dependencies
of the oscillations which are in general disfavored over the standard
oscillation mechanism by experimental data,
which also put strong bounds on the contribution of new physics to this
phenomenon~\cite{Fogli:2003th,GonzalezGarcia:2004wg}. This seems to indicate that our understanding
of neutrino oscillations as driven by mass differences between neutrino flavors
is indeed correct.
In this letter we want to argue that this might not be the case. We will
present an example of new physics to the SM able to generate neutrino
oscillations and which is essentially different from previously considered
models in one or several of the following aspects: it does not necessarily
add new fields to those present in the SM, it may be completely indistinguishable
from the standard oscillation mechanism in the energy ranges where the
phenomenon has been studied (for neutrinos of medium and high energies), so
that automatically satisfies all constraints which are already fulfilled by
the standard mechanism, and finally, it predicts new physics in the infrared,
so that future neutrino low energy experiments could
distinguish this mechanism from the standard one.
Our example will be based on the so-called theory of noncommutative
quantum fields, which has recently been proposed
as an specific scheme going beyond quantum field
theory~\cite{Carmona:2002iv,Carmona:2003kh,Mandanici:2004ht,Balachandran:2007ua}.
The consequences for neutrino oscillations of a simple model with
modified anticommutators for the neutrino fields, which can be
identified as an example of a SME in the neutrino sector, has very
recently been explored in Ref.~\cite{Arias:2007zz}.
We will see however that it is possible to introduce a generalization
of the anticommutation relations of fields in a more general way than
that studied in Ref.~\cite{Arias:2007zz}, going beyond the effective field
theory framework of the SME,
which is the key to reproduce the oscillation results without the
need to introduce a neutrino mass, and with new consequences at
infrared energies.
\section{Noncanonical fields and neutrino oscillations}
The theory of noncommutative fields was first considered in
Refs.~\cite{Carmona:2002iv,Carmona:2003kh}. It is an extension of the usual canonical
quantum field theory in which the procedure of quantification of a
classical field theory is changed in the following way:
the quantum Hamiltonian remains the same
as the classical Hamiltonian, but the canonical commutation
relations between fields are modified. In the case of the scalar
complex field this modification leads to the introduction of two
new energy scales (one infrared or low-energy scale, and another one
ultraviolet or high-energy scale), together with new observable
effects resulting from the modification of the dispersion relation
of the elementary excitations of the fields~\cite{Carmona:2003kh}. If one is
far away from any of these two scales the theory approaches the
canonical relativistic quantum field theory with corrections involving
Lorentz invariance violations which can be expanded in powers of the
ratios of the infrared scale over the energy and the energy over the
ultraviolet scale. By an appropriate choice of the two new energy
scales one can make the departures from the relativistic theory
arbitrarily small in a certain energy domain.
We will now explore the relevance of the extension of relativistic
quantum field theory based on noncanonical fields in neutrino
oscillations. In particular, we will show that it is possible to
obtain oscillations with the observed experimental properties just
by considering a modification of the anticommutators of the fields
appearing in the SM, without the need to introduce a right-handed
neutrino or a mass for this particle.
With the left-handed lepton fields of the SM
\begin{equation}
\Psi_{L \alpha}=\left(\begin{array}{c}\nu_{\alpha} \\
l_{\alpha}\end{array}\right)_L ,
\end{equation}
where $\alpha$ runs the flavor indices, $(\alpha=e,\mu,\tau)$, the
simplest way to consider an analog of the extension of the canonical
quantum field theory for a complex scalar field proposed in
Refs.~\cite{Carmona:2002iv,Carmona:2003kh} is to introduce the modified anticommutation
relations
\begin{equation}
\{\nu_{L\alpha}(\mathbf{x}),\nu^{\dagger}_{L\beta}(\mathbf{y})\} =
\{l_{L\alpha}(\mathbf{x}),l^{\dagger}_{L\beta}(\mathbf{y})\}
= \left[\delta_{\alpha\beta} + A_{\alpha\beta}\right]
\,\delta^3(\mathbf{x}-\mathbf{y}).
\label{leptoncomm}
\end{equation}
A particular choice for the matrix $A_{\alpha\beta}$ in flavor space which
parametrizes the departure from the canonical anticommutators
corresponds to the new mechanism for neutrino oscillations proposed
in Ref.~\cite{Arias:2007zz} which, however, is not compatible with the
energy dependence of the experimental data.
In order to reproduce the observed properties of neutrino
oscillations~\cite{GonzalezGarcia:2007ib} one has to go beyond this extension
and consider an anticommutator between fields at different
points. This can be made compatible with rotational and translational
invariance and with $SU(2)\times U(1)_Y$ gauge symmetry by making use
of the Higgs field
\begin{equation}
\Phi=\left(\begin{array}{c}\varphi^+ \\
\varphi^0 \end{array}\right),
\tilde{\Phi}=\left(\begin{array}{c}\varphi^{0*} \\
-\varphi^- \end{array}\right).
\end{equation}
The modified anticommutators of the left-handed lepton fields that we
consider in this work are
\begin{equation}
\{\Psi_{L\alpha}(\mathbf{x}), (\Psi_{L\beta})^\dagger(\mathbf{y})\} =
\delta_{\alpha\beta}\,\delta^3(\mathbf{x}-\mathbf{y})
+ \tilde{\Phi}(\mathbf{x})
\tilde{\Phi}^\dagger(\mathbf{y}) B_{\alpha\beta}(|\mathbf{x}-\mathbf{y}|),
\label{phicomm}
\end{equation}
where $B_{\alpha\beta}$ are now functions of $|\mathbf{x}-\mathbf{y}|$ instead
of constants. Note that Eq.~(\ref{phicomm}) is compatible with gauge
invariance since $\tilde{\Phi}(\mathbf{x})$ has the same $SU(2)\times U(1)_Y$
quantum numbers as $\Psi_{L\alpha}(\mathbf{x})$.
After introduction of spontaneous symmetry breaking ($\langle
\varphi^0 \rangle = v/\sqrt{2}$),
and neglecting effects coming from the fluctuation of the scalar field
which surely is a good approximation for neutrino oscillations, the
only anticommutators that are changed are those of the neutrino fields
\begin{equation}
\{\nu_{L\alpha}(\mathbf{x}),\nu^{\dagger}_{L\beta}(\mathbf{y})\}=
\delta^3(\mathbf{x}-\mathbf{y})\,\delta_{\alpha\beta}+C_{\alpha\beta}(|\mathbf{x}-\mathbf{y}|),
\label{neutrinocomm}
\end{equation}
where
\begin{equation}
C_{\alpha\beta}(|\mathbf{x}-\mathbf{y}|) = \frac{v^2}{2}
B_{\alpha\beta}(|\mathbf{x}-\mathbf{y}|).
\end{equation}
One can suspect that the new anticommutation relations Eq.~(\ref{neutrinocomm})
introduce a source of mixing between flavors that will affect neutrino oscillations.
We will see in the next Section that this is indeed the case.
\section{Solution of the free theory}
In order to study the neutrino oscillations induced by the modified
anticommutators of fields in the neutrino sector,
one needs to solve the free theory given by the Hamiltonian
\begin{equation}
H=\sum_{\alpha}\int d^3\mathbf{x} \left[i\,\nu_{L\alpha}^\dagger
\left(\bm{\sigma}\cdot\bm{\nabla}\right) \nu_{L\alpha}\right]
\label{ham}
\end{equation}
(where $\bm{\sigma}$ are the $2\times 2$ Pauli matrices),
and the anticommutation relations
showed in Eq.~(\ref{neutrinocomm}).
Let us introduce a plane wave expansion for the neutrino field
\begin{equation}
\nu_{L\alpha} (\mathbf{x}) = \int \frac{d^3\mathbf{p}}{(2\pi)^3} \, \frac{1}{\sqrt{2p}}\,
\sum_{i} \left[\, b_i(\mathbf{p})\, u_{L\alpha}^i (\mathbf{p}) \, e^{i
\mathbf{p}\cdot\mathbf{x}}\right. +
\,\left. d_i^\dagger(\mathbf{p})\, v_{L\alpha}^i (\mathbf{p}) \, e^{-i \mathbf{p}\cdot\mathbf{x}} \right],
\label{neutrino}
\end{equation}
where $p=|\mathbf{p}|$, and $(b_i(\mathbf{p}),d_i^\dagger(\mathbf{p}))$ are the annihilation
and creation operators of three types of particles and antiparticles
(expressed by subindex $i$) with momentum $\mathbf{p}$. We now use the
following \emph{ansatz} for the expression of the
Hamiltonian~(\ref{ham}) as a function of the creation-annihilation operators:
\begin{eqnarray}
H = \int \frac{d^3\mathbf{p}}{(2\pi)^3} \,\sum_{i}
\left[\ E_i(p)\, b_i^\dagger(\mathbf{p})\,b_i(\mathbf{p})\right. +
\left. \bar{E}_i(p)\, d_i^\dagger(\mathbf{p})\,d_i(\mathbf{p})\right].
\label{ham2}
\end{eqnarray}
This corresponds to the assumption that the free theory describes
a system of three types of free particles and antiparticles for each
value of the momentum, with energies $E_i(p), \bar{E}_i(p)$, respectively.
Now, computing $[H,\nu_{L\alpha}]$ by two procedures: firstly by using Eq.~(\ref{ham})
for the Hamiltonian and the anticommutators~(\ref{neutrinocomm}), and secondly, by using
the expressions~(\ref{neutrino}) and~(\ref{ham2}), and equalling both results, one
obtains the following simple result for the energies and the coefficients in the
plane wave expansion of the field:
\begin{eqnarray}
E_i(p) & = & \bar{E}_i(p) = p\,[1+\tilde{c}_i(p)], \label{E} \\
u_{L\alpha}^i(\mathbf{p}) & = & v_{L\alpha}^i(\mathbf{p})=e_\alpha^i(p)\,\chi^i(\mathbf{p}), \label{uv}
\end{eqnarray}
where $\chi^i(\mathbf{p})$ is the two component spinor solution
of the equation
\begin{equation}
(\bm{\sigma}\cdot \mathbf{p})\, \chi^i(\mathbf{p})=-p\, \chi^i(\mathbf{p})
\end{equation}
with the normalization condition
\begin{equation}
\chi^{i\dagger}(\mathbf{p})\, \chi^i(\mathbf{p})=2 E_i(p),
\end{equation}
$\tilde{c}_i(p)$ are the three eigenvalues of
$\tilde{C}_{\alpha\beta}(p)$,
the Fourier transform of $C_{\alpha\beta}(|\mathbf{x}-\mathbf{y}|)$ in Eq.~(\ref{neutrinocomm}),
and $e_{\alpha}^i(p)$ are the components of the normalized eigenvectors
of $\tilde{C}_{\alpha\beta}(p)$.
From Eq.~(\ref{E}), we see that the model presented here contains violation
of Lorentz invariance, but preserves $CPT$ symmetry.
\section{New IR physics and neutrino oscillations}
Since in the free theory solution
there are three types of particles and antiparticles with different energies,
and a mixing of creation and annihilation operators of different kinds of
particle-antiparticle in the expression of each field, it is clear that the
nonvanishing anticommutators of different fields will produce neutrino
oscillations, even for massless neutrinos. This observation was
already present in Ref.~\cite{Arias:2007zz}.
The probability of conversion of a neutrino of flavor $\alpha$
produced at $t=0$ to a neutrino of flavor $\beta$, detected at time
$t$, can be directly read from the propagator of the neutrino
field (\ref{neutrino}). This probability can be written as
\begin{equation}
{\cal P}\left(\nu_{\alpha}(0)\to \nu_{\beta}(t)\right) \,=\, \left|\sum_i
e^i_{\alpha}(p)^* \, e^i_{\beta}(p) \, e^{-i\,E_i(p)t}\right|^2 .
\end{equation}
This is the standard result for the oscillation between three states
with the unitary mixing matrix elements $U^i_{\alpha}$ replaced by the
coefficients $e^i_{\alpha}(p)$ of the plane wave expansion of the
noncanonical neutrino fields and the energy of a relativistic
particle $\sqrt{p^2 + m_i^2}$ replaced by the energy $E_i(p)$ of the
particle created by the noncanonical neutrino fields.
Let us now make the assumption that the modification of the anticommutators
is a footprint of new physics at low energies, parametrized
by an infrared scale $\lambda$. Although the introduction of corrections
to a quantum field theory parametrized by a low-energy scale has not been
so well explored in the literature as the corrections produced by
ultraviolet cutoffs, there are several phenomenological and
theoretical reasons that have recently lead to think on the necessity to
incorporate a new IR scale to our
theories~\cite{ArkaniHamed:1998rs,Antoniadis:1998ig,Sundrum:1998ns,ArkaniHamed:1998kx,Libanov:2005vu,Cohen:1998zx,Carmona:2000gd,Carmona:2005we}.
If the modifications of the anticommutation relations
are parametrized by an infrared scale $\lambda$ then it is reasonable
to assume an expansion in powers of $\lambda^2/p^2$ so that
\begin{equation}
\tilde C_{\alpha\beta}(p)\approx \tilde C_{\alpha\beta}^{(1)}
\,\frac{\lambda^2}{p^2} \text{ for } p^2\gg\lambda^2,
\label{approximation}
\end{equation}
and then
\begin{equation}
\tilde{c}_i(p)\approx \tilde{c}_i^{(1)}\,\frac{\lambda^2}{p^2},
\quad e_\alpha^i(p)\approx e_\alpha^{i(1)},
\label{eigenvalues}
\end{equation}
where $e_{\alpha}^{i(1)}$ ($\tilde{c}_i^{(1)}$) are eigenvectors
(eigenvalues) of $\tilde C_{\alpha\beta}^{(1)}$,
independent of $p$.
But in this approximation, the description of neutrino oscillations produced
by the new physics is completely undistinguishable from the conventional
description based on mass differences ($\Delta m_{ij}^2$) with a
mixing matrix ($U^i_{\alpha}$) between flavor and mass
eigenstates, just by making the correspondence
\begin{equation}
\Delta m_{ij}^2= 2 \,(\tilde c_i^{(1)}-\tilde c_j^{(1)})\,\lambda^2,
\quad U^i_{\alpha}=e_{\alpha}^{i(1)}.
\end{equation}
One should note that when $\lambda \neq 0$ the different energies for
different states select a basis in the Fock space and the mixing of
Fock space operators in the fermionic fields is unavoidable. It is
only when one considers the energy splitting of the
different particles that one has a physical consequence of the mixing
of different creation-annihilation Fock space operators in each
fermionic field. On the other hand, in the case $\lambda=0$
(corresponding to unmodified anticommutation relations) there is
an arbitrariness in the construction of the Fock space. One could make
use of this arbitrariness to choose a basis such that each fermionic
field contains only one annihilation and one creation operator (which
is equivalent to saying that the $e_{\alpha}^i(p)$ are indeed the
$\delta_{\alpha i}$) and therefore no oscillation phenomenon is
produced.
\section{Conclusions}
We have seen in the previous Section that the observations of neutrino
oscillations are compatible with their interpretation as a footprint of
new physics in the infrared. As far as we are aware, this is the first time
that an interpretation of neutrino oscillations coming from new physics,
without the need to introduce neutrino masses, and compatible with
all experimental results, is presented.
This is achieved because of the indistinguishability of the new mechanism
from the conventional one in the range of momenta $p^2 \gg \lambda^2$.
In order to reveal the origin of the oscillations it is necessary to go
beyond the approximation Eq.~(\ref{approximation}), which requires the
exploration of the region of small momenta ($p^2 \approx
\lambda^2$). To get this result it has been crucial to introduce a new
infrared scale through a nonlocal modification of the anticommutation
relations of the neutrino field. Gauge invariance forbids a similar nonlocal
modification for the remaining fields due to the choice of quantum
numbers for the fermion fields in the standard model. In fact the
possibility to have the modified anticommutators (\ref{neutrinocomm}) for the
neutrino fields is related to the absence of the right-handed neutrino
field.
The model of noncanonical fields presented in this work has to be considered
only as an example of the general idea that new infrared physics may be
present in, or be (partially) responsible of, neutrino oscillations, and that
the conventional interpretation may be incomplete. In fact
an extension of relativistic quantum field theory based on the
modification of canonical anticommutation relations
of fields might not be consistent. We have not examined the associated
problems of unitarity or causality beyond
the free theory. But it seems plausible
that the consequences that we have obtained in the neutrino sector will be
valid beyond this specific framework.
In conclusion, in this work we have shown that the experimentally observed
properties of neutrino oscillations do not necessarily imply the
existence of neutrino masses. In fact, future experiments
attempting to determine the neutrino mass, such as KATRIN~\cite{Osipowicz:2001sq},
may offer a window to the
identification of new physics beyond relativistic quantum field theory
in the IR. At this level it is difficult to predict specific
observational effects due to the lack of criteria to select a choice
for $\tilde C_{\alpha\beta}(p)$ in this specific model. A simple
example, however, would be the presence of negative eigenvalues
of this matrix, which could be reflected in an apparent negative
mass squared for the neutrino (see Eq.~(\ref{eigenvalues}))
in the fits from the tail of the tritium spectrum.
Effects on cosmology could also be possible, again depending
on the exact modification of the neutrino dispersion
relation in the infrared. All we can say is that if
the origin of neutrino oscillations is due to new physics in the
infrared then experiments trying to determine the absolute values of
neutrino masses and/or cosmological observations might have a
reflection of the generalized
energy-momentum relation Eq.~(\ref{E}) for neutrinos.
This work has been partially supported by CICYT (grant
FPA2006-02315) and DGIID-DGA (grant2006-E24/2). J.I. acknowledges a FPU
grant from MEC.
\chapter{The Theory of a Quantum Noncanonical Field in Curved Spacetimes}
Much attention has been recently devoted to the possibility that quantum gravity effects could lead to departures from Special Relativity in the form of a deformed Poincar\`e algebra. These proposals go generically under the name of Doubly or Deformed Special Relativity (DSR). In this article we further explore a recently proposed class of quantum field theories, involving noncanonically commuting complex scalar fields, which have been shown to entail a DSR-like symmetry.
An open issue for such theories is whether the DSR-like symmetry has to be taken as a physically relevant symmetry, or if in fact the ``true'' symmetries of the theory are just rotations and translations while boost invariance has to be considered broken.
We analyze here this issue by extending the known results to curved spacetime under both of the previous assumptions. We show that if the symmetry of the free theory is taken to be a DSR-like realization of the Poincar\'e symmetry, then it is not possible to render such a symmetry a gauge symmetry of the curved physical spacetime. However, it is possible to introduce an auxiliary spacetime which allows to describe the theory as a standard quantum field theory in curved spacetime. Alternatively, taking the point of view that the noncanonical commutation of the fields actually implies a breakdown of boost invariance, the physical spacetime manifold has to be foliated in surfaces of simultaneity and the field theory can be coupled to gravity by making use of the Arnowitt-Deser-Misner prescription.
\section{Introduction}
Quantum gravity has historically suffered from lack of observational support and problems with conceptual issues common to all of the most studied proposals \cite{GreSchWit,Polchinski2,Rovelli:1997yv,Thiemann:2001yy,Ashtekar:2001ir,Smolin:2003rk} (e.g.\/ the ``problem of time" and the ``background-independence problem" \cite{stachashtbook}). However, over the past few years there has been a growing interest in possible low energy, observable, effects of quantum gravity scenarios, ranging from TeV-scale quantum gravity to high energy departures from Local Lorentz invariance of spacetime (see e.g.\/ the discussion in Ref.~\cite{Mattingly:2005re,AmelinoCamelia:2008qg}). In particular, it was recently suggested that the Planck length $L_p$ ($L_p \sim 10^{-33} \text{cm} $) should be taken into account in describing the rotation/boost transformations between inertial observers.
In alternative to the standard approach of considering Planck suppressed Lorentz (and possibly CPT) invariance violations in effective field theory (EFT) (see e.g. Ref.~\cite{Mattingly:2005re,Mattingly:2007zz}), it was conjectured that the relativity principle could be preserved via a deformation of special relativity, in the form of the so-called ``Doubly or Deformed special relativity" (DSR) \cite{AmelinoCamelia:2000mn,AmelinoCamelia:2002wr,Judes:2002bw}. While this proposal is formulated in momentum space, a coordinate space definitive formulation is still lacking (and several open issues are still unresolved, see e.g.\/ Ref.~\cite{Schutzhold:2003yp,Liberati:2004ju}).
Although some interpretation of DSR in commutative spacetime have been proposed \cite{Liberati:2004ju,Aloisio:2006nd,Aloisio:2005qt}, there has been a growing activity \cite{KowalskiGlikman:2002we} toward connecting it with noncommutative spacetime \cite{Majid:1994cy, Lukierski:1993wx,KowalskiGlikman:2004qa}, which also arise in 3d-Quantum Gravity \cite{Freidel:2005me} and String Theories \cite{connes}.
However, the connection between quantum gravity and noncommutative spacetime is in principle problematic: in fact the latter seem to entail a ``background-independence problem" given that the commutation relations among spacetime coordinates promote them to the role of {operators}, despite they should be treated as mere labels in a background independent theory (see e.g.\/ Ref.~\cite{Westman:2007yx}).
An alternative to modify the commutation relations between coordinates, while preserving background independence, is to modify the commutation relations of field operators on a spacetime manifold. If the manifold is flat spacetime, the Quantum Theory of Noncanonical Fields (QNCFT) has been studied \cite{Carmona:2002iv,Carmona:2003kh}\footnote{{These references named the theory Quantum Theory of Noncommutative Fields, but Quantum Theory of Noncanonical Fields provides indeed a more precise description.}} and it can be seen as a natural extension to field theory of Non Commutative Quantum Mechanics \cite{Mezincescu:2000zq, Duval:2000xr, Gamboa:2000yq}. Noticeably, it has been shown that there is a link between the one particle sector of QNCFT in flat spacetime and DSR~\cite{Carmona:2009ra} .
Here we are going to use the results of Ref.~\cite{Carmona:2009ra} to study the case of the QNCFT in Curved Spacetimes. In the Section II we will review the main results of Ref.~\cite{Carmona:2003kh}. In the Section III we will discuss the connection between QNCFT and DSR as derived in Ref.~\cite{Carmona:2009ra}. In the Section IV we will derive the Lagrangian, the equation of motion, and the internal product of solutions of the QNCFT in Curved Spacetimes in different cases depending on {which are the} symmetries of the free theory in flat spacetime that we want to promote to gauge symmetries of the curved spacetime. The Section V is devoted to some closing remarks.
\section{QNCFT in Flat Spacetime}
Let us consider the theory of a complex scalar field, given by the Hamiltonian density:
\begin{equation}
\mathcal{H} = \Pi^\dagger(\mathbf{x})\Pi(\mathbf{x})+\mathbf{\nabla}\Phi^\dagger(\mathbf{x})\cdot\mathbf{\nabla}\Phi(\mathbf{x})+m^2\Phi^\dagger(\mathbf{x})\Phi(\mathbf{x}) \label{ham1} \, ,
\end{equation}
the Hamiltonian being the integral to spacelike coordinates of $\mathcal{H}$.
The field is quantized by imposing the noncanonical commutation relations:
\begin{eqnarray}
\left[\Phi(\mathbf{x}),\Phi^\dagger(\mathbf{x}')\right] & = & \theta \delta^{(3)}(\mathbf{x}-\mathbf{x}') \nonumber\, , \\
\left[\Phi(\mathbf{x}),\Pi^\dagger(\mathbf{x}')\right] & = & i \delta^{(3)}(\mathbf{x}-\mathbf{x}') \label{com1}\, ,\\
\left[\Pi(\mathbf{x}),\Pi^\dagger(\mathbf{x}')\right] & = & B \delta^{(3)}(\mathbf{x}-\mathbf{x}') \nonumber \, ,
\end{eqnarray}
where $\theta$ and $B$ are considered to be ``sufficiently small", so that the canonical quantum field theory should be seen as a good approximation to this theory for a certain range of energies or momenta. It should be noticed that $1/\theta$ plays the role of an UV scale and $B$ plays the role of an IR scale \cite{Carmona:2003kh}.
We will forget for the moment about the IR scale and set $B=0$ in our calculations (the effect of the scale $B$ turns out to be, when $\theta = 0$, a redefinition of the mass $m$ and a constant energy gap between n-particle and (n+1)-particle states)~\cite{Carmona:2003kh}.
We shall see now that the just given set of definitions completely specifies the theory.
Let us start by noticing that the momentum operator
\begin{equation}
\mathbf{P}=\int d^3\vx \left[\Pi^\dagger(\mathbf{x})\mathbf{\nabla}\left(\Phi(\mathbf{x})-\frac{i\theta}2\Pi(\mathbf{x})\right)+ \,h.c.\,\right]\, ,
\end{equation}
commutes with the Hamiltonian, and hence the field can be expanded in terms of creation-annihilation operators of particles and antiparticles with a given momentum.
\begin{eqnarray}
\Phi(x)& =& \int \frac{d^3\vp}{(2\pi)^3} \left(U_{\vp}(\mathbf{x},t) \, a_\mathbf{p} + V_{\vp}^*(\mathbf{x},t) \, b^\dagger_\mathbf{p}\right)\label{field}\, ,\\
U_{\vp}(\mathbf{x},t) & = & \sqrt\frac{E_a}{E_b\left(E_a+E_b\right)} e^{-i(E_a t- \mathbf{p} \cdot \mathbf{x})}\label{possol}\, ,\\
V_{\vp}(\mathbf{x},t) & = & \sqrt\frac{E_b}{E_a\left(E_a+E_b\right)} e^{-i(E_b t- \mathbf{p} \cdot \mathbf{x} )}\label{negsol} \, .
\end{eqnarray}
Equations (\ref{ham1}) and (\ref{com1}) describe a theory of particles with energy-momentum relation
\begin{equation}
\frac{E_a(\mathbf{p})}{\omega_{\bm{p}}}=\sqrt{1+\left(\frac{\theta \omega_{\bm{p}}}2\right)^2}+\frac{\theta \omega_{\bm{p}}}2 \label{disp_a}\, ,
\end{equation}
and their antiparticles, with energy-momentum relation
\begin{equation}
\frac{E_b(\mathbf{p})}{\omega_{\bm{p}}}=\sqrt{1+\left(\frac{\theta \omega_{\bm{p}}}2\right)^2}-\frac{\theta \omega_{\bm{p}}}2 \label{disp_b}\, ,
\end{equation}
where $\omega_{\bm{p}} = \sqrt{\mathbf{p}^2 + m^2}$. The difference of energies between particles and antiparticles with the same momentum is a consequence of the CPT violation of the theory.
Notice also that $E_a(-\theta) =E_b (\theta)$, as it should be, due to the symmetry $a\leftrightarrow b$, $\Phi\leftrightarrow \Phi^\dagger$, $\theta \rightarrow -\theta$. Therefore $V_{\vp}(\mathbf{x},t;\theta)=U_{\vp}(\mathbf{x},t;-\theta)$.
The equation of motion of the field can be obtained by using the Heisenberg equation for the evolution of operators in the Heisenberg picture. The result is
\begin{equation}
L(x,\partial) \Phi(x)=\left[\partial_0^2+(m^2-\Delta)(1+ i \theta \partial_0)\right] \Phi(x) = 0 \label{motion1}\, .
\end{equation}
If we define the following internal product in the space of solutions of (\ref{motion1}),
\begin{eqnarray}
\left(\varphi_1(\theta),\varphi_2(\theta)\right) & = & -i\int d^3\vx
\left(\varphi_1(\theta)\frac {\textstyle 1}{\textstyle 1-i\theta\partial_0}\partial_0 \varphi_2^*(\theta) \right.\nonumber \\
& & \left. - \varphi_2^*(-\theta)\frac {\textstyle 1}{\textstyle 1-i\theta\partial_0}\partial_0 \varphi_1(-\theta)
\right)\, ,
\label{prod}
\end{eqnarray}
then $U_{\vp}(\mathbf{x},t)$ and $V_{\vp}^*(\mathbf{x},t)$ form an orthonormal (in the sense of Ref.~\cite{BD}) basis of solutions of the equation of motion.
As a side remark let us notice that equations (\ref{disp_a}), (\ref{disp_b}), (\ref{motion1}), show for $m=0$ a striking similarity with the dispersion relation
\begin{equation}
\omega = \pm \sqrt{c^2 k^2 -\left(\frac{2 \nu k^2}{3} \right)^2} -i \frac{2\nu k^2}3\, ,
\end{equation}
and equation of motion
\begin{equation}
\partial_t^2 \psi_1 \, = \, c^2\mathbf{\nabla}^2 \psi_1 + \frac 43 \nu \partial_t \mathbf{\nabla}^2 \psi_1
\end{equation}
of sound waves propagating through a viscous fluid of viscosity $\nu$ at rest ($\vec{v}_0 = 0$) \cite{Visser:1997ux}. They turn out to be the same, taking into account that the frequency associated to a b-mode of energy $E_b$ is $-E_b$, and redefining $i\theta = \frac {4}{3} \nu$.
The meaning of a purely imaginary viscosity is far from clear, although the main consequence is obviously that the dissipative effects associated to viscosity are turned into dispersive ones, the dispersion relation becoming real.
\section{Symmetries of QNCFT in Flat Spacetime}
A symmetry of a theory is defined as a transformation which leaves the action of the theory invariant. A symmetry also has the property of transforming solutions of the equation of motion into solutions of the equations of motion. Thus, if we want to search for the symmetries of the theory, a good starting point would be to find the operators $\mathcal{O}(x,\partial,...)$ such that if $\Phi(x)$ is a solution of \eqref{motion1}, then also $\mathcal{O}(x,\partial,...)\Phi(x)$ is. {This implies that the action of the commutator of $L$ and $\mathcal{O}$ on solutions of the field equations has to vanish}
\begin{equation}
\left[L(x,\partial,...),\mathcal{O}(x,\partial,...)\right]\Phi(x) = 0 \label{prop}\, .
\end{equation}
The above program is carried out more easily if one Fourier transforms the field $\Phi(x)=\int \frac{d^4 p}{(2\pi)^4} e^{-i p\cdot x}\Phi(p)$.
Then, the equation of motion turns out to be \cite{MasterDiego},
\begin{equation}
L \Phi(p) = \left\{ -p_0^2+(m^2+\mathbf{p}^2)(1+\theta p_0) \right\}\Phi(p) = 0\, ,
\label{eq:disp}
\end{equation}
and the operators verifying the property (\ref{prop}) are only
\begin{eqnarray}
\mathcal{P}_i & = & p_i \, , \label{eq:sy1}\\
\mathcal{P}_0 & = & p_0\, , \label{eq:sy2}\\
\mathcal{M}_{ij} & = & p_i \frac \partial {\partial p^j}-p_j \frac \partial {\partial p^i}\, , \label{eq:sy3}\\
\mathcal{M}_{0i} & = & \frac {p_0}{\sqrt{1+\theta p_0}} \frac \partial {\partial p^i}-p_i \frac{(1+\theta p_0)^{3/2}}{1+\theta p_0/2} \frac \partial {\partial p^0}\, , \label{eq:sy4}
\end{eqnarray}
or arbitrary functions of them. Therefore the previous operators can be considered the generators of the symmetry group.
It is now evident that while the generators of translations $\mathcal{P}_i ,\mathcal{P}_0$ and rotations $\mathcal{M}_{ij}$ are left unchanged, a new set of generators for the boosts $\mathcal{M}_{0i}$ is induced by the noncanonical commutation of the field.
Notice however, that if one restricts $p_0$ so that $p_0>-1/\theta$, then under the change of variables
\begin{equation}
\tilde{p}_0 = \frac {p_0}{\sqrt{1+\theta p_0}}\, , \qquad \tilde{p}_i = p_i \, ,\label {aux}
\end{equation}
the generators of the symmetry group can be rewritten as:
\begin{eqnarray}
\mathcal{P}_i & = & \tilde{p}_i \, , \\
\mathcal{P}_0 & = & \tilde{p}_0\left(\sqrt{1+\theta^2 \tilde{p}_0^2/4}+ \theta \tilde{p}_0/2 \right)\, ,\\
\mathcal{M}_{ij} & = & \tilde{p}_i \frac \partial {\partial \tilde{p}^j}-\tilde{p}_j \frac \partial {\partial \tilde{p}^i}\, ,\\
\mathcal{M}_{0i} & = & \tilde {p}_0 \frac \partial {\partial \tilde{p}^i}-\tilde{p}_i \frac \partial {\partial \tilde{p}^0}\, \label{gens}.
\end{eqnarray}
This shows that the generators of rotations and deformed boosts are isomorphic to the generators of the Lorentz group and it is only the action of the boosts on the four-momentum that it is deformed due to the presence of $\theta$ (reducing to its standard form in the limit of $\theta\to 0$). This is in striking similarity with what is normally conjectured in DSR scenarios \cite{AmelinoCamelia:2000mn, AmelinoCamelia:2002wr, Judes:2002bw} \footnote{Note that in the DSR literature as well as in Ref.~\cite{Carmona:2009ra} Greek letters $(\pi_0,\mathbf{\pi})$ are used to refer to the auxiliary variables $(\tilde{p}_0,\tilde{p}_i)$. We chose here a different notation in order to render the expressions in the rest of the paper more understandable. } and hence it is no surprise that also in this case the whole group of symmetries can be made isomorphic to the Poincar\'e group when $\mathcal{P}_0$ is substituted by $\tilde{\mathcal{P}}_0 = \tilde{p}_0$, which is also the generator of a symmetry. This clearly hints at the possibility to recast the theory in an auxiliary spacetime associated with momenta $(\tilde{p}_i ,\tilde{p}_0)$ in which the action of the Poincar\'e group will be standard. We shall explore this possibility in what follows but before a few remarks about the nature of the just found symmetry are in order.
Symmetries in field theories can be of two types. Internal symmetries are transformations in the field space which do not mix fields at different spacetime points. These transformations will have the form $\Phi(x)\rightarrow\Phi'(x)$ with $\Phi'(x)=\mathcal{O}(x)\Phi(x)$. We will call instead external symmetries those symmetries which do mix fields at different spacetime points. For example this can happen if the transformation affects also spacetime itself. Indeed, spacetime symmetries are transformations in the spacetime background which affect the field. These transformations will have the form $x\rightarrow x'(x)$ and $\Phi(x)\rightarrow \Phi'(x')$ where, in the case of the scalar field, $\Phi'(x')=\Phi(x)$. Therefore spacetime symmetries are external symmetries. However, the opposite is not always true. In particular the symmetry associated to (\ref{eq:sy1}), (\ref{eq:sy2}), (\ref{eq:sy3}), (\ref{eq:sy4}), is not a proper spacetime symmetry because of its non-local nature.
Coming back to our study of the symmetries of the theory in flat spacetime we now want to explore the possibility to recast the whole field theory in a auxiliary spacetime in which the Poincar\'e group acts in the usual way. We shall do so by looking for transformations which leave the action invariant. While a rigorous treatment can be found in Ref.~\cite{Carmona:2009ra} we shall provide here a more concise derivation for the sake of simplicity.
The first step would be to find the action $S$, which will be the spacetime integral of some Lagrangian density $\mathcal{L}$. In order to achieve this, we will write (\ref{ham1}) in terms of field variables whose commutation relations are canonical,
\begin{equation}
\Phi_c = \Phi-\frac{i\theta}2 \Pi \label{phic}\, .
\end{equation}
Then, writing the Hamiltonian in these variables and performing the canonical transformation
\begin{eqnarray}
\bar{\Phi} & = & \left(1-\frac{i\theta}2 \sqrt{m^2-\Delta}\right)^{-1}\Phi_c \, ,\nonumber \\
\bar{\Pi} & = & \left(1-\frac{i\theta}2 \sqrt{m^2-\Delta}\right) \Pi \nonumber \, ,
\end{eqnarray}
the action can be written, up to boundary terms, as
\begin{eqnarray}
S & = & \int d^4 x \mathcal{L} (x) \nonumber \\
& = & \int d^4 x \left(\Pi^\dagger \dot{\Phi}_c + \dot{\Phi^\dagger}_c \Pi - \mathcal{H} (\Phi(\Phi_c),\Pi)\right) \nonumber \\
& = & \int d^4 x \left(\bar{\Pi}^\dagger \dot{\bar{\Phi}} + \dot{\bar{\Phi}}^\dagger \bar{\Pi} - \mathcal{H} (\Phi(\bar{\Phi}),\Pi(\bar{\Pi}))\right) \nonumber \\
& = & \int d^4 x \, \bar\Phi^\dagger \left( -\partial_0^2-(1+i\theta\partial_0) (m^2-\Delta)\right)\bar{\Phi} \label{lag1} \, ,
\end{eqnarray}
where in the last line we have substituted the momentum as a function of the field and its derivatives via the Hamilton-Jacobi equations.
We want now to show that the action \eqref{lag1} can be expressed as a standard scalar field action in flat spacetime. We can start by introducing the field redefinition
\begin{equation}
\phi(x) = \sqrt{1+i \theta \partial_0}\bar{\Phi}(x)\, ,
\end{equation}
so that the action takes the simpler form,
\begin{equation}
S = \int d^4 x \, \phi^\dagger(x) \left( \frac{-\partial_0^2}{1+i\theta\partial_0}-(m^2-\Delta)\right)\phi (x)\, .
\end{equation}
Let us look at this action in momentum space by using the Fourier transform of $\phi(x)$, $\phi(p)$
\begin{equation}
S = \int \frac{d^4 p}{(2\pi)^4} \phi^\dagger(p) \left( \frac{p_0^2}{1+\theta p_0}-(m^2+\mathbf{p}^2)\right)\phi (p)\, \label{lagfrac},
\end{equation}
where the integral over $p_0$ goes from $-\infty$ to $+\infty$. Now we can split the action into two terms $S^\theta$ and $\bar{S}^\theta$, with the integral over $p_0$ ranging respectively from $-1/\theta$ to $+\infty$ and from $-\infty$ to $-1/\theta$. All the solutions of the classical equations of motion lay in $S^\theta$, but the $\bar{S}^\theta$ term can affect the quantum behavior of the theory~\cite{Carmona:2009ra}. Changing in $S^\theta$ the integration variable $p_0$ for $\tilde{p}_0$, defined in (\ref{aux}), this can be rewritten as
\begin{equation}
S^\theta = \int \frac{d^3\vp}{(2\pi)^3} \frac{d\tilde{p}_0}{2 \pi} \frac{\partial p_0}{\partial \tilde{p}_0}\phi^\dagger(p(\tilde{p})) \left( \tilde{p}_0^2-(m^2+\mathbf{p}^2)\right)\phi (p(\tilde{p}))\, .
\end{equation}
Finally we can define the auxiliary field
%
\begin{equation}
\tilde{\phi}(\tilde{p})=\sqrt{\frac{\partial p_0}{\partial \tilde{p}_0}}\,\phi (p(\tilde{p})) \, ,
\label{auxfield}
\end{equation}
%
and inverse Fourier transform, so to arrive to the standard Klein-Gordon-action in some ``auxiliary spacetime".
\begin{equation}
S^\theta =\int d^3\vx \, d\tilde{x}_0\, \tilde{\phi}^\dagger (\tilde{x})\left(-\eta^{\mu\nu}\partial_\mu\partial_\nu - m^2 \right) \tilde\phi(\tilde{x})\, . \label{actaux}
\end{equation}
Now the properties of the symmetry group of the theory can be understood. When written in the auxiliary variables $(\tilde{x}_0,\mathbf{x})$ (auxiliary frame), the symmetry group is the Poincar\'e one associated to the flat auxiliary spacetime symmetries of $S^\theta$.
Reversely, if we assume that the Poincar\'e group acts trivially on $\bar{S}^\theta$, its action on the real spacetime and the original field variables (original frame) can be derived by pulling back the operations that we have applied above. In this case the action of the group is found to be realized in some non-local way. Indeed, this setting is similar to the one proposed in the DSR framework~\cite{AmelinoCamelia:2000mn}.\footnote{Note that the above treatment does not depend on the detailed form of the field dispersion relation nor on its origin. In principle, the discussion may be applied to more general dispersion relations.}
However, it has to be noticed that the necessary splitting of the action into two terms implies that our initial theory is not exactly equivalent to a relativistic field theory in flat space, at the quantum level \cite{Carmona:2009ra}. As a matter of fact, one might argue that the symmetry we have found is a fake one, because it invokes non-locality (see e.g.\/ the discussion about symmetries of Maxwell equations in \cite{Giulini:2006uy}).
Indeed spacetime symmetries are local because the transformed field in a spacetime point is a function of the value of the field in just one spacetime point. In the case we are considering, the $\mathcal{M}_{0i}$ are generating a family of external transformations which cannot be interpreted as spacetime symmetries, at least in the usual way, due to their intrinsic non-locality. Indeed such a symmetry transforms a whole field configuration, which is a solution of the equation of motion, into a different field configuration which is also a solution.
Finally, one must also take into account that the Hamiltonian of the theory, the space integral of Eq.~(\ref{ham1}), does not belong to the set of generators of this Poincar\'e-like symmetry group, although time translations are a symmetry of the free theory. The problem is that the commutator of the Hamiltonian with the generator of boosts would end up defining a new generator and hence the algebra would not be not closed.
An infinite set of generators of the symmetry group of the free theory in flat spacetime
may be obtained by iteration of this procedure. However, the whole symmetry group forbids
the addition of new terms in the action or the appearance of a spacetime dependence of
the coefficients of the existing terms. Therefore the symmetry group must be reduced
(broken into subgroups) in order to extend the theory.
These considerations force us to take a decision. We either try to generalize the theory to curved spacetime by promoting the non-local implementation of the Poincar\'e group to a gauge symmetry, and therefore time translations are taken as an accidental symmetry of the theory in flat spacetime (or at low energies where $\theta$ corrections are supposed to be negligible), or we can take an alternative point of view and consider as the physical symmetries only those realized in a local way while boost invariance is broken.
The first approach has been shown to suffice in the construction of a renormalizable interacting theory in flat spacetime~\cite{Carmona:2009ra}, but, as previously discussed, some of the intuitive notions of what a symmetry is can be spoiled by the non-locality of the implementation. While more intuitive, the second approach will imply a preferred time direction and preferred slices of simultaneity in spacetime.
\section{QNCFT in Curved Spacetimes}
The aforementioned alternative points of view regarding the actual symmetries imply that two different approaches in constructing a QNCFT in Curved spacetimes can be taken depending on which is the symmetry of the free field theory in flat spacetime we would like to promote to a gauge symmetry of the spacetime. Hence, we will perform them separately in the following subsections.
\subsection{A spacetime with a gauge DSR-like Invariance?}
One of the ways of formulating General Relativity (GR) consists in turning the global Poincar\'e symmetry of Minkowski spacetime into a gauge symmetry of any curved geometry.\footnote{ Indeed the Strong Equivalence Principle (which selects GR among relativistic theories of gravitation with spin-2 gravitons) entails Local Lorentz invariance as a fundamental postulate (together with universality of free fall and local position invariance of experiments, including gravitational ones)\cite{Will:2001mx}.} This procedure provides also a link between Quantum Field Theory (QFT) in flat spacetimes and QFT in curved backgrounds. We will try to follow the same steps by gauging the symmetry we have found for the noncanonical field.
A global symmetry in the noncanonical field has been found, whose generators are:
\begin{eqnarray}
\mathcal{P}_i & = & -i\partial_i \, ,\nonumber\\
\mathcal{\tilde{P}}_0 & = & i\frac{\partial_0}{\sqrt{1+i\theta\partial_0}} \, ,\nonumber\\
\mathcal{M}_{ij} & = & x_i\partial_j - x_j\partial_i \, ,\nonumber\\
\mathcal{M}_{0i} & = & x_i\frac{\partial_0}{\sqrt{1+i\theta\partial_0}}-x_0\partial_i \frac{(1+i\theta\partial_0)^{3/2}}{1+i\theta\partial_0/2} \label{gen}\, .
\end{eqnarray}
These operators satisfy the Poincar\'e Algebra. We want to generalize our considerations about the symmetry to curved spacetime without loosing this connection. We can do this by introducing the tetrad field formalism i.e.\/ by constructing a set of normal coordinates $y^a_X$ at each point $X$ of the spacetime manifold (for example for time-like observers one can pick up as a preferred axis the one associated to the tangent vector to the observer worldline and the other three axis as an orthogonal basis in the spacelike hypersurface orthogonal to such vector).
The relationship between these coordinates and the general coordinates in the manifold defines the tetrad field $e^a_\mu (x) = \frac{\partial y^a_X}{\partial x^\mu}\vert_{x=X}$. In terms of the $y^a_X$ the metric at $X$ is simply $\eta_{ab}$ (which is used to lower latin indices), and can be transformed to the metric expressed in terms of the general coordinates $g_{\mu\nu}(x)$ (which is used to lower greek indices) with the use of the tetrad.
In the standard case, we must introduce the covariant derivative, an operator which turns global Poincar\'e and Lorentz invariant tensors into local Poincar\'e and general coordinate transformations invariant tensor.
\begin{equation}
D_\mu = \partial_\mu - e^a_\mu (x) \mathcal{P}_a -\frac 12 \omega^{ab}_\mu (x) \mathcal{M}_{ab} \label{covar}\, ,
\end{equation}
where $e^a_\mu (x)$ and $\omega^{ab}_\mu (x)$ (the spin connection) play the role of gauge fields. This leads from QFT in Flat Spacetime to QFT in Curved Spacetimes when partial derivatives $\partial_a$ are replaced by covariant derivatives $D_a = e_a^\mu D_\mu$ and Lorentz tensors are replaced by generalized tensors built up with the tetrad.
However, if this procedure is applied to the case of the noncanonical field, changing the partial derivatives by covariant derivatives in the action (\ref{lag1}) and inserting the tetrad field does not make the gauge transformation (\ref{gen}) a gauge symmetry. The connection between the QNCFT and the relativistic QFT (\ref{actaux}) is broken.
In order to see this, let us try to follow the same procedure as in the standard case. The action will have the form
\begin{eqnarray}
S & = & \int d^4 x \sqrt{-g(x)} \left[\Pi^\dagger e_0^\mu(x) (D_\mu \Phi) + e_0^\mu(x) (D_\mu \Phi^\dagger) \Pi \right.\nonumber\\
&&- \Pi^\dagger\Pi - \delta^{ij}e_i^\mu(x) (D_\mu \Phi^\dagger) e_j^\nu(x) (D_\nu \Phi) - m^2\Phi^\dagger\Phi\nonumber\\
&&\left. -\frac{i\theta}{2} \left(\Pi^\dagger e_0^\mu(x) (D_\mu \Pi) - e_0^\mu(x) (D_\mu \Pi^\dagger) \Pi \right) \right].\label{eq:break}
\end{eqnarray}
This action is explicitly general covariant and in the flat spacetime limit does reproduce the second line of Eq.~(\ref{lag1}). However, it is easy to see that there is no transformation enabling one to write \eqref{eq:break} as the curved spacetime generalization of the last line of Eq.~(\ref{lag1}) due to the explicit dependence of the tetrad field on the spacetime coordinates. Indeed, the derivation of Eq.~(\ref{lag1}), required infinite integrations by parts. If these were done in the curved spacetime case, an infinite series of new terms would appear in the action, involving covariant derivatives of the tetrad fields. Furthermore, it is also possible to check that the infinitesimal gauge transformations generated either by $\mathcal{\tilde{P}}_0$ or by $\mathcal{M}_{0i}$ do not leave the Lagrangian associated with \eqref{eq:break} invariant even at linear order in $\theta$.
These evident problems lead us to the conclusion that there seems to be a fundamental obstruction in gauging external symmetries generated by non-local operators and in particular in making the symmetry generated by (\ref{gen}) a gauge symmetry of the field action.
However, in the previous section we have shown that the non-locally implemented symmetry can be seen as a proper (i.e.~local) spacetime symmetry of a new auxiliary $\tilde{x}$-spacetime. Though non-locality prevents the gauging of symmetries in $x$-spacetime, the procedure of gauging spacetime symmetries and its results are well known. Therefore we can argue that the proper way to proceed is to gauge global DSR-like symmetries in $\tilde{x}$-spacetime. Of course, the procedure of $\tilde{x}$-gauging the global Poincar\'e symmetry leads to the standard QFT in curved spacetime, but now the curved spacetime is the auxiliary one.
In those regions in which the auxiliary spacetime is asymptotically flat, the mapping between auxiliary and noncanonical fields (\ref{auxfield}) can be used and the propagator in the noncanonical field, in physical spacetime, can be found. However, a connection between the auxiliary field and the noncanonical field in a point of auxiliary spacetime in which the metric is not asymptotically flat seems precluded, at least in the present treatment.
Let us then reformulate all the theory in the auxiliary frame, in which everything is standard. The {part of the action which contains the solutions of the equations of motion} in this DSR-like approach would be
\begin{equation}
S^\theta=\int d^4 \tilde{x} \sqrt{\tilde{g}(\tilde{x})} \tilde{\phi}(\tilde{x})\left(-\tilde{g}^{\mu\nu}\tilde{\nabla}_\mu\tilde{\nabla}_\nu - m^2 +\xi \tilde{R}(\tilde{x}) \right) \tilde\phi(\tilde{x})\, ,
\end{equation}
where $\tilde{\nabla}$ means covariant derivative, $\tilde{R}(\tilde{x})$ is the Ricci scalar, and the line element in the $\tilde{x}$-spacetime is given by
\begin{equation}
d\tilde{s}^2 = \tilde{g}_{\mu\nu} d\tilde{x}_\mu d\tilde{x}_\nu \nonumber\, .
\end{equation}
The theory in the auxiliary variables is ordinary QFT in Curved Spacetime \cite{BD}. The equation of motion of the field in auxiliary spacetime is,
\begin{equation}
\left(\tilde{g}^{\mu\nu}\tilde{\nabla}_\mu\tilde{\nabla}_\nu + m^2 +\xi \tilde{R}(\tilde{x}) \right) \tilde\phi(\tilde{x})=0 \label{mov2}\, .
\end{equation}
Its solutions are spanned by a basis $\left\{\tilde{u}_\mathbf{p}(\tilde{x}),\tilde{u}^*_\mathbf{p}(\tilde{x})\right\}$ which is orthonormal under the internal product
\begin{equation}
\left(\varphi_1,\varphi_2\right)= -i\int_{\tilde{\Sigma}} d\tilde{\Sigma}\sqrt{-\tilde{g}_{\tilde{\Sigma}}(\tilde{x})}\tilde{n}_\mu
\left(\varphi_1 \tilde{\partial}_\mu \varphi_2^*-\varphi_2^*\tilde{\partial}_\mu \varphi_1 \right)\, ,
\end{equation}
where $\tilde{\Sigma}$ is some spacelike hypersurface, $\tilde{n}$ the future-oriented timelike vector orthonormal to it. The value of the internal product is independent of the choice of $\tilde{\Sigma}$.
From the above discussion the standard approach to QFT in curved spacetime straightforwardly follows. In particular standard problems like particle creation from the vacuum can be approached with the usual Bogoliubov techniques. Of course, one may wonder how far one could trust such calculations given that the correspondence between the original, physical, spacetime and the auxiliary one (in which one ends up working) is only via the $S^\theta$ part of the real action. In this sense it is important to notice that the virtual modes belonging to the $\bar{S}^\theta$ term of the action do not affect the Bogoliuvov transformation between modes in the $S^\theta$ term, as these virtual modes ``do not see" the auxiliary spacetime.
The conclusion is that working in this framework the resulting theory will give then the same results as the standard QFT in curved spacetime, at least insofar one is working in the auxiliary spacetime. It still remains open the issue of mapping back the results to the original spacetime which, as we said, seems doable only in asymptotically flat regions and hence not generically.
\subsection{Foliating spacetime with noncanonical commutation relations}
Let us adopt the alternative point of view that meaningful symmetries are local implementations of a certain symmetry group, in this case rotations in space and translations in space and time, {and that the DSR-like Poincar\'e group is just an accidental symmetry of the free action in flat spacetime}. Therefore, we would like to curve physical spacetime in a way such that just spatial rotations and spacetime translations are preserved. It should be clear that in this case, a preferred arrow of time emerges in which the fields appearing in the commutation relations are simultaneous. Each set of simultaneous events defines a spacelike hypersurface in the manifold of spacetime, foliating it, while the orthogonal, timelike direction defines the evolution in time. In those hypersurfaces no spacelike coordinates are preferred, and thus the commutation relations of the field in flat spacetime (\ref{com1}) have to be suitably generalized in order to make them covariant under general coordinate transformations on the spacelike hypersurfaces,
\begin{eqnarray}
\left[\Phi(\mathbf{x},t),\Phi^\dagger(\mathbf{x}',t)\right]& =&\frac \theta{\sqrt{h(t)}} \delta^{(3)}(\mathbf{x}-\mathbf{x}') \nonumber\\
\left[\Phi(\mathbf{x},t),\Pi^\dagger(\mathbf{x}',t)\right]&=&\frac {i\hbar}{\sqrt{h(t)}} \delta^{(3)}(\mathbf{x}-\mathbf{x}') \nonumber\\
\left[\Pi(\mathbf{x},t),\Pi^\dagger(\mathbf{x}',t)\right]&=&0\,
\end{eqnarray}
where $h(t)$ is the determinant of the three-metric $h_{ij}(t)$ associated to each spacelike hypersurface of simultaneity. This foliation of spacetime into slices of simultaneity naturally leads to the Arnowitt Deser Misner description of GR (ADM) \cite{Arnowitt:1960es,Arnowitt:1962hi}.
We briefly review it here for completeness and for fixing the notation.
In the ADM formalism one takes a foliation of spacetime in spacelike hypersurfaces, which implies a split of the whole spacetime metric, the four-metric, into its spacelike-spacelike components $g_{ij}(\mathbf{x},t)$ --- which coincide in this frame of reference with the ones of the induced metric on the hypersurfaces (the three-metric $h_{ij}$) --- a three-vector shift function $N^i(\mathbf{x},t)$ and a three-scalar lapse function $N(\mathbf{x},t)$. The components of the four-metric in this frame are then
\begin{equation}
\begin{array}{c}
g_{\mu\nu}
\end{array}
=
\left(
\begin{array}{cc}
g_{00} & g_{0j}\\
g_{i0} & g_{ij}
\end{array}
\right)
=
\left(
\begin{array}{cc}
N^2-N^i N^k h_{ik} & -h_{kj} N^k \\
-h_{ik} N^k & -h_{ij}
\end{array}
\right)\, .\label{ADM}
\end{equation}
It has to be noticed that the formalism is still covariant under general spacelike coordinate transformations. Spacelike indices $i$,$j$,$k$,... are lowered with the three-metric $h_{ij}$ and raised with the inverse of the three-metric, $h^{ij}$. The inverse of the four-metric turns out to be
\begin{equation}
\begin{array}{c}
g^{\mu\nu}
\end{array}
=
\left(
\begin{array}{cc}
1/N^2 & -N^j/N^2 \\
-N^i/N^2 & -h^{ij}+N^i N^j/N^2
\end{array}
\right)\, ,
\end{equation}
and the determinant of the four-metric is $g = -N^2 h$. The {timelike} vector which is orthonormal to the spacelike hypersurfaces is
\begin{equation}
n^\mu = (1/N, -N^i/N) \, . \label{time}
\end{equation}
The covariant four derivative is constructed in the usual way with the affine connections which are furthermore assumed to be the Christoffel symbols $\Gamma^{(4)\lambda}_{\mu\nu}$ associated with the four-metric $g_{\mu\nu}$.
The extrinsic curvature $K_{ij}$ is defined with the covariant four-derivative of the normal to the hypersurfaces ($K_{ij}= N \Gamma^{(4)0}_{ij})$. The covariant three-derivative is defined as the projection of the covariant four-derivative on the spacelike hypersurface, $\nabla^{(3)}_i V^j \equiv V^j_{\vert i} = \partial_i V^j + \Gamma^{(3)j}_{ik}V^k$, where $\Gamma^{(3)j}_{ik} = \Gamma^{(4)j}_{ik} + K_{ik}N^j/N$ coincides with the Christoffel symbols built up with the three-metric. (From now on the superscript $(3)$ can be dropped in order to simplify the notation.) The intrinsic curvature $R^i_{jkl}$ is built in the usual way with the three-metric $h_{ij}$ and its derivatives, and can be related to the four-curvature and the extrinsic curvature.
Finally, this notation has also a link to the tetrad notation introduced in the previous subsection. The tetrad $e_a^\mu$ associated with the metric (\ref{ADM}) is
\begin{equation}
\begin{array}{ccc}
e_0^0 \, = \, 1/N & & e_0^i \, = \,N^i/N \\
e_i^0 \, = \, 0 & & e_j^i \, = \,e^{(3)j}_i \\
\end{array}\, ,\label{tetradADM}
\end{equation}
where $e^{(3)i}_a$ is the dreibein associated to the $h_{ij}$ three-metric.
We have now to consider how can we couple the noncanonical field to the physical spacetime metric (\ref{ADM}). The result was already given in the previous subsection in Eq.~(\ref{eq:break}). Written in the notation of the ADM prescription with the use of (\ref{tetradADM}), the expression of the action, Lagrangian density and Hamiltonian density of the field are
\begin{eqnarray}
S_\Phi & = & {\textstyle \int d^4 x }\sqrt{h} \mathcal{L}_\Phi \, ,\label{actionADM}\\
\mathcal{L}_\Phi & = & \Pi^\dagger\dot{\Phi}_c + \dot{\Phi}^\dagger_c \Pi-N \mathcal{H}_\Phi \, , \label{lagADM} \\
\mathcal{H}_\Phi & = & \Pi^\dagger \Pi+h^{ij}\partial_i\Phi^\dagger\partial_j\Phi+m^2\Phi^\dagger\Phi \nonumber\\
&& +\left(\frac{N^i}N \Pi^\dagger \partial_i\Phi_c \,+\, h.c.\,\right)\, . \label{hamADM}
\end{eqnarray}
where $\Phi_c$ is given by (\ref{phic}). The stress energy tensor is defined as
\begin{equation}
T_{\mu\nu} = \frac 2{N\sqrt{h}}\frac{\delta S_\Phi}{\delta g^{\mu\nu}}\, ,
\end{equation}
and turns out to be
\begin{eqnarray}
T_{00} & = & N\Pi^\dagger\dot{\Phi}_c +N N^i\Pi^\dagger \partial_i\Phi_c+N^i N^j \partial_i \Phi^\dagger \partial_j \Phi\, +\,h.c.\nonumber\\
& & -(N^2-N_i N^i)\mathcal{L}_\Phi \nonumber\\
T_{0i} & = & N^j\partial_i\Phi^\dagger \partial_j\Phi+N\Pi^\dagger\partial_i\Phi_c \,+\,h.c.\,+N_i \mathcal{L}_\Phi \nonumber\\
T_{ij} & = & \partial_i\Phi^\dagger \partial_j \Phi \, +\, h.c.\, +h_{ij} \mathcal{L}_\phi\, .
\end{eqnarray}
Now we have all the information required about the coupling of the noncanonical field to the gravitational potential in curved physical spacetime. In particular, we can derive from the action (\ref{actionADM}) the equation of motion of the field. These equations can be rewritten in a more compact way if we write them in terms of the Lie derivative of the fields along the timelike direction (\ref{time}),
\begin{equation}
\pounds_n\Phi = \frac 1 N \dot{\Phi} - \frac{N^i} N \partial_i\Phi\, .
\end{equation}
Varying the action with respect to $\Pi^\dagger$ we get
\begin{equation}
N\pounds_n\Phi-i\theta N\pounds_n\Pi-N\Pi-\frac{i\theta}4 \frac{\dot h}h\Pi + \frac{i\theta}2 N^i_{\vert i}\Pi\,=\, 0 \label{PiADM}
\end{equation}
and varying the action with respect to $\Phi^\dagger$,
\begin{equation}
-N\pounds_n\Pi-\frac{\dot{h}}{2 h}\Pi-N m^2\Phi+(N \partial_i \Phi)^{\vert i}+N^i_{\vert i}\Pi\,=\, 0\, .\label{PhiADM}
\end{equation}
These equations define a system of coupled differential equations which should be solved provided the metric of spacetime is known and we neglect the back-reaction of the field on the metric. We should remark that the general covariance of the whole spacetime is broken by the preferred choice of a time variable, but the general covariance on the spacelike hypersurfaces is guaranteed by construction. This setting reminds the context of analogue gravity models \cite{Barcelo:2005fc}.
{The formulation of the quantum theory of a field in curved spacetime requires the definition of a scalar product in the space of solutions of the field equations. We are now able to define an inner product in the space of solutions. The inner product essentially carries the information of the commutation relations. Given two solutions, $\Phi = \varphi_{1,\theta}$ and $\Phi = \varphi_{2,\theta}$, of the equations of the field (\ref{PiADM}), (\ref{PhiADM}) we} can define the internal product $\left(\, , \,\right)$ as
\begin{equation}
\left(\varphi_1, \varphi_2\right) = -i\int d^3\vx \sqrt{h} \left[\varphi_{1,\theta}\varphi^{\Pi\,*}_{2,\theta}-\varphi^*_{2,-\theta}\varphi^\Pi_{1,-\theta}\right]\, ,
\label{intprod}
\end{equation}
where the integral is evaluated at any of the spacelike hypersurfaces in which spacetime becomes foliated and
\begin{equation}
\varphi_{r,\theta}^\Pi = \left[ 1+\frac{i\theta}N\left(N\pounds_n+\frac{\dot{h}}{4h}+\frac{N^i_{\vert i}}{2}\right) \right]^{-1} \pounds_n\varphi_r (\theta)\, ,
\end{equation}
is the conjugate momentum solution $\Pi = \varphi_{r,\theta}^\Pi$ associated to the field solution $\Phi = \varphi_{r,\theta}$ with $r=1,2$. This internal product preserves the symplectic structure of the theory and is independent of the choice of the hypersurface by construction \cite{Wald}.
Let $\{\Phi \, = \, U_{\vp}(\mathbf{x},t;\theta)\}$ be an orthonormal basis of solutions of the system of differential equations (\ref{PiADM}), (\ref{PhiADM}) with positive frequency with respect to the preferred time direction (\ref{time}), labeled by the index $\mathbf{p}$. Let $\{\Phi \, = \, V_{\vp}^*\}$ be an orthonormal basis of solutions of negative frequency of the same system of differential equations. Then the field can be expanded in terms of annihilation operators of the $U_{\vp}(\mathbf{x},t;\theta)$ and creation operators of the $V_{\vp}^*(\mathbf{x},t;\theta)$
\begin{equation}
\Phi(\mathbf{x},t)=\int \frac{d^3\vp}{(2\pi)^3}\left[U_{\vp}(\mathbf{x},t;\theta) \,a_\mathbf{p}\,+\,V_{\vp}(\mathbf{x},t;\theta)^* \,b^\dagger_\mathbf{p}\right]\, .
\end{equation}
If $V_{\vp}^*$ is a solution of negative frequency of the system of field equations (\ref{PiADM}), (\ref{PhiADM}), then $V_{\vp}$ is a solution of positive frequency of the complex conjugate of the system, which coincides with the result of changing $\theta\,\rightarrow\, -\theta$ in the system. Therefore $V_{\vp}(\mathbf{x},t;\theta) \propto U_{\vp} (\mathbf{x},t;-\theta)$. A more careful analysis of the commutation relations shows that $V_{\vp}(\mathbf{x},t;\theta) = U_{\vp} (\mathbf{x},t;-\theta)$. This is due to the symmetry $\theta\,\rightarrow \,-\theta, \, a\,\leftrightarrow \, b, \,\Phi \,\leftrightarrow \, \Phi^\dagger, \, \Pi\,\leftrightarrow \,\Pi^\dagger $.
As usual, there is non-uniqueness in the choice of an orthonormal basis $\{U_{\vp},V_{\vp}^*\}$ of solutions of the system of field equations if the metric induces a loss of the time translation symmetry. We must then resort to Bogoliubov techniques in order to relate the associated non equivalent vacua. Let us expand the field in terms of two different basis of solutions of the equation of motion
\begin{eqnarray}
\Phi(x) & = & \int \frac{d^3\vp}{(2\pi)^3} \left[ U_{\vp}(x)\, a_\mathbf{p} +V_{\vp}^*(x) \, b^\dagger_\mathbf{p} \right]\nonumber\\
& = & \int \frac{d^3 \mathbf{p}'}{(2\pi)^3} \left[ \bar{U}_{\mathbf{p}'}(x) \bar{a}_{\mathbf{p}'} +\bar{V}_{\mathbf{p}'}^*(x) \bar{b}_{\mathbf{p}'}^\dagger \right]
\label{expansion}\, .
\end{eqnarray}
As both sets of solutions of the equation of motion are basis, we can define the Bogoliubov transformation as a change of basis in the space of solutions of the equation of motion,
\begin{equation}
U_{\vp} = \int \frac{d^3 \mathbf{p}'}{(2\pi)^3} \left[ \bar{U}_{\mathbf{p}'}(x) \alpha_{\mathbf{p}' \mathbf{p}} +\bar{V}_{\mathbf{p}'}^*(x) \beta_{\mathbf{p}' \mathbf{p}} \right] \, ,\label{Bogo}
\end{equation}
where
\begin{eqnarray}
\alpha_{\mathbf{p}' \mathbf{p}} = \left( U_{\vp} , \bar{U}_{\mathbf{p}'} \right)\, , \qquad \beta_{\mathbf{p}' \mathbf{p}} & = & -\left( U_{\vp} , \bar{V}^*_{\mathbf{p}'} \right)
\, .
\end{eqnarray}
We define a state $\vert 0 \rangle$ as the vacuum in the $\{U_{\vp},V_{\vp}^*\}$ basis. We want to compute the expectation value in this state of the number operators $\bar{N}_{a,\mathbf{p}'}$, $\bar{N}_{b,\mathbf{p}'}$ defined in the $\{\bar{U}_{\mathbf{p}'},\bar{V}_{\mathbf{p}'}^*\}$ basis. Inserting (\ref{Bogo}) in (\ref{expansion}) in order to get the Bogoliubov transform of the creation-annihilation operators and plugging them in the expression of the number operators, the result will be
\begin{eqnarray}
\langle 0 \vert \bar{N}_{a,\mathbf{p}'} \vert 0 \rangle & = & \int d^3 \mathbf{p} \vert \beta_{\mathbf{p}' \mathbf{p}}(-\theta) \vert ^2\, ,\\
\langle 0 \vert \bar{N}_{b,\mathbf{p}'} \vert 0 \rangle & = & \int d^3 \mathbf{p} \vert \beta_{\mathbf{p}' \mathbf{p}}(\theta) \vert ^2\, .
\end{eqnarray}
Noticeably, the internal product (\ref{intprod}) has two terms which differ not only in sign but also in magnitude, unlike in the standard QFT in curved spacetime or the QNCFT in Curved auxiliary Spacetime. This two terms will in general be complex and oscillating with different phases. Hence, when computing the number of particles and antiparticles observed by a detector in the vacuum of some other detector (like for example in the case of Rindler observers in flat spacetime or static observers at infinity in a black hole spacetime), one then finds that the corresponding spectrum follows the pattern of interference of the terms in the Bogoliubov coefficient $\beta$. This result is essential for example in the studying black hole evaporation by Hawking radiation, something which we plan to address elsewhere.
\section{conclusions}
Recent works \cite{Carmona:2009ra} have offered a mechanism to identify new symmetries in the theory of QNCFT \cite{Carmona:2003kh}. This mechanism goes beyond this particular example and can be applied in principle to more general theories with modified dispersion relations. The new symmetry transformations will be non-local, but the action can be rewritten in terms of an auxiliary spacetime in which the symmetry transformations are local and the action is Poincar\'e invariant. There is a link between the conjugate four-momenta in physical and auxiliary spacetimes, which reminds the DSR paradigm \cite{Judes:2002bw}. As a result a non-local implementation of the Poincar\'e group is found to be a symmetry group of the theory in flat spacetime. However, the generators of the symmetries of this group do not form a closed algebra with the generator of the symmetry of translations in time. When the theory is tried to be extended, it is unclear which of the symmetries are to be kept and which will be treated as accidental. Some works \cite{Giulini:2006uy} do not give much credit to non-locally implemented symmetries. On the other hand, in the absence of the Poincar\'e symmetry, the possibility of constructing a perturbatively renormalizable interacting theory of the noncanonical field is also unclear \cite{Carmona:2009ra}.
Two different paths have been followed in order to derive a QNCFT in Curved spacetimes depending on the symmetries that are required to be gauged: either the DSR-like symmetries of the free theory, though some of its elements are implemented non-locally, or the group of symmetries of the free theory that are locally implemented, i.e.~translations and rotations.
We find that curving the physical spacetime breaks the DSR-like symmetry. This maybe due to an obstruction in gauging symmetries generated by non-local operators. Consequently, if one wants to preserve the DSR-like implementation of the Poincar\'e group, one must take the auxiliary spacetime as the one to be curved. This leads to a relativistic QFT in Curved Auxiliary Spacetime, which is very similar to the standard one. The correspondence between the fields in the auxiliary frame and fields in the physical spacetime can be trivially done in regions in which the auxiliary spacetime is asymptotically flat. Strictly speaking, the two theories are not really equivalent due to the presence of $\bar{S}^\theta$ part of the physical action. The role of the $\bar{S}^\theta$ modes, which cannot be mapped to auxiliary spacetime, is to fix time ordering in the original frame but they do not seem to have any observable influence on the auxiliary frame, for example they do not affect the Bogoliubov transformations.
On the other hand, if only the symmetries that are implemented locally are required to characterize spacetime at short scales, the physical spacetime of the QNCFT can be made curve using the ADM prescription. The commutation relations must be adapted to fit in a general covariant frame. The action of the field coupled to the three-metric, lapse and shift functions can be written unambiguously. General coordinate invariance is broken by the commutation relations into invariance under general coordinate transformations on the slices of simultaneity, and reparameterizations of time. The resulting QNCFT in Curved Physical Spacetime seems to have the same structure as the ordinary one, but new phenomena related to the UV scale may appear. The main difference between this theory and the standard QFT in Curved spacetimes is that, due of the different energies of particles and antiparticles, the two terms appearing in the internal product (\ref{intprod}) become different not only in sign, but also in magnitude.
We conclude that the theories introduced offer clearly distinguishable outcomes for related phenomenology. This will be further studied in future work.
We would like to thank C.~Barcel\'o, J.~M.~Carmona, J.~L.~Cort\'es, D.~Maz\'on, J.~Rubio, L.~Sindoni and M.~Visser for fruitful discussions. J.I. would like to thank the Scuola Internazionale Superiore di Studi Avanzati (SISSA) for the hospitality during the development of this work. Financial support was provided by CICYT (project FPA2003-02948), DGIID-DGA (project 2008-E24/2). J.I. acknowledges an FPU grant from the Spanish Ministerio de Ciencia e Innovaci\'on (MICIIN).
\chapter{An extension of the cosmological standard model \\ with a bounded
Hubble expansion rate}
The possibility of having an extension of the cosmological standard
model with a Hubble expansion rate $H$ constrained to a finite
interval is considered. Two periods of accelerated expansion arise
naturally when the Hubble expansion rate approaches to the two
limiting values. The new description of the history of the universe is
confronted with cosmological data and with several theoretical ideas
going beyond the standard cosmological model.
\section{Introduction}
According to General Relativity, if the universe is filled with the particles
of the Standard Model of particle physics, gravity should lead to a
deceleration of the expansion of the universe. However, in 1998 two
independent evidences of present accelerated expansion were presented
\cite{Riess:1998cb,Perlmutter:1998np} and later confirmed by different
observations \cite{Jaffe:2000tx,Spergel:2003cb,Peacock:2001gs}. On the other hand, measurements of large
scale structure \cite{Tegmark:2003ud} and CMB anisotropy \cite{Jaffe:2000tx} also
indicate that the universe evolved through a period of early accelerated
expansion (inflation).
There is no compelling explanation for any of these cosmic accelerations,
but many intriguing ideas are being explored. In the case of inflation,
the origin of the accelerated expansion can be either a modification of
gravity at small scales \cite{Starobinsky:1980te} or a coupling of the expansion of
the universe to the progress of phase transitions \cite{Kazanas:1980tx,Sato:1980yn,Guth:1980zm,Linde:1981mu,Albrecht:1982wi}.
In the case of the present accelerated expansion these ideas can be classified
into three main groups: new exotic sources of the gravitational field with
large negative pressure \cite{Turner:1998ex,Zlatev:1998tr,ArmendarizPicon:2000dh} (Dark Energy), modifications of
gravity at large scales \cite{Bekenstein:1984tv,Bekenstein:2004ne} and rejection of the spatial
homogeneity as a good approximation in the description of the present
universe \cite{Tomita:2000jj,Kolb:2005da,Wiltshire:2007zj}.
Different models (none of them compelling) of the source responsible
for each of the two periods of accelerated expansion have been
considered. Einstein equations admit a cosmological constant
$\Lambda$, which can be realized as the stress-energy tensor of empty
space. This $\Lambda$, together with Cold Dark Matter, Standard Model
particles and General Relativity, form the current cosmological model
($\Lambda$CDM). However, quantum field theory predicts a value of
$\Lambda$ which is 120 orders of magnitude higher than observed.
Supersymmetry can lower this value 60 orders of magnitude, which is
still ridiculous \cite{Weinberg:1988cp}.
In order to solve this paradox, dynamical Dark Energy models have been proposed.
This has also lead to explore the possibility that cosmic acceleration arises
from new gravitational physics. Also here several alternative
modifications of the Einstein-Hilbert action at large and small
curvatures \cite{Gasperini:1991ak,Carroll:2003wy,Sotiriou:2005hu,Hu:2007nk,Nojiri:2006ri,Nojiri:2003ft,Cognola:2007zu}, or even higher
dimensional models \cite{Deffayet:2000uy,Dvali:2000hr}, producing an accelerated
expansion have been identified.
All these analysis include an \textit{ad hoc} restriction to actions involving simple
functions of the scalar curvature and/or the Gauss-Bonnet tensor.
This discussion is sufficient to establish the point that cosmic acceleration can
be made compatible with a standard source for the gravitational field, but it is
convenient to consider a more general framework in order to make a systematic
analysis of the cosmological effects of a modification of general relativity.
In this paper we parametrize the evolution of the universe (considered isotropic and
homogeneous) with the Hubble parameter $H$. One finds that it is possible to restrict
the domain of $H$ to a bounded interval. This restriction naturally
produces an accelerated expansion when the Hubble expansion rate
approaches any of the two edges of the interval.
Therefore, we find a new way to incorporate two periods
of cosmic acceleration produced by a modification of general
relativity. But the dependence on the two limiting values
of $H$ can be chosen independently. One could even consider a Hubble
expansion rate constrained to a semi-infinite interval with a unique
period of accelerated expansion. In this sense the aim to
have a unified explanation of both periods of accelerated expansion is
only partially achieved. This simple phenomenological approach to the problem
of accelerated expansion in cosmology proves to be equivalent at
the homogeneous level to other descriptions based on modifications
of the Einstein-Hilbert action or the introduction of exotic components
in the matter Lagrangian.
In the next section we will review how a general modification of the gravitational
action leads to a generalized first Friedman equation at the homogeneous level. In
the third section we will present a specific model (ACM) based on the simplest
way to implement a bounded interval of $H$. In the fourth section we will contrast
the predictions of ACM for the present acceleration with astrophysical observations.
In the fifth section we will show that it is always possible to find modified
gravitational actions which lead to a given generalized first Friedman equation, and
present some simple examples. In the sixth section we will show that it is also always
possible to find Dark Energy models which are equivalent to a given generalized
first Friedman equation, and present some simple examples. The last section is devoted
to summary and conclusions.
\section{Action of the cosmological standard model extension}
The spatial homogeneity and isotropy allow to reduce the
gravitational system to a mechanical system with two
variables $a(t)$, $N(t)$ which parametrize the Robertson-Walker
geometry
\begin{equation}
ds^2 \,=\, N^2(t) dt^2 - a^2(t) \Delta_{ij} dx^i dx^j
\end{equation}
\begin{equation}
\Delta_{ij} \,=\, \delta_{ij} + \frac{k x_i x_j}{1 - k{\bf x}^2} \, .
\end{equation}
Invariance under parameterizations of the time variable
imply that the invariant time differential $N(t)dt$ must be used. Also
a rescaling of the spatial variables
$x^i\rightarrow\lambda x^i$ together with $a(t)\rightarrow\lambda^{-1}
a(t)$ and $k\rightarrow\lambda^{-2} k$ is a symmetry that must be kept
in the Lagrangian. The action of the reduced homogeneous gravitational
system can be then written as
\begin{equation}
I_g \,=\, \int dt N \, L\left(\frac{k}{a^2}, H,
\frac{1}{N}\frac{dH}{dt},
\frac{1}{N}\frac{d}{dt}\left(\frac{1}{N}\frac{dH}{dt}\right),
...\right) \, ,
\label{Ig}
\end{equation}
with $H \,=\, \frac{1}{aN}\frac{da}{dt}$.
If we keep the standard definition of the gravitational coupling and
the density ($\rho$) and pressure ($p$) of a cosmological homogeneous
and isotropic fluid as a source of the gravitational field, we have
the equations of the reduced system
\begin{equation}
\left(\frac{8\pi G_N}{3}\right) \rho \,=\, - \frac{1}{a^3} \frac{\delta
I_g}{\delta N(t)}
\end{equation}
\begin{equation}
- \left(8\pi G_N\right) p \,=\, - \frac{1}{N a^2} \frac{\delta
I_g}{\delta a(t)}\, .
\end{equation}
It is possible to choose a new time coordinate $t^{'}$ such that
\begin{equation}
\frac{dt^{'}}{dt} \,=\, N(t)\, .
\end{equation}
This is equivalent to set $N=1$ in the action (\ref{Ig}) of the gravitational
system and the evolution equations of the cosmological model reduce to a
set of equations for the scale factor $a(t)$. If we introduce the
notation
\begin{equation}
H^{(i)} \,=\, \left(\frac{d}{dt}\right)^i \, H
\end{equation}
\begin{equation}
\left(\delta_i L\right)^{(j)} \,=\, \left(\frac{d}{dt}\right)^j
\left[a^3 \partial_i L\right]\, ,
\end{equation}
with $\partial_i L$ denoting the partial derivatives of the Lagrangian
as a function of the variables $(k/a^2, H, H^{(1)}, H^{(2)}, ...)$ we
have
\begin{equation}
\left(\frac{8\pi G_N}{3}\right) \rho \,=\, - L +
\sum_{i=2}^{\infty} \sum_{j=0}^{i-2} \frac{ (-1)^{i-j}}{a^3} H^{(j)} \left(\delta_i
L\right)^{(i-j-2)}
\label{rho}
\end{equation}
\begin{equation}
- \left(8\pi G_N\right) p \,=\, - 3 L + \frac{2 k}{a^2} \partial_1 L +
\sum_{i=2}^{\infty} \frac{(-1)^i}{a^3} \left(\delta_i
L\right)^{(i-1)} \, .
\label{p}
\end{equation}
In the homogeneous and isotropic approximation, the vanishing of the
covariant divergence of the energy-momentum tensor leads to the
continuity equation
\begin{equation}
\frac{d}{dt} \left(\rho a^3\right) \,=\, - p \frac{d}{dt} a^3
\label{cont_eq}\, .
\end{equation}
In the radiation dominated era one has
\begin{equation}
\rho \,=\ 3 p \,=\, \frac{\sigma}{a^4}
\label{rad}\, ,
\end{equation}
where $\sigma$ is a constant parameterizing the general
solution of the continuity equation. In a period dominated by matter
one has a pressureless fluid and then
\begin{equation}
p \,=\, 0 \, ,{\hskip 2cm} \rho \,=\, \frac{\eta}{a^3}\, ,
\end{equation}
with constant $\eta$. When these
expressions for the energy density and pressure are plugged in
(\ref{rho}-\ref{p}), one ends up with two compatible differential equations for
the scale factor $a(t)$ which describe the evolution of the
universe. From now on we will use the more common notation
$\frac{dH}{dt}\equiv\dot{H}$,...
\section{The Asymptotic Cosmological Model}
Let us assume that the Lagrangian L of the gravitational system is
such that the evolution equations (\ref{rho}-\ref{p}) admit a solution
such that $a(t)>0$ (absence of singularities), $\dot{a} >0$ (perpetual
expansion) and $\dot{H}<0$. In that case, one has a different value of
the scale factor $a$ and the Hubble rate $H$ at each time and then one
has a one to one correspondence between the scale factor and the
Hubble rate. Since the continuity equation (together with the equation
of state) gives a relation between the scale factor and the density,
one can describe a solution of the evolution equations of the
generalized cosmological model through a relation between the energy
density and the Hubble rate, i.e. through a generalized first Friedman
equation
\begin{equation}
\left(\frac{8\pi G_N}{3}\right) \rho \,=\, g(H)\, ,
\label{rho_g}
\end{equation}
with $g$ a smooth function which parametrizes the different algebraic
relations corresponding to different solutions of different
cosmological models. Each choice for the function $g(H)$ defines a
phenomenological description of a cosmological model. Then one
can take it as a starting point trying to translate any observation
into a partial information on the function $g(H)$ which parametrizes
the cosmological model.
Eq. (\ref{rho_g}) is all one needs in order to reconstruct the
evolution of the universe at the homogeneous level. The generalized
second Friedman equation is
obtained by using the continuity equation (\ref{cont_eq}) and the
expression for $\rho$ as a function of $H$. One has
\begin{equation}
- \left(8\pi G_N\right) p \,=\, 3 g(H) + \frac{g^{'}(H)}{H} \dot{H}
\label{pressure}\, .
\end{equation}
In the matter dominated era one has
\begin{equation}
\frac{\dot{H}}{H^2} \,=\, - 3 \frac{g(H)}{H g^{'}(H)}
\label{matter}
\end{equation}
and then the assumed properties of the solution for the evolution
equations require the consistency conditions
\begin{equation}
g(H) > 0 \, ,{\hskip 1cm} g^{'}(H) >0\, .
\label{gcons}
\end{equation}
In the period dominated by radiation one has
\begin{equation}
\frac{\dot{H}}{H^2} \,=\, - 4 \frac{g(H)}{H g^{'}(H)}
\label{radiation}
\end{equation}
instead of (\ref{matter}) and the same consistency conditions
(\ref{gcons}) for the function $g(H)$ which defines the generalized
first Friedman equation.
We introduce now a phenomenological cosmological model
defined by the condition that the Hubble rate has an upper bound $H_+$
and a lower bound $H_-$. This can be implemented through a function
$g(H)$ going to infinity when $H$ approaches $H_+$ and going to
zero when $H$ approaches $H_-$. We will also assume that there is an interval
of $H$ in which the behavior of the energy density with the Hubble parameter is,
to a good approximation, scale-free i.e. $g(H)\propto H^2$. The source
of the gravitational field will be a homogeneous and isotropic fluid
composed of relativistic and non-relativistic particles.
The Cosmological Standard Model without curvature is recovered in the
limit $H_-/H\rightarrow 0$ and $H_+/H\rightarrow\infty$ which is a
good approximation for the period of decelerated expansion.
Notice that this interpretation is independent of the underlying
theory of gravitation.
The Hubble parameter can be used to parametrize the history of
universe as long as $\dot{H}\neq0\,\forall \,t$. The total density can be
thus expressed as a function of $H$. If $H$ is bounded, then $\rho(H)$
will have a pole at $H=H_+$ and a zero at $H=H_-$. Far from these scales,
the behavior of $\rho(H)$ can be assumed to be approximately scale-free.
Under these conditions, we can parametrize the dependence of the
cosmological model on the lower bound $H_-$ by
\begin{equation}
g(H) \,=\, H^2 h_{-}\left(\frac{H_-^2}{H^2}\right)
\label{g-}
\end{equation}
and similarly for the dependence on the upper bound $H_+$
\begin{equation}
g(H) \,=\, \frac{H^2}{
h_{+}\left(\frac{H^2}{H_+^2}\right)}
\label{g+}\, ,
\end{equation}
where the two functions $h_{\pm}$ satisfy the conditions
\begin{equation}
\lim_{x\to 0} h_{\pm} (x) \,=\, \beta^{\pm 1}\, , {\hskip 1cm}
\lim_{x\to 1} h_{\pm} (x) \,=\, 0\, .
\end{equation}
$\beta$ is a constant allowed in principle by dimensional
arguments. If $\beta \neq 1$ then it can be moved to the lhs of the
Friedman equation, turning $G_N\rightarrow\beta G_N$, and can be
interpreted as the ratio between an effective cosmological value of
the gravitational coupling and the value measured with local
tests. But $\beta\neq 1$ would be in conflict with Nucleosynthesis,
through the relic abundances of $^4 He$ and other heavy elements (for
3 neutrino species) \cite{Mukh}, so we set $\beta= 1$ . Therefore
\begin{equation}
\lim_{x\to 0} h_{\pm} (x) \,=\, 1 \, ,{\hskip 1cm}
\lim_{x\to 1} h_{\pm} (x) \,=\, 0\, .
\label{hpm}
\end{equation}
The consistency conditions (\ref{gcons}) result in
\begin{equation}
h_{\pm} (x) > 0 \, ,{\hskip 1cm} h_{\pm} (x) > x h_{\pm}^{'}(x)
\label{hcons}
\end{equation}
for the two functions $h_{\pm}$ defined in the interval $0<x<1$. Thus,
we can divide the cosmic evolution history into three periods. In the
earliest, relativistic particles dominate the energy density of the
universe and the generalized first Friedman equation shows a
dependence on the upper bound $H_+$. There is also a transition period
in which the effect of the bounds can be neglected and the rhs of the
first Friedman equation is scale-free; this period includes the
transition from a radiation dominated universe to a matter dominated
universe. In the third present period, non-relativistic particles
dominate the energy density of the universe but the dependence on the
lower bound $H_-$ must be accounted for in the generalized first
Friedman equation.
\indent From the definition of the Hubble parameter, one has
\begin{equation}
\frac{\dot{H}}{H^2} \,=\, -1 + \frac{a \ddot{a}}{\dot{a}^2}\, .
\end{equation}
Then, in order to see if there is an accelerated or decelerated
expansion, one has to determine whether $\dot{H}/H^2$ is greater or smaller
than $-1$.
In the period dominated by radiation one has
\begin{equation}
\frac{\dot{H}}{H^2} \,=\, - 2 \, \frac{1}{1 - \frac{x h_+^{'}(x)}{
h_+(x)}}\, ,
\end{equation}
with $x=H^2/H_{+}^2$, where we have used (\ref{g+}) assuming that only
the dependence on the upper bound ($H_{+}$) of the Hubble parameter is
relevant. A very simple choice for this dependence is given by
\begin{equation}
g(H) \,=\, \frac{H^2}{\left(1 - \frac{H^2}{H_{+}^2}\right)^{\alpha_+}}
\label{alpha+}\, ,
\end{equation}
with $\alpha_+$ a (positive) exponent which parametrizes the departure
from the standard cosmological model when the Hubble rate approaches
its upper bound. With this simple choice one has a transition from an
accelerated expansion for $H^2 > H_+^2/(1+\alpha_+)$ into a decelerated
expansion when $H^2 < H_+^2/(1+\alpha_+)$, which includes the domain of
validity of the standard cosmological model ($H\ll H_+$).
In the period dominated by matter (which corresponds to lower values
of the Hubble rate) we assume that only the dependence on the lower
bound ($H_-$) of the Hubble rate is relevant. Then one has
\begin{equation}
\frac{\dot{H}}{H^2} \,=\, - \frac{3}{2} \, \frac{1}{1 - \frac{x
h_-^{'}(x)}{h_-(x)}}\, ,
\end{equation}
with $x=H_{-}^2/H^2$. We can also consider a dependence on $H_-$
parametrized simply by an exponent $\alpha_-$
\begin{equation}
g(H) \,=\, H^2 \, \left(1 - \frac{H_-^2}{H^2}\right)^{\alpha_-}
\label{alpha-}\, .
\end{equation}
With this choice one has a transition from a decelerated expansion
for $H^2 > H_{-}^2(1+\alpha_-/2)$ leaving the domain of validity of
the standard cosmological model and entering into an accelerated
expansion when the Hubble rate approaches its lower bound for $H^2 <
H_{-}^2(1+\alpha_-/2)$.
The possibility of describing $\rho$ as a function of $H$ is independent
of the existence of spatial curvature $k$. However, in the kinematics of
observables in the expanding universe we do need to specify the value of
$k$. In the rest of the paper we will assume that the universe is flat
($k=0$), although the same analysis could be done for arbitrary $k$.
The properties of the expansion obtained in this simple example (a
period of decelerated expansion separating two periods of accelerated
expansion) are general to the class of phenomenological models with a
generalized first Friedman equation (\ref{rho_g}) with $g(H)$
satisfying the consistency conditions (\ref{gcons}) and a Hubble rate
constrained to a finite interval. The specific part of the example
defined by (\ref{alpha+},\ref{alpha-}) is the simple dependence on the
Hubble rate bounds and the values of the Hubble rate at the
transitions between the three periods of expansion. From now on we
will name this description the Asymptotic Cosmological Model (ACM).
With respect to the late accelerated expansion, ACM can be seen as a
generalization of $\Lambda$CDM, which can be recovered by setting
$\alpha_-=1$. It also includes an early period of exponential
expansion which can be seen as a phenomenological description of the
evolution of the universe at inflation in the homogeneous approximation.
\subsection{Horizon problem}
One can see that the horizon of a radiation-dominated universe can be
made arbitrarily large as a consequence of an upper bound on the
Hubble parameter and in this way one can understand the observed
isotropy of the cosmic microwave background at large angular scales.
Let us consider the effect of the modification of the cosmological
model on the calculation of the distance $d_h(t_f,t_i)$ of a source of
a light signal emitted at time $t_i$ and observed at time $t_f$
\begin{equation}
d_h(t_f, t_i) \,=\, c \, a(t_f) \int_{t_i}^{t_f} \frac{dt}{a(t)}\, .
\end{equation}
We have
\begin{equation}
\frac{dt}{a} \,=\, \frac{da}{a^2 H}\,=\, \left(\frac{8 \pi G_N
\sigma}{3}\right)^{-1/4} \frac{g^{'}(H) dH}{4H g(H)^{3/4}} \, ,
\end{equation}
where in the first step we have used the definition of the Hubble
expansion rate $H$ and in the second step we have used the relation
between the scale factor $a$ and $H$ as given by
(\ref{rad}-\ref{rho_g}). We are considering both times $t_i$ and $t_f$
in the radiation dominated period.
The distance $d_h$ is then given by
\begin{equation}
d_h(H_f, H_i) \,=\, \frac 1{4 g(H_f)^{1/4}} \int_{H_f}^{H_i} \frac{dH
g^{'}(H)}{H g(H)^{3/4}} \label{dist}\, .
\end{equation}
If $H_i$ is very close to $H_+$ (i.e. if we choose the time $t_i$ when
the light signal is emitted well inside the period of accelerated
expansion) then the integral is dominated by the
region around $H_i$ which is very close to $H_+$. Then one can
approximate in the integrand
\begin{equation}
g(H) \approx \frac{H_+^2}{\left(1-\frac{H^2}{H_+^2}\right)^{\alpha_+}} \, .
\end{equation}
On the other hand if the observation is made at a time $t_f$ within the
domain of validity of the cosmological standard model ($H_-^2\ll
H_f^2\ll H_+^2$) then the factor $g(H_f)^{1/2}$ in
front of the integral can be approximated by $H_f$ and then one has
\begin{equation}
d_h(H_f, H_i) \approx \frac{\alpha_+}{4 \sqrt{H_f H_i}}
B_{\frac{H^2_i}{H^2_+}}(1/2,-\alpha/4) \, ,
\end{equation}
where $B_z(m,n)$ is the incomplete Beta function, and it can be made
arbitrarily large by choosing $H_i$ sufficiently close
to $H_+$. In this way we see that a cosmological model with a finite
interval of variation for $H$ solves the horizon problem.
\section{Constraints of ACM by Observations}
In this section we will carry out a more technical analysis about how
the astrophysical observations constrain the parameter space of ACM in
the matter dominance period. This analysis is based on the use of
(assumed) standard candles, basically Type Ia Supernovae \cite{Riess:2006fw}
and CMB \cite{Spergel:2006hy}. These observations constrain the parameter
space to confidence regions in which the combination
$\Omega_m\equiv(1-H_-^2/H_0^2)^{\alpha_-}$ is constrained to be around
one quarter.
The consideration of both Type Ia SNe and CMB together favor
$\alpha_->1.5$. The results of this analysis can be seen in
figures (FIG. 1-3).
A reader not interested in technical details might well skip this section.
We center our discussion of the experimental tests of the Asymptotic
Cosmological Model
in the late accelerated expansion produced when $H$ approaches its
lower bound $H_-$. The vast amount of supernovae data collected by
\cite{Riess:2006fw} and \cite{Astier:2005qq}, the data from the SDSS Baryon
Acoustic Oscillation \cite{Eisenstein:2005su}, the mismatch between total energy
density and total matter energy density seen at CMB anisotropies
\cite{Spergel:2006hy} and the measurements of present local mass density by
2dF and SDSS \cite{Tegmark:2003ud,Cole:2005sx} compared with the measurements of $H_0$
from the HST Cepheids \cite{Freedman:2000cf} show that the universe undergoes
a surprising accelerated expansion at the present time.
We will firstly confront the model with the Supernovae Ia data from
Riess \textit{et al.} and SNLS collaboration.
The usefulness of the Supernovae data as a test of Dark Energy models
relies on the assumption that Type Ia SNe behave as standard candles,
i.e., they have a well defined environment-independent luminosity
$\mathcal{L}$ and spectrum. Therefore, we can use measured bolometric
flux $\mathcal{F}=\frac{\mathcal{L}}{4\pi d_L^2}$ and frequency to
determine luminosity distance $d_L$ and redshift $z$.
The luminosity distance is given now by
\begin{equation}
d_L(z) \, = \, c(1+z)\int_0^z \frac{dz'}{H(z')}
\, = \, \frac{c(1+z)}{3}\int^{H(z)}_{H_0}\frac{g'(H)dH}{H g^{1/3}(H_0)g^{2/3}(H)}\, .
\end{equation}
The computed value must be compared with the one obtained
experimentally from the measured extinction-corrected
distance moduli ($\mu_0=5 log_{10}(\frac{d_L}{1 Mpc})+25$) for each
SN. The SNe data have been compiled in references \cite{Riess:2006fw}, and
we have limited the lowest redshift at $cz<7000 km/s$ in order to
avoid a possible ``Hubble Bubble'' \cite{Jha:2002af,Zehavi:1998gz}. Therefore our
sample consists on 182 SNe. We will determine the likelihood of the
parameters from a $\chi^2$ statistic,
\begin{equation}
\chi^2(H_0,H_-,\alpha_-)=
\sum_i\frac{(\mu_{p,i}(z_i;H_0,H_-,\alpha_-)-\mu_{0,i})^2}{\sigma^2_{\mu_{0,i}}+\sigma^2_v}\,,
\end{equation}
where $\sigma_v$ is the dispersion in supernova redshift due to
peculiar velocities (we adopt $\parallel\mathbf{v_p}\parallel = 400
km/s$ in units of distance moduli), $\sigma_{\mu_{0,i}}$
is the uncertainty in the individual measured distance moduli
$\mu_{0,i}$, and $\mu_{p,i}$ is the value of $\mu_0$ at $z_i$
computed with a certain value of the set of parameters
${H_0,H_-,\alpha_-}$. This $\chi^2$ has been marginalized over
the nuisance parameter $H_0$ using the adaptive method in reference
\cite{Wang:2004xz}. The resulting likelihood
distribution function $e^{-\chi^2/2}$ has been explored using Monte
Carlo Markov Chains. We get a best fit
of ACM at $\alpha_-=0.36$ and $\frac{H_-}{H_0}=0.95$, for which
$\chi^2=157.7$. In contrast, fixing $\alpha_-=1$, we get
$\Omega_\Lambda\equiv\frac{H_-^2}{H_0^2}=0.66$ and
$\chi^2=159.1$ for the best fit $\Lambda$CDM. The confidence regions
are shown in Fig. 1 (top).
\begin{figure}
\centerline{\includegraphics[scale=0.6,
width=9cm]{sinrestriccion.png}}
\end{figure}
\begin{figure}
\centerline{\includegraphics[scale=0.6,
width=9cm]{restricciondos.png}}
\label{confidence}
\caption{Confidence regions in parameter space of the Asymptotic
Cosmological Model (ACM) with (bottom) and without (top) priors from
CMB and estimations of present matter energy density at $1\sigma$
(green), $2\sigma$ (red) and $3\sigma$ (blue). The $\Lambda$CDM is
inside the 1$\sigma$ region if we consider no priors, but is moved
outside the $2\sigma$ region when we confront the SNe data with CMB
and estimations from present matter energy density. In this case a
value $\alpha_-> 1.5$ is favored.}
\end{figure}
We can add new constraints for the model coming from measurements of
the present local matter energy density from the combination of 2dF
and SDSS with HST Cepheids, rendering $\Omega_m = 0.28 \pm 0.03$; and
the distance to the last scattering surface from WMAP, which leads to
$r_{CMB}\equiv\sqrt{\Omega_m}\int^{1089}_0 \frac{H_0
dz'}{H(z')}=1.70\pm0.03$ \cite{Riess:2006fw}.
Including these priors we get a best fit of ACM at $\alpha_-=3.54$ and
$\frac{H_-}{H_0}=0.56$ (our simulation explored the region with
$\alpha_-<3.6$), for which $\chi^2=164.1$. In contrast, we get
$\Omega_\Lambda=0.73$ and $\chi^2=169.5$ for the best fit
$\Lambda$CDM, which is outside the $2\sigma$ confidence region shown
in Fig. 1 (bottom). The fits of the best fit $\Lambda$CDM and ACM
taking into account the priors to the SNe data are compared in Fig. 2.
The information which can be extracted from the data is limited. This
can be seen in Fig. 3, in which it is explicit that the data
constrain mainly the value of the present matter energy density,
$\Omega_m\equiv\frac{\rho_m}{\rho_C}=(1-\frac{H_-^2}{H_0^2})^{\alpha_-}$.
\begin{figure}
\centerline{\includegraphics[scale=0.6, width=9cm]{bestfitACM.png}}
\label{bestfit}
\caption{Distance Moduli vs. Redshift comparison between the best
fits of ACM ($H_0 = 66 \, Km/s MPc$, red) and $\Lambda$CDM ($H_0 =
65 \, Km/s MPc$, blue) to the SNe data with priors. ACM fits clearly
better the medium redshift SNe. }
\end{figure}
\begin{figure}
\centerline{\includegraphics[scale=0.6, width=9cm]{omega.png}}
\label{omegabestfit}
\caption{$\Omega_m$ vs. $\alpha$ confidence regions with the CMB prior
only. The confidence regions show $\Omega_m = 0.25 \pm 0.03$ and
$\alpha > 1.15$ at $2\sigma$ level. }
\end{figure}
\section{Generalized First Friedman Equation and $f(R)$ gravity}
The extension of the cosmological model considered here could be
compared with recent works on a modification of gravity at large
or very short distances \cite{Gasperini:1991ak,Carroll:2003wy,Sotiriou:2005hu,Hu:2007nk,Nojiri:2006ri,Nojiri:2003ft,Cognola:2007zu}.
It has been shown that, by considering a correction to the
Einstein Hilbert action including positive and negative powers of the
scalar curvature, it is possible to reproduce an accelerated expansion
at large and small values of the curvature in the cosmological
model. Some difficulties to make these modifications of the
gravitational action compatible with the solar system tests of general
relativity have lead to consider a more general gravitational action,
including the possible scalars that one can construct with the
Riemann curvature tensor \cite{Carroll:2004de}, although the Gauss-Bonnet scalar is the only
combination which is free from ghosts and other pathologies. In fact, there
is no clear reason to restrict the extension of general relativity
in this way. Once one goes beyond the derivative expansion, one should
consider scalars that can be constructed with more than
two derivatives of the metric and then one does not have a good
justification to restrict in this way the modification of the
gravitational theory. It does not seem difficult to find an
appropriate function of the scalar curvature or the Gauss-Bonnet
scalar, which leads to a cosmology with a bounded Hubble expansion rate.
One may ask if a set of metric f(R) theories which include ACM as an
homogeneous and isotropic solution exists.
The answer is that a bi-parametric family of f(R) actions which lead to
an ACM solution exists. In the following section we will derive them
and we will discuss some examples. The derivation follows the same
steps of modified f(R)-gravity reconstruction from any FRW cosmology
\cite{Nojiri:2006gh}.
We start from the action
\begin{equation}
S=\frac 1{16 \pi G_N}\int d^4 x\sqrt{-g}[f(R) + 16\pi G L_m]\, ,
\end{equation}
which leads to the ``generalized first Friedman equation''
\begin{equation}
-18(H \ddot H +4 H^2 \dot H)f''(R)
-3(\dot H +H^2)f'(R)
-\frac 12 f(R) = \left( 8\pi G_N \right) \rho \, ,
\end{equation}
where
\begin{equation}
R = -6(\dot{H}+2H^2)\, .
\end{equation}
If the energy density of the universe is mainly due to matter (as in
the late accelerated expansion), we can use (\ref{matter}) and
(\ref{rho_g}) to express $\rho$, $R$, $\dot{H}$ and $\ddot{H}$ as functions of
$H$. Thus we get
\begin{equation}
[H^3 g g'^2-3H^2 g^2 g'+3H^3 g^2 g'']f''(R)
+\frac{[3H g g'^2-H^2 g'^3]}{18}f'(R)-\frac{g'^3}{108}f(R)=\frac{g g'^3}{18} \, , \label{difR}
\end{equation}
where
\begin{equation}
R(H) = 18 H g(H)/g'(H)-12 H^2 \label{RH}\, .
\end{equation}
$H(R)$ can be obtained from (\ref{RH}) and set into (\ref{difR}); then
we get an inhomogeneous second order linear differential equation with
non-constant coefficients. Therefore, there will always be a
bi-parametric family of $f(R)$ actions which present ACM as their
homogeneous and isotropic solution. The difference between these
actions will appear in the behavior of perturbations, which is not
fixed by (\ref{rho_g}). Some of these actions are particularly easy to
solve. If $g(H)=H^2$ then $R=-3H^2$ and the differential equation
becomes
\begin{equation}
6 R^2 f''(R) -R f'(R)-f(R) +2 R=0\, .
\end{equation}
Its general solution is
\begin{equation}
f(R)=R+c_1 R^{\frac 1{12}(7-\sqrt{73})}+c_2 R^{\frac 1{12}(7+\sqrt{73})}\, ,
\end{equation}
which will give (\ref{rho_g}) as First Friedman equation as long as
the radiation energy density can be neglected. In general the
differential equation (\ref{difR}) will not be solvable
analytically, and only approximate solutions can be found as power
series around a certain singular point $R_0$ (a value of $R$ such that
$H(R)$ cancels out the coefficient of $f''(R)$ in (\ref{difR})). These
solutions will be of the form $f(R)=f_p(R) +c_1 f_1(R) +c_2 f_2(R)$
with
\begin{equation}
\frac{f_i(R)}{R_0}=\sum_{m=0}^\infty a_m^{(i)} (\frac{R-R_0}{R_0})^{s_i+m} \, ,
\end{equation}
where $i=p,1,2$, $a_0^{(1)}=a_0^{(2)}=1$ and the series will converge
inside a certain radius of convergence. An interesting choice of $R_0$
is $R_0 = -12 H_-^2 \equiv R_-$, which is the value of $R$ at
$H_-$. One can also find the approximate solution of the differential
equation for $\mid R\mid\gg\mid R_-\mid$, which will be of the form
\begin{equation}
\frac{f_i(R)}{R}=\sum_{m=0}^\infty a_m^{(i)} (\frac{R_-}{R})^{s_i+m} \label{fR-}\, .
\end{equation}
One can in principle assume that the action (\ref{fR-}) could be
considered as valid also in the region in which radiation begins to
dominate, but this action reproduces (\ref{rho_g}) only if matter
dominates. However, some of the new terms appearing in (\ref{fR-})
will be negligible against $R$ when radiation begins to dominate. The
others can be canceled out by setting to zero the appropriate integration
constant and therefore the Cosmological Standard Model will be
recovered as a good approximation when radiation begins to dominate.
$f(R)$-theories are not the only modified gravity theories studied in
the literature; $f(G)$-theories \cite{Nojiri:2005jg} are also a popular field of research.
In these theories, the Einstein-Hilbert action is supplemented by a
function of the Gauss-Bonnet scalar
$G=R^2-4R_{\mu\nu}R^{\mu\nu}+R_{\mu\nu\rho\sigma}R^{\mu\nu\rho\sigma}$
(not to be confused with the gravitational coupling $G_N$).The same
approach can be used to answer what $f(G)$ actions are able to
reproduce ACM homogeneous evolution, once more in analogy with the
modified f(G)-gravity reconstruction of a FRW cosmology
\cite{Nojiri:2006je}. The form of the action will be
\begin{equation}
S=\frac 1{16 \pi G_N}\int d^4 x\sqrt{-g}[R+f(G) + 16\pi G_N L_m]\, ,
\end{equation}
from which we derive the modified Friedman equation
\begin{equation}
\left(8\pi G_N\right) \rho = 3 H^2-\frac 12 G f'(G)+\frac 12
f(G)+12f''(G)\dot{G}H^3 \, .
\end{equation}
Following the same procedure as in the case of $f(R)$-theories we
arrive to the differential equation
\begin{equation}
3 g(H) = 3 H^2-\frac 12 G f'(G)+\frac 12 f(G)
+\frac{864 H^6 g}{g'^3}[9 g g'-H g'^2-3 H g g''] f''(G) \label{difG} \, ,
\end{equation}
where
\begin{equation}
G=24 H^2(H^2-3H g/g')\label{GH}\, .
\end{equation}
For a given $g(H)$, one can use (\ref{GH}) to get $H(G)$, then set it
into (\ref{difG}) and solve the second order linear differential
equation. As in the previous case, there will be a bi-parametric family
of solutions for $f(G)$ which will have (\ref{rho_g}) as their
homogeneous isotropic solution as long as matter dominates. The
parameters will need to be fixed in order to make the contribution of
undesired terms in the Friedman equation to be negligible when
radiation dominates. Again $g(H)=H^2$ is an example which can be
solved analytically. The differential equation turns to be
\begin{equation}
12 G^2 f''(G)-G f'(G)+f(G)=0\, ,
\end{equation}
which has the solution
\begin{equation}
f(G)=c_1 G + c_2 G^{1/12}\, .
\end{equation}
In most cases this procedure will not admit an analytical solution
and an approximate solution will need to be found numerically.
A similar discussion can be made to explain the early accelerated
expansion (inflation) as a result of an $f(R)$ or $f(G)$ action,
taking into account that in this case the dominant contribution to the
energy density is radiation instead of matter. The analogue to
(\ref{fR-}) will be now a solution of the form
\begin{equation}
\frac{f_i(R)}{R}=\sum_{n=0}^\infty a_n^{(i)} (\frac{R}{R_+})^{r_i+n} \label{fR+}\, ,
\end{equation}
for $\vert R \vert \ll \vert R_+\vert$ with $R_+ = -12 H_+^2$.
Terms in the action which dominate over $R$ when $R$ becomes small
enough should be eliminated in order to recover the Standard
Cosmological model before matter starts to dominate.
Both accelerated expansions can be described together in the
homogeneous limit by an $f(R)$ action with terms
$(\frac{R_-}{R})^{s_i+m}$ with $s_i>0$ coming from (\ref{fR-}) and
terms $(\frac{R}{R_+})^{r_i+n}$ with $r_i>0$ coming from
(\ref{fR+}). The action will contain a term, the Einstein-Hilbert
action $R$, which will be dominant for $H_-\ll H\ll H_+$ including the
period when matter and radiation have comparable energy densities.
\section{Alternative Descriptions of the ACM}
In general, by virtue of the gravitational field equations, it is
always possible to convert a modification in the gravitational term of
the action to a modification in the matter content of the universe. In
particular, at the homogeneous level, it is possible to convert a
generalized first Friedman equation of the type (\ref{rho_g}) to an
equation in which, apart from the usual matter term, there is a dark
energy component with an unusual equation of state $p=p(\rho)$ and in
which General Relativity is not modified.
The trivial procedure is the following. Assume that the source of the
gravitational field in the modified gravity theory behaves as
$p_d=\omega \rho_d$ (matter or radiation). This is the case when the
modification of gravity is relevant. We can then define a dark energy
or effective gravitational energy density as
\begin{equation}
\rho_g =\frac{3}{8\pi G_N} \left( H^2-g(H)\right) \label{dedensity}\, .
\end{equation}
The continuity equation (\ref{cont_eq}) allows to define a pressure
for this dark energy component as $p_g=-\rho_g-\dot{\rho}_g/3 H$. The
equation of state of the dominant content of the universe leads to
\begin{equation}
\dot{H}=-3(1+\omega)\frac{H g(H)}{g'(H)} \label{dH}\, ,
\end{equation}
which can be used to express $\dot{\rho}_g$ as a function of $H$, and
one gets the final expression for $p_g$,
\begin{equation}
p_g=\frac{-3}{8\pi G_N}\left(
H^2-2(1+\omega)H\frac{g(H)}{g'(H)}+\omega g(H)\right)
\label{depressure}\, .
\end{equation}
Then, for any given $g(H)$ we can use (\ref{dedensity}) to get
$H=H(\rho_g)$ and substitute in (\ref{depressure}) to get an
expression of $p_g(\rho_g;H_\pm,\omega)$ which can be interpreted as
the equation of state of a dark energy component. In the simple case
of a matter dominated universe with ACM, $\alpha_-=1$ gives obviously
a dark energy component verifying $p_g=-\rho_g$. Another simple
example is $\alpha_-=2$, for which
\begin{equation}
\left(\frac{8 \pi G_N}{3 H_-^2} \right) p_g=\frac{2}{\frac{8 \pi
G_N}{3 H_-^2}\rho_g -3}\, .
\end{equation}
The inverse procedure is also straightforward. Suppose we have a
universe filled with a standard component $p_d=\omega\rho_d$ and a
dark energy fluid $p_g=p_g(\rho_g)$. Using the continuity equation of
both fluids one can express their energy densities as a function of
the scale factor $a$ and then use this relation to express $\rho_g$ as
a function of $\rho_d$. Then the Friedman equation reads
\begin{equation}
H^2=\frac{8\pi G_N}{3}(\rho_d+\rho_g(\rho_d))\, ,
\end{equation}
which, solving for $\rho_d$, is trivially equivalent to (\ref{rho_g}).
This argument could be applied to reformulate any cosmological model
based on a modification of the equation of state of the dark energy
component \cite{Nojiri:2005sr} as a generalized first Friedman equation
(\ref{rho_g}).
Until now we have considered descriptions in which there is an exotic
constituent of the universe besides a standard component (pressureless
matter or radiation). In these cases, pressureless matter includes both
baryons and Dark Matter. However, there are also descriptions in which
Dark Matter is unified with Dark Energy in a single constituent of the
universe. One of these examples is the Chaplygin gas $p_g=-1/\rho_g$
\cite{Kamenshchik:2001cp,Alvarenga:2001nm}. The use of the continuity equation leads to
\begin{equation}
\rho_g=\sqrt{A+B a^{-6}}\, ,
\end{equation}
where $A$ and $B$ are integration constants. The previous method can
be used to find the generalized first Friedman equation for the baryon
density in this model,
\begin{equation}
\frac{8\pi
G_N}{3}\rho_b=\frac{H^2}{k}\left(\sqrt{1+k(1-\frac{H_-^4}{H^4}})-1
\right) \label{Chap}\, ,
\end{equation}
where $H_-=\sqrt{\frac{8\pi G}{3}}A^{1/4}$, $k=B \rho_{b0}^{-2}-1$,
and $\rho_{b0}$ is the present value of $\rho_b$. In the $H\gg H_-$
limit this model can be interpreted as a universe filled with baryons
and dark matter or as a universe filled with baryons and with a higher
effective value of $G_N$. Equation (\ref{Chap}) does not fulfill
(\ref{hpm}) because it describes the behavior of just baryon
density. In the $H\gtrsim H_-$ limit, the model can be interpreted as
a universe filled with baryons and a cosmological constant or as an
ACM model with $\alpha_-=1$ and filled only with baryons.
A similar example in the early period of accelerated expansion would
be a universe filled with a fluid which behaves as
a fluid of ultra-relativistic particles if the energy density is low
enough but whose density has an upper bound
\begin{equation}
\rho_X=\frac\sigma{a^4+C}\, .
\end{equation}
This dependence for the energy density in the scale factor follows
from the equation of state
\begin{equation}
p_X=\frac 13 \rho_X-\frac {4 C}{3 \sigma}\rho_X^2 \, .
\end{equation}
This model turns out to be equivalent to a universe filled with
radiation following eq. (\ref{alpha+}) with $\alpha_+=1$ and
$H_+^2=(\frac{8 \pi G_N}{3})\frac{\sigma}{C}$.
Another equivalent description would be to consider that the universe
is also filled with some self interacting scalar field $\varphi$ which
accounts for the discrepancy between the standard energy-momentum and
GR Einstein tensors. Given an arbitrary modified Friedman equation
(\ref{rho_g}) a potential $V(\varphi)$ can be found such that the
cosmologies described by both models are the same. The procedure is
similar to the one used to reconstruct a potential from a
given cosmology \cite{Nojiri:2005pu}. From the point of
view of the scalar field, the cosmology is defined by a set of three
coupled differential equations
\begin{eqnarray}
\ddot{\varphi}+3 H \dot{\varphi}+V'(\varphi)&=0 \label{klein}\\
\frac{8\pi G_N}{3}(\rho_d+\frac{\dot{\varphi}^2}2 +V(\varphi))&=H^2
\label{quint}\\ -3(1+\omega)H\rho_d&=\dot{\rho_d} \label{contw}\, .
\end{eqnarray}
The solution of these equations for a certain $V(\varphi)$ will give
$H(t)$ and $\rho_d(t)$, and therefore $\rho_d(H)$ which is $g(H)$ up
to a factor $\frac{8 \pi G_N}{3}$. In this way one finds the
generalized first Friedman equation associated with the introduction
of a self interacting scalar field.
Alternatively, given a function $g(H)$ in a generalized first Friedman
equation (\ref{rho_g}) for the density $\rho_d$, one can find a scalar
field theory leading to the same cosmology in the homogeneous
limit. By considering the time derivative of (\ref{quint}) and using
(\ref{klein}) and (\ref{contw}) one gets
\begin{equation}
\dot{\varphi}^2=-(1+\omega)\rho_d-\frac{\dot{H}}{4\pi G_N} \label{dtvarphi}\, ,
\end{equation}
where we can use (\ref{dH}) and (\ref{rho_g}) to get $\dot{\varphi}^2$
as a function of $H$. Setting this on (\ref{quint}) we get
$V(\varphi)$ as a function of $H$. On the other hand,
$\varphi'(H)=\dot{\varphi}/\dot{H}$, so
\begin{equation}
\frac{8\pi G_N}{3}(\varphi'(H))^2=\frac{(2H
-g'(H))g'(H)}{9(1+\omega) H^2 g(H)}
\label{phi}
\end{equation}
and
\begin{equation}
\frac{8\pi G_N}{3}V(\varphi(H))=H^2-g(H)+\frac{(1+\omega)(2H-g'(H))g(H)}{2
g'(H)}\, .
\label{V(phi)}
\end{equation}
If $g(H)$ is such that the rhs of (\ref{phi}) is positive definite, it
can be solved and the solution $\varphi(H)$ inverted and substituted
into (\ref{V(phi)}) in order to get $V(\varphi)$. This is the case of
a function of the type (\ref{alpha-}) with $\alpha_->1$. For the
particular case of an ACM expansion with $\alpha_-=1$, the solution is
a flat potential $V=V_0$ and a constant value of $\varphi=\varphi_0$.
If $g(H)$ is such that the rhs of (\ref{phi}) is negative definite, as
it happens in (\ref{alpha-}) with $\alpha_-<1$ or in (\ref{alpha+}),
the problem can be solved by changing the sign of the kinetic term in
(\ref{quint}). The result is a phantom quintessence in the case of
(\ref{alpha-}) with $\alpha_-<1$. The case (\ref{alpha+}) is more
complicated because the energy density of the associated inflaton
turns out to be negative. Moreover, it is of the same order as the
energy density of ultra-relativistic particles during the whole period
of accelerated expansion. Therefore, it seems that this model is
inequivalent to other inflation scenarios previously studied.
In summary, there are many equivalent ways to describe the discrepancy
between observed matter content of the universe and Einstein's General
Relativity. At the homogeneous level, it is trivial to find relations
among them. The possibility to establish the equivalence of different
descriptions is not a peculiarity of the description of this discrepancy
in terms of a generalized first Friedman equation (\ref{rho_g}). The
same relations can be found among generalized equations of state,
scalar-tensor theories and f(R) modified gravity \cite{Capozziello:2005mj}.
\section{Summary and discussion}
It may be interesting to go beyond $\Lambda$CDM in the description of
the history of the universe in order to identify the origin of the
two periods of accelerated expansion. We have proposed to use the
expression of the energy density as a function of the Hubble parameter
as the best candidate to describe the history of the universe. In this
context the late time period of accelerated expansion and the early
time inflation period can be easily parametrized.
We have considered a simple modification of the cosmological equations
characterized by the appearance of an upper and a lower bound on the
Hubble expansion rate. A better fit of the experimental data can be obtained
with this extended cosmological model as compared with the
$\Lambda$CDM fit. Once more precise data are available, it will be
possible to identify the behavior of the energy density as a function
of the Hubble parameter and then look for a theoretical derivation of
such behavior.
We plan to continue with a systematic analysis of different
alternatives incorporating the main features of the example considered
in this work. Although the details of the departures from the standard
cosmology can change, we expect a general pattern of the effects induced
by the presence of the two bounds on $H$. We also plan to go further,
considering the evolution of inhomogeneities looking for new consequences
of the bounds on $H$.
The discussion presented in this work, which is based on a new description
of the periods of accelerated expansion of the universe, can open a new way
to explore either modifications of the theory of gravity or new components
in the universe homogeneous fluid. Lacking theoretical criteria to select
among the possible ways to go beyond $\Lambda$CDM, we think that the
phenomenological approach proposed in this work is justified.
We are grateful to Paola Arias and Justo L\'opez-Sarri\'on for
discussions in the first stages of this work. We also acknowledge
discussions with Julio Fabris, Antonio Segu\'{\i}, Roberto Empar\'an,
Sergei Odintsov, Aurelio Grillo and Fernando M\'endez.
This work has been partially supported by CICYT (grant
FPA2006-02315) and DGIID-DGA (grant2008-E24/2). J.I. acknowledges a
FPU grant from MEC.
\chapter{Linearized Treatment of Scalar perturbations in the Asymptotic Cosmological Model}
In this paper the implications of a recently proposed
phenomenological model of cosmology, the Asymptotic Cosmological
Model (ACM), on the behavior of scalar perturbations are
studied. Firstly we discuss new fits of
the ACM at the homogeneous level, including fits to the Type Ia
Supernovae UNION dataset, first CMB peak of WMAP5 and BAOs. The
linearized equations of scalar perturbations in the FRW metric are
derived. A simple model is used to compute the CMB
temperature perturbation spectrum. The results are compared with the
treatment of perturbations in other approaches to the problem of the
accelerated expansion of the universe.
\section{Introduction}
According to General Relativity (GR), if the universe is filled with the particles
of the Standard Model of particle physics, gravity should lead to a
decelerated expansion of the universe. However, in 1998 two
independent evidences of present accelerated expansion were presented
\cite{Riess:1998cb,Perlmutter:1998np} and later confirmed by different
observations \cite{Jaffe:2000tx,Spergel:2003cb,Peacock:2001gs}.
There is no compelling explanation for this cosmic acceleration,
but many intriguing ideas are being explored. These ideas can be classified
into three main groups: new exotic sources of the gravitational field with
large negative pressure \cite{Turner:1998ex,Zlatev:1998tr,ArmendarizPicon:2000dh} (Dark Energy), modifications of
gravity at large scales \cite{Bekenstein:1984tv,Bekenstein:2004ne} and rejection of the spatial
homogeneity as a good approximation in the description of the present
universe \cite{Tomita:2000jj,Kolb:2005da,Wiltshire:2007zj}.
Different models (none of them compelling) for the source responsible
of acceleration have been considered. Einstein
equations admit a cosmological constant $\Lambda$, which can be realized as
the stress-energy tensor of empty space. This $\Lambda$ together with Cold Dark
Matter, Standard Model particles and General Relativity form the current
cosmological model, $\Lambda$CDM. However, quantum field theory predicts a
value of $\Lambda$ which is 120 orders of magnitude higher than observed.
Supersymmetry can lower this value 60 orders of magnitude, which is
still ridiculous \cite{Weinberg:1988cp}.
In order to solve this paradox, dynamical Dark Energy models have been proposed.
This has also lead to explore the possibility that cosmic acceleration arises
from new gravitational physics. Again here several alternatives for a
modification of the Einstein-Hilbert action at large and small
curvatures \cite{Faraoni:2008mf}, or even higher
dimensional models \cite{Deffayet:2000uy,Dvali:2000hr}, producing an accelerated
expansion have been identified.
All these analyses include an {\it ad hoc} restriction to actions involving simple
functions of the scalar curvature and or the Gauss-Bonnet tensor.
This discussion is sufficient to establish the point that cosmic acceleration can
be made compatible with a standard source for the gravitational field but it is
convenient to consider a more general framework in order to make a systematic
analysis of the cosmological effects of a modification of general relativity.
The Asymptotic Cosmological Model (ACM) was presented
\cite{Cortes:2008fy} as a strictly phenomenological
generalization of the Standard Cosmological Model including a
past and a future epoch of accelerated expansion. It follows from the
assumptions that GR is not a fundamental theory, but only a good
approximation when the Hubble rate $H$
is between but far away from two
fundamental scales $H_-$ and $H_+$, which act as bounds on
$H$. A general covariant metric theory of gravity without
spatial curvature is assumed. The model is well defined in the
homogeneous approximation
and includes $\Lambda$CDM as a particular case.
In next section we review the ACM and we provide new fits
to the Type Ia Supernovae UNION dataset, first acoustic peak of CMB of
WMAP5 and BAOs. In the third section we
derive the linearized equations of the scalar perturbations of the
metric, following from
general covariance and a single new assumption on the
perturbations. In the fourth section
we will consider how to solve the system of equations for adiabatic
perturbations in a given fluid. In the fifth section we derive the
CMB spectrum in the ACM. In the sixth
section we will compare the treatment of the scalar perturbations in
the ACM with other models
which are equivalent in the homogeneous approximation. The last
section is devoted to the summary and conclusions.
\section{The Asymptotic Cosmological Model: Homogeneous Background}
The Asymptotic Cosmological Model (ACM) was introduced in
Ref. \cite{Cortes:2008fy}. In this model the universe is filled with
photons and neutrinos (massless particles), baryons (electrically
charged massive particles) and Dark Matter (electrically neutral
massive particles), but General Relativity (GR) is only a good
approximation to the gravitational interaction in a certain range of
the Hubble rate $H$, between but far from its
two bounds, $H_-$ and $H_+$. General Covariance and the absence of
spatial curvature are assumed.
The gravitational part of the action might include derivatives of the
metric of arbitrarily high order, and therefore arbitrarily high
derivatives of the scale factor should appear in the Friedman
Equations. We should start by considering a generalized first Friedman
equation
\begin{equation}
8\pi G \rho(t) = 3 f(H(t),\dot{H}(t),...,H^{(n)}(t),...) \label{fh} \, ,
\end{equation}
and the corresponding equation for the pressure will be derived using the continuity
equation
\begin{equation}
\dot{\rho} = -3 H(\rho + p)\, . \label{cont}
\end{equation}
However, as the resulting differential equations should be solved and only one of
its solutions deserves interest (the one describing the
evolution of the universe), we can use the one to one
correspondence between time $t$ and Hubble parameter $H$ (assuming $\dot{H} < 0$)
to write the modified First Friedman Equation as a bijective map linking
the total energy density $\rho$ with $H$ in our universe
\begin{equation}
8\pi G \rho = 3 g(H) \label{gh} \, .
\end{equation}
The use of the continuity equation \eqref{cont} enables us to write the modified
Friedman Equation for the pressure evaluated at the solution
corresponding to the cosmic evolution
\begin{equation}
-8\pi G p = 3 g(H) + g'(H)\dot{H}/H \label{ph} \, .
\end{equation}
The evolution of the universe will be determined by the concrete form of
the function $g(H)$, which we assume to be smooth. However, the most significant
features of the evolution at a given period can be described by some
simple approximation to $g(H)$ and the matter content. Those are a
pole at $H = H_+$ of order $\alpha_+$ and a zero of order $\alpha_-$ at $H =
H_-$. The energy density of the universe has a contribution from
both massless particles (radiation) and massive ones (matter).
For the sake of simplicity, the history of the universe can be then divided into
three periods. In the first period, $H \gg H_-$ so the effect of the lower bound
can be neglected and the universe is radiation dominated. The universe undergoes
(and exits from) an accelerated expansion which we call inflation. The simplest
parametrization is
\begin{equation}
g(H) \approx H^2 \left( 1 - \frac{H^2}{H_+^2} \right)^{-\alpha_+} \, . \label{past}
\end{equation}
In the second period we cannot neglect neither the effect of radiation nor the
effect of nonrelativistic matter, but $H_+ \gg H \gg H_-$. Then GR
offers a good description
of the gravitational interaction, $g(H) \approx H^2$ in this region, and the
universe performs a decelerated expansion.
In the third period, $H_+ \gg H$ so the effect of the upper bound can be
neglected, and the universe is matter dominated. This period corresponds to
the present time in which the universe also undergoes an accelerated expansion. The
simplest choice is
\begin{equation}
g(H) \approx H^2 \left( 1 - \frac{H_-^2}{H^2} \right)^{\alpha_-} \, . \label{future}
\end{equation}
We will assume that the transitions between these three periods are smooth
and that the details about these transitions are unimportant.
This model preserves the successes of the Cosmological Standard Model, while
giving a description of the early accelerated expansion
(inflation) and of the present one, including $\Lambda$CDM as a particular case
($\alpha_-= 1$).
Without any knowledge of the evolution of perturbations in this model, the
background evolution at late times can be used to fit the Type Ia Supernovae
UNION dataset \cite{Kowalski:2008ez}, the first acoustic peak in the Cosmic
Microwave (CMB) Background \cite{Dunkley:2008ie}, and the Baryon Acoustic
Oscillations (BAOs) \cite{Eisenstein:2005su}, via the parameters $H_0$
(the present Hubble parameter), $H_-$ and $\alpha_-$.
We use Monte Carlo Markov Chains to explore the likelihood of the fit of the
supernovae UNION dataset \cite{Kowalski:2008ez} to the ACM. The dataset provides
the luminosity distance
\begin{equation}
d_l(z) = c(1+z)\int_0^z dz'/H(z')
\end{equation}
and redshift of 307 supernovae. A $\chi^2$ analysis have been performed, where
$\chi^2$ has been marginalized over the nuisance parameter $H_0$ using the method
described in \cite{Wang:2004xz}. The resulting parameter space is spanned by the values
of $\alpha_-$ and $H_-/H_0$ (FIG. 1).
\begin{figure}
\centerline{\includegraphics[scale=0.6, width=8cm]{UNION.png}}
\label{U}
\caption{Confidence regions in parameter space of the Asymptotic
Cosmological Model (ACM) from the fit of the supernovae UNION dataset
without priors at $1\sigma$, $2\sigma$ and $3\sigma$. The $\Lambda$CDM is
inside the 1$\sigma$ region. The best fit to ACM lies in
$\alpha_- = 0.35$, $H_-/H_0 = 0.97$
($\chi^2 = 310.5$) in contrast to the best fit to $\Lambda$CDM, which lies in
$\alpha_- \equiv 1$ , $H_-/H_0 = 0.84$
($\chi^2 = 311.9$).}
\end{figure}
The constraints from the CMB data follow from the reduced distance to the surface
of last scattering at $z=1089$. The reduced distance $R$ is often written as
\begin{equation}
R = \Omega_m^{1/2} H_0 \int_0^{1089} dz/H(z) \, .
\end{equation}
The WMAP-5 year CMB data alone yield $R_0 = 1.715\pm 0.021$ for a fit assuming a
constant equation of state $\omega$ for the dark energy \cite{Komatsu:2008hk}. We
will take this value as a first approximation to the fit assuming ACM. We can
define the corresponding $\chi^2$ as $\chi^2 = [(R-R_0)/\sigma_{R_0}]^2$, and find
the confidence regions of the joint constraints (FIG. 2).
\begin{figure}
\centerline{\includegraphics[scale=0.6, width=8cm]{UNION+WMAP5.png}}
\label{UW}
\caption{Confidence regions in parameter space of the Asymptotic
Cosmological Model (ACM) from the fit of the supernovae UNION
dataset (red), of the distance to the surface of last scattering from WMAP-5
(blue) and from the joint fit (magenta) at $1\sigma$, $2\sigma$ and $3\sigma$.
Significantly, the $\Lambda$CDM is still inside the 1$\sigma$ region, unlike in
our previous study. This is due to the change in the Supernovae dataset. The
best fit to ACM lies in $\alpha_- = 1.50$,
$H_-/H_0 = 0.77$ ($\chi^2 = 312.5$) in contrast to the best fit to $\Lambda$CDM,
which lies in $\alpha_- \equiv 1$, $H_-/H_0 = 0.86$
($\chi^2 = 313.2$).}
\end{figure}
BAO measurements from the SDSS data provide a
constraint on the distance parameter
$A(z)$ at redshift $z=0.35$,
\begin{equation}
A(z)= \Omega_m^{1/2} H_0 H(z)^{-1/3} z^{-2/3} \left[\int_0^z dz'/H(z')\right]^{2/3} \, .
\end{equation}
Ref.~\cite{Eisenstein:2005su} gives $A_0 = 0.469 \pm 0.17$. We can define the
corresponding $\chi^2$ as $\chi^2 = [(A(z\,=\,0.35)-A_0)/\sigma_{A_0}]^2$ . The
confidence regions resulting from adding this constraint are shown in FIG. 3.
\begin{figure}
\centerline{\includegraphics[scale=0.6, width=8cm]{UNION+WMAP5+BAO.png}}
\label{UWB}
\caption{Confidence regions in parameter space of the Asymptotic
Cosmological Model (ACM) from the fit of the supernovae UNION dataset (red),
of the distance to the surface of last scattering from WMAP-5 (blue), of
the Baryon Acoustic Oscillations peak (green) and from the joint fit
(grey) at $1\sigma$, $2\sigma$ and $3\sigma$. Measurements of the
BAOs peak do not add significant information with their present
precision. The best fit to ACM lies in
$\alpha_- = 1.50$, $H_-/H_0 = 0.77$
($\chi^2 = 312.7$) in contrast to the best fit to $\Lambda$CDM, which lies
in $\alpha_- \equiv 1$, $H_-/H_0 = 0.86$
($\chi^2 = 313.7$).}
\end{figure}
In our previous work the use of the supernovae Gold dataset
\cite{Riess:2006fw} and of the WMAP-3 data \cite{Spergel:2006hy} led us to the
conclusion that the $\Lambda$CDM was at $3\sigma$ level in the
parameter space of the ACM. The position of the confidence regions
seems to depend very tightly on the dataset that is being
used. However, the value of the combination of parameters
\begin{equation}
\Omega_m \equiv \left(1-\frac{H_-^2}{H_0^2}\right)^{\alpha_-} \label{omegam}
\end{equation}
does not depend much neither on the value of $\alpha_-$ or the
dataset used (FIG. 4).
Moreover, we can conclude from the figures that
BAO's do not provide much information in order to constrain the
confidence regions of the ACM, unlike in other models such as
$\Lambda$CDM with nonzero spatial curvature.
\begin{figure}
\centerline{\includegraphics[scale=0.6, width=8cm]{omega2.png}}
\label{omega}
\caption{Confidence regions in parameter space of the Asymptotic
Cosmological Model (ACM) from the fit of the supernovae UNION dataset (red),
of the distance to the surface of last scattering from WMAP-5
(blue), of the Baryon Acoustic Oscillations peak (green) and from
the joint fit (grey) at $1\sigma$, $2\sigma$ and $3\sigma$. The plot
shows $\alpha_-$ in the horizontal axis and $\Omega_m$ in the
vertical axis. Noticeably, $\Omega_m = 0.26 \pm 0.04$ almost
independently of the value of $\alpha_-$.}
\end{figure}
\section{Scalar Perturbations}
The lack of an action defining the ACM is a serious obstacle in
the derivation of the equations governing the behavior of the
perturbations. Given a background behavior described by the ACM,
what can be said about the evolution of perturbations on top of this
background? We will find that general covariance together with an
additional assumption fixes completely the set of equations for the
scalar perturbations in the linearized approximation and in the region
close to $H_-$.
The key point will be the following. Knowing the exact Friedman
equations in the homogeneous approximation gives us a clue on the form
of the equations for the scalar perturbations. In particular, we can
formally describe perturbations over the FRW metric which do not depend
on the spatial coordinates, $\{\phi(\mathbf{x},t)=\phi(t),
\psi(\mathbf{x},t)=\psi(t)\}$. Thus the perturbed
metric becomes the FRW metric
written in a new coordinate frame. With a change
of coordinates one can derive, starting from the Friedman equations,
the terms containing only time derivatives of the scalar
perturbations. Next an assumption on the validity of the GR
description of scalar perturbations when $H_-\ll H \ll H_+$,
together with the relations, valid for any general covariant theory,
between terms with time derivatives and those involving spatial
derivatives, allow to derive the evolution of scalar perturbations.
After this short sketch, we will perform the derivation of the
equations for the scalar perturbations in detail. We can write the
metric of spacetime with scalar perturbations in the Newtonian gauge,
\begin{equation}
ds^2 = (1+2\phi(\mathbf{x},t))dt^2 - a^2(t) (1-2\psi(\mathbf{x},t)) d\mathbf{x}^2 \, ,\label{newt}
\end{equation}
and the stress-energy tensor of the source fields will be
\begin{equation}
\begin{array}{ccl}
T^0_0 & = & \rho_{(0)}(t) + \delta \rho(\mathbf{x},t) \, , \\
T^0_i & = & (\rho_{(0)}(t) + p_{(0)}(t)) \partial_i \theta(\mathbf{x},t) \, ,\\
T^i_j & = & -(p_{(0)}(t) +\delta p(\mathbf{x},t))\delta^i_j\,+\,\partial^i\partial_j\Pi(\mathbf{x},t)\, ,
\end{array} \label{set}
\end{equation}
where $\theta$ is the velocity potential of the fluid, $\Pi$ is the
anisotropic stress tensor (shear) potential, $\delta\rho$ and $\delta
p$ are small perturbations on top of the background homogeneous
density $\rho_{(0)}$ and pressure $p_{(0)}$, respectively, and
$\phi\sim\psi\ll 1$. We have used the Newtonian gauge because the
gauge invariant scalar perturbations of the metric ($\Phi$ and $\Psi$)
and gauge invariant perturbations in the stress energy tensor
coincide with the perturbations explicitly written in this gauge.
The general form of the equations for the perturbations in a metric
theory with arbitrarily high derivatives is
\begin{eqnarray}
8 \pi G \delta\rho & = & \sum_{n=0}^\infty \sum_{m=0}^\infty
\left[\frac{a_{nm}}{H^{2n+m-2}}\left(\frac\Delta{a^2}\right)^n
\partial_t^m \phi +
\frac{b_{nm}}{H^{2n+m-2}}\left(\frac\Delta{a^2}\right)^n
\partial_t^m \psi \right] \label{rho1} \, ,\\
8 \pi G (\rho_{(0)} + p_{(0)}) \partial_i \theta & = &
\partial_i\sum_{n=0}^\infty \sum_{m=0}^\infty
\left[\frac{c_{nm}}{H^{2n+m-1}}\left(\frac\Delta{a^2}\right)^n
\partial_t^m \phi +
\frac{d_{nm}}{H^{2n+m-1}}\left(\frac\Delta{a^2}\right)^n
\partial_t^m \psi \right] \label{vel1} \, ,\\
8 \pi G (-\delta p \,\delta^i_j + \partial_i \partial_j \Pi) & = &
\delta^i_j\sum_{n=0}^\infty \sum_{m=0}^\infty
\left[\frac{e_{nm}}{H^{2n+m-2}}\left(\frac\Delta{a^2}\right)^n
\partial_t^m \phi +
\frac{f_{nm}}{H^{2n+m-2}}\left(\frac\Delta{a^2}\right)^n
\partial_t^m \psi \right] \nonumber \\
& & +\sum_{n=0}^\infty \sum_{m=0}^\infty \frac 1{a^2}(\delta^i_j
\Delta -\partial_i
\partial_j)\left[\frac{g_{nm}}{H^{2n+m}}\left(\frac\Delta{a^2}\right)^n
\partial_t^m \phi +
\frac{h_{nm}}{H^{2n+m}}\left(\frac\Delta{a^2}\right)^n \partial_t^m
\psi \right] \label{p1} \, .
\end{eqnarray}
The coefficients $a_{nm},...,h_{nm}$ are adimensional functions of the
Hubble parameter an its time derivatives, and can be turned into
functions of just the Hubble parameter using the bijection explained
in the beginning of the previous section. We have mentioned we are
going to be able to determine exactly the terms with only time derivatives, that
is, the precise form of the coefficients $a_{0m},b_{0m},e_{0m}$ and
$f_{0m}$. This shows that there is much freedom of choosing a
covariant linearized theory of cosmological perturbations, even for a
given solution of the homogeneous equations.
However, under a single assumption, it is possible to greatly reduce
this freedom. The assumption is that the standard linearized equations
of the perturbations of General Relativity are effectively recovered
in the limit $H_+ \gg H \gg H_-$ for all the Fourier modes with $H_+
\gg k > H_-$ (subhorizon modes in the present time).
Let us work out the implications of this assumption. In the equation
for the perturbation of the energy density in General relativity,
\begin{equation}
8\pi G \delta \rho = -6 H^2 \phi -6 H \dot \psi +\frac{2}{a^2}\Delta \psi\, ,
\end{equation}
there is only a term proportional to $\Delta \psi$. However, for
subhorizon modes ($k\gg H$) the dominant terms are those with the
highest number of spatial derivatives. If we demand
the terms with spatial derivatives of order greater than $2$ not to
spoil the behavior of these modes in the period in which GR is a good
approximation, they must be negligible at least for the observable modes.
For instance, they should
be negligible for the modes responsible of the acoustic peaks of the
CMB spectrum, and therefore for the modes which have entered the
horizon after recombination (which have even
lower $k$). It is possible that these terms
are suppressed by inverse powers of the UV scale
$H_+$, becoming irrelevant for
current tests of gravity, although they may be relevant for the
physics of quantum fluctuations in the very early universe.
Therefore, at times when $H\ll H_+$, the equations can be approximated by
\begin{eqnarray}
8 \pi G \delta\rho & = & \sum_{m=0}^\infty
\left[\frac{a_{0m}}{H^{m-2}}\partial_t^m \phi +
\frac{b_{0m}}{H^{m-2}}\partial_t^m \psi +
\frac{a_{1m}}{H^{m}}\frac\Delta{a^2} \partial_t^m \phi +
\frac{b_{1m}}{H^{m}}\frac\Delta{a^2}\partial_t^m \psi \right]
\label{rho2} \, ,\\
8 \pi G (\rho_{(0)} + p_{(0)}) \partial_i \theta & = &
\partial_i\sum_{m=0}^\infty \left[\frac{c_{0m}}{H^{m-1}}\partial_t^m
\phi + \frac{d_{0m}}{H^{m-1}}\partial_t^m \psi \right] \label{vel2}
\, ,\\
8 \pi G (-\delta p \,\delta^i_j + \partial_i \partial_j \Pi) & = &
\delta^i_j \sum_{m=0}^\infty \left[\frac{e_{0m}}{H^{m-2}}\partial_t^m
\phi + \frac{f_{0m}}{H^{m-2}}\partial_t^m \psi \right] \nonumber \\
& & +\sum_{m=0}^\infty \frac 1{a^2}(\delta^i_j \Delta -\partial_i
\partial_j)\left[\frac{g_{0m}}{H^{m}} \partial_t^m \phi +
\frac{h_{0m}}{H^{m}}\partial_t^m \psi \right] \label{p2} \, ,
\end{eqnarray}
where general covariance has been used to exclude the terms with
coefficients $e_{1m}$, $f_{1m}$ which are not compatible with
(\ref{rho2}), (\ref{vel2}).
We must also take into account that the modes which are responsible of
the acoustic peaks of the CMB undergo a phase in which they oscillate
as sound waves, i.e.: $\partial_t\phi~\sim~k\phi$. In order to
explain the acoustic peaks of the spectrum of the CMB it is required
that the acoustic oscillations of the modes of the gravitational
potentials, $\phi_\mathbf{k}$ and $\psi_\mathbf{k}$, which lead to the acoustic
peaks of the CMB spectrum, have a frequency $\sim k$. When radiation
dominates and we consider modes well inside the horizon, two of the
solutions of the system of differential equations of arbitrary order
have this property \footnotemark[1].
However, in order for nondecaying superhorizon modes to evolve
into oscillating subhorizon modes with frequency $\sim k$, a fine-tuning of
coefficients is required unless the equations are of order two in time
derivatives. Then the acoustic peaks in the CMB spectrum are
reproduced when terms with more than two derivatives in the equations
for the perturbations can be neglected. Then
\begin{eqnarray}
8 \pi G \delta\rho & = & \sum_{m=0}^2
\left[a_{0m}H^{2-m}\partial_t^m \phi +
b_{0m}H^{2-m}\partial_t^m \psi \right]+
a_{10}\frac\Delta{a^2}\phi + b_{10}\frac\Delta{a^2}\psi \label{rho3}
\, ,\\
8 \pi G (\rho_{(0)} + p_{(0)}) \partial_i \theta & = &
\partial_i\sum_{m=0}^1 \left[c_{0m}H^{1-m}\partial_t^m \phi +
d_{0m}H^{1-m}\partial_t^m \psi \right] \label{vel3} \, ,\\
8 \pi G (-\delta p \,\delta^i_j + \partial_i \partial_j \Pi) & = &
\delta^i_j \sum_{m=0}^2 \left[e_{0m}H^{2-m}\partial_t^m \phi
+ f_{0m}H^{2-m}\partial_t^m \psi \right] \nonumber \\
& & +\frac 1{a^2}(\delta^i_j \Delta -\partial_i
\partial_j)\left[g_{00}\phi + h_{00}\psi \right] \label{p3} \, ,
\end{eqnarray}
\footnotetext[1]{
As we will see
below, if the term with the highest time derivative is of order
$D$, $a_{1,D-2} = -e_{0D}$ as a consequence of general
covariance. Well inside the horizon these are the dominant terms
appearing in the wave equation resulting when calculating the
adiabatic perturbations in the radiation domination
epoch.}
\noindent and the arbitrariness in the evolution equations for the scalar
perturbations has been reduced to twenty undetermined dimensionless
coefficients at this level.
The equations \eqref{rho3},\eqref{vel3},\eqref{p3} are in principle
valid for any perturbation mode $\{\phi_\mathbf{k},\psi_\mathbf{k}\}$ as long as $k\ll H_+$.
In particular, it must be valid for the mode $\mathbf{k} = \mathbf{0}$, which
corresponds formally to a perturbation with no spatial dependence. In practice,
we will be able to neglect the spatial dependence of perturbation whose spatial
dependence is sufficiently smooth, i.e.: its wavenumber $k$ is sufficiently low.
In GR it suffices for a mode to be superhorizon $k\ll H$ in order to neglect
its spatial dependence in a first approximation.
If we consider the FRW metric perturbed by one of these modes we have,
\begin{equation}
ds^2 = (1+2\phi(t))dt^2 - a^2(t) (1-2\psi(t)) d\mathbf{x}^2 \, . \label{superh}
\end{equation}
By means of the invariance under time reparameterizations we can introduce
a new time variable $dt'= (1+\phi(t))dt$, a new scale factor
$a'(t')=a(t)(1-\psi(t))$ and a new energy density
$\rho'(t')=\rho_0(t)+\delta\rho(t)$
leading us back to the ACM in a homogeneous background. The Hubble
rate $H = \frac{\dot a}{a}$ and its time derivatives $H^{(n)}$ will change as
\begin{equation}
\begin{array}{l}
H^{(n)'} =
\frac{d^n}{dt^{'n}}\left[\frac{1}{a'(t')}\frac{da'}{dt'}\right] =
H^{(n)} +\delta H^{(n)}, \\
\delta H^{(n)} = -\psi^{(n+1)}
-\sum_{m=0}^n
\left(\begin{array}{c}
n+1 \\ m+1
\end{array}
\right) H^{(n-m)}\phi^{(m)}.\label{Hn}
\end{array}
\end{equation}
Noticeably, this procedure must be applied to the generalized Friedman
equation before one of its homogeneous solutions is used to write the
time variable $t$ as a function of $H$. Therefore we should start by
considering a generalized first Friedman equation in primed
coordinates
\begin{equation}
8\pi G \rho'(t') = 3 f(H'(t'),\dot{H}'(t'),...,H^{(n)'}(t'),...) \label{fh'} \, ,
\end{equation}
and the corresponding equation for the pressure will be derived using
the continuity equation. The linearized equations for the sufficiently smooth
scalar perturbations in the Newtonian gauge~(\ref{superh}) can then be
derived with the help of~(\ref{Hn}). The result is
\begin{eqnarray}
8\pi G \delta \rho(t) & = & \sum_{n = 0}^\infty g_{\vert n}(H(t))\delta H^{(n)}(t),
\label{rho4} \\
-8\pi G \delta p & = & \frac 1H \sum_{n = 0}^\infty\sum_{m = 0}^\infty
g_{\vert nm}(H)H^{n+1}\delta H^{(m)}\nonumber\\
& &+\sum_{n = 0}^\infty g_{\vert n}(H) \left(\frac{\delta
H^{(n+1)}}H-\frac{H^{(n+1)}}{H^2}\delta H \right)\nonumber\\
& &+ 3\sum_{n = 0}^\infty g_{\vert n} (H)\delta H^{(n)},\label{p4}
\end{eqnarray}
where
\begin{eqnarray}
g_{\vert n} (H) & = & \left[ \frac{\partial f}{\partial
H^{(n)}}\right]_{H^{(n)} = H^{(n)}(t(H))}\, ,\\
g_{\vert nm}(H) & = & \left[ \frac{\partial^2 f}{\partial
H^{(n)}\partial H^{(m)}}\right]_{H^{(n)} = H^{(n)}(t(H))} \, ,
\end{eqnarray}
and the time dependence of $H$ is derived from the
homogeneous evolution \eqref{gh}. This procedure fixes exactly the
coefficients $a_{0m},b_{0m},e_{0m}$ and $f_{0m}$ in
\eqref{rho2},\eqref{vel2},\eqref{p2}, i.e. all the terms with no
spatial derivatives in the equations for the scalar perturbations, as
functions of $f(H,\dot{H},...)$ and its partial derivatives.
We can now count the number of functional degrees of freedom in the
superhorizon modes of the linearized theory. Let us assume
that there is a maximum number $D$ of time derivatives in the equations
for the perturbations~\footnotemark[2].\footnotetext[2]{We are
forgetting here the restriction on the
number of derivatives required in order to reproduce the acoustic peaks of
CMB.} The number of derivatives, with the use of
\eqref{Hn},\eqref{rho4},\eqref{p4}, fixes $f$ to be a function of at
most $H^{(D-2)}$. When particularized to a solution of the homogeneous
Friedman equations, $f$ and its first and second partial derivatives
become functions of the Hubble rate $H$: $g(H)$, $g_{\vert n}(H)$ and
$g_{\vert nm}(H)$ respectively. That makes $1+(D-1)+D(D-1)/2=D(D+1)/2$
functional degrees of freedom. The first derivative of $g(H)$ can be
written in terms of the $g_{\vert n}(H)$, and the first derivative of the
later can be also written in terms of the $g_{\vert nm}(H)$. That makes $1 +
(D-1)$ conditions, so the result depends on $D(D-1)/2$ independent
functions of the Hubble parameter $H$ in the terms without any spatial
derivatives.
Notice however that if we restrict the number of derivatives appearing
in the equations for the perturbations to two as in
Eqs.~\eqref{rho3},\eqref{vel3},\eqref{p3}, then
Eqs.~\eqref{Hn},\eqref{rho4},\eqref{p4} tell us that we must consider
just functions $f(H,\dot H, \ddot H,...)\, =\, f(H)\, = \, g(H)$, at
least as an approximation at times $H\ll H_+$ and modes $k \ll
H_+$. Therefore the number of functional degrees of freedom in the
terms with no spatial derivatives is just one: the homogeneous
evolution $f(H) = g(H)$.
Our assumption might be relaxed and we could impose that general
relativity should be valid for all the modes $H_-<k<k_{obs}$ that have
been observed in the spectrum of CMB and matter perturbations. This
could lead to new terms in the equations for the perturbations
(suppressed not necessarily by the UV scale $H_+$) which have been
negligible for the observed modes but that could lead to ultraviolet
deviations from the spectrum derived in the general relativistic
cosmology which have not yet been observed. However, terms with more
than two spatial derivatives will be very tightly constrained by solar
system experiments.
In this article we will restrict ourselves to the system of second order
differential equations (\ref{rho3}), (\ref{vel3}),
(\ref{p3}).
Let us now derive the linearized equations for
the rest of the modes under this assumption. In the $(t',\mathbf{x})$
coordinate system, the metric is Friedman Robertson Walker, and we
know that for this metric, we can use equation (\ref{gh}). Thus,
\begin{equation}
8 \pi G \rho' = 3 g(H') \, ,
\end{equation}
with
\begin{equation}
H' = \frac 1 {a'} \frac{d a'}{d t'} = H(1-\phi)-\dot{\psi}
\end{equation}
at linear order in perturbations. Therefore, we can deduce from
(\ref{gh}) that in the $(t,\mathbf{x})$ coordinate system
\begin{eqnarray}
8 \pi G \delta\rho & = & -3 g'(H)\left(H \phi\right) +\dot{\psi})+
a_{10}\frac\Delta{a^2}\phi + b_{10}\frac\Delta{a^2}\psi \label{rho5}
\, ,\\
8 \pi G (\rho_{(0)} + p_{(0)}) \partial_i \theta & = &
\partial_i\sum_{m=0}^1 \left[\frac{c_{0m}}{H^{m-1}}\partial_t^m \phi +
\frac{d_{0m}}{H^{m-1}}\partial_t^m \psi \right] \label{vel5} \, ,\\
8 \pi G (-\delta p \,\delta^i_j + \partial_i \partial_j \Pi) & = &
\delta^i_j \left[g'(H)\left(3 H \phi +3
\dot{\psi}-\frac{\dot{H}}{H^2}\dot{\psi}+\frac{\dot{H}}H \phi
+\dot{\phi} + \frac 1 H \ddot{\psi}\right) \right. \nonumber\\
& & \left. +\frac{\dot{H}}H g''(H)(H\phi+\dot{\psi}) \right] \nonumber \\
& & +\frac 1{a^2}(\delta^i_j \Delta -\partial_i
\partial_j)\left[g_{00}\phi + h_{00}\psi \right] \label{p5} \, ,
\end{eqnarray}
and the twelve coefficients of terms with no spatial derivatives
are fixed by the function $g(H)$ which defines the homogeneous cosmological
model.
The last requirement comes again from general covariance. If the
tensor $T_{\mu\nu}$ comes from the variation of a certain matter
action $S_m$ with respect to the inverse of the metric $g^{\mu\nu}$,
and $S_m$ is a scalar under general coordinate transformations, then
$T^\mu_\nu$ must be a 1-covariant, 1-contravariant divergenceless
tensor, i.e. $\nabla_\mu T^\mu_\nu = 0$. As the stress-energy tensor
is proportional to $G^\mu_\nu$, the later must also be
divergenceless. In the linearized approximation $G^\mu_\nu =
G^\mu_{(0)\nu}+\delta G^\mu_\nu$ and the Christoffel symbols
$\Gamma^\lambda_{\mu\nu} = \Gamma^\lambda_{(0)\mu\nu}+\delta
\Gamma^\lambda_{\mu\nu}$ , and we can write the linearized version of
this requirement as
\begin{equation}
\begin{array}{c}
\delta G^0_{\mu\vert 0} +\delta G^i_{\mu\vert i} + \delta
G^\lambda_\mu \Gamma^\nu_{(0)\lambda\nu} - \delta G^\lambda_\nu
\Gamma^\nu_{(0)\lambda\mu} \\
+ G^\lambda_{(0)\mu} \delta\Gamma^\nu_{\lambda\nu}
- G^\lambda_{(0)\nu} \delta\Gamma^\nu_{\lambda\mu} = 0 \, .
\end{array}
\end{equation}
Let us study this condition in order to see if we can further limit
the number of independent coefficients of
the equations for scalar perturbations. For $\mu = i$ the term
$\delta G^0_{i\vert 0}$ will give at most a term proportional to
$d_{01}\ddot{\psi}_{\vert i})$, which will not be present in other
terms except $\delta G^j_{i \vert j}$. This term will give at most a
term $g'\ddot{\psi}_{\vert i}/H$ which fixes the
exact value of $d_{01}$ for all $H \ll H_+$ . For the same reason,
the term $c_{01}\ddot{\phi}_{\vert i}$ in $\delta G^0_{i\vert 0}$
can not be canceled and the term $c_{00}H\dot{\phi}_{\vert i}$ can
only be canceled by the term $g'\dot{\phi}_{\vert i}$ in
$\delta G^j_{i \vert j}$, which means that $c_{01}=0$ and fixes
the exact value of $c_{00}$ for all $H\ll H_+$. The value
of $d_{00}$ can be fixed in
terms of $d_{01}$ and the terms proportional to $\dot{\psi}$ in
(\ref{p5}). This fixes (\ref{vel5}) completely in terms of the
function describing the homogeneous evolution, $g(H)$:
\begin{equation}
8 \pi G (\rho_0+p_0)\partial_i\bar{\theta} =
\frac{g'(H)}{H}\partial_i \left (H \Phi + \dot{\Psi} \right )
\label{vdef} \, .
\end{equation}
In the previous equation and from now on we will refer directly to the
gauge invariant counterparts of the variables in the Newtonian gauge,
which we will denote by $\Phi,\Psi,\bar{\theta},\bar{\delta\rho}$ and
$\bar{\delta p}$ ($\Pi$ is already gauge invariant). Now we can use
the divergenceless condition for $\mu=0$. The term $\delta G^0_{0\vert
0}$ will give at most a term proportional to $a_{10}\Delta
\dot{\phi}$ and a term proportional to $b_{10}\Delta \dot{\psi}$,
which will not be present in other terms except $\delta G^j_{0 \vert
j}$. This term will give at most a term proportional to $\Delta
\phi$ and a term proportional to $\Delta \dot{\psi}$. Therefore
$a_{10}=0$ and the value of $b_{10}$ is set completely in terms of
$g(H)$:
\begin{equation}
8 \pi G \bar{\delta \rho} = -3 g'(H)(H \Phi +\dot{\Psi})+ \frac
{g'(H)}{a^2 H}\Delta \Psi
\label{rhodef} \, .
\end{equation}
The complete knowledge of (\ref{rhodef}) and (\ref{vdef}) when $H \ll
H_+$ gives a complete knowledge of (\ref{p5}) in terms of $g(H)$ when
$H \ll H_+$:
\begin{equation}
\begin{array}{ccl}
8 \pi G (\bar{\delta p} \, \delta_{ij} + \partial_i \partial_j \bar{\Pi})
& = &
\delta_{ij}\left\{
g'(H)\left[3 H \Phi +3
\dot{\Psi}-\frac{\dot{H}}{H^2}\dot{\Psi}+\frac{\dot{H}}H \Phi
+\dot{\Phi} + \frac 1 H \ddot \Psi \right]
+\frac{\dot H}H g''(H)\left[H \Phi + \dot \Psi \right]
\right\}\\
& & +\frac 1{a^2}\left(\delta_{ij}\Delta -\partial_i\partial_j\right)
\left( \frac{g'(H)}{2 H}(\Phi-\Psi) + \frac{\dot H}{2 H^3}(g'(H) - H
g''(H))\Psi \right)
\end{array}\label{pdef} \, ,
\end{equation}
and we have finally a set of equations for the scalar perturbations
(\ref{rhodef}), (\ref{vdef}), (\ref{pdef}) which are determined by the
function $g(H)$ which defined the ACM in the homogeneous approximation.
Had we relaxed our assumption, we could work with the system of
equations~\eqref{rho1},\eqref{vel1},\eqref{p1} restricted to a maximum
number of derivatives $D$ as an approximation. Together with the
condition $G^\mu_{\nu;\mu} =0$, the system of equations for the
perturbations define a set of coupled differential equations for the
coefficients of the terms with at least one spatial derivative. The
set of equations coming from $G^\mu_{i;\mu}=0$ define the terms
proportional to $e_{mn}$ and $f_{mn}$ in \eqref{p1} as a function of the
terms in \eqref{vel1}. The set of equations coming from
$G^\mu_{0;\mu}$ define the rest of the terms in \eqref{p1} as a
function of the terms in \eqref{vel1} and \eqref{rho1}. Therefore, the
only freedom, for what concerns linearized perturbations, is that of
choosing the set of functions $\{a_{nm}, b_{nm}, c_{nm}, d_{nm}\}$,
with some of them fixed by the homogeneous dynamics \eqref{fh}. An
interesting case is found if $D=4$ is imposed. This includes a
description of $f(R)$-theories \cite{Faraoni:2008mf}, bigravity
theories \cite{Damour:2002wu}, and other of the most studied modified
gravity theories. We will study further this case
in the subsection devoted to the comparison of the model with $f(R)$
theories.
The rhs of the equations derived are the equivalent to the components
of the linearized Einstein tensor derived in a general covariant
theory whose Friedman equation is exactly (\ref{gh}). Aside from
General Relativity with or without a cosmological constant, i.e. for
$\alpha_-\neq 1, 0$, it will be necessary to build an action depending
on arbitrarily high derivatives of the metric in order to derive a
theory such that both the equations in the homogeneous approximation
and the linearized equations for the scalar perturbations are of
finite order.
Noticeably, the nonlinear equations for the perturbations will include
derivatives of arbitrary order of the metric perturbations, with the
exceptional case of General Relativity with a cosmological
constant. This makes problematic the consistency of the ACM beyond the
linear approximation. This issue will be further studied in a future
work.
To summarize, what we have shown in this section is
that the behavior of scalar perturbations in a general covariant
theory is intimately connected to the background evolution.
Except in the very early universe, the
linearized equations for the scalar perturbations are determined by
the equations in the homogeneous approximation if one assumes that
there are no terms with more than two spatial derivatives. This
assumption could be relaxed in order to include more general theories.
There will be a subset of general covariant theories with a background
evolution given by (\ref{gh}) that verify these
equations for the perturbations. We may be able to distinguish among
these at the linearized level by means of the vector and tensor
perturbations.
\section{Hydrodynamical Perturbations}
Vector perturbations represent rotational flows which decay very
quickly in the General Relativistic theory. As we expect a small
modification of the behavior of perturbations just in the vicinity of
the lower bound $H_-$, we will assume that the vector perturbations
have decayed to negligible values when the scale $H_-$ begins to play
a role and therefore they will be ignored.
Tensor perturbations correspond to gravitational waves, which at the
present have been not observed, their effect being far beyond the
resolution of current observations.
Present measurements restrict their attention to perturbations in the
photon sector (CMB) and the matter sector (matter power spectrum, both
dark and baryonic). These observations can be computed taking into
account only the effect of scalar perturbations. Therefore, we can use
the result of the previous section to study the deviations from
$\Lambda$CDM in the spectrum of CMB. A comprehensive study of the
matter power spectrum predicted by the ACM would require a detailed
knowledge of the nonlinear regime, which we lack at present.
Let us start by deriving the behavior of scalar perturbations in a
universe whose gravitation is described by the ACM and filled with a
perfect fluid ($\Pi = 0$) in (\ref{set}). Therefore, (\ref{pdef}) with
$i\neq j$ fixes $\Phi$ as a function of $\Psi$,
\begin{equation}
\Phi = \left\{1\,+\,\frac{\dot{H}}{2 H^2}\left(1 - \frac{H
g''(H)}{g'(H)}\right)\right\}\Psi \label{Phi}\, .
\end{equation}
Then, this expression can be used to substitute $\Phi$ in the
remaining three differential equations for $\bar{\delta\rho}$,
$\bar{\delta p}$ and $\bar{\theta}$. Given the equation of state of
the fluid $p = p(\rho, S)$, with $S$ the entropy density, the
perturbation of the pressure can be written as
\begin{equation}
\bar{\delta p} = c_s^2 \bar{\delta \rho} + \tau \delta S \, \label{ppert},
\end{equation}
where $c_s^2 \equiv (\partial p/\partial \rho)_S$ is the square of the
speed of sound in the fluid and $\tau \equiv (\partial p/\partial
S)_\rho$. Substituting the pressure and the energy density
perturbations for the corresponding functions of $\Psi$ and its
derivatives into (\ref{ppert}), we arrive at the equation for the
entropy density perturbation $\delta S$. The perturbations we are
interested in are adiabatic, i.e.: $\delta S = 0$ and therefore, the
equation for the entropy perturbations turns into an equation of
motion of the adiabatic perturbations of the metric. If we want to
consider the effect of the lower bound of the Hubble rate, $H_-$, on
the evolution of perturbations, we must consider a matter dominated
universe ($c_s^2 = 0$),
\begin{equation}
\begin{array}{c}
g'(H)\left[3 H \Phi +3
\dot{\Psi}-\frac{\dot{H}}{H^2}\dot{\Psi}+\frac{\dot{H}}H \Phi
+\dot{\Phi} + \frac 1 H \ddot{\Psi}\right] \\
+\frac{\dot{H}}H g''(H)\left[H\Phi+\dot{\Psi}\right] \; = \; 0 \, .
\end{array} \label{Psi}
\end{equation}
This equation is a second order linear differential equation with
non-constant coefficients. Thus it is useful to work with the Fourier
components of the metric perturbation
$\Psi_k$. The equation of motion for the Fourier components is just a
second order linear ODE with non-constant coefficients, which will be
analytically solvable just for some simple choices of $g(H)$.
However, in general, it will be mandatory to perform a numerical
analysis.
\section{The CMB Spectrum in ACM versus $\Lambda$CDM}
The main observation that can be confronted with the predictions of
the theory of cosmological perturbations at the linearized level is
the spectrum of temperature fluctuations of the Cosmic Microwave
Background (CMB), from which WMAP has recorded accurate measurements
for five years \cite{Komatsu:2008hk}.
The temperature fluctuations $\frac{\delta T}T$ are connected to the
metric perturbations via the Sachs-Wolfe effect \cite{Sachs:1967er},
which states that, along the geodesic of a light ray
$\frac{dx^i}{dt}=l^i(1+\Phi+\Psi)$, characterized by the unit
three-vector $l^i$, the temperature fluctuations evolve according to
\begin{equation}
(\frac \partial{\partial t} + \frac {l^i} a \frac \partial{\partial
x^i})(\frac{\delta T}T+\Phi) = \frac \partial {\partial t}(\Psi +
\Phi) \, .
\end{equation}
Neglecting a local monopole and dipole contribution, taking
recombination to be instantaneous at a certain time and assuming
that the reionization optical depth is negligible, the present
temperature fluctuation of a distribution of photons coming from a
given direction of the sky $l^i$ can be related to the temperature
perturbation and metric perturbation $\Psi$ in the last scattering
surface plus a line integral along the geodesic of
the photons of the derivatives of $\Phi$ and $\Psi$ (Integrated Sachs
Wolfe effect, ISW).
These relations are purely kinematical and remain unchanged in the
ACM. In order to take into account all the effects involved in the
calculation of the CMB spectrum the code of CAMB \cite{CAMB} has been
adapted to the ACM. (see the appendix for details).
We have compared the late time evolution of the perturbations
described by the ACM \eqref{future} for several values of $\alpha_-$ (taking into
account that $\alpha_- = 1$ corresponds to $\Lambda$CDM) for constant
$H_0$ and $\Omega_m$ defined in Eq.~\eqref{omegam}. We have
assumed a Harrison-Zel'dovich scale invariant spectrum of scalar
perturbations as initial condition ($n_s = 0$) in the region in
which the universe is radiation dominated but $H\ll H_+$ . Thus the study of the
effect of the scale $H_+$ is postponed. We have taken $H_0 = 72 \pm 8\,
km/sMpc$ from the results of the Hubble Key Project
\cite{Freedman:2000cf}\footnotemark[3] and $\Omega_m = 0.26 \pm 4$ from our previous
analysis of the evolution at the homogeneous level. The effect of ACM
as compared to $\Lambda$CDM is twofold (FIG. 5).
\begin{figure}
\centerline{\includegraphics[scale=0.6, width=9cm]{spectrum_om_26_h_72.png}}
\label{spectrum}
\caption{Spectrum of temperature fluctuations of the Cosmic Microwave
Background for $\Omega_m = 0.26$, $H_0 = 72 km/sMpc$ and varying
$\alpha_- = 0.3$ (red), $0.6$, $1.0$, $1.5$, $2.0$, $3.0$, $4.0$
(blue) . The lowest values of $\alpha_-$ show extreme ISW effects
and therefore will be ruled out also by the CMB
spectrum. Unfortunately, the cosmic variance masks the effect of
$1.0 < \alpha_- < 4.0$ at large angular scales and therefore we can
obtain little information about the ACM from the late ISW effect.}
\end{figure}
There is a shift in the peak positions due to the difference in the
distance to the last scattering surface for different values of
$\alpha_-$. This shift increases with the peak number. In particular,
more precise measurements of the position of the third peak could be
used to estimate the value of $\alpha_-$.
There is also an increase of the lower multipoles due to the late
Integrated Sachs-Wolfe effect, which is particularly extreme for
$\alpha_- \lesssim 1.0$ (which corresponds to $H_-\simeq H_0$ for
constant $\Omega_m$). This large deviation is clearly not present in
the experimental data. The much smaller deviation for $\alpha_-
\gtrsim 1.0$ is within the error bands due to the cosmic variance, and
therefore the value of the lower multipoles cannot be used to exclude
values of $\alpha_-$ greater than but of the order of one.
\section{Comparison with the treatment of perturbations in other models}
\subsection{Perturbations in f(R) theories}
In our previous work \cite{Cortes:2008fy} we found that given an
expansion history parameterized by a modified Friedman equation in a universe
filled by a given component, it was always possible to find a biparametric
family of $f(R)$ theories which had the same homogeneous evolution as
a solution of their equations of motion. We
wonder now if these theories have also the same linearized equations
for the scalar perturbations of the metric.
An $f(R)$ theory is defined by its action and therefore the
equations for the perturbations are uniquely determined. The action is
given by
\begin{equation}
S=\frac 1{16 \pi G}\int d^4 x\sqrt{-g}[f(R) + 16\pi G L_m] \, ,
\end{equation}
where $R$ is the curvature scalar and $f(R)$ is an arbitrary function
of this scalar. The equations governing the evolution of the
perturbations will be derived from the Einstein equations of the
$f(R)$ theory,
\begin{equation}
f'(R)R^\mu_\nu-\frac12
f(R)\delta^\mu_\nu+(\delta^\mu_\nu\square-\nabla^\mu\nabla_\nu)f'(R)=8\pi G
T^\mu_\nu \, .\label{EinfR}
\end{equation}
At the homogeneous level, it is possible to write the Friedman
equations of the $f(R)$ theory,
\begin{eqnarray}
8 \pi G \rho_{(0)} & = & -3(\dot{H}+H^2)f'(R_{(0)})-\frac 12 f(R_{(0)}) -3 H
f''(R_{(0)})\dot{R}_{(0)} \label{fRrho} \, ,\\
-8 \pi G p_{(0)} & = & -(\dot{H} + 3 H^2)f'(R_{(0)})-\frac 12 f(R_{(0)})
+f^{(3)}(R_{(0)})\dot{R}_{(0)}^2 + f''(R_{(0)})\ddot{R}_{(0)} \label{fRp} \, ,
\end{eqnarray}
\footnotetext[3]{ Recently the SHOES team has provided a
more precise measurement, $H_0 = 74.2 \pm 3.6 \, km/sMpc$,\cite{Riess:2009pu}}
\noindent where
\begin{equation}
R_{(0)}(t) = -6(\dot{H} +2H^2)
\end{equation}
is the curvature scalar of the Friedman-Robertson-Walker (FRW) metric
(zeroth order in perturbations). The solution of the system of
equations \eqref{fRrho},\eqref{fRp} together with the equation of
state of the dominant component of the stress-energy tensor, defines
$R_{(0)}$ and its derivatives as functions of the Hubble parameter $H$
\cite{Cortes:2008fy}.
If we expand \eqref{EinfR} in powers of the
scalar perturbations of the FRW metric \eqref{newt}, the first order
term gives the linearized equations for the
perturbations. This will be a system of differential equations of
fourth order, but it will be possible to turn it into a system of
differential equations of second order, as we will see below.
If the stress-energy tensor has no shear, the Einstein equation for
$\mu = i\neq \nu = j$ reduces to
\begin{equation}
f''(R_{(0)}) \delta R +f'(R_0) (\Phi-\Psi) = 0 \label{fRGij} \, ,
\end{equation}
where $\delta R$ is the perturbation of the curvature scalar
\begin{equation}
\begin{array}{c}
\delta R \,=\, 12(\dot{H} + 2 H^2) \Phi + 6 H \dot{\Phi} + 24 H \dot{\Psi}
\\ + 6 \ddot{\Psi} + \frac 2{a^2}\Delta(\Phi-2\Psi))\, .
\end{array} \label{delR}
\end{equation}
Eq.~\eqref{fRGij} is a second order differential equation for the scalar
perturbations except in the case of General Relativity ($f''=0$)
where it becomes an algebraic equation ($\Phi=\Psi$).
Eq.~\eqref{fRGij} can be used to turn the remaining components
of the equations of $f(R)$ theories \eqref{EinfR} into second order
equations.
In the case of adiabatic perturbations in a matter dominated epoch one has
\begin{equation}
\begin{array}{c}
-8 \pi G \bar{\delta p} = f'(R_{(0)})(a^{-2}\Delta(\Phi-2\Psi)+\ddot
\Psi +6 H \dot\Psi + H\dot\Phi +2(\dot{H} + 3 H^2)\Phi -\frac 12 \delta R)
\\ + f''(R_{(0)})(-(\dot{H}+3H^2)\delta R -2\dot{R}_{(0)} \dot \Psi -8 H \dot{R}_{(0)}
\Psi -4 H \dot{R}_{(0)} \Phi -2 \ddot{R}_{(0)} \Phi) - 2
f^{(3)}(R_{(0)})\dot{R}_{(0)}^2\Phi \\
-\partial^2_t\left[ f'(R_{(0)})(\Phi-\Psi) \right]-2 H
\partial_t\left[ f'(R_{(0)})(\Phi-\Psi) \right] = 0 \, .
\end{array} \label{fRadiab}
\end{equation}
If we compare the equations governing the behavior of adiabatic
perturbations in a $f(R)$ theory \eqref{fRGij}, \eqref{fRadiab} with
those coming from the ACM, \eqref{Phi},\eqref{Psi}, we find that they
describe very different
behaviors. In one case we have a system
of two coupled second order differential equations for $\Phi$ and
$\Psi$. In the ACM case we got a single second order differential
equation and an algebraic equation between the two gravitational
potentials. For a given $f(R)$ theory with an associated $g(H)$
homogeneous behavior, we find that the behavior of linearized
perturbations differs from the one defined by the linearized
perturbations in the ACM. This means that $f(R)$ theories break in
general our assumption in the third section (there are terms with more than
two derivatives in the equations for the perturbations, and therefore
the behavior of perturbations in General Relativity
is not recovered when $H_+ \gg H \gg H_-$ for all modes).
Let us see this in a simple example. We will choose the easiest
biparametric family of $f(R)$ theories: the one that, under matter
domination, gives the same background evolution as General Relativity
without a cosmological constant \cite{Cortes:2008fy},
\begin{equation}
f(R)=R+c_1 \vert R \vert^{\frac 1{12}(7-\sqrt{73})}+c_2 \vert R
\vert^{\frac 1{12}(7+\sqrt{73})} \, \label{fRGR}.
\end{equation}
These kind of models were introduced in Ref.~\cite{Nojiri:2003ft}.
Let us first check if the solutions of the equations for the
perturbations in General Relativity in the presence of pressureless
matter,
\begin{equation}
\begin{array}{ccl}
\Psi = \Phi & = & const \, ,\\
\Psi = \Phi & \propto & t^{-5/3} \, , \label{solsGR}
\end{array}
\end{equation}
are solutions of the equations of motion in the case of the
biparametric family of $f(R)$ theories (\ref{fRGR}).
The easiest is to verify if Eq.~\eqref{EinfR} with $\mu = i \neq \nu = j$,
\begin{equation}
\begin{array}{c}
2 f''(R_{(0)}) (3 H^2 \Phi + 3 H \dot{\Phi} + 12 H \dot{\Psi} + 3
\ddot{\Psi} + \frac 1{a^2}\Delta(\Phi-2\Psi))\\
+f'(R_{(0)}) (\Phi-\Psi) = 0 \label{fRGijGR} \, ,
\end{array}
\end{equation}
where $R_{(0)} (H) = -3 H^2$ is the value of the homogeneous curvature
scalar as a function of $H$ in this family of theories, is also
fulfilled by the solutions (\ref{solsGR}). The result is obviously
not.
The second order differential equations for the perturbations in the
$f(R)$ theories (\ref{fRGR}) in the presence of matter can be found by
substituting (\ref{fRGijGR}) and its derivatives into the other
equations of the system \eqref{EinfR}. The equation for the pressure
perturbations then gives
\begin{equation}
\begin{array}{c}
-f'(R_{(0)})[\ddot{\Psi}+\ddot{\Phi}+ 4 H
\dot{\Psi}+4H\dot{\Phi}+3H^2\Phi] \\
-27H^3f''(R_{(0)})\dot{\Phi}+\frac{f(R_{(0)})}2(\Psi-3\Phi)\,=\,8\pi G
\bar{\delta p} \, = \, 0 \, \label{fRGiiGR} \, .
\end{array}
\end{equation}
It is obvious that the $i\neq j$ equation of the cosmologic
perturbations in General Relativity,
\begin{equation}
\Phi - \Psi = 0
\end{equation}
is not recovered from (\ref{fRGij}) in the $H \gg H_-$ limit for all
$k\ll H_+$. In fact the deviation would be significant for modes with
$k \gtrsim H a \left(\frac{H}{H_-}\right)^{\frac 1{12}(7-\sqrt{73})}
\sim H a$. Therefore, the assumption we have made in order to derive
the equations for the perturbations in the ACM is broken.
It is a known problem that $f(R)$ theories are unable to pass
cosmological and astrophysical tests involving perturbations \cite{Chiba:2003ir},
unless the function involved is properly fine-tuned. In particular, if the
conformal equivalence between $f(R)$ theories and scalar-tensor theories is
used, it is necessary that the effective mass acquired by the new scalar
degree of freedom is unnaturally large \cite{Faulkner:2006ub}.
Some $f(R)$ theories that pass cosmological and solar system tests have been
proposed \cite{Hu:2007nk,Nojiri:2007as,Cognola:2007zu}.
However, it is also subject of debate if the conformal equivalence can
be used in order to extract physical predictions from this models, especially
predictions which involve perturbations \cite{Carloni:2009gp}. A mathematically
rigorous treatment of perturbations in $f(R)$-gravity can be found in
Ref.~\cite{Carloni:2007yv}.
\subsection{Quintessence Models}
Another class of models used to describe the acceleration of the
universe are those in which a so called quintessence field, typically
of scalar type $\varphi$, is added as a component of the universe
\cite{Wetterich:1987fm}. The only functional degree of freedom in most
models is just the scalar potential $V(\varphi)$, which can be tuned
to fit the homogeneous expansion of the universe. In our previous work
\cite{Cortes:2008fy} we found the correspondence between a given
homogeneous evolution parameterized by $g(H)$ and the potential for
the quintessence which drives under General Relativity this
evolution. We now wonder if the behavior of perturbations in
quintessence models also resembles the behavior in ACM.
The field action is given by
\begin{equation}
S_\varphi = \int d^4 x \sqrt{-g} \left[ g^{\mu\nu}\partial_\mu \varphi
\partial_\nu \varphi -V(\varphi) \right] \, . \label{lagquint}
\end{equation}
Let us assume that the field can be split in two components: one which
is only time dependent and which drives the homogeneous evolution of
the universe, $\varphi_{(0)} (t)$, and a small perturbation which is
inhomogeneous, $\delta\varphi({\mathbf{x}},t)$. We will consider linearized
perturbations of the metric, the field, and the other components of
the universe. The resulting linearized stress-energy tensor of the
field is
\begin{eqnarray}
\delta T^\varphi_{00} & = & \dot{\varphi}_{(0)}
\dot{\delta\varphi}-\Phi V(\varphi_{(0)}) + \frac 12
V'(\varphi_{(0)})\delta\varphi \, , \\
\delta T^\varphi_{0i} & = & \dot{\varphi}_{(0)} \partial_i \delta\varphi \, ,\\
\delta T^\varphi_{ij} & = & a^{-2}
\delta_{ij}\left[\dot{\varphi}_{(0)}^2 (\Phi-\Psi)
+\dot{\varphi}_{(0)} \dot{\delta\varphi} \right. \nonumber \\
& & \left. +\Psi V(\varphi_{(0)}) - \frac 12 V'(\varphi_{(0)})\delta\varphi
\right] \, .
\end{eqnarray}
By virtue of the Einstein's equations, it is always possible to turn a
modification in the Einstein tensor into a new component of the
stress-energy momentum tensor of the sources of the gravitational
field. We wonder if it is possible to account for the modification of
the effective Einstein tensor described in the third
section with a new component described by this quintessence
field. However, it is not hard to see that the effective Einstein
tensor that we are proposing has a modified $i\neq j$ component, while
the $i\neq j$ component of the stress-energy tensor of the quintessence field
$\varphi$ is zero. Therefore, it is not possible to describe the
evolution of perturbations in the ACM as driven by an effective scalar
field component~(\ref{lagquint}).
\subsection{Dark Fluid Models}
A deformation of the gravitational physics can be also made equivalent
to the addition of a non-standard fluid component to the cosmic pie at
the homogeneous level. The fluid component used to explain the present
accelerated expansion of the universe is typically taken to be a
perfect fluid with large negative pressure.
One of the most popular parameterizations of this fluid is the so
called equation of state $\omega = p / \rho$ \cite{Turner:1998ex}. The
simplest models are $\omega = -1$, which is equivalent to a
cosmological constant, or a constant $\omega$, but recently a possible
time dependence of $\omega$ has been considered
\cite{Bassett:2004wz}. Needless to say, these {\it ad hoc}
parameterizations are subsets of the more general equation of state of
a perfect fluid, $p = p(\rho,s)$, where $s$ is the entropy density. It
is also questionable why must we restrict ourselves to perfect fluids
and not include a possible anisotropic stress tensor, possibly
depending on the energy and entropy densities
\cite{Capozziello:2005pa}.
As in the previous subsection, it is always possible to use the
Einstein's equations to turn a modification in the gravitational
physics into a new fluid component of the universe. In the case of the
equations derived in the third section, the
equivalent Dark Energy Fluid would have the following properties:
{\small
\begin{eqnarray}
8\pi G \bar{\delta \rho}_X & = & (2H-g')\left[\frac 1{a^2 H}\Delta
\Psi -3(H\Phi + \dot{\Psi})\right], \\
8\pi G \bar{\theta}_X & = & \frac{(2H-g')}H(H\Phi + \dot{\Psi}), \\
8\pi G \Pi_X & = & \frac
1{a^2}\left\{\frac{2H-g'}{2H}(\Phi-\Psi)-\frac{\dot{H}}{2H^3}(g'-Hg'')\Psi\right\},
\\
8\pi G \bar{\delta p}_{X} & = & (2H-g')(3H\Phi +3\dot\Psi-\frac{\dot
H}{H^2}\dot \Psi + \frac{\dot H}{H}\Phi + \dot \Phi + \frac 1H
\ddot\Psi)\nonumber \\
& & +\frac{\dot H}{H}(2-g'')(H\Phi+\dot\Psi)-8\pi G \Delta\Pi_X.
\end{eqnarray}
}
The resulting fluid is not an ideal fluid ($\Pi_X = 0$) or even a
Newtonian fluid ($\Pi_X \propto \bar\theta_X$). With the use of
\eqref{gh}, \eqref{Phi} and \eqref{Psi}, it will be possible to write
$\bar{\delta\rho}_X$ and $\bar{\delta p}_X$ as a combination of $\bar
\theta_X$, $\Pi_X$ and their derivatives,
\begin{eqnarray}
\bar{\delta\rho}_X & = & f_1(\rho_{(0)X}) \Delta \Pi_X +
f_2(\rho_{(0)X}) \bar\theta_X \, ,\\
\bar{\delta p}_X & = & f_3(\rho_{(0)X})\Pi_X -\Delta\Pi_X +
f_4(\rho_{(0)X})\bar\theta_X \, .
\end{eqnarray}
These very non-standard properties show that, although it is formally
possible to find a fluid whose consequences mimic the ones of such a
modification of the gravitational physics, this fluid is very exotic.
\section{Summary and Conclusions}
In view of recent data, an updated comparison of cosmological
observations with a phenomenological model proposed in a recent work
has been presented.
An extension of this phenomenological model (ACM)
beyond the homogeneous approximation has been introduced allowing us to
describe the evolution of scalar perturbations at the linear level.
A comparison with the
spectrum of thermal fluctuations in CMB has been used to explore the
possibility to determine the parameters of the ACM through its role in
the evolution of scalar perturbations. The results of this comparison
does not further restrict the parameters of the model, due to the masking
of the associated late ISW effect by the cosmic variance. However, better
measurements of the position of the third acoustic peak should improve
the constraints significatively.
It has been shown that the equivalence of different formulations of
the accelerated expansion of the universe in the homogeneous
approximation is lost when one considers inhomogeneities. In
particular we have shown that the general structure of the evolution
equations for scalar perturbations in the ACM differs from the
structure of the equations corresponding
to modified $f(R)$ theories of gravity, to quintessence models or to a
dark fluid with standard properties.
The possibility of going beyond
the linearized approximation for the scalar perturbations and to
consider vector and tensor perturbations will be the subject of a
future work.
We would like to thank Marco Bruni for inspiration. We would also like
to thank Valerio Faraoni, Daniel G. Figueroa, Troels Hagub\o{}lle,
Sergei Odintsov and
Javier Rubio for fruitful discussions. J.I. also thanks the
hospitality of the Benasque Center for Science and the Scuola
Internazionale Superiore di Studi Avanzati (SISSA) during the
development of this work.
This work has been partially supported by CICYT (grant
FPA2006-02315) and DGIID-DGA (grant2008-E24/2). J.I. acknowledges a FPU
grant from MEC.
\section*{Appendix: Covariant Perturbation Equations for the ACM}
The code of CAMB \cite{CAMB} makes use of the equations for the
perturbations of the metric in the covariant approach. The quantities
can be computed in a given ``frame'', labeled by a 4-velocity
$u^\mu$. In particular, CAMB uses the dark matter frame, in which the
velocity of the dark matter component is zero (the dark matter
frame).
Furthermore, it parameterizes the time evolution with the conformal
time $a(\tau)d\tau = dt$.
In order to apply CAMB to the ACM model it is necessary to
identify frame invariant quantities, and then relate them to
gauge invariant quantities \cite{Gordon:2002gv}. The following
comoving frame quantities are used in the CAMB code: $\eta$ (the
curvature perturbation), $\sigma$ (the shear scalar), $z$ (the
expansion rate perturbation), $A$ (the acceleration, $A=0$ in the dark
matter frame), $\phi$ (the Weyl tensor perturbation), $\chi^{(i)}$
(the energy density perturbation of the species i), $q^{(i)}$ (the
heat flux of the species i), and $\Pi^{(i)}$ (the anisotropic
stress of the species i). All of these quantities are defined in
\cite{Challinor:1998xk}. These variables are related to the gauge
invariant variables via the following dictionary,
\begin{equation}
\begin{array} {rcl}
-\frac \eta 2 -\frac{\mathcal{H}\sigma}{k} & \equiv & \Psi \, ,\\
- A + \frac{\sigma'+\mathcal{H}\sigma}{k} & \equiv & \Phi \, ,\\
\mathcal{\chi} + \frac{\rho'\sigma}{k} & \equiv & \bar{\delta\rho} \, ,\\
\rho q + (\rho + p)\sigma & \equiv & \frac k a (\rho + p) \bar{\theta}\, ,
\end{array}
\end{equation}
where prime denotes derivatives with respect to conformal time,
except when acting on $g(H)$ where it denotes a derivative with
respect to the Hubble rate $H$, and $\mathcal{H} = a'/a = aH$. On the other hand
$z = \sigma + \frac{3}{2k}(\eta'+2\mathcal{H} A)$ and $\phi = (\Phi +
\Psi)/2$. The anisotropic stress $\Pi$ is already frame invariant.
Written in the dark matter frame, the equations for the scalar perturbations read
\begin{equation}
\begin{array}{rcl}
g'\left(\frac{k^2 \eta}{2 \mathcal{H}} + k z\right) & = & 8\pi G a \Sigma_i \mathcal{\chi}^{(i)} \, ,\\
\frac{g' k^2}{3\mathcal{H}}(\sigma - z) = -\frac{g'k}{2\mathcal{H}} \eta' & = & 8\pi G
a \Sigma_i \rho_i q^{(i)}\, ,\\
-\mathcal{H} k g'(z' + \mathcal{H} z) & = & 8 \pi G a \Sigma_i\left[
\mathcal{H}^2 (1+3c_s^{(i)2}) - (\mathcal{H}'-\mathcal{H})^2(1-\frac{\mathcal{H} g''}{a g'}) \right]\mathcal{\chi}^{(i)}\, ,
\\
\frac {g'}\mathcal{H} \left(\frac{\sigma'+\mathcal{H} \sigma}{k}-\phi
\right)-\frac{\mathcal{H}'-\mathcal{H}^2}{2 \mathcal{H}^3}(g'-\mathcal{H} g''/a) \left(\frac\eta
2+\frac{\mathcal{H}\sigma}k \right)& = & - 8\pi G a \Sigma_i\Pi^{(i)}/k^2 \, .
\end{array}
\end{equation}
The following combination of the constraint equations is also useful:
\begin{equation}
k^2\phi = -\frac{8\pi G a \mathcal{H}}{g'}\Sigma_i\left[ \Pi^{(i)}
+(1-\frac{\mathcal{H}'-\mathcal{H}^2}{2 \mathcal{H}^2}(1-\frac{\mathcal{H} g''}{a g'}))(\mathcal{\chi}^{(i)}
+3\mathcal{H} \rho_i q^{(i)}/k)\right]\, .
\end{equation}
These equations are plugged into the Maple files provided with the
CAMB code and run to get the ISW effect \cite{CAMB}.
\chapter{Summary and Conclusions}
The General Theory of Relativity (GR) and the Standard Model of particle physics (SM) have offered a very accurate description of all known phenomena with two remarkable exceptions, the neutrino oscillations and the accelerated expansion of the universe in the early universe and the present age. The neutrino oscillations and the present acceleration of the universe accept a quite conventional explanation in terms of neutrino masses and a cosmological constant/vacuum energy, respectively. However, these explanations are not by themselves very satisfactory because:
\begin{itemize}
\item
they introduce an enormous naturalness problem,
\item
they introduce a new ingredient in the theory that is hard or impossible to measure aside from in the phenomena that caused their proposal,
\item
they are unable to explain the striking coincidence in scale between the effective energy density attached to the cosmological constant ($\rho_{vac}\sim (10^{-3} \, eV)^4 $) and the lowest neutrino mass difference ($\Delta m^2 \sim (10^{-3} \, ev)^2$),
\item
the present experimental and observational data are still unable to rule out some of the alternative proposals to these explanations.
\end{itemize}
Despite these questions that have been raised, the neutrino masses and the cosmological constant have been taken as a confirmed fact by many because they are the simplest explanation in agreement with the experimental and observational data, and because they do not compromise any of the basic principles underlying GR or the SM.
However nothing is sacred in science, aside from the proof of experiment. Simplicity is also important, but aesthetics cannot be used as a substitute of experimental verification of a prediction. The physical principles that are taken as basic must be rejected in case they conflict a single experiment or observation, and thus must always be subject of scrutiny.
Therefore we would like to build models that are alternatives to neutrino masses and the cosmological constant, which are consistent with all the known phenomena, and that could be contrasted with the more conventional explanations in the light of future experiments and surveys. We will not fear to violate the basic principles underlying GR and the SM, as long as the departures from them is small enough to be consistent with the available data.
For instance, neutrino masses are the only explanation to neutrino oscillations in vacuum within Special Relativity (SR), and thus we have considered a small Lorentz Invariance Violation (LIV) in neutrinos. On the other side, the too high quantum corrections to a vacuum energy are computed in the framework of Effective Field Theory (EFT), which is based in the Decoupling Theorem, which in turn is based in locality, so we have considered a small departure from locality.
Not always it is required to violate the underlying principles in a theory in order to go beyond it, as is the case of GR. We have considered more general actions than the Einstein-Hilbert action in the search for new physics that can explain the accelerated expansion.
In both cases, the departure from the standard framework has been characterized by energy scales which determine the onset of new physics. These scales can be either ultraviolet (UV) or infrared (IR), depending on the requirements of the phenomena we want to explore. Neutrino oscillations as well as the present accelerated expansion of the universe require the introduction of a new IR scale, whereas the explanation of the horizon and flatness problems requires the introduction of a UV scale. It is possible that the two IR scales are the same, as it is suggested by the coincidence mentioned above, but we have not been able to explain this coincidence within a certain consistent model.
Instead of resorting to a top-down approach like String Theories (ST), Loop Quantum Gravity (LQG) or Grand Unified Theories (GUT), which have grown far from the contact with experiment, we have prefer to adopt a bottom-up approach in which we have started from simple models that offer a description of the data and may grow in the future if they are not ruled out by future data first.
In order to explain the two epochs of accelerated expansion of the universe, we have introduced a phenomenological extension of the Cosmological Standard Model (CSM) based on GR. The model is similar to other phenomenological descriptions that have been proposed, which relate the expansion of the universe with more ``anthropocentric'' parameters such as the redshift $z$. The Asymptotic Cosmological Model (ACM) is based (as GR) in a metric theory of spacetime in the homogeneous approximation, and it is defined through the relation between the Hubble rate $H$ and the energy density of the matter content of the universe defined by the SM particles and the Dark Matter. The effect of new physics is encoded in the dependence of these relation on an UV and an IR scales which act as bounds of the open interval to which the Hubble rate is constrained. For dimensional arguments, in the limit in which the Hubble rate is far from these bounds, the standard behavior predicted by GR is recovered, whereas the universe tends asymptotically to a deSitter universe in the infinite past and the infinite future when the bounds are approached. The CSM supplied with a cosmological constant ($\Lambda$CDM) is included as a particular case.
The ACM is built with four free parameters, two of them characterizing the early accelerated expansion and two of them characterizing the present accelerated expansion. We have constrained the two of them relevant for the present accelerated expansion, contrasting the results derived from simulations using Monte Carlo-Markov Chains with the observational data from Type Ia Supernovae (SNe), the distance to the recombination surface deduced from the Cosmic Microwave Background (CMB), direct measurements of the energy density and the distance parameter of Baryon Acoustic Oscillations (BAO). We have found that there are better fits to the expansion of the universe than $\Lambda$CDM, but that the quality of the fit depends too crucially on the dataset used. Thus an improvement on the datasets is important in order to favor $\Lambda$CDM or one of its competitors.
The homogeneous behavior of the ACM can be realized in the framework of modified gravity (for instance $f(R)$-theories), scalar-tensor theories or phenomenological Dark Energy models indistinguishably. Differences between these descriptions may be found in the behavior of perturbations.
The lack of an action for the ACM is a serious obstacle in building a theory of perturbations in the ACM. However, under an additional assumption based on phenomenological consistency (the differential equations being of order two), the homogeneous behavior of the ACM determines univocally the linearized behavior of scalar perturbations.
With the assumption of a Harrison-Zel'dovich scale invariant spectrum from the very early universe, this linear treatment suffices to derive the CMB spectrum deduced from the ACM. The comparison between this prediction and the observed spectrum imposes similar bounds on parameter space than the homogeneous analysis. This is due to the fact that the Integrated Sachs-Wolfe effect (ISW) of the ACM for the parameters lying within the confidence regions of our homogeneous analysis is masked by the cosmic variance, i.e.: the intrinsic statistical error associated to the lowest multipoles (cosmic censorship).
The treatment of linear perturbations is able to distinguish the ACM from the most na\"{\i}ve Dark Energy forms, such as scalar-tensor theories or f(R)-theories. These examples usually present fine-tuning problems such as the requirement of a chameleon mechanism for a fifth long-range interaction to disappear.
In the future we plan to explore further the consequences of this model, such as the nonlinear behavior of perturbations, the behavior of vector and tensor modes, or the weak gravity limit.
It is often stated that the neutrino oscillation in vacuum experiments prove that neutrinos are massive particles. We have shown that this statement is false. Neutrino oscillations can be also explained in terms of oscillations between particles with different dispersion relations, wether this difference relies on different masses or not.
Neutrinos with different constant velocities or dispersion relations modified by a UV scale are certainly ruled out, with the exception of a single very fine-tuned model. However, neutrino oscillation experiments only probe the high energy (few MeV to few GeV) regime of neutrinos, and therefore {\it any} modification of the dispersion relation in the IR, such that the high energy effect of the modification amounts to a term proportional to the square of the IR scale, can naturally explain neutrino oscillations in vacuum.
It is difficult to find such a modification of the dispersion relation different from a mass term in QFT. We have found a model in which this is possible also respecting the gauge symmetry and lepton number conservation of the standard model. However, we have had to sacrifice not only Lorentz invariance but also locality. We plan to study a modification of this model allowing for lepton number violation in the future.
In order to distinguish between massive neutrinos or other IR modification of the standard model, it will be necessary to probe the low energy regime (few meV to few eV) of the neutrinos. The Cosmic Neutrino Background (C$\nu$B) predicted by cosmology offers the possibility of measuring such neutrinos in the future.
However, in order to study the effect of a Lorentz non-invariant theory in cosmology it is important to build a quantum field theory in curved spacetime which preserves the symmetry properties of such a theory. Although this is a very complicated task, outside the scope of this thesis, as a first step we have explored the possibility of building the scalar Quantum Theory of Noncanonical Fields (QNCFT) in a curved spacetime.
Respecting either the non-locally implemented Poincar\'e group of symmetries or the locally implemented Euclidean group of symmetries of the theory in flat spacetime leads to two different ways of defining the theory in curved spacetime.
If one chooses to gauge the non-locally implemented Poincar\'e group of symmetries by the standard procedure, an obstruction is found. The solution is to introduce curvature in the auxiliary spacetime in which the Poincar\'e symmetry can be interpreted as a spacetime symmetry, instead of curving the physical spacetime. The resulting theory is nonlocal in time. In the limit in which the parameter $\theta$ tends to zero both spacetimes coincide and one recovers the standard QFT in curved spacetime. The consequences of this theory will be similar to the ones of standard QFT in curved spacetime but the particles will have a modified dispersion relation.
Alternatively, one may curve the physical spacetime, but then the non-locally implemented symmetry of the theory in flat spacetime is broken. In the resulting theory a preferred reference frame arises, and spacetime is foliated in spacelike simultaneity hypersurfaces. New phenomena related to the UV scale $\theta$ may arise, which were not present in the standard QFT in curved spacetime.
We plan to study how the Hawking radiation is affected by the noncanonical commutation relation in both theories. We would also like to go further in determining what are the symmetries of in the model of noncanonical neutrinos and how they can be coupled to gravity.
With a na\"{\i}ve assumption of about how the modified dispersion relation of the noncanonical neutrinos can be introduced in cosmology, we have computed the effect of a IR modified dispersion which mimics a cosmological constant in cosmology. Using the bounds on `squared neutrino mass differences' and on the `spatial curvature', the resulting ``effective cosmological constant'' turns out to be three orders of magnitude above the observed value.
If there exist a connection between the IR scale pointed out by neutrino oscillations and by the present acceleration of the universe, much work is still to be done to uncover this relation. We hope that this PhD thesis will serve as inspiration for future research in this direction.
We will now proceed to enunciate the results of this thesis
\begin{itemize}
\item
We have built a cosmological model (ACM) defined by the relation between the energy density and the Hubble parameter in the homogeneous approximation. The model contains the $\Lambda$CDM as a particular case. This model suffices to explain the horizon problem and the present accelerated expansion. Through a simulation, the late time behavior of the model has been constrained. Although a cosmological constant is permitted at $1\sigma$, the observations favor other regions of the parameter space. However, the parameter determination depends strongly on the dataset which is being used.
\item
Assuming order two differential equations for the perturbations, we have found that it is possible to find a unique set of equations for the scalar perturbations over a FRW universe whose homogeneous behavior departs from GR. In contrast to common belief, this is doable without resorting to higher order equations or new degrees of freedom, at least at the linearized level. The model is distinguishable from other proposals such as $f(R)$-gravity, scalar tensor theories or the most na\"{\i}ve dark fluid models.
\item
We have used the equations for the scalar perturbations associated to the ACM to derive the spectrum of the CMB in a universe following an expansion governed by the ACM. The fit to observations does not add new information with respect to the homogeneous fits, due to a masking of the late ISW effect by the cosmic variance at low multipoles.
\item
It is often stated that neutrino oscillation experiments imply that neutrinos have mass. We have found a counter example to this statement. We have built a model in which the effect of new physics governed by an IR scale is able to explain all the neutrino oscillation experiments. Thus neutrino oscillations imply that neutrinos are sensitive to new physics in the IR. The most promising observation that would be able to distinguish between massive and noncanonical neutrinos is the cosmological effect of neutrinos, and in particular the C$\nu$B.
\item
As a first step towards determining the cosmological effect of noncanonical neutrinos, we have built two possible quantum theories of scalar noncanonical fields in curved spacetimes, depending on which symmetry of the theory in flat spacetime is broken when curving spacetime. In one of them there is no preferred reference frame and particles with modified dispersion relations are allowed, but it has an unclear interpretation. The other involves a preferred time direction and new effects may arise, but the renormalizability of the QFT is not guaranteed.
\item
Our work has tried to explore two phenomena which seem to point to new physics in the IR: neutrino oscillations and the present cosmic acceleration. If these phenomena are related, there is still much work to do to uncover this relation.
\end{itemize}
|
\section{Introduction}
Some of the most fundamental properties of molecular clouds are the
mass and size of these clouds and their substructure. Today, these
properties are well constrained: we know the masses and sizes of dense
cores in molecular clouds ($\lesssim 0.1 ~ \rm pc$ size; e.g.\
\citealt{motte1998:ophiuchus}, \citealt{johnstone2000:rho_ophiuchi},
\citealt{hatchell2005:perseus}, \citealt{enoch2007:cloud_comparison}),
and those of clumps (some $0.1 ~ \rm pc$) and clouds
($\gtrsim 10 ~ \rm pc$) containing the cores (e.g.,
\citealt{williams1994:clumpfind}, \citealt{cambresy1999:extinction},
\citealt{kirk2006:scuba-perseus}; see \citealt{williams2000:pp_iv} for
definitions of cores, clumps, and clouds). We do, however, not know
much about the \emph{relation} between the masses and sizes of cores,
clumps, and clouds: traditionally, every domain is characterized and
analyzed separately. As a result, it is still not known how the core
densities (and thus star-formation properties) relate to the state
of the surrounding cloud.
In principle, the relation between the mass in cloud structure at
large and small spatial scales is described by mass-size relations.
\citet{larson1981:linewidth_size} presented one of the first studies
of such relations. He concluded (in his Eq.\ 5) that the mass
contained within the radius $r$ obeys a power-law,
\begin{equation}
m(r) = 460 \, M_{\sun} \, (r / {\rm pc})^{1.9} \, .
\label{eq:mass-size-larson}
\end{equation}
Most subsequent work refers to this relation as ``Larson's
3$^{\rm rd}$ law'', and replaces the original result with
$m(r) \propto r^2$ (e.g., \citealt{mckee2007:review}). This ``law of
constant column density'' (with respect to scale, $r$) is now
considered one of the fundamental properties of molecular cloud
structure (e.g., reviews by \citealt{ballesteros-paredes2007:ppv},
\citealt{mckee2007:review}, \citealt{bergin2007:dense-core-review}).
This relation has, however, never been re-examined comprehensively on
the basis of up-to-date data. It is, e.g., not clear whether recent
dust extinction and emission work is consistent with
$m(r) \propto r^2$.
Further, the limitations of available structure identification schemes
(such as CLUMPFIND; \citealt{williams1994:clumpfind}) forced past
studies to break cloud structure maps up into discrete fragments.
These fragments typically have a size slightly larger than the map
resolution. As a consequence, the cloud structure is only probed in a
narrow spatial domain; the largest spatial features in a given map
are, for example, usually not characterized. Today, approaches
permitting automatic examination of a continuous range of spatial
scales are available. \citet{rosolowsky2008:dendrograms}, in
particular, provide software for such studies (their dendrogram
analysis); our work would be impossible without the work by
\citeauthor{rosolowsky2008:dendrograms}. Such software permits
derivation of spatially more comprehensive mass-size relations than
possible in the past.
In this series of papers, we combine contemporary column density
observations of high sensitivity with a new data analysis technique to
examine the mass-size relation in molecular clouds for a continuous
range of spatial scales of order $0.01 ~ {\rm to} ~ 10 ~ \rm pc$. We
rely on extinction maps of molecular clouds (here:
\citealt{ridge2006:complete-phase_i}), as well as maps of dust
emission \citep{enoch2006:perseus}. Following the terminology of
\citet{peretto2009:irdc-catalogue}, we define ``cloud fragments'' in
the maps as regions enclosed by a continuous column density contour,
and derive their mass and size at various contour levels. This is
implemented using algorithms introduced by
\citet{rosolowsky2008:dendrograms}.\medskip
The first two papers in this series establish our analysis approach
(part I, the present paper) and explore several clouds in the solar
neighborhood ($\lesssim 500 ~ \rm pc$; part II). Section
\ref{sec:method} of the present paper describes our cloud fragment
extraction and characterization scheme. In Sec.\
\ref{sec:mass-size-properties} we provide a first idea how basic
physical properties and observational limitations affect the mass-size
measurements. This includes a comparison to results obtained using the
CLUMPFIND algorithm. A detailed discussion of analysis uncertainties
is presented in Sec.\ \ref{sec:all-scales_perseus}. These are explored
using data for the Perseus Molecular Cloud. We also explain how dust
emission and extinction data can be combined for a given
cloud.
Section \ref{sec:purpose} briefly describes how the new map analysis
scheme might help to advance star formation research. As we describe
there, it will help to jointly analyse data taken at different spatial
resolution. This is a key feature in the age of multi-wavelength and
multi-resolution studies. We conclude with a summary in Sec.\
\ref{sec:summary}.
\begin{figure*}
\begin{tabular}{ll}
\large \textsf{a) input map} & \large\textsf{b) mass-size data}\vspace{-0.4cm}\\
\begin{minipage}{0.45\linewidth}
\includegraphics[width=0.7\linewidth,angle=-52.7]{./SchemeMap.eps}
\end{minipage} &
\begin{minipage}{0.45\linewidth}
\includegraphics[width=0.9\linewidth,bb=15 6 360 395,clip]{./mass_perseus-children-27.eps}
\end{minipage}
\end{tabular}
\caption{Fundamental concept of mass-size measurements. Panel (a)
shows an example column density map for the Perseus molecular cloud,
where labels refer to individual star-forming regions. Such maps
usually contain several local maxima (\emph{numbered crosses} in
map). We pick one of these maxima, and draw contours at constant
column density around this peak (\emph{map inset}). For each
contour, we measure mass and size. These measurements can then be
placed in a mass-size plot, as marked by \emph{crosses} in panel
(b). When a map contains several maxima, it can be divided into
cloud fragments where contours just contain a single maximum
(\emph{light green} and \emph{yellow} boundaries in panel [a]), two
of these (\emph{dark green} boundary), or even more (\emph{magenta}
contour). In this sense, mass-size measurements for contours
contained in one of these areas are related. These relations are
highlighted by \emph{colored lines} in panel (b); the \emph{color}
refers to the map region from where the measurements are taken, and
\emph{numbers} indicate the central column density peak of each
fragment. When two fragments blend into a combined one that contains
both, mass and size measurements jump discontinuously from those of
the individual fragments to those for the combined one. These jumps
are indicated by \emph{dotted lines}.\label{fig:processing-scheme}}
\end{figure*}
\section{Method}
\label{sec:method}
\subsection{Processing of Contour Maps}
Consider some column density map containing a number of local maxima,
as sketched in Fig.\ \ref{fig:processing-scheme} (a). In this map,
cloud fragments can be identified as structures bound by some
continuous contour. To characterize these, we pick one of
the column density maxima, and measure the mass, $m$, and area, $A$,
contained within each column density contour containing this
peak. Here, we use the effective radius,
\begin{equation}
r = (A / \pi)^{1/2} \, ,
\end{equation}
to quantify $A$. By following the trends from contour to contour, as
shown in Fig.\ \ref{fig:processing-scheme} (b), one can construct a
mass-size relation for every cloud fragment. Strictly speaking, we do
thus construct mass-area diagrams. In practice, however, $r$ is
arguably a more intuitive variable than $A$.
Some contours may contain several local maxima. They represent
composite fragments. To give an example based on Fig.\
\ref{fig:processing-scheme}, the region bound by the dark-green
contour consists of the two regions marked in lighter shades of green,
and the magenta contour contains the fragments marked in yellow and
dark-green. Merging of two fragments occurs at the first column
density contour containing both of the fragments. In our analysis, we
enforce that only two fragments can merge at a time.
During a merger, mass and size jump discontinuously from the
pre-merger situation, $m_i$ and $r_i$, to their post-merger value,
$m = m_{\rm A}+ m_{\rm B}$ and $r = (r_{\rm A}^2 + r_{\rm B}^2)^{1/2}$. This
yields gaps in the mass-size relation (Fig.\
\ref{fig:processing-scheme} [b]). The merger information is preserved
during processing, so that it is possible to look up which fragments
are contained within others.
Our map characterization scheme is thus closely related to the one
employed by \citet{peretto2009:irdc-catalogue}. Like us, these authors
measure sizes and masses for regions bound by lines of constant column
density. The schemes differ in the number of contours considered: we
use a very large number ($10^2$ to $10^3$), where
\citeauthor{peretto2009:irdc-catalogue} only consider two contours per
cloud.\medskip
In practice, we use the dendrogram (i.e., tree analysis) code
presented by \citet{rosolowsky2008:dendrograms} for automatic
processing of the maps. A minimum significant contour has to be set
for every region; emission below this limit is not characterized here.
It is further necessary to specify a minimum contrast between peaks,
in order to identify significant central maxima for the objects. These
threshold column densities and contrasts are listed separately for
every region in the following. The identified maxima are required to
be spaced by more than one spatial resolution element. Again, this
parameter is noted separately for every map. In this work, all sources
with an effective diameter (i.e., $2 r$) smaller than twice the map
resolution are rejected and are treated as parts of enveloping
objects.
In the terminology of \citeauthor{rosolowsky2008:dendrograms}, our
mass measurement approach (i.e., integrate column density within some
contour) corresponds to their ``bijection paradigm''. Such mass
measurements, e.g.\ made towards a dense core, always give a sum over
several spatially overlapping components (e.g., some fraction of the
dense core, plus some fraction of its envelope). This is not a
problem, if properly taken into account in the later analysis. Other
choices are possible (e.g., the ``clipping paradigm'' of
\citeauthor{rosolowsky2008:dendrograms}), but they are less intuitive
and even harder to model. Also, most previously existing data has been
published effectively adopting the bijection paradigm.
\subsection{Mass Estimates\label{sec:mass-estimates}}
As a first example, below we present a mass-size analysis of Perseus,
based on column densities derived from 2MASS-derived extinction
data. As explained by \citet{ridge2006:complete-phase_i}, the map is
derived in terms of magnitudes of visual extinction, $A_V$. We convert
this to $\rm H_2$ column densities using the relation
\begin{equation}
N_{\rm H_2} = 9.4 \times 10^{20} ~ {\rm cm^{-2}} \, (A_V / {\rm
mag})
\label{eq:coldens-dust-extinction}
\end{equation}
\citep{bohlin1978:av_conversion}. Mass surface densities, $\Sigma$,
can then be derived as
\begin{equation}
\Sigma = \mu_{\rm H_2} \, m_{\rm H} \, N_{\rm H_2} \, ,
\end{equation}
where $\mu_{\rm H_2} = 2.8$ is the mean molecular weight per $\rm H_2$
molecule \citep{kauffmann2008:mambo-spitzer} and $m_{\rm H}$ is the
weight of the hydrogen molecule. In practice,
$\Sigma = 0.047 ~ {\rm g \, cm^{-2}} \,
(N_{\rm H_2} / 10^{22} ~ {\rm cm^{-2}}) =
226 ~ M_{\sun} \, {\rm pc^{-2}} \,
(N_{\rm H_2} / 10^{22} ~ {\rm cm^{-2}})$, or
$A_V = 227.5 ~ {\rm mag} \, (\Sigma / {\rm g \, cm^{-2}})$. The mass
is then derived by integrating the mass surface density,
$m = \int \Sigma \, {\rm d} A$. For this we adopt a Perseus distance
of $260 ~ \rm pc$ \citep{cernis1993:distance-perseus}.
\citet{goodman2008:column-density} present a comparison of column
density tracers for the Perseus region. After studying column density
maps based on dust extinction (from 2MASS data), dust emission (from
IRAS imaging), and line emission data (from large field $\rm ^{13}CO$
[1--0] mapping), they conclude that extinction-based maps provide the
best available information on a cloud's spatial mass
distribution. Extinction-based column densities deviate by $\sim 25\%$
from those derived from other tracers (after removing global offsets
between estimates, e.g.\ due to the choice of dust opacities). The
true column density is supposedly in between these estimates. If every
tracer has a similar scatter with respect to the true column density,
this scatter would then be $\sim 25\% / 2^{1/2} \approx 18\%$ for all
tracers. Extinction-based estimates of the column density do probably
deviate by a lower amount from the true value. Here, we thus adopt a
systematic uncertainty of $\lesssim 15\%$ for extinction-based mass
estimates in Perseus.
Young stars embedded in the clouds can further bias extinction
observations, given their red intrinsic colors. This bias is
particularly significant towards clusters, such as NGC1333 and IC348
in Perseus. We do not exclude these regions from our study, but
measurements towards the clusters should be interpreted with
particular caution.
\subsection{Example Map\label{sec:example-map}}
Figure \ref{fig:example_map} shows the aforementioned extinction map
for Perseus. The two most prominent star-forming regions in Perseus,
NGC1333 and IC348, manifest as extended extinction structures in this
map. Fragments of very small size (e.g., $\lesssim 0.1 ~ \rm pc$), like
dense cores around individual young stellar objects, are not visible in
the map, due to its too poor spatial resolution ($5 \arcmin$,
corresponding to $\approx 0.4 ~ \rm pc$). Section
\ref{sec:combine-masses} shows how such structures can still be
bootstrapped into mass-size studies.
\begin{figure}
\includegraphics[height=\linewidth,angle=-90]{./MapPerseus.eps}
\caption{Example column density map for Perseus, presented in terms of
visual extinction, $A_V$. The map is taken from
\citet{ridge2006:complete-phase_i}. \emph{Contours} are drawn in steps of
$1 ~ \rm mag$, starting at $2 ~ \rm mag$. \emph{Labels} indicate the
rough position of the star forming regions NGC1333 and
IC348.\label{fig:example_map}}
\end{figure}
Figure \ref{fig:example_mr-relation} presents the mass-size data
derived from the Perseus extinction map. This diagram reveals that the
Perseus complex has a total (effective) radius of $8 ~ \rm pc$ and a
total mass of $1 \times 10^4 ~ \rm M_{\odot}$. The properties from
contours containing the two major stellar clusters, NGC1333 and IC348,
are highlighted by bold black lines. As one may naively expect, the
cloud fragments enclosing these clusters are, at given radius, the
most massive fragments within the cloud complex. This analysis also
reveals fragments that are, again at given radius, much less massive
than the regions containing the clusters. As discussed in the next
paragraphs, the column density sensitivity of the map sets a
radius-dependent lower limit to the masses that can be detected in a
given map. Fragments of a mass much lower than those shown here may
thus well exist in Perseus. Eventually, at sufficiently large radius,
most of the contour-bound objects found merge into a single fragment
containing essentially all of the Perseus molecular cloud.
\begin{figure}
\includegraphics[width=\linewidth,bb=20 54 361 386,clip]{./mass_perseus-extinction-1.eps}\\
\includegraphics[width=\linewidth,bb=20 10 361 386,clip]{./mass_perseus-extinction-2.eps}
\caption{Mass-size relation for the Perseus molecular cloud. Compared
to Fig.\ \ref{fig:processing-scheme}, we only keep the lines
connecting related measurements. The \emph{top panel} describes the
data in detail. \emph{Bold solid lines} indicate data for fragments
either containing NGC1333 or IC348. \emph{Light solid lines} show
observations for regions not associated with these clusters. The
\emph{dashed line} gives the sensitivity limit of the analysis. In
the \emph{bottom panel} we present the data in the context of
reference mass-size relations. \emph{Dotted lines} indicate mean
$\rm H_2$ column densities, $\langle N_{\rm H_2} \rangle$,
respectively the mass-size relation for a singular hydrostatic
equilibrium sphere supported by isothermal pressure from gas at
$10 ~ \rm K$ temperature (see Sec.\ \ref{sec:reference-relations}
for both).\label{fig:example_mr-relation}}
\end{figure}
In the processing of the map, we have used a minimum threshold
extinction of $2 ~ \rm mag$. This limit is much larger than the noise
level of $0.4 ~ \rm mag$ \citep{ridge2006:complete-phase_i} and
rejects the signal from background extinction structures not related
to the cloud. Correspondingly, we cannot detect fragments with a mean
extinction below $2 ~ \rm mag$. Given the relation between mass, size
and mean column density (Eq.\ \ref{eq:mr_column-density}), this minimum
column density sets a lower limit to the detectable mass. Substitution
of the minimum column density, $N_{\rm H_2, min}$ (in Eq.\
\ref{eq:mr_column-density}; note that
$\langle N_{\rm H_2} \rangle \ge N_{\rm H_2, min}$), gives sensitivity
limits like the one shown in Fig.\ \ref{fig:example_mr-relation}.
The minimum contrast between peaks is required to be the noise level
times a factor 3, i.e.\ $1.2 ~ \rm mag$. The spatial resolution of the
map is $5 \arcmin$; we do therefore require the maxima to be separated
by at least 3 pixels ($7 \farcm 5$), and objects with a radius smaller
than $5 \arcmin$ are rejected as being unphysical.
\section{Properties of Mass-Size Data}
Some mathematical and physical laws governing the properties of
mass-size data have to be heeded, if a meaningful interpretation of
the observations is desired. Here we list the most fundamental of
these.\label{sec:mass-size-properties}
\subsection{Reference Relations\label{sec:reference-relations}}
A number of reference mass-size relations can help to navigate within
the observational data more intuitively. They are in part derived
assuming a spherical geometry for cloud fragments. The assumption of
spherical symmetry may not be appropriate, though. This caveat should
be kept in mind when using the following reference relations.
Mass and size measurements can be used to calculate the mean mass
surface (or column) density of a fragment,
$\langle \Sigma \rangle = m(r) / A(r)$. Conversely, one can draw lines
of constant mass surface density,
$m(r) = \langle \Sigma \rangle \, \pi \, r^2$. Conversion to column
density, and substitution of the aforementioned constants, yields
\begin{equation}
m(r) = 71 \, M_{\sun} \,
(\langle N_{\rm H_2} \rangle / [10^{21} ~ {\rm cm^{-2}}]) \,
(r / {\rm pc})^2 \, .
\label{eq:mr_column-density}
\end{equation}
These lines are drawn in most mass-size plots presented here (e.g.,
Fig.\ \ref{fig:example_mr-relation}). Equation
(\ref{eq:mr_column-density}) implies one of the most important
properties of mass-size data: since column density decreases with
increasing radius\footnote{By definition, this is always the case for
our source characterization scheme. We start off from a column
density peak and then consider, by design, regions that increase
with size when lowering the threshold column density.}, the observed
mass-size relations must be flatter than $m(r) \propto r^2$.
Equation (\ref{eq:mr_column-density}) can actually be used to
transform our results into the analytical diagram presented by
\citeauthor{tan2007:massive-sf} (\citeyear{tan2007:massive-sf};
column-density vs.\ mass). Our study goes beyond the work by
\citet{tan2007:massive-sf} in that it systematically populates the
parameter space with observational data.
For spherical clouds, the mean density is
$\langle \varrho \rangle = m(r) / (4/3 \, \pi \, r^3)$. The
corresponding mass-size relation reads
$m(r) = 4/3 \, \pi \, \langle \varrho \rangle \, r^3$, or
\begin{equation}
m(r) = 282 \, M_{\sun} \, (\langle n_{\rm H_2} \rangle / [100 ~ {\rm cm^{-3}}]) \,
(r / {\rm pc})^3 \, ,
\end{equation}
where we substitute the density of $\rm H_2$ molecules,
$n_{\rm H_2} = \varrho / (\mu_{\rm H_2} \, m_{\rm H})$.
In part II of this series, we shall study spherical power-law
density profiles, $\varrho(s) \propto s^{-k}$ (where $s$ is the
radius), as models for the observed mass-size relations. We show that
\begin{equation}
\varrho(s) \propto s^{-k} \quad \Leftrightarrow \quad
m(r) \propto r^{3-k} \, .
\label{eq:slope-vs-density}
\end{equation}
The slope of the mass-size relation is therefore related to the slope
of the density law. A density profile $\varrho(s) \propto s^{-2}$ is
often adopted to describe dense cores (see
\citealt{dapp2009:density-profile} for a discussion). This gives a
mass-size relation $m(r) \propto r$. In hydrostatic spheres supported
by isothermal pressure, mass, size, and gas temperature are related by
\begin{equation}
m(r) = 2.6 \, M_{\sun} \,
\left( \frac{T_{\rm g}}{10 ~ \rm K} \right) \,
\left( \frac{r}{0.1 ~ \rm pc} \right)
\label{eq:mass-size-sis}
\end{equation}
(see \citealt{kauffmann2008:mambo-spitzer}, Eq.\ 13). For gas
temperatures $T_{\rm g} = 10 ~ \rm K$, one obtains the model relation
drawn in most mass-size diagrams of this paper (e.g., Fig.\
\ref{fig:example_mr-relation}).
We stress that we obtain two-dimensional mass-size relations from
column density maps. These are related to, but not identical with,
mass-size laws obtained from three-dimensional density maps. This is
illustrated by the experiments conducted by
\citeauthor{shetty2009:ppp-ppv} (\citeyear{shetty2009:ppp-ppv},
submitted) who use the fragment identification technique also used by
us. Their analysis is based on three-dimensional numerical simulations
of turbulent clouds. As part of their experiments, they fit power-laws
(similar to Eq.\ \ref{eq:mass-size-larson}) to their mass-size
data. For their particular set of simulations, the exponent derived in
this fashion is similar to the number of dimensions used for mass
measurements (i.e., 3 when based on density, and 2 when based on
column density). This underlines that the number of dimensions
considered has to be kept in mind. Note, though, that these details do
not compromise mass-size measurements as a tool for cloud structure
analysis. \emph{Observed} mass-size laws unambiguously summarize
actual cloud structure. Only their \emph{interpretation} is sensitive
to the assumed geometry.
\subsection{Relation to CLUMPFIND-like Results}
The approach chosen here constitutes one of several possible choices
to measure the mass and size of objects in maps. Another popular
approach is to use the CLUMPFIND algorithm
\citep{williams1994:clumpfind}. This method uses contours to break
emission up into several objects, just as done here.
CLUMPFIND-extracted boundaries do, however, not necessarily follow
contours. This is a major difference to our method, where objects are
always bound by some column density contour. Based on this fact alone,
CLUMPFIND and our approach will thus extract very different
objects. \citet{goodman2008:nature} illustrate this problem
comprehensively.
Further, CLUMPFIND does not allow for hierarchical structure, i.e.,
fragments containing other fragments. For example, CLUMPFIND will
never determine the integral properties of the entire cloud, since the
cloud is usually broken up into many independent fragments. The
relation of CLUMPFIND-results to cloud hierarchy is studied by
\citet{pineda2009:clumpfind}.
\begin{figure}
\includegraphics[width=\linewidth,bb=15 6 360 395,clip]{./mass_comp-clumpfind.eps}
\caption{CLUMPFIND (\emph{crosses}) and dendrogram mass-size results
(\emph{lines}) for the extinction map presented in Fig.\
\ref{fig:example_map}. The two approaches yield broadly similar,
but not directly related results.\label{fig:comp-clumpfind}}
\end{figure}
Figure \ref{fig:comp-clumpfind} compares mass and size measurements
from CLUMPFIND to those from our approach\footnote{The CLUMPFIND
results were kindly provided by J.E.\ Pineda.}. Both
characterizations are based on our Perseus extinction map. To initiate
CLUMPFIND, we choose contours as close as possible to the parameters
used in our contour-based segmentation (Sec.\ \ref{sec:example-map});
the contour spacing is $1.2 ~ \rm mag$, with the lowest contour at
$2 ~ \rm mag$. Still, this CLUMPFIND segmentation yields objects that
are only remotely related to the regions extracted by our method. To
start with, CLUMPFIND extracts 52 peaks enclosed by individual clumps,
where our method only identifies 17 significant peaks. This
discrepancy is a consequence of CLUMPFIND's relaxed peak
identification criterion: every local peak encircled by a continuous
contour is considered significant, independent of the depth of the
column density dip towards the next peak. There is thus no
well-defined correspondence between regions identified by the
different algorithms.
To characterize the relation between objects extracted by different
methods in some sense, we inspect the structure around the 17
significant column density peaks found by our method. For every such
peak, we identify the CLUMPFIND object containing this peak. The mass
of the latter clump can then be compared to our results, taken at the
clump's radius. Because of merging of objects, our method does not
provide a mass measurement for every possible radius (consider, e.g.,
the evolution of mass-size measurements starting from peak 3 in Fig.\
\ref{fig:processing-scheme}[b]). For those cases where mass
measurements exist for the CLUMPFIND-derived radius, we find that
CLUMPFIND gives masses of order 55\% to 95\% of the masses derived by
our approach.
In summary, mass and size measurements from CLUMPFIND are thus broadly
compatible with our results. By this we mean that the
CLUMPFIND-derived mass-size measurements reside in the space spanned
by our own measurements. There is, however, no good correspondence on
a detailed level.
\section{Perseus in Detail}
Given the tools derived above, it is now possible to characterize the
mass-size relation for Perseus. While doing so, we also evaluate
uncertainties affecting our analysis. In a final step, we extend the
analysis to clouds other than Perseus.\label{sec:all-scales_perseus}
\subsection{Mass Uncertainty\label{sec:uncertainty-mass}}
To explore the impact of noise on mass measurements, we run
Monte-Carlo experiments in which we add artificial Gaussian noise of
root-mean-square (RMS) amplitude $\sigma(m_i)$ to the map
before structure characterization (where $m_i$ is the mass per
pixel). Such trials are presented in Fig.\
\ref{fig:comparison-noise-resolution}(a). Comparison of the derived
mass-size data to the one from the original map does then reveal the
impact of noise. In our experiments, we find that
\begin{equation}
\frac{\sigma(m)}{m} = 6 \,
\frac{\sigma(m_i)}{m} \, \frac{r}{r_{\rm beam}}
\label{eq:uncertainty-noise}
\end{equation}
is an upper limit to the noise-induced mass changes, $\sigma(m)$. In
this, $r_{\rm beam}$ is the beam radius and the numerical constant
does slightly depend on the number of pixels per beam.
Equation (\ref{eq:uncertainty-noise}) follows from Gaussian error
propagation of $m = \Sigma_i m_i$. The numerical constant is increased
by a factor 3, though, to provide a strict upper limit to
uncertainties. To test for non-Gaussian sources of error, we validate
Eq.\ (\ref{eq:uncertainty-noise}) by comparing the original data with
those derived from maps with additional noise. For any given
structural branch with $r < 2 ~ \rm pc$, Eq.\
(\ref{eq:uncertainty-noise}) is indeed found to set an upper limit to
the mass deviation at given radius, following the noise experiments
depicted in Fig.\ \ref{fig:comparison-noise-resolution}(a). For larger
radii, a small (but significant) additional uncertainty of $\sim 1 \%$
has to be included to capture non-Gaussian sources of error.
In our study, the relative uncertainty of column density estimates
(i.e., $\sigma(m_i) / m_i$) is $\le 1/5$ in characterized regions
(given extinction noise levels and selection thresholds of
$0.4 ~ \rm mag$ and $2.0 ~ \rm mag$, respectively). Further, the mass
is at least as large as the one at the sensitivity limit (Fig.\
\ref{fig:example_mr-relation}[a]),
$m > \pi \, m_{i, \rm min} \, (r / r_{\rm beam})^2$ (in our map, the beam
contains, by area, $\pi$ pixels). Substitution of these parameters in
Eq.\ (\ref{eq:uncertainty-noise}) give uncertainties of 19\% for the
smallest extracted features (for which $r = 2 r_{\rm beam}$) located
just at the sensitivity limit. Since
$\sigma(m) / m \propto \langle N_{\rm H_2} \rangle^{-1} \, r^{-1}$
(via substitution of Eq.\ [\ref{eq:mr_column-density}]),
larger regions well above the detection threshold will suffer lesser
uncertainties, $\ll 10\%$.\medskip
\begin{figure}
\begin{tabular}{lc}
\includegraphics[width=0.85\linewidth,bb=15 31 360 395,clip]{./mass_comp-noise.eps} &
\begin{sideways}
\textsf{\hspace{1.8cm}\large{}(a) impact of noise}
\end{sideways}\\
\includegraphics[width=0.85\linewidth,bb=15 6 360 390,clip]{./mass_comp-resolution.eps} &
\begin{sideways}
\textsf{\hspace{2.5cm}\large{}(b) impact of smoothing}
\end{sideways}
\end{tabular}
\caption{Impact of noise (\emph{top}) and limited spatial resolution
(\emph{bottom}) on mass-size measurements. \emph{Note that the size
and mass scale differ between the panels.} Noise and resolution
are explored using artificial maps created by adding noise to
(\emph{top}), respectively smoothing of (\emph{bottom}), the
observed data for Perseus (Sec.\ \ref{sec:uncertainty-mass}).
Differences between the input data and the properties of the
artificial maps are due to these biases. The original input data is
drawn in \emph{black}. \emph{yellow lines in the top panel}
show mass-size relations from maps with additional noise similar
to the observed one ($0.4 ~ \rm mag$). \emph{Red and yellow
lines in the bottom panel} present mass-size relations for maps
with a resolution worsened to
$r_{\rm beam, f} / r_{\rm beam, i} = 3/2$ and $5/2$, respectively
(and features with $r < 2 \, r_{\rm beam, f}$ removed). At given
radius, both biases just induce moderate changes in mass (Sec.\
\ref{sec:uncertainty-mass}).\label{fig:comparison-noise-resolution}}
\end{figure}
In later papers, we will compare the properties of clouds that are
located at different distances. Since the angular resolution of the
observations is about constant, an increase in distance implies a
decrease in physical resolution. To first order, this decrease in
physical resolution corresponds to smoothing of the map. We explore
the impact of smoothing by extracting mass-size data from smoothed
maps of the Perseus cloud, as demonstrated in Fig.\
\ref{fig:comparison-noise-resolution}(b).
Smoothing does effectively mean that mass is transferred to larger
scales. For fixed radius, smoothing does thus imply a reduction in
mass. To estimate this reduction, consider a region of effective
radius $r$. After smoothing, the mass contained in this
region will be smeared our over an area of radius
$r' \sim r_{\rm ini} + r_{\rm k}$. In this,
$r_{\rm k} = (r_{\rm beam, f}^2 - r_{\rm beam, i}^2)^{1/2} /
(\ln(2))^{1/2}$ is the effective radius of the smoothing kernel
required to go from the initial to the final beam radius of the map
during smoothing, $r_{\rm beam, i} \to r_{\rm beam, f}$. Very
approximately, after smoothing the mass retained in the initial area
will be of order of the fraction of the initial area to the one after
smoothing, $r^2 / (r + r_{\rm k})^2$.
After rearrangement (and using $r_{\rm k} \ll r$), we find that the
smoothing-induced mass bias should obey a relation of the form
\begin{equation}
\frac{\Delta m}{m} \approx 0.2 ~ {\rm to} ~ 0.3 \cdot
\left[ 1 -
\frac{1}{(1 + r_{\rm k} / r)^2} \right] \, .
\label{eq:uncertainty-resolution}
\end{equation}
To allow for more realistic source geometries, the numerical constant
must be derived from experiments with actual data. We use the
smoothing experiments shown in Fig.\
\ref{fig:comparison-noise-resolution}(b) for this purpose. With these
parameters, Eq.\ (\ref{eq:uncertainty-resolution}) describes the
relative mass bias for all structure branches in our example map.
In our study, we only consider regions with radii larger than twice
the effective beam radius, $r > r_{\rm beam}$. Substitution of this
(with $r_{\rm beam, i} = 0$, since the true distribution on the sky
has infinite resolution) in Eq.\
(\ref{eq:uncertainty-resolution}) shows that the mass of regions
with a radius similar to the beam diameter may be underestimated by
not more that 20\%, and by less for larger fragments.
\subsection{Combining Dust Emission and Extinction
Data\label{sec:combine-masses}}
Our Perseus extinction map has a spatial resolution of $5 \arcmin$
($0.4 ~ \rm pc$). It cannot be used to characterize much smaller
fragments, such as dense cores of $\lesssim 0.1 ~ \rm pc$ size. This
limitation can, however, be overcome when including mass-size data
from additional data sets.
\subsubsection{Basic Concept}
For Perseus, \citet{enoch2006:perseus} present a Bolocam map of dust
emission in Perseus at $1.1 ~ \rm mm$ wave length. The beam width at
half sensitivity is $31 \arcsec$, corresponding to $0.04 ~ \rm pc$. At
these wave lengths, the continuum emission of dust at temperature
$T_{\rm d}$ is optically thin, and so the observed dust continuum
emission intensity, $I_{\nu}$, is a direct measure of the column
density along a given line of sight,
\begin{equation}
I_{\nu} = \mu_{\rm H_2} \, m_{\rm H} \, \kappa_{\nu} \,
N_{\rm H_2} \, B_{\nu}(T_{\rm d})
\label{eq:coldens-dust-emission}
\end{equation}
(see \citealt{kauffmann2008:mambo-spitzer} for full details), where
$B_{\nu}$ is the Planck function, and the dust opacity, $\kappa_{\nu}$,
is evaluated per total gas mass. Column density maps from dust
emission can then be contoured and analyzed as described before (Sec.\
\ref{sec:method}) to derive mass-size data for the fragments in the
dust emission map.
\begin{figure*}
\includegraphics[scale=0.55,bb=20 10 361 386,clip]{./mass_perseus-bw-1.eps}
\includegraphics[scale=0.55,bb=76 10 361 386,clip]{./mass_perseus-bw-2.eps}
\includegraphics[scale=0.55,bb=76 10 361 386,clip]{./mass_perseus-bw-3.eps}
\caption{Mass-size data for Perseus (\emph{solid lines}). The data are
from dust extinction maps probing large spatial scales, and Bolocam
dust emission observations sensitive to small ones. The \emph{left
panel} highlights the nature of the data; \emph{dashed lines} give
the instrument-dependent sensitivity limits. \emph{Circles and the
dotted line} indicate calculation of the global mass-size slope
(Sec.\ \ref{sec:slope_global-perseus}). The \emph{middle panel}
highlights slopes derived by matching power-laws to parts of the
data. Tangential slopes (Sec.\ \ref{sec:slopes-tangential}) are
derived in the same fashion, but for infinitesimal radius
ranges. The \emph{right panel} presents Perseus in the context of
the reference mass-size relations from Sec.\
\ref{sec:reference-relations}. Here, \emph{black solid lines}
highlight Bolocam-detected fragments in NGC1333, as well as the
extinction-probed fragments containing this cluster. Other data are
drawn using \emph{grey solid lines}.\label{fig:emission-extinction}}
\end{figure*}
Figure \ref{fig:emission-extinction} combines the Bolocam-derived
mass-size measurements with the extinction-based ones. Spatially, they
are separated by a gap, since Bolocam ($\lesssim 2 \arcmin$) and 2MASS
extinction maps ($\gtrsim 5 \arcmin$) probe different spatial
scales. The extinction-identified fragments do, however, contain the
Bolocam-derived ones. For example, shading in Fig.\
\ref{fig:emission-extinction} highlights Bolocam-detected fragments
contained in the extinction peak harboring NGC1333.
\subsubsection{Relative Calibration of Dust Opacities}
\citet{bianchi2003:dust-emissivity} studied B68 to examine the
relation of mass estimates from dust extinction and emission. They
find relative differences by factors of $1 ~ {\rm to} ~ \sim 2.35$,
when adopting opacities for \citet{ossenkopf1994:opacities} dust
grains with thin ice mantles that coagulate for $10^5 ~ \rm yr$ at
$10^6 ~ \rm cm^{-3}$ density (see
\citealt{kauffmann2008:mambo-spitzer}, Table A.1, for numerical
values). This is consistent with previous studies
of this subject, including sources as diverse as diffuse clouds and
galaxies (see references in \citealt{bianchi2003:dust-emissivity}). If
one fixes the opacity used for extinction observations, it thus
appears that the the \citet{ossenkopf1994:opacities} model opacities
near $1 ~ \rm mm$ wave length are too large by an average factor
$1.5 \approx 2.35^{1/2}$.
In our emission-based column density estimates, we do therefore
basically adopt the aforementioned \citet{ossenkopf1994:opacities}
opacities for dust emission observations near $1 ~ \rm mm$ wave
length, but scale these down by a further factor 1.5 to bring mass
estimates from extinction into harmony with those from dust
emission. As seen in Fig.\ \ref{fig:emission-extinction}, this
procedure yields reasonable results, since the dust emission and
extinction observations for NGC1333 match within less than a factor 2
(i.e., when comparing the masses of the most massive Bolocam-detected
fragment to the one of the least massive 2MASS-identified
fragment). Incorrect assumptions about dust temperatures may cause
most of this offset (see below).
To some extent, correction factors between masses from dust emission
and extinction are also influenced by spatial filtering affecting
bolometer observations (Sec.\
\ref{sec:mass-uncertainties_bolo-filtering}). Future studies need to
address this problem in more detail. Still, the aforementioned scaling
factor aligns dust emission and extinction studies, which is the only
aspect needed in our present study.
\subsubsection{Dust Temperatures}
For Perseus, \citet{rosolowsky2007:perseus_nh3} estimate gas
temperatures between $9 ~ \rm K$ and $18 ~ \rm K$. Assuming that dust
and gas temperatures are similar, we do therefore adopt a temperature
of $(12.5 \pm 2.5) ~ \rm K$ in our mass estimates. This temperature
uncertainty results in a relative mass uncertainty of $\sim 20\%$ for
Perseus. A few sources may have dust temperatures, and mass biases,
outside of this range.
\subsubsection{Impact of Spatial
Filtering\label{sec:mass-uncertainties_bolo-filtering}}
Like all other upcoming, present, and past ground-based
bolometer-derived dust emission maps
\citep{kauffmann2008:mambo-spitzer}, the Bolocam maps of Perseus are
not sensitive to structure larger than some instrument-dependent
spatial scale. In its impact, this problem is similar to the spatial
filtering in interferometric imaging. For Bolocam,
\citet{enoch2006:perseus} show the filtering scale to be of order
$1 \arcmin$ to $2 \arcmin$ radius.
Quantitatively, this removal of large-scale structure has an influence
opposite to the impact of smoothing (Sec.\
\ref{sec:uncertainty-mass}): here, the relative mass-loss increases
with spatial scale. This bias has to be considered carefully when
using emission-based mass-size measurements for analysis. Obviously,
the true mass will be larger than the observed value. Filtering in
bolometer maps is unfortunately too complex to be characterized in a
compact fashion; see \citet{kauffmann2008:mambo-spitzer} for a few
rules of thumb. For the particular case of the Bolocam maps of
Perseus, \citet{enoch2006:perseus} report losses $\le 10\%$ for
radii $\le 1 \arcmin$, but do not sufficiently explore larger
objects.
\subsection{Global Trends in Maximum Mass for given
Size\label{sec:slope_global-perseus}}
The mass-size tendencies seen, e.g., in Fig.\
\ref{fig:emission-extinction} suggest to describe these trends with
power laws of the form
\begin{equation}
m(r) = m_0 \, (r / r_0)^b \, .
\label{eq:power-law}
\end{equation}
In this, $b$ is the slope of the relation, and $m_0$ is the
intercept. As shown in Fig.\ \ref{fig:emission-extinction}, such laws
trace, e.g., the mass-size relation of the most massive fragments in
the $0.04 \le r / {\rm pc} \le 0.3$ and $1 \le r / {\rm pc} \le 8$
radius intervals with maximum deviations of $\pm 6\%$.
Beyond such detailed descriptions of individual features, power-laws
can be used to capture more global aspects of a cloud's
structure. Consider, for example, the relation between the maximum
fragment masses, $m_{\rm max}$, observed at radii of
$r_{\rm sm} = 0.05 ~ \rm pc$ and $r_{\rm lg} = 3.0 ~ \rm pc$. (These
radii are chosen to permit comparison with other clouds, as becomes
more obvious below.) Based on these, we can define a global slope,
\begin{equation}
b_{\rm glob} = \frac{
\ln[m_{\rm max}(r_{\rm lg}) / m_{\rm max}(r_{\rm sm})]
}{
\ln[r_{\rm lg} / r_{\rm sm}]
} \, .
\label{eq:slope-global}
\end{equation}
As illustrated in Fig.\ \ref{fig:emission-extinction}, this slope is
defined such that Eq.\ (\ref{eq:power-law}) connects the mass
and size measurements for $b = b_{\rm glob}$. In Perseus,
$b_{\rm glob} = 1.21 \pm 0.14$, where we evaluate the
uncertainty very conservatively by scaling the emission-based mass in
Eq.\ (\ref{eq:slope-global}) up and down by a factor 2.
In Perseus, at the chosen scale of $r = 0.05 ~ \rm pc$ (i.e.,
$40\arcsec$), the Bolocam dust emission map is not significantly
affected by spatial filtering. Similarly, at $r = 3 ~ \rm pc$
($40\arcmin$), the extinction map is not biased by presence of stellar
clusters. The measurements of $b_{\rm glob}$ are thus immune to these
influences.
\subsection{Local Trends in Mass\label{sec:slopes-tangential}}
We now turn to slopes of infinitesimal tangents. As we explain below,
the Bolocam data are not suited for this analysis, since they suffer
from too strong spatial filtering.
\subsubsection{Method and Uncertainties}
In some cases, one may wish to use Eq.\ (\ref{eq:power-law}) to
describe tangents to the data on spatial scales smaller than the one
for which we calculate global slopes. This is demonstrated in Fig.\
\ref{fig:emission-extinction} using tangents to the most massive
fragments seen at given radius. In particular, it can be desirable to
fit infinitesimal tangents to mass-size trends. For these, the slope
reads
\begin{equation}
b(r) =
\left.
\frac{{\rm d} \, \ln(m[r'])}{{\rm d} \, \ln(r')}
\right|_{r' = r} \, .
\label{eq:slopes_tangential}
\end{equation}
Figure \ref{fig:comp-slopes-smoothing}(a) shows slopes derived from
mass and size differences between consecutive contours. As seen in the
figure, these data are rather noisy. We do therefore smooth the
measurements. To do this, at radius $r$, we replace slope and radius
by their respective unweighted arithmetic means, as derived within a
smoothing kernel of width $[r, 1.15 \cdot r]$ (for a
given cloud fragment, not permitting mergers). This yields the data
shown in Fig.\ \ref{fig:comp-slopes-smoothing}(b). Sometimes, however,
the smoothing kernel is not filled well, and the data are still
noisy. Thus, we do finally remove all data where the kernel is not
filled to at least $2/3$. Figure \ref{fig:comp-slopes-smoothing}(c)
shows this final result.
\begin{figure}
\begin{tabular}{cc}
\includegraphics[width=0.92\linewidth,bb=20 55 360 200,clip]{./slopes_comp-noise-unsmoothed.eps} &
\begin{sideways}
\textsf{\hspace{0.4cm}a) plain differences}
\end{sideways}\\
\includegraphics[width=0.92\linewidth,bb=20 55 360 188,clip]{./slopes_comp-noise-unfiltered.eps} &
\begin{sideways}
\textsf{\hspace{0.7cm}b) smoothed}
\end{sideways}\\
\includegraphics[width=0.92\linewidth,bb=20 10 360 188,clip]{./slopes_comp-noise.eps} &
\begin{sideways}
\textsf{\hspace{1.cm}c) smoothed and filtered}
\end{sideways}
\end{tabular}
\caption{Calculation scheme for slopes (\emph{top to bottom}) and
impact of noise (\emph{black vs.\ yellow lines}). We start with slopes
directly calculated from mass and size differences for
successive contours (panel a). These measurements are subsequently
smoothed (panel b), and data are removed where the smoothing kernel
is not filled (panel c). The \emph{black lines} hold for the
observed Perseus extinction data. As in Fig.\
\ref{fig:comparison-noise-resolution}, \emph{yellow and black lines}
indicate results from maps with and without artificial noise,
respectively. The \emph{dotted line} indicates the upper limit on
slopes inherent to our method, $b < 2$ (Eq.\
\ref{eq:mr_column-density}).\label{fig:comp-slopes-smoothing}}
\end{figure}
The impact of noise can be estimated by propagating the mass
uncertainties due to noise (Eq.\ \ref{eq:uncertainty-noise}) within
the slope calculations. This yields
\begin{equation}
\frac{\sigma(b)}{b} \le 20 \,
\frac{\sigma(m_i)}{m} \, \frac{r}{r_{\rm beam}} \, .
\label{eq:uncertainty-smooting-filtering_slopes}
\end{equation}
Because of the aforementioned smoothing, the numerical constant must
be calibrate with our noise experiments (as done for Eq.\
[\ref{eq:uncertainty-noise}]). To obtain an estimate of the expected
uncertainties, we can repeat the mass and size substitutions done in
the discussion of Eq.\ (\ref{eq:uncertainty-noise}). This gives
maximum uncertainties $\sim 60\%$. Since
$\sigma(b) / b \propto \langle N_{\rm H_2} \rangle^{-1} \, r^{-1}$
(see discussion of Eq.\ [\ref{eq:uncertainty-noise}]), the
uncertainties are small for larger regions well above the detection
threshold. In practice, uncertainties $< 10\%$ are a reasonable
estimate for well-detected regions warranting detailed study.
The slope difference due to smoothing is given by the first derivative
of the mass bias due to smoothing (Eq.\
\ref{eq:uncertainty-resolution}) with respect to the radius. Including
the usual calibration of numerical constants, we constrain the
absolute smoothing-induced error to
\begin{equation}
\Delta b \le 0.4 ~ {\rm to} ~ 0.6 \,
\frac{r_{\rm k} / r}{(1 + r_{\rm k} / r)^3} \, .
\label{eq:uncertainty-resolution_slope}
\end{equation}
In a few cases, however, the error can be larger by a factor 2. In
this paper, we reject regions with a diameter smaller than twice the
beam diameter. Substitution of this limit (i.e.,
$r_{\rm k} / r < 1/2$) into Eq.\
(\ref{eq:uncertainty-resolution_slope}) implies that slopes are
overestimated by a number smaller than 0.1 due to resolution.
\begin{figure}
\includegraphics[width=\linewidth,bb=20 10 360 200,clip]{./slopes_comp-resolution.eps}
\caption{Impact of limited spatial resolution on slope
measurements. See Figs.\ \ref{fig:comparison-noise-resolution}(b)
and \ref{fig:comp-slopes-smoothing} for explanations of
mark-up.\label{fig:comp-slopes-resolution_noise-multi}}
\end{figure}
As mentioned in Sec.\ \ref{sec:mass-uncertainties_bolo-filtering},
bolometer maps suffer from spatial filtering. This has an impact
opposite to the influence of smoothing, and artificially shallow
slopes are measured from such maps. Since the filtering is very strong
in Bolocam maps, we do not use these for the derivation of tangential
slopes.
\subsubsection{Results for Perseus\label{sec:slopes-infinitesimal-perseus}}
Figures \ref{fig:comp-slopes-smoothing}(c) and
\ref{fig:comp-slopes-resolution_noise-multi} show tangential slopes
for Perseus. The tangential slopes in the $1 \le r / {\rm pc} \le 8$
radius interval are in the range $b = 1.2 ~ {\rm to} ~ 1.7$.
At a given radius, the slopes for different fragments do often differ
by more than their uncertainty. Also, in a given fragment the slope
can change significantly with respect to radius. This means that it is
not possible to describe an entire cloud by a single tangential
slope. The observed tangential slopes bear no obvious relation to the
global slope $\sim 1.56$ derived for the same radius range (Fig.\
\ref{fig:emission-extinction}).
A slope of $\ge 2$ is actually not possible for a mass-size relation
as defined by us, since this would require that the column density
increases with radius (Eq.\ \ref{eq:mr_column-density}). Since our
search algorithm proceeds by decreasing the column density threshold,
such fragments are not identified by our identification scheme. A
slope $b < 0$ would mean that the mass decreases with increasing
radius. This is impossible in all but the most jolly insane models of
cloud structure.
\section{Summary \& Outlook}
\subsection{Utility of and Outlook for Mass-Size
Studies\label{sec:purpose}}
As we have shown above, our map characterization scheme yields
reliable measurements of mass and size. From these, mass-size slopes
and intercepts can be derived. Below, we describe how these data
contribute to critical fields of star formation research.\medskip
First of all, this approach permits a \emph{continuous}
characterization of cloud structure across a large range of spatial
scales. This is just a desirable feature of any data analysis method,
independent of the exact nature of the later analysis. The need for
such a procedure led \citet{rosolowsky2008:dendrograms} to the
development of the ``dendrogram technique''.
In star formation research, the basic mass-size measurements permit to
compare fragment masses at a given spatial scale. Consider the
classical case in order to see the advantage: usually, ``cores'' and
``clumps'' extracted from maps differ in their size. In this case, it
is not clear what differences in masses mean, even if just a single
cloud is considered.
Spatially continuous cloud characterizations become particularly
useful when comparing observations for different molecular
clouds. Usually, every cloud is studied at a different physical
resolution (i.e., pc per pixel). In the classical case, mass
measurements will thus usually refer to vastly different spatial
scales. With our approach, however, all scales are probed, and
different clouds can be compared at the same physical scale. This is
extensively employed in part II, where we study a sample of
clouds.\medskip
The general utility of measurements of mass and size is known since
long. For example, one can compare the actual to virial masses or,
more generally, masses predicted by theoretical cloud models. Equation
(\ref{eq:mass-size-sis}), for example, relates model fragment masses
and gas temperatures. We shall not discuss such considerations here in
detail.
A property \emph{uniquely} constrained by our method are mass-size
slopes; these can \emph{only} be measured via a scale-independent
method. In particular, this gives access to the density structure of
molecular clouds. For simple models of cloud structure, the mass-size
slope is, e.g., directly related to the slope of the density profile
(Eq.\ \ref{eq:slope-vs-density}). Such work on large-scale structure
in molecular clouds is urgently needed, since cloud density profiles
are presently not known on scales $\gtrsim 0.1 ~ \rm pc$. Part II
presents slope measurements for various molecular clouds, as well as
several model mass-size relations.
\subsection{Summary\label{sec:summary}}
This work studies the internal structure of molecular clouds by
breaking individual cloud complexes up into several nested
fragments. For these, we derive masses and sizes, as e.g.\ outlined in
Fig.\ \ref{fig:processing-scheme}. Effectively, we perform a
``dendrogram analysis'' of a two-dimensional map, as introduced by
\citet{rosolowsky2008:dendrograms}.
The present paper establishes the method via a detailed analysis of
the Perseus Molecular Cloud. Other solar neighborhood molecular clouds
($\lesssim 500 ~ \rm pc$; the Pipe Nebula, Taurus, Ophiuchus, and
Orion) are discussed in the next paper of this series (part II).
Power-laws of the form $ m(r) = m_0 \, (r / {\rm pc})^b$,
with slope $b$ and intercept $m_0$, prove useful to quantify
the relations between mass, $m$, and size, $r$ (i.e., the
effective radius). Sections \ref{sec:slope_global-perseus} and
\ref{sec:slopes-tangential} discuss two different approaches to define
and measure the slope. We use \emph{global slopes} to measure the
relation between structure at small and large scales. This is done by
connecting the mass-size measurements of the most massive fragments at
small ($0.05 ~ \rm pc$) and large radius ($3.0 ~ \rm pc$) by a
power-law (Eq.\ \ref{eq:slope-global}). \emph{Tangential slopes}, on
the other hand, are calculated infinitesimally at a given spatial
scale (Eq.\ \ref{eq:slopes_tangential}). The uncertainties in these
properties are examined in Secs.\ \ref{sec:uncertainty-mass} and
\ref{sec:combine-masses} (for mass and intercept), respectively Secs.\
\ref{sec:slope_global-perseus} and \ref{sec:slopes-tangential} (for
slopes).
We conclude that our mass, slope, and intercept measurements provide a
reliable method to characterize cloud structure. Our approach enables
a continuous and reliable characterization of cloud structure in the
$0.05 \lesssim r / {\rm pc} \lesssim 10$ spatial range. This is not
possible using previous methods, since these are usually biased
towards a particular spatial scale (see, e.g., the CLUMPFIND analysis
in Fig.\ \ref{fig:comp-clumpfind}). Such comprehensive pictures of
star-forming regions can be used to develop a more complete
theoretical understanding of global cloud structure (Sec.\
\ref{sec:purpose}).
A first observational and theoretical exploitation of this method is
presented in part II of this series. We characterize, for example, the
typical parameter space for solar neighborhood molecular clouds not
forming massive stars. Based on this, we chart a potential mass-size
threshold for the formation of massive stars. Mass-size slopes are
used to constrain large-scale density gradients within molecular
clouds.
\acknowledgements{We are grateful to Nicolas Peretto, who served as a
considerate and knowledgeable referee who helped to significantly
improve the quality of the paper. This project would not have been
possible without help from Erik Rosolowsky. His dendrogram analysis
code \citep{rosolowsky2008:dendrograms} was instrumental for our
analysis. We thank Jaime Pineda for his help with the CLUMPFIND
experiments presented in Fig.\ \ref{fig:comp-clumpfind}.
\citet{enoch2006:perseus} contributed maps to the present study. We
are grateful for their help. This work was in part made possible
through Harvard Interfaculty Initiative funding to the Harvard
Initiative in Innovative Computing (IIC).}
\bibliographystyle{apj}
|
\section*{Introduction}
The aim of this paper is to give a unifying description of various constructions (subanalytic, semialgebraic, o-minimal site) using a slight modification of the notion of $\mathcal{T}$-topology introduced by Kashiwara and Schapira in \cite{KS01}. Sheaf theory on locally semialgebraic spaces was studied in \cite{De91}. Subanalytic sheaf theory was introduced in \cite{KS01} and then developed in \cite{Pr1}. O-minimal sheaf theory was introduced in \cite{ejp} in the definable case but it can be applied also to the locally definable case. Note that since locally semialgebraic spaces are locally definable spaces in a real closed field and real closed fields are o-minimal structures and, relatively compact subanalytic sets are definable sets in the o-minimal expansion of the field of real numbers by restricted globally analytic functions, both the semialgebraic and subanalytic sheaf theory are special cases of the o-minimal sheaf theory. The idea of all of these constructions is the same: on a topological space $X$ one chooses a subfamily $\mathcal{T}$ of the open subsets of $X$ satisfying some suitable hypothesis, and for each $U \in \mathcal{T}$ one defines the category of coverings of $U$ as the topological covering $\{U_i\}_{i \in I} \subset \mathcal{T}$ of $U$ admitting a finite subcover. In this way one defines a site $X_\mathcal{T}$. This idea was already present in \cite{KS01}. So it is natural to study the category of sheaves on $X_\mathcal{T}$ (called $\mod(k_\mathcal{T})$) and treat in a unifying way all the previous construction. Moreover, the exigence to treat some non Hausdorff cases (as conic subanalytic sheaves which are related to the extension of the Fourier-Sato transform \cite{Pr07}) and the non-standard setting which appears naturally in the o-minimal context, motivates a slight modification of the definition of \cite{KS01}.\\
We call a topological space $X$ with such a subfamily $\mathcal{T}$ of open subsets a $\mathcal{T}$-space and the associated Grothendick topology the $\mathcal{T}$-topology. Given a $\mathcal{T}$-space $X$ we study the category of sheaves on the site $X_\mathcal{T}$. The natural functor of sites $\rho:X \to X_\mathcal{T}$ induces relations between the categories of sheaves on $X$ and $X_\mathcal{T}$, given by the functors $\rho_*$ and $\imin \rho$. Moreover when $X$ is locally weakly quasi-compact (a generalization of locally compact spaces which includes also Noetherian, spectral and locally spectral topological spaces) there is a right adjoint to the functor $\imin \rho$, denoted by $\rho_!$. We introduce the category of $\mathcal{T}$-flabby sheaves (known as $sa$-flabby in \cite{De91} and as quasi-injective in \cite{Pr1}): $F \in \mod(k_\mathcal{T})$ is $\mathcal{T}$-flabby if the restriction $\Gamma(U;F) \to \Gamma(V;F)$ is surjective for each $U,V \in \mathcal{T}$ with $U \supseteq V$. We prove that $\mathcal{T}$-flabby sheaves are acyclic with respect to the functor $\Gamma(U;\bullet)$, for $U \in \mathcal{T}$. More generally, if one introduces the category $\mathrm{Coh}(\mathcal{T}) \subset \mod(k_X)$ of coherent sheaves (i.e. sheaves admitting a finite resolution consisting of finite sums of $k_{U_i}$, $U_i \in \mathcal{T}$), then $\mathcal{T}$-flabby sheaves are acyclic with respect to $\mathrm{Hom}_{k_\mathcal{T}}(\rho_*G,\bullet)$, for $G \in \mathrm{Coh}(\mathcal{T})$. Coherent sheaves also give a description of sheaves on $X_\mathcal{T}$: for each $F \in \mod(k_\mathcal{T})$ there exists a filtrant inductive family $\{F_i\}_{i\in I}$ such that $F \simeq \lind i \rho_*F_i$. Then we prove that as in \cite{ejp} the category of sheaves on $X_\mathcal{T}$ is equivalent to the category of sheaves on a locally quasi-compact space $\widetilde{X}_\mathcal{T}$, the $\mathcal{T}$-spectrum of $X$, which generalizes the notion of o-minimal spectrum.\\
Our theory can then be specialized to each of the examples we mentioned above: when $\mathcal{T}$ is the category of semialgebraic open subsets of a locally semialgebraic space $X$ we obtain the constructions and results of \cite{De91}, when $\mathcal{T}$ is the category of relatively compact subanalytic open subsets of a real analytic manifold $X$ we obtain the constructions and results of \cite{KS01,Pr1}, when $\mathcal{T}$ is the category of definable open subsets of a locally definable space $X$ we obtain in the definable case the constructions of \cite{ejp}. Moreover, when $\mathcal{T}$ is the category of conic subanalytic open subsets of a real analytic manifold $X$ we obtain a suitable category of conic subanalytic sheaves.\\
The paper is organized in the following way. In Section \ref{Sec sheaves on lwqcs} we introduce the locally weakly quasi-compact spaces and study some properties of sheaves on such spaces. The results of this section will be used in two crucial ways on the theory of sheaves on $\mathcal{T}$-spaces, they are required to show that: (i) when a $\mathcal{T}$-space $X$ is locally weakly quasi-compact, then there is a right adjoint $\rho_!$ to the functor $\imin \rho$ induced by the natural functor of sites $\rho:X \to X_\mathcal{T}$; (ii) for a $\mathcal{T}$-space $X$, category of sheaves on $X_\mathcal{T}$ is equivalent to the category of sheaves on a locally quasi-compact space $\widetilde{X}_\mathcal{T}$, the $\mathcal{T}$-spectrum of $X$. In Section \ref{Sec tspaces} we introduce the $\mathcal{T}$-spaces and develop the theory of sheaves on such spaces as already described above. Finally on Section \ref{Sec examples} we apply our theory to the main examples.
\section{Sheaves on locally weakly quasi-compact spaces}\label{Sec sheaves on lwqcs}
Let $X$ be a non necessarily Hausdorff topological space. One denotes by $\mathrm{Op}(X)$ the category whose
objects are the open subsets of $X$ and the morphisms are the inclusions. In this section we generalize some classical results about sheaves on locally compact spaces. For classical sheaf theory our basic reference is \cite{KS90}. We refer to \cite{Ta94} for an introduction to sheaves on Grothendieck topologies.
\subsection{Locally weakly quasi-compact spaces}
\begin{df} An open subset $U$ of $X$ is said to be relatively weakly quasi-compact in $X$ if, for any covering $\{U_i\}_{i \in I}$ of $X$, there
exists $J \subset I$ finite, such that
$U \subset \bigcup_{i \in J} U_i.$\\
\end{df}
We will write for short $U \subset\subset X$ to say that $U$ is a relatively
weakly quasi-compact open set in $X$, and we will call $\mathrm{Op}^c(U)$ the subcategory
of $\mathrm{Op}(U)$ consisting of open sets $V \subset\subset U$. Note
that, given $V,W \in \mathrm{Op}^c(U)$, then $V \cup W \in \mathrm{Op}^c(U)$.
\begin{df}
A topological space $X$ is locally weakly quasi-compact if satisfies the following hypothesis for every $U,V\in \mathrm{Op}(X)$
\begin{itemize}
\item[LWC1.]
Every $x \in U$ has a fundamental neighborhood system $\{V_i\}$ with $V_i \in \mathrm{Op}^c(U)$.
\item[LWC2.]
For every $U' \in \mathrm{Op}^c(U)$ and $V' \in \mathrm{Op}^c(V)$ one has $U' \cap V' \in \mathrm{Op}^c(U \cap V)$.
\item[LWC3.]
For every $U' \in \mathrm{Op}^c(U)$ there exists $W \in \mathrm{Op}^c(U)$ such that $U' \subset\subset W$.
\end{itemize}
\end{df}
Of course an open subset $U$ of a locally weakly quasi-compact space $X$ is also a locally weakly quasi-compact space.
\begin{es}
{\em
Let us consider some examples of locally weakly quasi-compact
spaces:
\begin{enumerate}
\item
A locally compact topological space $X$ is a locally weakly quasi-compact.
In this case, for $U,V \in \mathrm{Op}(X)$ we have $V\subset \subset U$ if and only if $V$ is relatively compact subset of $U$.
\item
Let $X$ be a topological space with a basis of quasi-compact open subsets closed under taking finite intersections. Then $X$ is locally weakly quasi-compact and, for $U,V \in \mathrm{Op}(X)$ we have $V\subset \subset U$ if and only if $V$ is contained in a quasi-compact subset of $U$. In this situation we have the following particular cases:
\begin{itemize}
\item
$X$ is a Noetherian topological space (each open subset of $X$ is quasi-compact). This includes in particular: (a) algebraic varieties over algebraically closed fields; (b) compact complex varieties (reduced, irreducible complex analytic spaces) with the Zariski topology.
\item
$X$ is a spectral topological space (in addition: (i) $X$ is quasi-compact; (ii) $T_0$; (iii) every irreducible closed subset is the closure of a unique point). This includes in particular: (a) real algebraic varieties over real closed fields; (b) the o-minimal spectrum of a definable space in some o-minimal structure.
\item
$X$ is an increasing union of open spectral topological spaces $X_i$'s, i.e. $X$ is the space $\bigcup _{i\in I}X_i$. This space $X$ has a basis of quasi-compact open subsets closed under taking finite intersections and in addition is: (i) not quasi-compact in general unless $I$ is finite; (ii) $T_0$. This includes in particular: (a) the semialgebraic spectrum of locally semialgebraic space; (b) more generally, the o-minimal spectrum of a locally definable space in some o-minimal structure.
\end{itemize}
\item
Let $E$ be a real vector bundle over a locally compact space $Z$ endowed with the
natural action $\mu$ of $\mathbb{R}^{{\scriptscriptstyle{+}}}$ (the multiplication on the fibers). Let $\overset{\hspace{0.1cm}}{\dot{E}}=E\setminus Z$,
and for $U \in \mathrm{Op}(E)$ set $U_Z=U\cap Z$ and $\overset{\hspace{0.1cm}}{\dot{U}}=U \cap \overset{\hspace{0.1cm}}{\dot{E}}$. Let
$E_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$ denote the space $E$ endowed with the conic topology
i.e. open sets of $E_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$ are open sets of $E$ which are
$\mu$-invariant. With this topology $E_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$ is a locally
weakly quasi-compact space and, for $U,V \in \mathrm{Op}(E_{\mathbb{R}^{{\scriptscriptstyle{+}}}})$ we have $V
\subset\subset U$ if and only if $V_Z \subset \subset U_Z$ in $Z$
and $\overset{\hspace{0.1cm}}{\dot{V}} \subset\subset \overset{\hspace{0.1cm}}{\dot{U}}$ in $\overset{\hspace{0.1cm}}{\dot{E}}_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$ (the later is $\overset{\hspace{0.1cm}}{\dot{E}}$ with the induced conic topology).
\end{enumerate}
}
\end{es}
\subsection{Sheaves on locally weakly quasi-compact spaces}
Recall that $X$ is a non necessarily Hausdorff topological space.
\begin{df} Let $\mathcal{U}=\{U_i\}_{i \in I}$ and $\mathcal{U}'=\{U'_j\}_{j \in J}$
be two families of open subsets of $X$. A refinement $\mathcal{U}' \to \mathcal{U}$ consists
of a map $\varepsilon:J \to I$ of the index sets and a family of
inclusions $U'_j \subset U_{\varepsilon(j)}$.
\end{df}
One denotes by $\mathrm{Cov}(U)$ the category whose objects are the
coverings of $U \in \mathrm{Op}(X)$ and the morphisms are the refinement
maps, and by $\mathrm{Cov}^f(U)$ its full subcategory consisting of finite
coverings of $U$.\\
Given $V \in \mathrm{Op}(U)$ and $S \in \mathrm{Cov}(U)$, one sets $S \cap V=\{U
\cap V\}_{U \in S} \in \mathrm{Cov}(V)$.
\begin{df} The site $X^f$ on the topological space $X$ is the category $\mathrm{Op}(X)$ endowed with
the fol\-lowing topology: $S \subset \mathrm{Op}(U)$ is a covering of $U$
if and only if it has a refinement $S^f \in \mathrm{Cov}^f(U)$.
\end{df}
\begin{df} Let $U,V \in \mathrm{Op}(X)$ with $V \subset U$. Given $S=\{U_i\}_{i
\in I} \in \mathrm{Cov}(U)$ and $T=\{V_j\}_{j\in J} \in \mathrm{Cov}(V)$, we write
$T \subset\subset S$ if $T$ is a refinement of $S \cap V$, and
$V_j \subset U_i$ if and only if $V_j \subset\subset U_i$.
\end{df}
\begin{df} A presheaf of $k$-modules on $X$ is a functor
$\mathrm{Op}(X)^{op} \to \mod(k)$. A morphism of presheaves is a morphism
of such functors. One denotes by $\mathrm{Psh}(k_X)$ the category of
presheaves of $k$-modules on $X$.
\end{df}
Let $F \in \mathrm{Psh}(k_X)$, and let $S \in \mathrm{Cov}(U)$. One sets
$$F(S)=\ker \Big(\prod_{W \in S}F(W)
\rightrightarrows \prod_{W',W'' \in S}F(W' \cap W'')\Big).$$
\begin{df} A presheaf $F$ is separated (resp. is a
sheaf) if for any $U \in \mathrm{Op}(X)$ and for any $S \in \mathrm{Cov}(U)$ the
natural morphism $F(U) \to F(S)$ is a monomorphism (resp. an
isomorphism). One denotes by $\mod(k_X)$ the category of sheaves
of $k$-modules on $X$.
\end{df}
Let $F \in \mathrm{Psh}(k_X)$, one defines the presheaf $F^+$ by setting
$$F^+(U) = \lind {S \in \mathrm{Cov}(U)} F(S).$$
One can show that $F^+$ is a separated presheaf and if $F$ is a
separated presheaf, then $F^+$ is a sheaf. Let $F \in \mathrm{Psh}(k_X)$,
the sheaf $F^{++}$ is called the sheaf associated to the presheaf
$F$.\\
\begin{lem}\label{fac+} For $F \in \mathrm{Psh}(k_X)$, and let $U \in \mathrm{Op}(X)$. If $F$ is a sheaf on $X^f$, then for any $V \in
\mathrm{Op}^c(U)$ the morphism
\begin{equation}\label{factors+}
F^+(U) \to F^+(V)
\end{equation}
factors through $F(V)$.
\end{lem}
\dim\ \ Let $S \in \mathrm{Cov}(U)$, and set $S \cap V = \{W \cap V\}_{W
\in S}$. Since $V \in \mathrm{Op}^c(U)$, there is a finite refinement $T^f \in
\mathrm{Cov}^f(V)$ of $S \cap V$. Then the morphism (\ref{factors+}) is
defined by
\begin{eqnarray*}
F^+(U) & \simeq & \lind {S \in \mathrm{Cov}(U)} \hspace{-3mm} F(S)\\
& \to & \lind {S \in \mathrm{Cov}(U)} \hspace{-3mm}F(S \cap V )\\
& \to & \lind {T^f \in \mathrm{Cov}^f(V)}\hspace{-3mm} F(T^f)\\
& \to & \lind {T \in \mathrm{Cov}(V)}\hspace{-3mm}F(T)\\
& \simeq & F^+(V).
\end{eqnarray*}
The result follows because $F(T^f) \simeq F(V)$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{cor}\label{fac|} With the hypothesis of Lemma \ref{fac+}, we consider two co\-ve\-rings $S \in \mathrm{Cov}(U)$ and $T \in
\mathrm{Cov}(V)$. If $T \subset\subset S$, then the morphism
\begin{equation}\label{factorscov+}
F^+(S) \to F^+(T) \end{equation}
factors through $F(T)$. In particular, if $T$ is finite, then the
morphism \eqref{factorscov+} factors through $F(V)$.
\end{cor}
From now on we will assume the following hypothesis:
\begin{equation}\label{hyplqc}
\text{the topological space $X$ is locally weakly quasi-compact.}
\end{equation}
\begin{lem} Let $U \in \mathrm{Op}(X)$, and consider a subset
$V \subset\subset U$. Then for any $S^f \in \mathrm{Cov}^f(U)$ there
exists $T^f \in \mathrm{Cov}^f(V)$ with $T^f \subset\subset S^f$.
\end{lem}
\dim\ \ Let $S^f=\{U_i\}$. For each $x \in U$ and $U_i \ni x$,
consider a $V_{x,i} \in \mathrm{Op}^c(U_i)$ containing $x$. Set
$V_x=\bigcap_i V_{x,i}$, the family $\{V_x\}$ forms a covering of
$U$. Then there exists a finite subfamily $\{V_j\}$ containing
$V$. By construction $V_j \cap V \subset\subset U_i$ whenever $V_j
\subset U_i$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{lem}\label{fac++} Let $F \in \mathrm{Psh}(k_X)$, and let $U \in \mathrm{Op}(X)$. If $F$ is a sheaf on $X^f$, then for any $V \in
\mathrm{Op}^c(U)$ the morphism
\begin{equation}\label{factors++}
F^{++}(U) \to F^{++}(V)
\end{equation}
factors through $F(V)$.
\end{lem}
\dim\ \ Since $X$ is locally weakly quasi-compact, there exists $W \in
\mathrm{Op}^c(U)$ with $V \subset\subset W$. As in Lemma \ref{fac+} we
obtain a diagram
$$
\xymatrix{F^{++}(U) \ar[r] \ar@ <-2pt> [d] & F^{++}(W) \ar[r]
\ar[d]
& F^{++}(V) \\
\hspace{-15mm}\lind {S^f \in \mathrm{Cov}^f(W)}\hspace{-3mm}
F^+(S^f) \ar@ <7pt> [r] \ar[ur] & \hspace{-4mm}\lind {T^f \in
\mathrm{Cov}^f(V)}\hspace{-3mm} F^+(T^f). \ar[ur]. & }
$$
Since $X$ is locally weakly quasi-compact then for any $S^f \in
\mathrm{Cov}^f(W)$ there exists $T^f \in \mathrm{Cov}^f(V)$ with $T^f
\subset\subset S^f$. By Corollary \ref{fac|} the morphism
$$F^+(S^f) \to F^+(T^f)$$
factors through $F(T^f) \simeq F(V)$. Then the morphism
$$\lind {S^f \in \mathrm{Cov}^f(W)}\hspace{-3mm}F^+(S^f) \to \hspace{-3mm}\lind {T^f \in
\mathrm{Cov}^f(V)}\hspace{-3mm}F^+(T^f)$$ factors through $F(V)$ and the
result follows. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{cor}\label{faclimlim} Let $F \in \mathrm{Psh}(k_X)$. If $F$ is a sheaf on $X^f$, then:
\begin{itemize}
\item[(i)] for any $V \in \mathrm{Op}^c(X)$ one has the isomorphism $\lind {U \supset\supset V}F(U)
\stackrel{\sim}{\to} \lind {U \supset\supset V}F^{++}(U)$. \\
\item[(ii)] for any $U \in \mathrm{Op}(X)$ one has the isomorphism $\lpro {V \subset\subset U}F(V)
\stackrel{\sim}{\to} \lpro {V \subset\subset U}F^{++}(V)$.
\end{itemize}
\end{cor}
\dim\ \ (i) By Lemma \ref{fac++} for each $U \in \mathrm{Op}(X)$ with $U
\supset\supset V$ we have a commutative diagram
$$
\xymatrix{ F^{++}(U) \ar[r] \ar[rd] & F^{++}(V) \\
F(U) \ar[u] \ar[r] & F(V) \ar[u] .}
$$
This implies that the identity morphism of $\lind {U
\supset\supset V}F(U)$ factors through $\lind {U \supset\supset V}
F^{++}(U)$. On the other hand this also implies that the identity
morphism of $\lind {U \supset\supset V}F^{++}(U)$ factors through
$\lind {U \supset\supset V} F(U)$. Then $\lind {U \supset\supset
V}F(U) \stackrel{\sim}{\to} \lind {U \supset\supset V}F^{++}(U)$.
The proof of (ii) is similar. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{cor}\label{facKK} Let $X$ be a quasi-compact and locally weakly quasi-compact
space, and let $F \in \mathrm{Psh}(k_X)$. If $F$ is a sheaf on $X^f$, then
the natural morphism
\begin{equation}\label{factorsK}
F(X) \to F^{++}(X)
\end{equation}
is an isomorphism.
\end{cor}
\dim\ \ It follows immediately from Corollary \ref{faclimlim} (i)
with $V=X$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
Let $\{F_i\}_{i \in I}$ be a filtrant inductive system in
$\mod(k_X)$. One sets
\begin{equation*}
\begin{array}{l}
\text{$\indl i F_i$ = inductive limit in the category of presheaves,} \\
\text{$\lind i F_i$ = inductive limit in the category of sheaves.}
\end{array}
\end{equation*}
Recall that $\lind i F_i=(\indl i F_i)^{++}$.
\begin{prop}\label{fac} Let $\{F_i\}_{i \in \mathrm{I}}$ be a filtrant inductive system in $\mod(k_X)$ and let $U \in \mathrm{Op}(X)$.
Then for any $V \in \mathrm{Op}^c(U)$ the morphism
\begin{equation*}
\Gamma(U;\lind i F_i) \to \Gamma(V;\lind i F_i)
\end{equation*}
factors through $\lind i \Gamma(V;F_i)$.
\end{prop}
\dim\ \ By Lemma \ref{fac++} it is enough to show that $\indl i
F_i$ is a sheaf on $X^f$. Let $U \in \mathrm{Op}(X)$ and $S \in
\mathrm{Cov}^f(U)$. Since $\lind i$ commutes with finite projective limits
we obtain the isomorphism $(\indl i F_i)(S) \simeq \lind i
F_i(S)$. The result follows because $F_i \in \mod(k_X)$ for each
$i \in I$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{cor}\label{faclim} Let $\{F_i\}_{i \in \mathrm{I}}$ be a filtrant inductive system in $\mod(k_X)$.
\begin{itemize}
\item[(i)] For any $V \in \mathrm{Op}^c(X)$ one has the isomorphism $\lind {U
\supset\supset V,i}\Gamma(U;F_i) \stackrel{\sim}{\to} \lind {U \supset\supset
V}\Gamma( U; \lind i F_i)$. \\
\item[(ii)] For any $U \in \mathrm{Op}(X)$ one has the isomorphism $\lpro {V \subset\subset
U}\lind i \Gamma(V;F_i) \stackrel{\sim}{\to} \lpro {V \subset\subset
U}\Gamma(V;\lind i F_i)$.
\end{itemize}
\end{cor}
\dim\ \ It follows from Corollary \ref{faclimlim} with $F=\indl i
F_i$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{cor}\label{facK} Let $X$ be a quasi-compact and locally weakly quasi-compact
space. Then the natural morphism
$$ \lind i \Gamma(X;F_i) \to \Gamma(X;\lind i F_i)$$
is an isomorphism.
\end{cor}
\dim\ \ It follows from Corollary \ref{facKK} with $F=\indl i
F_i$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{es}\label{exiUVi} Let us consider the formula
\begin{equation}\label{iUVi}
\lind {U \supset\supset V,i}\Gamma(U;F_i) \stackrel{\sim}{\to} \lind {U
\supset\supset V}\Gamma( U; \lind i F_i)
\end{equation}
\begin{itemize}
\item[(i)] Let $X$ be a Noetherian space and let $V \in \mathrm{Op}(X)$. Then $\Gamma(V;F) \simeq \lind {U \supset\supset
V}\Gamma(U;F)$, since every open set is quasi-compact and
\eqref{iUVi} becomes $\lind i \Gamma(V;F_i) \simeq \Gamma(V;\lind
i F_i)$.
\item[(ii)] Assume that $X$ has a basis of quasi-compact open subsets and let $V \in \mathrm{Op}^c(X)$. Then $V$ is contained in a quasi-compact open subset of $X$ and $\lind {U \supset\supset V}\Gamma(U;F) \simeq \lind {W \supset V}\Gamma(W;F)$, where $W$ ranges through the family of quasi-compact subsets of $X$.
\item[(iii)] Let $X$ be a locally compact space and let $V \in \mathrm{Op}^c(X)$. Then $\Gamma(\overline{V};F) \simeq \lind {U \supset\supset
V}\Gamma(U;F)$, and \eqref{iUVi} becomes $\lind
i\Gamma(\overline{V};F_i) \simeq \Gamma(\overline{V};\lind
iF_i)$.
\item[(iv)] Let $E_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$ be a vector bundle endowed with the
conic topology, and let $V \in \mathrm{Op}^c(E_{\mathbb{R}^{{\scriptscriptstyle{+}}}})$. Then $\lind {U
\supset\supset V}\Gamma(U;F) \simeq \Gamma(K;F)$, where $K$ is the
union of the closures of $V_Z$ in $Z$ and $\overset{\hspace{0.1cm}}{\dot{V}}$ in $\overset{\hspace{0.1cm}}{\dot{E}}_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$,
and \eqref{iUVi} becomes $\lind i\Gamma(K;F_i) \simeq
\Gamma(K;\lind iF_i)$.
\end{itemize}
\end{es}
\begin{lem}\label{lproc} Let $F \in \mathrm{Psh}(k_X)$. Then we have the isomorphism
$$
\lpro {V\subset\subset X} \lind {V\subset\subset W} F(W) \stackrel{\sim}{\to} \lpro {V \subset\subset X} F(V).
$$
\end{lem}
\dim\ \ The result follows since for each $V \in \mathrm{Op}^c(X)$ there exists $W \in \mathrm{Op}^c(X)$ such that $V \subset\subset W$ since $X$ is locally weakly compact. Let $U,V \subset\subset X$ such that $U\supset\supset V$. The restriction morphism $F(U) \to F(V)$ factors through $\lind {W \supset\supset V}F(W)$. Taking the projective limit we obtain the result. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{oss} The notion of locally weakly quasi-compact can be extended to the case of a site, by generalizing the hypothesis LWC1-LWC3. For our purpose we are interested in the topological setting and we refer to \cite{Phd} for this approach.
\end{oss}
\subsection{c-soft sheaves on locally weakly quasi-compact spaces}
Let $X$ be a locally weakly quasi-compact space, and consider the
category $\mod(k_X)$.
\begin{df} We say that a sheaf $F$ on $X$ is c-soft if the restriction
morphism $\Gamma(W;F) \to \lind{U \supset\supset V}\Gamma(U;F)$ is
surjective for each $V,W \in \mathrm{Op}^c(X)$ with $V \subset\subset W$.
\end{df}
It follows from the definition that injective sheaves and flabby
sheaves are c-soft. Moreover, it follows from Corollary
\ref{faclim} that filtrant inductive limits of c-soft sheaves are
c-soft.
\begin{prop}\label{cs1} Let $\exs{F'}{F}{F''}$ be an exact sequence in
$\mod(k_X)$, and assume that $F'$ is c-soft. Then the sequence
$$\exs{\lind{U \supset\supset V}\Gamma(U;F')}{\lind{U \supset\supset V}\Gamma(U;F)}{\lind{U \supset\supset V}\Gamma(U;F'')}$$
is exact for any $V \in \mathrm{Op}^c(X)$.
\end{prop}
\dim\ \ Let $s'' \in \lind{U \supset\supset V}\Gamma(U;F'')$. Then
there exists $U \supset\supset V$ such that $s''$ is represented
by $s''_U \in \Gamma(U;F'')$. Let $\{U_i\}_{i \in I} \in \mathrm{Cov}(U)$
such that there exists $s_i \in \Gamma(U_i;F)$ whose image is
$s''_U|_{U_i}$ for each $i$. There exists $W \in \mathrm{Op}^c(U)$ with
$W\supset\supset V$, a finite covering $\{W_j\}_{j=1}^n$ of $W$
and a map $\varepsilon:J \to I$ of the index sets such that $W_j
\subset\subset U_{\varepsilon(j)}$. We may argue by induction on
$n$. If $n=2$, set $U_i=U_{\varepsilon(i)}$, $i=1,2$. Then
$(s_1-s_2)|_{U_1 \cap U_2}$ belongs to $\Gamma(U_1 \cap U_2;F')$,
and its restriction defines an element of $\lind {W' \supset\supset W_1 \cap
W_2}\Gamma(W';F')$, hence it extends to $s' \in \Gamma(U;F')$. By
replacing $s_1$ with $s_1-s'$ on $W_1$ we may assume that
$s_1=s_2$ on $W_1 \cap W_2$. Then there exists $s \in \Gamma(W_1
\cup W_2;F)$ with $s|_{W_i}=s_i$. Thus the induction proceeds.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop}\label{www} Let $\exs{F'}{F}{F''}$ be an exact sequence in
$\mod(k_{X})$, and assume $F',F$ c-soft. Then $F''$
is c-soft.
\end{prop}
\dim\ \ Let $V,W \in \mathrm{Op}^c(X)$ with $V\subset\subset W$ and let us
consider the diagram below
$$ \xymatrix{\Gamma(W;F) \ar[d]^\alpha \ar[r] & \Gamma(W;F'') \ar[d]^\gamma \\
\lind {U\supset\supset V}\Gamma(U;F) \ar[r]^\beta & \lind {U\supset\supset V}\Gamma(U;F'').} $$ The morphism
$\alpha$ is surjective since $F$ is c-soft and $\beta$ is
surjective by Proposition \ref{cs1}. Then $\gamma$ is
surjective. \nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\begin{prop} The family of c-soft sheaves is injective respect
to the functor $\lind{U \supset\supset V}\Gamma(U;\bullet)$ for each
$V \in \mathrm{Op}^c(X)$.
\end{prop}
\dim\ \ The family of c-soft sheaves contains injective sheaves, hence it is cogenerating. Then the result follows from Propositions \ref{cs1} and \ref{www}.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
Assume the following hypothesis
\begin{equation}\label{Xcountable}
\text{$X$ has a countable cover $\{U_n\}_{n\in\mathbb{N}}$ with $U_n\in\mathrm{Op}^c(X)$, $\forall n\in\mathbb{N}$.}
\end{equation}
\begin{lem}\label{X has a cov} Assume \eqref{Xcountable}. Then there exists a covering $\{V_n\}_{n \in \mathbb{N}}$ of $X$ such
that $V_n \subset \subset V_{n+1}$ and $V_n\in \mathrm{Op} ^c(X)$ for each $n \in \mathbb{N}$.
\end{lem}
\dim\ \ Let $\{U_n\}_{n\in\mathbb{N}}$ be a countable cover of $X$ with $U_n \in\mathrm{Op}^c(X)$ for each $n\in\mathbb{N}$. Set $V_1=U_1$. Given $\{V_i\}_{i=1}^n$ with $V_{i+1}\supset\supset V_i$, $i=1,\dots,n-1$, let us construct $V_{n+1} \supset\supset V_n$. Consider $x \notin V_n$. Up to take a permutation of $\mathbb{N}$ we may assume $x \in U_{n+1}$. Since $X$ is locally weakly quasi-compact there exists $V_{n+1} \in \mathrm{Op}^c(X)$ such that $V_n \cup U_{n+1} \subset\subset V_{n+1}$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop}\label{cs2} Assume \eqref{Xcountable}. Then the category of c-soft sheaves is injective respect
to the functor $\Gamma(X; \bullet)$.
\end{prop}
\dim\ \ Take an exact sequence $\exs{F'}{F}{F''}$ , and suppose
$F'$ c-soft. Consider a covering $\{V_n\}_{n \in \mathbb{N}}$ of $X$ such
that $V_n \subset \subset V_{n+1}$ (and $V_n\in \mathrm{Op} ^c(X)$) for each $n \in \mathbb{N}$. All the
sequences
$$\exs{\lind{U_n
\supset\supset V_n}\Gamma(U_n;F')}{\lind{U_n \supset\supset
V_n}\Gamma(U_n;F)}{\lind{U_n \supset\supset V_n}\Gamma(U_n;F'')}$$
are exact by Proposition \ref{cs1}, and the morphism
$\lind{U_{n+1} \supset\supset V_{n+1}}\Gamma(U_{n+1};F') \to
\lind{U_n \supset\supset V_n}\Gamma(U_n;F')$ is surjective for all
$n$. Then by Proposition 1.12.3 of \cite{KS90} the sequence
$$\exs{\lpro n\lind{U_n
\supset\supset V_n}\Gamma(U_n;F')}{\lpro n\lind{U_n \supset\supset
V_n}\Gamma(U_n;F)}{\lpro n\lind{U_n \supset\supset
V_n}\Gamma(U_n;F'')}$$ is exact. By Lemma \ref{lproc} $\lpro n\lind{U_n
\supset\supset V_n}\Gamma(U_n;G) \simeq \Gamma(X;G)$ for any $G
\in \mod(k_X)$ and the result follows. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{es} Let us consider some particular cases
\begin{itemize}
\item[(i)] When $X$ is Noetherian c-soft sheaves are flabby
sheaves.
\item[(ii)] When $X$ has a basis of quasi-compact open subsets $F \in \mod(k_X)$ is c-soft if the
restriction morphism $\Gamma(U;F) \to \Gamma(V;F)$ is
surjective, for any quasi-compact open subsets $U,V$ of $X$ with $U \supseteq V$.
\item[(iii)] When $X$ is a locally compact space countable at
infinity, then we recover c-soft sheaves as in chapter II of \cite{KS90}.
\item[(iv)] When $E_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$ is a vector bundle endowed with the
conic topology, then $F \in \mod(k_{E_{\mathbb{R}^{{\scriptscriptstyle{+}}}}})$ is c-soft if the
restriction morphism $\Gamma(E_{\mathbb{R}^{{\scriptscriptstyle{+}}}};F) \to \Gamma(K;F)$ is
surjective, where $K$ is defined as in Example \ref{exiUVi}.
\end{itemize}
\end{es}
\section{Sheaves on $\mathcal{T}$-spaces.}\label{Sec tspaces}
In the following we shall assume that $k$ is a field and $X$ is a topological space.
Below we give the definition of $\mathcal{T}$-space, adapting the construction of Kashiwara and Schapira \cite{KS01}. We study the category of sheaves on $X_\mathcal{T}$ generalizing results already known in the case of subanalytic sheaves. Then we prove that as in \cite{ejp} the category of sheaves on $X_\mathcal{T}$ is equivalent to the category of sheaves on a locally weakly-compact topological space $\widetilde{X}_\mathcal{T}$, the $\mathcal{T}$-spectrum, which generalizes the notion of o-minimal spectrum.
\subsection{$\mathcal{T}$-sheaves}
Let $X$ be a topological space and let us consider a family $\mathcal{T}$ of open subsets of $X$.
\begin{df} The topological space $X$ is a $\mathcal{T}$-space if the family $\mathcal{T}$ satisfies the hypotheses below
\begin{equation}\label{hytau}
\begin{cases}
\text{(i) $\mathcal{T}$ is a basis for the topology of $X$, and $\varnothing \in \mathcal{T}$}, \\
\text{(ii) $\mathcal{T}$ is closed under finite unions and intersections},\\
\text{(iii) every $U \in \mathcal{T}$ has finitely many $\mathcal{T}$-connected components,}\\
\end{cases}
\end{equation}
where we define:
\begin{itemize}
\item
a $\mathcal{T}$-subset is a finite Boolean combination of elements of $\mathcal{T}$;
\item
a closed (resp. open) $\mathcal{T}$-subset is a $\mathcal{T}$-subset which is closed (resp. open) in $X$;
\item
a $\mathcal{T}$-connected subset is a $\mathcal{T}$-subset which is not the disjoint union of two proper clopen $\mathcal{T}$- subsets.
\end{itemize}
\end{df}
Let $X$ be a $\mathcal{T}$-space. One can endow the category $\mathcal{T}$ with a Gro\-then\-dieck topology, called the $\mathcal{T}$-topology, in the following way: a
family $\{U_i\}_i$ in $\mathcal{T}$ is a covering of $U \in \mathcal{T}$ if it admits a finite subcover. We denote by $X_{\mathcal{T}}$ the associated site, write for short $k_{\mathcal{T}}$ instead of $k_{X_{\mathcal{T}}}$, and let $\rho: X \to X_{\mathcal{T}}$ be the natural morphism of sites.
We have functors
\begin{equation}\label{rho}
\xymatrix{\mod(k_X)
\ar@ <2pt> [r]^{\mspace{0mu}\rho_*} &
\mod(k_\mathcal{T}) \ar@ <2pt> [l]^{\mspace{0mu}\imin \rho}. }
\end{equation}
\begin{prop}\label{fufu} We have $\imin \rho \circ \rho_* \simeq \mathrm{id}$. In particular the functor $\rho_*$ is fully faithful.
\end{prop}
\dim\ \ Let $V \in \mathrm{Op}(X)$ and let $G \in \mod(k_\mathcal{T})$. Then $\imin \rho G = (\rho^\gets F)^{++}$, where $\rho^\gets G \in \mathrm{Psh}(k_X)$ is defined by
$$
\mathrm{Op}(X) \ni V \mapsto \lind {U \supseteq V, U \in \mathcal{T}}G(U).
$$
In particular, when $U \in \mathcal{T}$, $\rho^\gets G(U)=G(U)$.
Let $F \in \mod(k_X)$ and denote by $\iota:\mod(k_X)\to\mathrm{Psh}(k_X)$ the forgetful functor.
The adjunction morphism $\rho^\gets\circ\rho_*\to\mathrm{id}$ in $\mathrm{Psh}(k_X)$ defines $\rho^\gets\rho_*F \to \iota F$. This morphism is an isomorphism on $\mathcal{T}$, since $\rho^\gets\rho_*F(U)\simeq\rho_*F(U)\simeq F(U) \simeq \iota F(U)$ when $U \in \mathcal{T}$. By \eqref{hytau} (i) $\mathcal{T}$ forms a basis for the topology of $X$, hence we get an isomorphism
$$
\imin\rho\rho_*F \simeq (\rho^\gets\rho_*F)^{++} \simeq (\iota F)^{++} \simeq F
$$
and the result follows. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop}\label{UlimU} Let $\{F_i\}_{i \in I}$ be a filtrant inductive
system in $\mod(k_\mathcal{T})$ and let $U \in \mathcal{T}$. Then
$$\lind i\Gamma(U;F_i) \stackrel{\sim}{\to} \Gamma(U;\lind i F_i).$$
\end{prop}
\dim\ \ Denote by $\indl i F_i$ the
presheaf $V \mapsto \lind i \Gamma(V;F_i)$ on $X_\mathcal{T}$. Let $U
\in \mathcal{T}$ and let $S$ be a finite covering of $U$. Since
$\lind i$ commutes with finite projective limits we obtain the
isomorphism $(\indl i F_i)(S) \stackrel{\sim}{\to} \lind i F_i(S)$ and $F_i(U)
\stackrel{\sim}{\to} F_i(S)$ since $F_i \in \mod(k_\mathcal{T})$ for each $i$.
Moreover the family of finite coverings of $U$ is cofinal in
$\mathrm{Cov}(U)$. Hence $\indl i F_i \stackrel{\sim}{\to} (\indl i F_i)^+$. Applying once
again the functor $(\cdot)^+$ we get
$$\indl i F_i \simeq (\indl i F_i)^+ \simeq (\indl i F_i)^{++} \simeq \lind i F_i.$$
Hence applying the functor $\Gamma(U;\cdot)$ we obtain
the isomorphism $\lind i\Gamma(U;F_i) \stackrel{\sim}{\to} \Gamma(U;\lind
i F_i)$ for each $U \in \mathcal{T}.$ \nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\begin{prop}\label{isiuei} Let $F$ be a presheaf on $X_\mathcal{T}$ and
assume that
\begin{itemize}
\item[(i)] $F(\varnothing)=0,$
\item[(ii)] For any $U,V \in \mathcal{T}$ the sequence $\lexs{F(U\cup V)}{F(U) \oplus F(V)}{F(U \cap
V)}$is exact.
\end{itemize}
Then $F \in \mod(k_\mathcal{T})$.
\end{prop}
\dim\ \ Let $U \in \mathcal{T}$ and let $\{U_j\}_{j=1}^n$ be a
finite covering of $U$. Set for short $U_{ij}=U_i \cap U_j$. We
have to show the exactness of the sequence
$$\lexs{F(U)}{\oplus_{1\leq k\leq n}F(U_k)}{\oplus_{1\leq i<j \leq
n}F(U_{ij})},$$ where the second morphism sends $(s_k)_{1\leq
k\leq n}$ to $(t_{ij})_{1\leq i<j \leq n}$ by
$t_{ij}=s_i|_{U_{ij}}-s_j|_{U_{ij}}$. We shall argue by induction
on $n$. For $n=1$ the result is trivial, and $n=2$ is the
hypothesis. Suppose that the assertion is true for $j \leq n-1$
and set $U'=\bigcup_{1\leq k <n}U_k$. By the induction hypothesis
the following commutative diagram is exact
$$
\xymatrix{&& 0 \ar[d] & 0 \ar[d] \\
0 \ar[r] & F(U) \ar[r] & F(U') \oplus F(U_n) \ar[d] \ar[r]
& F(U'\cap U_n) \ar[d] \\
&& \bigoplus_{i<n}F(U_i) \oplus F(U_n) \ar[d] \ar[r] &
\bigoplus_{i<n}F(U_{in}) \\
&& \bigoplus_{i<j<n}F(U_{ij}).}
$$
Then the result follows. \nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\subsection{$\mathcal{T}$-coherent sheaves}
Let us consider the category $\mod(k_X)$ of sheaves of
$k_X$-modules on $X$, and denote by $\mathcal{K}$ the subcategory whose
objects are the sheaves $F=\oplus_{i \in I} k_{U_i}$ with $I$
finite and $U_i \in \mathcal{T}$ for each $i$.
\begin{df} Let $\mathcal{T}$ be a subfamily of $\mathrm{Op}(X)$ satisfying
$(\ref{hytau})$, and let $F \in \mod(k_X)$.
\begin{itemize}
\item[(i)] $F$ is $\mathcal{T}$-finite if there exists an epimorphism
$G \twoheadrightarrow F$ with $G \in \mathcal{K}$.
\item[(ii)] $F$ is $\mathcal{T}$-pseudo-coherent if for any morphism
$\psi:G \to F$ with $G \in \mathcal{K}$, $\ker \psi$ is $\mathcal{T}$-finite.
\item[(iii)] $F$ is $\mathcal{T}$-coherent if it is both $\mathcal{T}$-finite
and $\mathcal{T}$-pseudo-coherent.
\end{itemize}
\end{df}
Remark that (ii) is equivalent to the same condition with ``$G$ is
$\mathcal{T}$-finite" instead of ``$G \in \mathcal{K}$". One denotes by
$\mathrm{Coh}(\mathcal{T})$ the full subcategory of $\mod(k_X)$ consisting of
$\mathcal{T}$-coherent sheaves. It is easy (see \cite{KS}, Exercise 8.23) to prove that $\mathrm{Coh}(\mathcal{T})$ is additive and stable by kernels.
\begin{prop}\label{rhoUsa} Let $U \in \mathcal{T}$ and consider the constant
sheaf $k_{U_{X_\mathcal{T}}} \in \mod(k_\mathcal{T})$. We have
$k_{U_{X_\mathcal{T}}} \simeq \rho_*k_U$.
\end{prop}
\dim\ \ Let $F$ be the presheaf on $X_\mathcal{T}$ defined by $F(V)=k$ if
$V \subset U$, $F(V)=0$ otherwise. This is a separated presheaf
and $k_{U_{X_\mathcal{T}}}=F^{++}$. Moreover there is an injective arrow
$F(V) \hookrightarrow \rho_*k_U(V)$ for each $V \in \mathrm{Op}(X_\mathcal{T})$.
Hence $F^{++} \hookrightarrow \rho_*k_U$ since the functor
$(\cdot)^{++}$ is exact. Let $\mathcal{S}\subseteq \mathcal{T}$ be the sub-family of $\mathcal{T}$-connected elements. Then $\mathcal{S}$ forms a basis for
the Grothendieck topology of $X_\mathcal{T}$. For each $W \in \mathcal{S}$ we have $F(W) \simeq
\rho_*k_U(W) \simeq k$ if $W \subset U$ and $F(W)=0$ otherwise.
Then $F^{++} \simeq \rho_*k_U$. \nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\begin{prop} \label{rho exact} The restriction of $\rho_*$ to $\mathrm{Coh}(\mathcal{T})$ is exact.
\end{prop}
\dim\ \
Let us consider an epimorphism
${G}\twoheadrightarrow{F}$ in $\mathrm{Coh}(\mathcal{T})$, we have to
prove that $\psi:{\rho_*G}\to{\rho_*F}$ is an epimorphism. Let $U
\in \mathcal{T}$ and let $0 \neq s \in \Gamma(U;\rho_*F) \simeq
\mathrm{Hom}_{k_X}(k_U,F)$ (by adjunction). Set $G'=G \times_F k_U=\ker(G \oplus k_U \rightrightarrows F)$. Then $G' \in
\mathrm{Coh}(\mathcal{T})$ and moreover $G'\twoheadrightarrow k_U$. There
exists a finite $\{U_i\}_{i \in I} \subset \mathcal{T}$ of $\mathcal{T}$-connected elements such that
$\oplus_ik_{U_i}\twoheadrightarrow G'$. The composition $k_{U_i}
\to G' \to k_U$ is given by the multiplication by $a_i \in k$. Set
$I_0=\{k_{U_i};\; a_i \neq 0\}$, we may assume $a_i=1$. We get a
diagram
$$
\xymatrix{\oplus_{i \in I_0}k_{U_i} \ar@{->>} [dr] \ar[r] & G' \ar@{->>} [d] \ar[r] & G \ar@{->>} [d] \\
& k_U \ar[r]^s & F.}
$$
The composition $k_{U_i} \to G' \to G$ defines $t_i \in
\mathrm{Hom}_{k_X}(k_{U_i},G)\simeq \Gamma(U_i;\rho_*G)$. Hence for each $s \in
\Gamma(U;\rho_*F)$ there exists a finite covering $\{U_i\}$ of $U$
and $t_i \in \Gamma(U_i;\rho_*G)$ such that $\psi(t_i)=s|_{U_i}$.
This means that $\psi$ is surjective.
\nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\begin{nt} Since the functor $\rho_*$ is fully faithfull and exact on $\mathrm{Coh}(\mathcal{T})$, we will often identify $\mathrm{Coh}(\mathcal{T})$ with its image in $\mod(k_\mathcal{T})$ and write $F$ instead of $\rho_*F$ for $F \in \mathrm{Coh}(\mathcal{T})$.
\end{nt}
\begin{teo}\label{teo coh stable}
The following hold:
\begin{itemize}
\item[(i)]
The category $\mathrm{Coh}(\mathcal{T})$ is stable by finite sums, kernels, cokernels and extensions in $\mod(k_\mathcal{T})$.
\item[(ii)]
The category $\mathrm{Coh}(\mathcal{T})$ is stable by $\bullet \otimes_{k_\mathcal{T}} \bullet $ in $\mod(k_\mathcal{T})$.
\end{itemize}
\end{teo}
\dim\ \
(i) The result follows from a general result of homological algebra of \cite{KS96}, Appendix A.1. With the notations of \cite{KS96} let ${\bf P}$ be the set of finite families of elements of $\mathcal{T}$, for ${\cal U}=\{U_i\}_{i\in I} \in {\bf P}$ set
$$
L({\cal U})=\oplus_ik_{U_i},
$$
for ${\cal V}=\{V_j\}_{j\in J} \in {\bf P}$ set
$$
\mathrm{Hom}_{{\bf P}}({\cal U},{\cal V})=\mathrm{Hom}_{k_\mathcal{T}}(L({\cal U}),L({\cal V}))=\oplus_i\oplus_j\mathrm{Hom}_{k_\mathcal{T}}(k_{U_i},k_{V_j})
$$
and for $F \in \mod(k_\mathcal{T})$ set
$$
H({\cal U},F)=\mathrm{Hom}_{k_\mathcal{T}}(L({\cal U}),F)=\oplus_i\mathrm{Hom}_{k_\mathcal{T}}(k_{U_i},F).
$$
By Proposition A.1 of \cite{KS96} in order to prove (i) it is enough to prove the properties (A.1)-(A.4) below:
\begin{itemize}
\item[(A.1)] For any ${\cal U}=\{U_i\} \in {\bf P}$ the functor $H({\cal U},\bullet)$ is left exact in $\mod(k_\mathcal{T})$.
\item[(A.2)] For any morphism $g:{\cal V} \to {\cal W}$ in ${\bf P}$, there exists a morphism $f:{\cal U} \to {\cal V}$ in ${\bf P}$ such that ${\cal U} \stackrel{f}{\to} {\cal V} \stackrel{g}{\to} {\cal W}$ is exact.
\item[(A.3)] For any epimorphism $f:F \to G$ in $\mod(k_\mathcal{T})$, ${\cal U} \in {\bf P}$ and $\psi \in H({\cal U},G)$, there exists ${\cal V} \in {\bf P}$ and an epimorphism $g \in \mathrm{Hom}_{\bf P}({\cal V},{\cal U})$ and $\varphi \in H({\cal V},F)$ such that $\psi \circ g = f \circ \varphi$.
\item[(A.4)] For any ${\cal U},{\cal V} \in {\bf P}$ and $\psi \in H({\cal U},L({\cal V}))$ there exists ${\cal W} \in {\bf P}$ and an epimorphism $f \in \mathrm{Hom}_{\bf P}({\cal W},{\cal U})$ and a morphism $g \in \mathrm{Hom}_{\bf P}({\cal W},{\cal U})$ such that $L(g)=\psi \circ f$ in $\mathrm{Hom}_{k_\mathcal{T}}(L({\cal W}),L({\cal V}))$.
\end{itemize}
It is easy to check that the axioms (A.1)-(A.4) are satisfied.
(ii) Let $F \in \mathrm{Coh}(\mathcal{T})$. Then $F$ has a resolution
$$
\oplus_{j \in J}k_{U_j} \to \oplus_{i\in I}k_{U_i} \to F \to 0
$$
with $I$ and $J$ finite. Let $V \in \mathcal{T}$. The sequence
$$
\oplus_{j \in J}k_{V \cap U_j} \to \oplus_{i \in T}k_{V \cap U_i} \to F_V \to 0
$$
is exact. Then it follows from (i) that $F_V$ is coherent. Let $G \in \mathrm{Coh}(\mathcal{T})$. The sequence
$$
\oplus_{j \in J}G_{U_j} \to \oplus_{i \in I}G_{U_i} \to G \otimes_{k_\mathcal{T}} F \to 0
$$
is exact. The sheaves $G_{U_i}$ and $G_{U_j}$ are coherent for each $i \in I$ and each $j \in J$. Hence it follows by (i) that $G \otimes_{k_\mathcal{T}} F$ is coherent as required.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{cor}\label{cor coh stable}
The following hold:
\begin{itemize}
\item[(i)]
The category $\mathrm{Coh}(\mathcal{T})$ is stable by finite sums, kernels, cokernels in $\mod(k_X)$.
\item[(ii)]
The category $\mathrm{Coh}(\mathcal{T})$ is stable by $\bullet \otimes_{k_X} \bullet $ in $\mod(k_X)$.
\end{itemize}
\end{cor}
\dim\ \ (i) The stability under finite sums and kernels is easy, see \cite{KS}, Exercise 8.23. Let $F,G \in \mathrm{Coh}(\mathcal{T})$ and let $\varphi:F \to G$ be a morphism in $\mod(k_X)$. Then $\rho_*(\varphi)$ is a morphism in $\mod(k_\mathcal{T})$ and
$\mathrm{coker}(\rho_*\varphi) \in \mathrm{Coh}(\mathcal{T})$ by Theorem \ref{teo coh stable}. We have
$\mathrm{coker}(\rho_*\varphi) \simeq \rho_*\mathrm{coker}\varphi$ since $\rho_*$ is exact on $\mathrm{Coh}(\mathcal{T})$ by Proposition \ref{rho exact}. Composing with $\imin \rho$ and applying Proposition \ref{fufu} we obtain
$\mathrm{coker}\varphi \in \mathrm{Coh}(\mathcal{T})$.
(ii) The proof of the stability by $\bullet\otimes_{k_X}\bullet$ is similar to that of Theorem \ref{teo coh stable}.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{teo}\label{eqlambda} (i) Let $G \in \mathrm{Coh}(\mathcal{T})$ and let
$\{F_i\}$ be a filtrant inductive system in $\mod(k_\mathcal{T})$.
Then we have an isomorphism
$$\lind i \mathrm{Hom}_{k_\mathcal{T}}(\rho_*G,F_i) \stackrel{\sim}{\to}
\mathrm{Hom}_{k_\mathcal{T}}(\rho_*G,\lind i F_i).$$
(ii) Let $F \in \mod(k_\mathcal{T})$. There exists a small filtrant
inductive system $\{F_i\}_{i \in I}$ in $\mathrm{Coh}(\mathcal{T})$ such
that $F \simeq \lind i \rho_*F_i$.
\end{teo}
\dim\ \ (i) There exists an exact sequence $\rexs{G_1}{G_0}{G}$
with $G_1,G_0$ finite direct sums of constant sheaves $k_U$ with
$U \in \mathcal{T}$. Since $\rho_*$ is exact on
$\mathrm{Coh}(\mathcal{T})$ and commutes with finite sums, by Proposition
\ref{rhoUsa} we are reduced to prove the isomorphism $\lind i
\Gamma(U;F_i) \stackrel{\sim}{\to} \Gamma(U;\lind i F_i)$. Then the result follows
from Proposition \ref{UlimU}.
(ii) Let $F \in \mod(k_\mathcal{T})$, and define
\begin{eqnarray*}
I_0 & := & \{(U,s);\; U \in \mathcal{T},\; s \in \Gamma(U;F)\} \\
G_0 & := & \oplus_{(U,s) \in I_0}\rho_*k_U
\end{eqnarray*}
The morphism $\rho_*k_U \to F$, where the section $1 \in
\Gamma(U;k_U)$ is sent to $s \in \Gamma(U;F)$ defines un
epimorphism $\varphi:G_0 \to F$. Replacing $F$ by $\ker\varphi$ we
construct a sheaf $G_1=\oplus_{(V,t) \in I_1}\rho_*k_V$ and an
epimorphism $G_1\twoheadrightarrow \ker\varphi$. Hence we get an
exact sequence $\rexs{G_1}{G_0}{F}$. For $J_0\subset I_0$ set for
short $G_{J_0}=\oplus_{(U,s)\in J_0}\rho_*k_U$ and define
similarly $G_{J_1}$. Set
$$
J=\{(J_1,J_0);\; J_k\subset I_k,\; J_k \text{ is finite and } {\rm
im} \varphi|_{G_{J_1}} \subset G_{J_0}\}.
$$
The category $J$ is filtrant and $F \simeq \lind {(J_1,J_0)\in J}
\mathrm{coker}(G_{J_1}\to G_{J_0})$. \nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\begin{cor} Let $G \in \mathrm{Coh}(\mathcal{T})$ and let
$\{F_i\}$ be a filtrant inductive system in $\mod(k_\mathcal{T})$.
Then we have an isomorphism
$$\lind i \mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F_i) \stackrel{\sim}{\to}
\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,\lind i F_i).$$
\end{cor}
\dim\ \ Let $U \in \mathcal{T}$. We have the chain of isomorphisms
\begin{eqnarray*}
\Gamma(U;\lind i \mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F_i)) & \simeq & \lind i \Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F_i)) \\
& \simeq & \lind i \mathrm{Hom}_{k_\mathcal{T}}(G_U,F_i)) \\
& \simeq & \mathrm{Hom}_{k_\mathcal{T}}(G_U,\lind iF_i)) \\
& \simeq & \Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,\lind i F_i)),
\end{eqnarray*}
where the first and the third isomorphism follow from Theorem \ref{eqlambda} (i). the fact that $G_U \in \mathrm{Coh}(\mathcal{T})$ follows from Theorem \ref{teo coh stable} (ii). \nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\subsection{$\mathcal{T}$-sheaves on locally weakly quasi-compact spaces}
Assume that $X$ is a locally weakly quasi-compact space.
\begin{lem}\label{UcssVTc} For each $U \in \mathrm{Op}^c(X)$ there exists $V \in \mathcal{T}$ such that $U \subset\subset V \subset\subset X$.
\end{lem}
\dim\ \ Since $X$ is locally weakly quasi-compact we may find $W \in \mathrm{Op}^c(X)$ such that $U\subset\subset W$. By \eqref{hytau} (i) we may find a covering $\{W_i\}_{i\in I}$ of $X$ with $W_i\in\mathcal{T}$ and $W_i\subset\subset X$ for each $i\in I$. Then there exists a finite family $\{W_j\}_{j=1}^\ell$ whose union $V=\bigcup_{j=1}^\ell W_j$ contains $W$. Then $V \in \mathcal{T}$ and $U\subset\subset V\subset\subset X$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
When $X$ is locally weakly quasi-compact we can construct a left adjoint to the functor $\imin \rho$.
\begin{prop}\label{imineta} Let $F \in \mod(k_\mathcal{T})$, and let $U \in \mathrm{Op}(X)$. Then
$$\Gamma(U;\imin \rho F) \simeq \lpro {V \subset\subset U, V \in
\mathcal{T}} \Gamma(V;F)$$
\end{prop}
\dim\ \ By Theorem \ref{eqlambda} we may assume $F=\lind i
\rho_*F_i$, with $F_i \in \mathrm{Coh}(\mathcal{T})$. Then $\imin \rho F
\simeq \lind i \imin \rho \rho_* F_i \simeq \lind i F_i$. We
have the chain of isomorphisms
$$
\begin{array}{ccccc}
\Gamma(U;\imin \rho F)
& \hspace{-0.2cm}\simeq & \hspace{-1.3cm}\lpro{V\subset\subset U, V \in \mathcal{T}
}\lind{V\subset\subset W} \hspace{-0.2cm}\Gamma(W;\imin \rho F)
& \hspace{-0.2cm}\simeq & \hspace{-0.7cm}\lpro{V\subset\subset U, V \in \mathcal{T}
}\lind {V \subset\subset W} \hspace{-0.2cm}\Gamma(W; \lind i \imin \rho\rho_*F_i)\\
& \hspace{-0.2cm}\simeq & \hspace{-0.8cm}\lpro{V\subset\subset U, V \in
\mathcal{T}}\lind{V\subset\subset W,i}\hspace{-0.25cm}\Gamma(W;\imin \rho
\rho_*F_i)
& \hspace{-0.2cm}\simeq & \hspace{-0.6cm}\lpro{V\subset\subset U, V \in \mathcal{T}}\lind i \hspace{1mm}\Gamma(V;\imin \rho \rho_*F_i)\\
& \hspace{-0.2cm}\simeq & \hspace{-0.6cm}\lpro{V\subset\subset U, V \in \mathcal{T}}\lind i \hspace{1mm}\Gamma(V;\rho_*F_i)
& \hspace{-0.2cm}\simeq & \hspace{-1.8cm}\lpro{V\subset\subset U, V \in \mathcal{T}}\Gamma(V;F),
\end{array}
$$
where the first and the fourth isomorphisms follow from Lemma \ref{lproc}, the third isomorphism is a consequence of Corollary
\ref{faclim}, and the last isomorphism follows from Proposition \ref{UlimU}. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop}\label{eta!} The functor $\imin \rho$ admits a left adjoint,
denoted by $\rho_!$. It satisfies
\begin{itemize}
\item[(i)] for $F \in \mod(k_X)$ and $U \in \mathcal{T}$, $\rho_! F$ is the sheaf associated to the presheaf $U
\mapsto \lind{U\subset\subset V}
\hspace{-0.2cm}\Gamma(V;F)$, \\
\item[(ii)] For $U \in \mathrm{Op}(X)$ one has $\rho_!k_U \simeq \lind{V\subset\subset U, V \in \mathcal{T}
}k_V$.
\end{itemize}
\end{prop}
\dim\ \ Let $\widetilde{F} \in \mathrm{Psh}(k_{\mathcal{T}})$ be the presheaf $U
\mapsto \lind {U \subset\subset V} \Gamma(V;F)$, and let $G \in
\mod(k_\mathcal{T})$. We will construct morphisms
$$\xymatrix{\mathrm{Hom}_{\mathrm{Psh}(k_\mathcal{T})}(\widetilde{F},G) \ar@ <2pt>
[r]^{\xi} &
\mathrm{Hom}_{k_X}(F,\imin \rho G) \ar@ <2pt> [l]^{\vartheta}}.$$
To define $\xi$, let $\varphi:\widetilde{F} \to G$ and $U \in
\mathrm{Op}(X)$. Then the morphism $\xi(\varphi)(U):F(U) \to \imin \rho
G(U)$ is defined as follows
$$F(U) \simeq \lpro{V\subset\subset U, V \in \mathcal{T}
}\lind{V\subset\subset W} F(W) \stackrel{\varphi}{\longrightarrow}
\lpro{V\subset\subset U, V \in \mathcal{T} } G(V) \simeq \imin \rho
G(U).$$
On the other hand, let $\psi:F \to \imin \rho G$ and $U \in \mathcal{T}$.
Then the morphism $\vartheta(\psi)(U):\widetilde{F}(U) \to G(U)$
is defined as follows
$$\widetilde{F}(U) \simeq \lind {U \subset\subset V \in \mathcal{T}} F(V)
\stackrel{\psi}{\longrightarrow} \lind {U \subset\subset V \in
\mathcal{T}} \imin \rho G(V) \to G(U).$$
By construction one can check that the morphism $\xi$ and
$\vartheta$ are inverse to each others. Then (i) follows from the
chain of isomorphisms
$$\mathrm{Hom}_{\mathrm{Psh}(k_\mathcal{T})}(\widetilde{F},G) \simeq
\mathrm{Hom}_{k_{\mathcal{T}}}(\widetilde{F}{}^{++},G) \simeq
\mathrm{Hom}_{k_\mathcal{T}}(\widetilde{F}{}^{++},G).$$
To show (ii), consider the following sequence of isomorphisms
\begin{eqnarray*}
\mathrm{Hom}_{k_\mathcal{T}}(\rho_!k_U,F) & \simeq & \mathrm{Hom}_{k_X}(k_U,\imin \rho F)\\
& \simeq & \lpro {V \subset\subset U, V \in \mathcal{T}}
\mathrm{Hom}_{k_\mathcal{T}}(k_V,F)\\
& \simeq & \mathrm{Hom}_{k_\mathcal{T}}(\lind {V \subset\subset U, V \in
\mathcal{T}}k_V,F),
\end{eqnarray*}
where the second isomorphism follows from Proposition
\ref{imineta}. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop} The functor $\rho_!$ is exact and commutes with
$\Lind$ and $\otimes$.
\end{prop}
\dim\ \ It follows by adjunction that $\rho_!$ is right exact and
commutes with $\Lind$, so let us show that it is also left exact.
With the notations of Proposition \ref{eta!}, let $F \in
\mod(k_X)$, and let $\widetilde{F} \in \mathrm{Psh}(k_{\mathcal{T}})$ be the
presheaf $U \mapsto \lind {U \subset\subset V} \Gamma(V;F)$. Then
$\rho_!F \simeq \widetilde{F}{}^{++}$, and the functors $F \mapsto
\widetilde{F}$ and $G \mapsto G^{++}$ are left exact.\\
Let us show that $\rho_!$ commutes with $\otimes$. Let $F,G \in
\mod(k_X)$, the morphism
$$\lind {U \subset\subset V}F(V) \otimes_k \hspace{-0.1cm} \lind {U
\subset\subset V} G(V) \to \lind {U \subset\subset V} (F(V)
\otimes_k G(V))$$ defines a morphism in $\mod(k_\mathcal{T})$
$$\rho_!F \otimes_{k_\mathcal{T}} \rho_!G \to \rho_!(F \otimes_{k_X} G)$$
by Proposition \ref{eta!} (i). Since $\rho_!$ commutes with
$\Lind$ we may suppose that $F=k_U$ and $G=k_V$ and the result
follows from Proposition \ref{eta!} (ii).
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop} The functor $\rho_!$ is fully faithful. In particular
one has $\imin \rho \circ \rho_! \simeq \mathrm{id}$. Moreover, for $F \in
\mod(k_X)$ and $G \in \mod(k_\mathcal{T})$ one has
$$\imin \rho \mathcal{H}\mathit{om}_{k_\mathcal{T}}(\rho_!F,G) \simeq \mathcal{H}\mathit{om}_{k_X}(F,\imin \rho G).$$
\end{prop}
\dim\ \ For $F,G \in \mod(k_X)$ by adjunction we have
$$\mathrm{Hom}_{k_X}(\imin \rho \rho_!F,G) \simeq \mathrm{Hom}_{k_X}(F,\imin \rho \rho_*G)
\simeq \mathrm{Hom}_{k_X}(F,G).$$ This also implies that $\rho_!$ is fully
faithful, in fact
$$\mathrm{Hom}_{k_\mathcal{T}}(\rho_!F,\rho_!G) \simeq \mathrm{Hom}_{k_X}(F,\imin \rho \rho_!G) \simeq
\mathrm{Hom}_{k_X}(F,G).$$ Now let $K,F \in \mod(k_X)$ and $G \in \mod(k_\mathcal{T})$, we
have
\begin{eqnarray*}
\mathrm{Hom}_{k_X}(K,\imin \rho \mathcal{H}\mathit{om}_{k_\mathcal{T}}(\rho_! F,G)) & \simeq & \mathrm{Hom}_{k_\mathcal{T}}(\rho_! K,\mathcal{H}\mathit{om}_{k_\mathcal{T}}(\rho_! F,G))\\
& \simeq & \mathrm{Hom}_{k_\mathcal{T}}(\rho_! K \otimes_{k_\mathcal{T}} \rho_! F,G)\\
& \simeq & \mathrm{Hom}_{k_\mathcal{T}}(\rho_! (K \otimes_{k_X} F),G)\\
& \simeq & \mathrm{Hom}_{k_X}(K \otimes_{k_X} F,\imin \rho G)\\
& \simeq & \mathrm{Hom}_{k_X}(K,\mathcal{H}\mathit{om}_{k_X}(F,\imin \rho G)).
\end{eqnarray*}
\nopagebreak \begin{flushright}$\Box$\end{flushright}
Finally let us consider sheaves of rings in $\mod(k_\mathcal{T})$. If $\mathcal{A}$
is a sheaf of rings in $\mod(k_X)$, then $\rho_*\mathcal{A}$ and $\rho_!\mathcal{A}$
are sheaves of rings in $\mod(k_\mathcal{T})$.
Let $\mathcal{A}$ be a sheaf of unitary $k$-algebras on $X$, and let
$\widetilde{\mathcal{A}}\in\mathrm{Psh}(k_\mathcal{T})$ be the presheaf defined by the correspondence $\mathcal{T}
\ni U \mapsto \lind {U \subset\subset V} \Gamma(V;\mathcal{A})$. Let $F
\in \mathrm{Psh}(k_\mathcal{T})$, and assume that, for $V \subset U$, with $U,V \in
\mathcal{T}$, the following diagram is commutative:
$$\xymatrix{\Gamma(U;\widetilde{\mathcal{A}}) \otimes_k \Gamma(U;F)
\ar[d] \ar[r] & \Gamma(U;F) \ar[d] \\
\Gamma(V;\widetilde{\mathcal{A}}) \otimes_k \Gamma(V;F) \ar[r] &
\Gamma(V;F).}$$ In this case one says that $F$ is a presheaf of
$\widetilde{\mathcal{A}}$-modules on $\mathcal{T}$.
\begin{prop}\label{eta!mod} Let $\mathcal{A}$ be a sheaf of $k$-algebras on $X$, and let
$F$ be a presheaf of $\widetilde{\mathcal{A}}$-modules on $X_\mathcal{T}$. Then
$F^{++} \in \mod(\rho_!\mathcal{A})$.
\end{prop}
\dim\ \ Let $U \in \mathcal{T}$, and let $r \in
\lind {U\subset\subset V} \Gamma(V;\mathcal{A})$. Then $r$ defines a morphism
$\lind{U\subset\subset V} \Gamma(V;\mathcal{A}) \otimes_k \Gamma(W;F) \to \Gamma(W;F)$ for each $W \subseteq U$, $W \in \mathcal{T}$, hence an endomorphism of
$(F^{++})|_{U_{X_\mathcal{T}}}\simeq (F|_{U_{X_\mathcal{T}}})^{++}$. This
morphism defines a morphism of presheaves $\widetilde{\mathcal{A}} \to
{\cal E}nd(F^{++})$ and $\widetilde{\mathcal{A}}^{++}\simeq\rho_!\mathcal{A}$ by
Proposition \ref{eta!}. Then $F^{++} \in \mod(\rho_!\mathcal{A})$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\subsection{$\mathcal{T}$-flabby sheaves}\label{quinj}
\begin{df}
We say that an object $F \in \mod(k_\mathcal{T})$ is $\mathcal{T}$-flabby if for each $U,V \in \mathcal{T}$ with $V \supseteq U$
the restriction morphism $\Gamma(V;F) \to \Gamma(U;F)$ is
surjective.
\end{df}
It follows from the definition that injective sheaves are $\mathcal{T}$-flabby.
\begin{prop}
The following hold:
\begin{itemize}
\item[(i)]
Let $F_i$ be a filtrant inductive system of $\mathcal{T}$-flabby sheaves. Then $\lind i F_i$ is $\mathcal{T}$-flabby.
\item[(ii)] Products of $\mathcal{T}$-flabby objects are $\mathcal{T}$-flabby.
\end{itemize}
\end{prop}
\dim\ \
We will only prove (i) since the proof of (ii) is similar since taking products is exact and commutes with taking sections. Let $U \in \mathcal{T}$. Then for each $i$ the restriction morphism $\Gamma(V;F_i) \to \Gamma(U;F_i)$ is surjective. Applying the exact $\lind i$ and using Proposition \ref{UlimU}, the morphism
$$
\Gamma(V;\lind i F_i) \simeq \lind i \Gamma(V;F_i) \to \lind i \Gamma(U;F_i) \simeq \Gamma(U;\lind i F_i)
$$
is surjective.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop} \label{qinjU}
The full additive subcategory of $\mod (k_\mathcal{T})$ of $\mathcal{T}$-flabby object is $\Gamma (U;\bullet )$-injective for every $U \in \mathcal{T}$, i.e.:
\begin{itemize}
\item[(i)]
For every $F\in \mod (k_\mathcal{T})$ there exists a $\mathcal{T}$-flabby object $F'\in \mod (k_\mathcal{T})$ and an exact sequence $0\rightarrow F\rightarrow F'.$
\item[(ii)]
Let $\exs{F'}{F}{F''}$ be an exact sequence in
$\mod(k_\mathcal{T})$ and assume that $F'$ is $\mathcal{T}$-flabby.
Then the sequence
$$\exs{\Gamma(U;F')}{\Gamma(U;F)}{\Gamma(U;F'')}$$
is exact.
\item[(iii)]
Let $F',F,F'' \in \mod(k_\mathcal{T})$, and consider the
exact sequence
$$
\exs{F'}{F}{F''}.
$$
Suppose that $F'$ is $\mathcal{T}$-flabby. Then $F$ is $\mathcal{T}$-flabby if and only if $F''$ is $\mathcal{T}$-flabby.
\end{itemize}
\end{prop}
\dim\ \
(i) It follows from the definition that injective sheaves are $\mathcal{T}$-flabby. So (i) holds since it is true for injective sheaves. Indeed, as a Grothendieck category, $\mod(k_\mathcal{T})$ admits enough injectives.
(ii) Let $s'' \in \Gamma(U;F'')$, and let $\{V_i\}_{i=1}^n\in \mathrm{Cov} (U)$ be such that there exists $s_i \in
\Gamma(V_i;F)$ whose image is $s''|_{V_i}$. For $n \geq 2$ on $V_1
\cap V_2$ $s_1-s_2$ defines a section of $\Gamma(V_1 \cap V_2;F')$
which extends to $s' \in \Gamma(U;F')$ since $F'$ is $\mathcal{T}$-flabby. Replace $s_1$ with
$s_1-s'$ (identifying $s'$ with it's image in $F$). We may suppose that $s_1=s_2$ on $V_1 \cap V_2$. Then
there exists $t \in \Gamma(V_1 \cup V_2,F)$ such that
$t|_{V_i}=s_i$, $i=1,2$. Thus the induction proceeds.
(iii) Let $U,V\in \mathcal{T}$ with $V \supseteq U$ and let us
consider the diagram below
$$ \xymatrix{0 \ar[r] & \Gamma(V;F') \ar[d]^\alpha \ar[r] & \Gamma(V;F) \ar[d]^\beta \ar[r] & \Gamma(V;F'') \ar[d]^\gamma \ar[r] & 0\\
0 \ar[r] & \Gamma(U;F') \ar[r] & \Gamma(U;F) \ar[r] & \Gamma(U;F'') \ar[r] & 0} $$
where the row are exact by (ii)
and the morphism
$\alpha$ is surjective since $F'$ is $\mathcal{T}$-flabby. It follows from the five lemma that $\beta$ is surjective if and only if $\gamma$ is surjective.
\nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\begin{teo}\label{teo flabby hom and Hom}
Let $F \in \mod(k_\mathcal{T})$. Then the following hold:
\begin{itemize}
\item[(i)]
$F$ is $\mathcal{T}$-flabby if and only if the functor $\mathrm{Hom}_{k_\mathcal{T}}(\bullet ,F)$ is exact on $\mathrm{Coh}(\mathcal{T})$.
\item[(ii)]
If $F$ is $\mathcal{T}$-flabby then the functor $\mathcal{H}\mathit{om} _{k_\mathcal{T}}(\bullet ,F)$ is exact on $\mathrm{Coh}(\mathcal{T})$.
\end{itemize}
\end{teo}
\dim\ \
(i) is a consequence of a general result of homological algebra (see Theorem 8.7.2 of \cite{KS}). For (ii), let $F \in \mod(k_\mathcal{T})$ be $\mathcal{T}$-flabby. There is
an isomorphism of functors
$$\Gamma(U;\mathcal{H}\mathit{om} _{k_\mathcal{T}}(\bullet ,F)) \simeq \mathrm{Hom}_{k_\mathcal{T}}((\bullet )_U,F)$$
for each $U \in \mathcal{T}$. By Theorem \ref{teo coh stable} and (i) the
functor $\mathrm{Hom}_{k_\mathcal{T}}((\bullet )_U,F)$
is exact on $\mathrm{Coh}(\mathcal{T})$ and so the functor $\mathcal{H}\mathit{om} _{k_\mathcal{T}}(\bullet ,F)$ is also exact on $\mathrm{Coh}(\mathcal{T})$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{teo}\label{rcinj}
Let $G \in \mathrm{Coh}(\mathcal{T})$. Then the following hold:
\begin{itemize}
\item[(i)]
The family of $\mathcal{T}$-flabby sheaves is injective with respect to
the functor $\mathrm{Hom}_{k_\mathcal{T}}(G,\bullet )$.
\item[(ii)]
The family of $\mathcal{T}$-flabby sheaves is injective with
respect to the functor $\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,\bullet )$.
\end{itemize}
\end{teo}
\dim\ \
(i) Let $G \in \mathrm{Coh}(\mathcal{T})$. Let $\exs{F'}{F}{F''}$ be an exact sequence in
$\mod(k_\mathcal{T})$ and assume that $F'$ is $\mathcal{T}$-flabby.
We have to show that the sequence
$$\exs{\mathrm{Hom}_{k_\mathcal{T}}(G,F')}{\mathrm{Hom}_{k_\mathcal{T}}(G,F)}{\mathrm{Hom}_{k_\mathcal{T}}(G,F'')}$$
is exact.
There is an epimorphism $\varphi:\oplus_{i \in I}
k_{U_i} \to G$ where $I$ is finite and $U_i \in \mathcal{T}$ for
each $i \in I$. The sequence $\exs{\ker \varphi}{\oplus_{i \in I} k_{U_i}}{G}$ is
exact. We set for short $G_1=\ker\varphi$ and $G_2=\oplus_{i \in
I} k_{U_i}$. We get the following diagram where the first column is
exact by Theorem \ref{teo flabby hom and Hom} (i)
$$
\xymatrix{ & 0 \ar[d] & 0 \ar[d] & 0 \ar[d] & \\
0 \ar[r] & \mathrm{Hom}_{k_\mathcal{T}}(G,F') \ar[d] \ar[r] &
\mathrm{Hom}_{k_\mathcal{T}}(G,F)
\ar[d] \ar[r] & \mathrm{Hom}_{k_\mathcal{T}}(G,F'') \ar[d] \ar[r] & 0 \\
0 \ar[r] & \mathrm{Hom}_{k_\mathcal{T}}(G_2,F') \ar[d] \ar[r] &
\mathrm{Hom}_{k_\mathcal{T}}(G_2,F)
\ar[d] \ar[r] & \mathrm{Hom}_{k_\mathcal{T}}(G_2,F'') \ar[d] \ar[r] & 0\\
0 \ar[r] & \mathrm{Hom}_{k_\mathcal{T}}(G_1,F') \ar[d] \ar[r] &
\mathrm{Hom}_{k_\mathcal{T}}(G_1,F)
\ar[d] \ar[r] & \mathrm{Hom}_{k_\mathcal{T}}(G_1,F'') \ar[d] \ar[r] & 0\\
& 0 & 0 & 0 & }
$$
The second row is exact by Proposition \ref{qinjU} (ii), hence the top row is exact by the snake lemma. \\
(ii) Let $G \in \mathrm{Coh}(\mathcal{T})$. It is enough to check that
for each $U \in \mathcal{T}$ and each exact sequence
$\exs{F'}{F}{F''}$ with $F'$ $\mathcal{T}$-flabby, the
sequence
$$\exs{\Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F'))}{\Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F))}{\Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F''))}$$
is exact. We have
$$\Gamma(U,\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,\bullet )) \simeq \mathrm{Hom}_{k_\mathcal{T}}(G_U,\bullet ),$$ and, by (i) and the fact that $G_U\in \mathrm{Coh}(\mathcal{T})$ (Theorem \ref{teo coh stable} (ii)), $\mathcal{T}$-flabby objects are
injective with respect to the functor
$\mathrm{Hom}_{k_\mathcal{T}}(G_U,\bullet )$ for each $G \in \mathrm{Coh}(\mathcal{T})$,
and for each $U \in \mathcal{T}$.
\nopagebreak \begin{flushright}$\Box$\end{flushright} \\
\begin{prop}\label{hoinj}
Let $F \in \mod(k_\mathcal{T})$. Then $F$ is
$\mathcal{T}$-flabby if and only if $\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F)$ is $\mathcal{T}$-flabby for
each $G \in \mathrm{Coh}(\mathcal{T})$.
\end{prop}
\dim\ \
Suppose that $F$ is $\mathcal{T}$-flabby, and let $G \in
\mathrm{Coh}(\mathcal{T})$.
We have
$$\mathrm{Hom}_{k_\mathcal{T}}(\bullet ,\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F)) \simeq \mathrm{Hom}_{k_\mathcal{T}}(\bullet \otimes_{k_\mathcal{T}} G,F)$$
and $\mathrm{Hom}_{k_\mathcal{T}}(\bullet \otimes_{k_\mathcal{T}} G,F)$ is exact on $\mathrm{Coh}(\mathcal{T})$ by Theorems \ref{teo coh stable} (ii) and \ref{teo flabby hom and Hom} (i).
Suppose that $\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F)$ is $\mathcal{T}$-flabby for each $G \in
\mathrm{Coh}(\mathcal{T})$. Let $U,V \in \mathcal{T}$ with $V \supseteq U$. For each $W \in \mathcal{T}$ the morphism $\Gamma(V;\Gamma_WF) \to \Gamma(U;\Gamma_WF)$ is surjective. Hence the morphism
\begin{eqnarray*}
\Gamma(V;F) & \simeq & \Gamma(V;\Gamma_VF) \\
& \to & \Gamma(U;\Gamma_VF) \\
& \simeq & \Gamma(U;F)
\end{eqnarray*}
is surjective. \nopagebreak \begin{flushright}$\Box$\end{flushright}
Let us consider the following subcategory of $\mod(k_\mathcal{T})$:
$$
\mathcal{P}_{X_\mathcal{T}} := \{G \in \mod(k_\mathcal{T});\; G \text{ is
$\mathrm{Hom}_{k_\mathcal{T}}(\bullet ,F)$-acyclic for each $F\in\mathcal{F}_{X_\mathcal{T}}$}\},
$$
where ${\cal F}_{X_\mathcal{T}}$ is the family of $\mathcal{T}$-flabby objects of $\mod(k_\mathcal{T})$.
This category is generating, in fact if $\{U_j\}_{j\in J} \in \mathcal{T}$, then $\oplus_{j\in J} k_{U_j} \in \mathcal{P}_{X_\mathcal{T}}$ by Theorem \ref{rcinj} (and the fact that
$$\Pi \mathrm{Hom} _{k_\mathcal{T}}(\bullet , \bullet ) \simeq \mathrm{Hom} _{k_\mathcal{T}}(\oplus \bullet , \bullet )$$
and products are exact). Moreover $\mathcal{P}_{X_\mathcal{T}}$ is stable by $\bullet
\otimes_{k_\mathcal{T}} K$, where $K \in \mathrm{Coh}(\mathcal{T})$. In fact if $G \in \mathcal{P}_{X_\mathcal{T}}$ and $F \in \mathcal{F}_{X_\mathcal{T}}$ we have
$$\mathrm{Hom}_{k_\mathcal{T}}(G \otimes _{k_\mathcal{T}}K,F) \simeq \mathrm{Hom}_{k_\mathcal{T}}(G,\mathcal{H}\mathit{om}_{k_\mathcal{T}}(K,F))$$
and $\mathcal{H}\mathit{om}_{k_\mathcal{T}}(K,F)$ is $\mathcal{T}$-flabby by Proposition \ref{hoinj}. In particular, if $G\in \mathcal{P}_{X_\mathcal{T}}$ then $G_U\in \mathcal{P}_{X_\mathcal{T}}$ for every $U\in \mathrm{Op} (X_\mathcal{T}).$
\begin{teo}\label{mathcalJP}
The category $(\mathcal{P}^{op}_{X_\mathcal{T}},\mathcal{F}_{X_\mathcal{T}})$
is injective with respect to
the functors $\mathrm{Hom}_{k_\mathcal{T}}(\bullet ,\bullet )$ and $\mathcal{H}\mathit{om}_{k_\mathcal{T}}(\bullet ,\bullet )$.
\end{teo}
\dim\ \
(i) Let $G \in \mathcal{P}_{X_\mathcal{T}}$ and consider an exact sequence
$\exs{F'}{F}{F''}$ with $F'$ $\mathcal{T}$-flabby. We have to
prove that the sequence
$$\exs{\mathrm{Hom}_{k_\mathcal{T}}(G,F')}{\mathrm{Hom}_{k_\mathcal{T}}(G,F)}{\mathrm{Hom}_{k_\mathcal{T}}(G,F'')}$$
is exact. Since the functor $\mathrm{Hom}_{k_\mathcal{T}}(G,\bullet )$ is acyclic
on $\mathcal{T}$-flabby sheaves we obtain the result.
Let $F$ be $\mathcal{T}$-flabby, and let $\exs{G'}{G}{G''}$ be
an exact sequence on $\mathcal{P}_{X_\mathcal{T}}$. Since the objects of
$\mathcal{P}_{X_\mathcal{T}}$ are $\mathrm{Hom}_{k_\mathcal{T}}(\bullet ,F)$-acyclic the sequence
$$\exs{\mathrm{Hom}_{k_\mathcal{T}}(G'',F)}{\mathrm{Hom}_{k_\mathcal{T}}(G,F)}{\mathrm{Hom}_{k_\mathcal{T}}(G',F)}$$
is exact.
(ii) Let $G \in \mathcal{P}_{X_\mathcal{T}}$, and let
$\exs{F'}{F}{F''}$ be an exact sequence with $F'$ $\mathcal{T}$-flabby. We shall show that for each $U \in
\mathcal{T}$ the sequence
$$\exs{\Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F'))}{\Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F))}{\Gamma(U;\mathcal{H}\mathit{om}_{k_\mathcal{T}}(G,F''))}$$
is exact. This is equivalent to show that for each $U \in
\mathcal{T}$ the sequence
$$\exs{\mathrm{Hom}_{k_\mathcal{T}}(G_U,F')}{\mathrm{Hom}_{k_\mathcal{T}}(G_U,F)}{\mathrm{Hom}_{k_\mathcal{T}}(G_U,F'')}$$
is exact. This follows since $G_U \in \mathcal{P}_{X_\mathcal{T}}$ as we saw above. The proof of
the exactness in $\mathcal{P}^{op}_{X_\mathcal{T}}$ is similar.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop}\label{prop gamma z and flabby}
Let $F \in \mod(k_\mathcal{T})$. The following assumptions are equivalent
\begin{itemize}
\item[(i)] $F$ is $\mathcal{T}$-flabby,
\item[(ii)] $F$ is $\mathrm{Hom}_{k_\mathcal{T}}(G,\bullet )$-acyclic for each $G \in \mathrm{Coh}(\mathcal{T})$,
\item[(iii)] $R^1\mathrm{Hom}_{k_\mathcal{T}}(k_{V \setminus U},F)=0$ for each $U,V \in \mathcal{T}$.
\end{itemize}
\end{prop}
\dim\ \
(i) $\Rightarrow$ (ii) follows from Theorem \ref{rcinj}, (ii) $\Rightarrow$ (iii) setting $G=k_{V \setminus U}$ with $U,V \in \mathcal{T}$, (iii) $\Rightarrow$ (i) since if $R^1\mathrm{Hom}_{k_\mathcal{T}}(k_{V \setminus U},F)=0$ for each $U,V \in \mathcal{T}$ with $V \supseteq U$, then the restriction $\Gamma(V;F) \to \Gamma(U;F)$ is surjective.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
Let $X,Y$ be two topological spaces and let $\mathcal{T} \subset \mathrm{Op}(X)$, $\mathcal{T}' \subset \mathrm{Op}(Y)$ satisfy \eqref{hytau}. Let $f:X \to Y$ be a continuous map. If $\imin f(\mathcal{T}') \subset \mathcal{T}$ then $f$ defines a morphism of sites $f:X_\mathcal{T} \to Y_{\mathcal{T}'}$.
\begin{prop}\label{prop f star and flabby}
Let $f:X_\mathcal{T} \to Y_{\mathcal{T}'}$ be a morphism of sites. $\mathcal{T}$-flabby sheaves are injective with
respect to the functor $f_*$. The functor $f_*$ sends
$\mathcal{T}$-flabby sheaves to $\mathcal{T}'$-flabby sheaves.
\end{prop}
\dim\ \
Let us consider $V \in \mathcal{T}'$. There is an
isomorphism of functors $\Gamma(V;f_*\bullet ) \simeq \Gamma(\imin
f(V);\bullet )$. It follows from Proposition \ref{qinjU} that
$\mathcal{T}$-flabby are injective with respect to the functor
$\Gamma(\imin f(V);\bullet )$ for any $V \in \mathcal{T}'$.
Let $F$ be $\mathcal{T}$-flabby and let $U,V \in \mathcal{T}'$ with $V \supset U$. Then the morphism
$$\Gamma(V;f_*F) = \Gamma(\imin f(V);F) \to \Gamma(\imin f(U);F) = \Gamma(U;f_*F)$$
is surjective.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop} Assume that $X$ is locally weakly quasi-compact. Let $F \in \mod(k_\mathcal{T})$ be $\mathcal{T}$-flabby. Then $\imin \rho F$ is c-soft.
\end{prop}
\dim\ \ Recall that if $U \in
\mathrm{Op}(X)$ then $\Gamma(U;\imin \rho F) \simeq \lpro {V
\subset\subset U} \Gamma(V;F)$, where $V \in \mathcal{T}$. Let $W \in \mathrm{Op}(X)$, $W \subset\subset X$. It follows from
Lemma \ref{UcssVTc} that every $U' \supset\supset W$, $U' \in \mathrm{Op}(X)$ contains $U \in \mathcal{T}$ such that $U \supset\supset W$. Hence
$$
\lind {U'}\Gamma(U';F) \simeq \lind U\Gamma(U;F),
$$
where $U' \supset\supset W$, $U' \in \mathrm{Op}(X)$ and $U \in \mathcal{T}$ such that $U \supset\supset W$.
We have
the chain of isomorphisms
\begin{eqnarray*}
\lind U \Gamma(U;\imin \rho F)
& \simeq & \lind U \lpro {V \subset\subset U} \Gamma(V;F) \\
& \simeq & \lind U \Gamma(U;F)
\end{eqnarray*}
where $U \in \mathcal{T}$, $U \supset\supset W$ and $V \in
\mathcal{T}$. The first isomorphism follows from Proposition \ref{imineta} and second one follows since for each $U \supset\supset W$, $U \in \mathcal{T}$, there exists $V \in \mathcal{T}$ such that $U \supset\supset V \supset\supset W$.
Let $V,W \in \mathrm{Op}^c(X)$ with $V \subset\subset W$.
Since $F$ is $\mathcal{T}$-flabby and filtrant inductive limits are
exact, the morphism $\lind {W'} \Gamma(W';\imin \rho F) \simeq \lind {W'} \Gamma(W';F) \to
\lind U \Gamma(U;F) \simeq \lind U \Gamma(U;\imin \rho F)$, where $W',U\in\mathcal{T}$, $W' \supset\supset W$, $U \supset\supset V$, is surjective. Hence
$\Gamma(W;\imin \rho F) \to \lind {U \supset\supset V}\Gamma(U;\imin \rho F)$ is surjective. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\subsection{$\mathcal{T}_{loc}$-sheaves}
Let $X$ be a $\mathcal{T}$-space and let
\begin{equation}\label{Tloc}
\mathcal{T} _{loc}=\{U\in \mathrm{Op}(X): U \cap W \in \mathcal{T}\,\,\, \textrm{for every} \,\,\, W \in \mathcal{T}\}.
\end{equation}
Clearly, $\varnothing, X \in \mathcal{T}_{loc}$, $\mathcal{T}\subseteq \mathcal{T}_{loc}$ and $\mathcal{T}_{loc}$ is closed under finite intersections.
\begin{df}
We make the following definitions:
\begin{itemize}
\item
a subset $S$ of $X$ is a $\mathcal{T}_{loc}$-subset if and only if $S\cap V$ is a $ \mathcal{T}$-subset for every $V\in \mathcal{T}$;
\item
a closed (resp. open) $\mathcal{T}_{loc}$-subset is a $\mathcal{T}_{loc}$-subset which is closed (resp. open) in $X$;
\item
a $\mathcal{T}_{loc}$-connected subset is a $\mathcal{T}_{loc}$-subset which is not the disjoint union of two proper clopen $\mathcal{T}_{loc}$-subsets.
\end{itemize}
\end{df}
Observe that if $\{S_i\}_i$ is a family of $\mathcal{T}_{loc}$-subsets such that $\{i:S_i\cap W\neq \emptyset \}$ is finite for every $W\in \mathcal{T}$, then the union and the intersection of the family $\{S_i\}_i$ is a $\mathcal{T}_{loc}$-subset. Also the complement of a $\mathcal{T}_{loc}$-subset is a $\mathcal{T}_{loc}$-subset. Therefore the $\mathcal{T}_{loc}$-subsets form a Boolean algebra.\\
One can endow $\mathcal{T}_{loc}$ with a Grothendieck topology in the following way: a
family $\{U_i\}_i$ in $\mathcal{T}_{loc}$ is a covering of $U \in \mathcal{T}_{loc}$ if for any
$V\in\mathcal{T}$, there exists a finite subfamily covering $U \cap V$. We denote by $X_{\mathcal{T}_{loc}}$ the associated site, write for short $k_{\mathcal{T}_{loc}}$ instead of $k_{X_{\mathcal{T}_{loc}}}$, and let
\[
\xymatrix{
& X \ar[ld]^{\rho _{loc}} \ar[rd]^{\rho } &\\
X_{\mathcal{T}_{loc}} \ar[rr] && X_{\mathcal{T}}
}
\]
be the natural morphisms of sites.
\begin{oss}\label{TlocT}
The forgetful functor, induced by the natural morphism of sites $X_{\mathcal{T}_{loc}} \rightarrow X_{\mathcal{T}}$, gives an equivalence of categories
$$\mod(k_{\mathcal{T}_{loc}}) \stackrel{\sim}{\to} \mod(k_{\mathcal{T}}).$$
The quasi-inverse to the forgetful functor sends $F \in \mod(k_{\mathcal{T}})$ to $F_{loc}\in\mod(k_{\mathcal{T}_{loc}})$ given by $F_{loc}(U)=\lpro {V\in\mathcal{T}}F(U \cap V)$ for every $U \in \mathcal{T}_{loc}$.
Therefore, we can and will identify $\mod(k_{\mathcal{T}_{loc}})$ with $\mod(k_{\mathcal{T}})$ and apply the previous results for $\mod(k_{\mathcal{T}})$ to obtain analogues results for $\mod(k_{\mathcal{T}_{loc}}).$
\end{oss}
Recall that $F \in \mod(k_{\mathcal{T}})$ is $\mathcal{T}$-flabby if the restriction $\Gamma(V;F) \to \Gamma(U;F)$ is surjective for any $U,V \in \mathcal{T}$ with $V \supseteq U$. Assume that
\begin{equation}\label{covT'}
\textrm{$X_{\mathcal{T}_{loc}}$ has a countable cover $\{V_n\}_{n \in \mathbb{N}}$ with $V_n \in \mathcal{T},$ $\forall n\in\mathbb{N}$.}
\end{equation}
\begin{prop}\label{tflc} Let $F \in \mod(k_{\mathcal{T}})$. Then $F$ is $\mathcal{T}$-flabby if and only if the restriction $\Gamma(X;F) \to \Gamma(U;F)$ is surjective for any $U \in \mathcal{T}_{loc}$.
\end{prop}
\dim\ \ Suppose that $F$ is $\mathcal{T}$-flabby. Consider a covering $\{V_n\}_{n \in \mathbb{N}}$ of $X_{\mathcal{T}_{loc}}$ satisfying \eqref{covT'}.
Set $U_n=U \cap V_n$ and $S_n=V_n \setminus U_n$. All the
sequences
$$
\exs{k_{U_n}}{k_{V_n}}{k_{S_n}}
$$
are exact. Since $F$ is $\mathcal{T}$-flabby the sequence
$$
\exs{\mathrm{Hom}_{k_{\mathcal{T}}}(k_{S_n},F)}{\mathrm{Hom}_{k_{\mathcal{T}}}(k_{V_n},F)}{\mathrm{Hom}_{k_{\mathcal{T}}}(k_{U_n},F)}
$$
is exact. Moreover the morphism $\mathrm{Hom}_{k_{\mathcal{T}}}(k_{S_{n+1}},F) \to \mathrm{Hom}_{k_{\mathcal{T}}}(k_{S_{n}},F)$ is surjective for all $n$ since $S_n=S_{n+1} \cap V_n$ is open in $S_{n+1}$. Then by Proposition 1.12.3 of \cite{KS90} the sequence
$$
\exs{\lpro n\mathrm{Hom}_{k_{\mathcal{T}}}(k_{S_n},F)}{\lpro n\mathrm{Hom}_{k_{\mathcal{T}}}(k_{V_n},F)}{\lpro n\mathrm{Hom}_{k_{\mathcal{T}}}(k_{U_n},F)}
$$
is exact. The result follows since $\lpro n\Gamma(U_n;G) \simeq \Gamma(U;G)$ for any $G \in \mod(k_{\mathcal{T}})$ and $U \in \mathcal{T}_{loc}$. The converse is obvious. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{prop}\label{tflcU} The full additive subcategory of $\mod (k_\mathcal{T})$ of $\mathcal{T}$-flabby object is $\Gamma (U;\bullet )$-injective for every $U \in \mathcal{T}_{loc}.$
\end{prop}
\dim\ \ Take an exact sequence $\exs{F'}{F}{F''}$, and suppose that
$F'$ is $\mathcal{T}$-flabby. Consider a covering $\{V_n\}_{n \in \mathbb{N}}$ of $X_{\mathcal{T}_{loc}}$ satisfying \eqref{covT'}.
Set $U_n=U \cap V_n$. All the
sequences
$$\exs{\Gamma(U_n;F')}{\Gamma(U_n;F)}{\Gamma(U_n;F'')}$$
are exact by Proposition \ref{qinjU}, and the morphism
$\Gamma(U_{n+1};F') \to
\Gamma(U_n;F')$ is surjective for all
$n$. Then by Proposition 1.12.3 of \cite{KS90} the sequence
$$\exs{\lpro n\Gamma(U_n;F')}{\lpro n\Gamma(U_n;F)}{\Gamma(U_n;F'')}$$ is exact. Since $\lpro n\Gamma(U_n;G) \simeq \Gamma(U;G)$ for any $G
\in \mod(k_{\mathcal{T}})$ the result follows. \nopagebreak \begin{flushright}$\Box$\end{flushright}
Let $X,Y$ be two topological spaces and let $\mathcal{T} \subset \mathrm{Op}(X)$, $\mathcal{T}' \subset \mathrm{Op}(Y)$ satisfy \eqref{hytau}. Let $f:X \to Y$ be a continuous map. If $\imin f(\mathcal{T}'_{loc}) \subseteq \mathcal{T}_{loc}$ then $f$ defines a morphism of sites $f:X_{\mathcal{T}_{loc}} \to Y_{\mathcal{T}'_{loc}}$.
\begin{cor} Let $f:X_{\mathcal{T}_{loc}} \to Y_{\mathcal{T}'_{loc}}$ be a morphism of sites. $\mathcal{T}$-flabby sheaves are injective with respect to the functor $f_*$. The functor $f_*$ sends $\mathcal{T}$-flabby sheaves to $\mathcal{T}'$-flabby sheaves.
\end{cor}
\dim\ \
Let us consider $V \in \mathcal{T}'_{loc}$. There is an isomorphism of functors $\Gamma(V;f_*\bullet ) \simeq \Gamma(\imin f(V);\bullet )$. It follows from Proposition \ref{tflcU} that $\mathcal{T}$-flabby are injective with respect to the functor $\Gamma(\imin f(V);\bullet )$ for any $V \in \mathcal{T}'_{loc}$.
Let $F$ be $\mathcal{T}$-flabby and let $U,V \in \mathcal{T}'$ with $V \supset U$. Then the morphism
$$\Gamma(V;f_*F) = \Gamma(\imin f(V);F) \to \Gamma(\imin f(U);F) = \Gamma(U;f_*F)$$
is surjective by Proposition \ref{tflc}.
\nopagebreak \begin{flushright}$\Box$\end{flushright}
\begin{oss}\label{Tc} An interesting case is when $X$ is a locally weakly quasi-compact space and there exists $\mathcal{S} \subseteq \mathrm{Op}(X)$ with $\mathcal{T}=\{U \in \mathcal{S}: \,U \subset\subset X\}$ satisfying \eqref{hytau}.
Assume that $X$ satisfies \eqref{Xcountable}.
Then $X$ has a covering $\{V_n\}_{n \in \mathbb{N}}$ of $X$ such that $V_n \in \mathcal{T}$ and
$V_n \subset \subset V_{n+1}$ for each $n \in \mathbb{N}$.
By Lemma \ref{X has a cov} we may find a covering $\{U_n\}_{n \in \mathbb{N}}$ of $X$ such that $U_n \in \mathrm{Op}^c(X)$ and
$U_n \subset \subset U_{n+1}$ for each $n \in \mathbb{N}$. By Lemma \ref{UcssVTc} for each $n \in \mathbb{N}$ there exists $V_n \in \mathcal{T}$ such that $U_n \subset\subset V_n \subset\subset U_{n+1}$.
In this situation Proposition \ref{tflc} and \ref{tflcU} are satisfied.
\end{oss}
\subsection{$\mathcal{T}$-spectrum}
Let $X$ be a topological space and let $\P(X)$ be the power set of $X$. Consider a subalgebra $\mathcal{F}$ of the power set Boolean algebra $\langle \P(X),\subseteq \rangle$. Then $\mathcal{F}$ is closed under finite unions, intersections and complements. We refer to \cite{BS71} for an introduction to this subject.
The Boolean algebra $\mathcal{F}$ has an associated topological space, that we denote by $S(\mathcal{F})$, called its Stone space. The points in $S(\mathcal{F})$ are the ultrafilters on $\mathcal{F}$.
Recall that an ultrafilter of
is a subset $\alpha$ of $\mathcal{F}$ such that
\begin{itemize}
\item[(i)] $X \in \alpha$.
\item[(ii)] $A \cap B \in \alpha$ if and only if $A \in \alpha$ and $B \in \alpha$.
\item[(iii)] $\varnothing \notin \alpha$.
\item[(iv)] $A \cup B \in \alpha$ if and only if $A \in \alpha$ or $B \in \alpha$.
\end{itemize}
In order to put a topology on $S(\mathcal{F})$, we consider $S(\mathcal{F})$ as a subset of the powerset $2^\mathcal{F}$, identifying an ultrafilter $\alpha$ with its characteristic function ${\bf 1}_\alpha : \mathcal{F} \to 2=\{0, 1\}$. We equip 2 with the discrete topology and $2^\mathcal{F}$ with the product topology. By Tychonoff theorem, $2^\mathcal{F}$ is compact Hausdorff. The topology on $S(\mathcal{F})$ is the topology induced by the topology of $2^\mathcal{F}$.
The topology on $S(\mathcal{F})$ is generated by a basis of open and closed sets consisting of all sets of the form
$$
\widetilde{A} = \{\alpha \in S(\mathcal{F}): A \in \alpha \},
$$
where $A\in \mathcal{F}$. The space
$S(\mathcal{F})$ is a compact totally disconnected Hausdorff space, such spaces are called Stone spaces. Moreover, for each $A \in \mathcal{F}$, the subspace $\widetilde{A}$ is Hausdorff and compact.
\begin{df}
Let $X$ be a $\mathcal{T}$-space and let $\mathcal{F}$ be the Boolean algebra of $\mathcal{T}_{loc}$-subsets of $X$ (i.e. Boolean combinations of elements of $\mathcal{T}_{loc}$).
The topological space $\widetilde{X}_\mathcal{T}$ is the data of:
\begin{itemize}
\item the points of $S(\mathcal{F})$ such that $U \in \alpha$ for some $U \in \mathcal{T}$,
\item a basis for the topology is given by the family of subsets $\{\widetilde{U}:U\in\mathcal{T}\}$.
\end{itemize}
We call $\widetilde{X}_\mathcal{T}$ the $\mathcal{T}$-spectrum of $X$.
\end{df}
With this topology, for $U \in \mathcal{T}$, the set $\widetilde{U}$ is quasi-compact in $\widetilde{X}_\mathcal{T}$ since it is quasi-compact in $S(\mathcal{F})$. Hence $\widetilde{X}_\mathcal{T}$ is locally weakly quasi-compact with a basis of quasi-compact open subsets given by $\{\widetilde{U}:U\in\mathcal{T}\}$. Note that if $X\in \mathcal{T}$, then $\widetilde{X}_\mathcal{T}=\widetilde{X}$ which is a spectral topological space.
\begin{prop}\label{isosites} Let $X$ be a $\mathcal{T}$-space. Then there is an equivalence of categories $\mod(k_\mathcal{T}) \simeq \mod(k_{\widetilde{X}_\mathcal{T}})$.
\end{prop}
\dim\ \ Let us consider the functor
\begin{eqnarray*}
\zeta^t:\mathcal{T} & \to & \mathrm{Op}(\widetilde{X}_\mathcal{T}) \\
U & \mapsto & \widetilde{U}.
\end{eqnarray*}
This defines a morphism of sites $\zeta:\widetilde{X}_\mathcal{T} \to X_\mathcal{T}$. Indeed, if $V\in \mathcal{T},$ $S\in \mathrm{Cov} (V)$, then $\widetilde{S}=\{\widetilde{V}_i: V_i\in S\} \in \mathrm{Cov}(\widetilde{V})$. Let $F \in \mod(k_\mathcal{T})$ and consider the presheaf $\zeta^\gets F \in \mathrm{Psh}(k_{\widetilde{X}_\mathcal{T}})$ defined by $\zeta^\gets F(U) = \lind {U \subseteq \widetilde{V}} F(V)$. In particular, if $U=\widetilde{V}$, $V \in \mathcal{T}$, $\zeta^\gets F(U) \simeq F(V)$. In this case, by Corollary \ref{facKK} we have the isomorphisms
$$
\imin \zeta F(\widetilde{V}) = (\zeta^\gets F)^{++}(\widetilde{V}) \simeq \zeta^\gets F(\widetilde{V}) \simeq F(V).
$$
Then for $V \in \mathcal{T}$ we have
$$
\zeta_*\imin\zeta F(V) \simeq \imin \zeta F(\widetilde{V}) \simeq F(V).
$$
This implies $\zeta_*\circ\imin\zeta\simeq\mathrm{id}$. On the other hand, given $\alpha \in \widetilde{X}_\mathcal{T}$ and $G \in \mod(k_{\widetilde{X}_\mathcal{T}})$,
\begin{eqnarray*}
(\imin \zeta \zeta_* G)_\alpha & \simeq & \lind {\widetilde{U} \ni \alpha, U \in \mathcal{T}} \imin \zeta \zeta_* G(\widetilde{U})\\
& \simeq & \lind {\widetilde{U} \ni \alpha, U \in \mathcal{T}} \zeta_*G(U)\\
& \simeq & \lind {\widetilde{U} \ni \alpha, U \in \mathcal{T}}G(\widetilde{U}) \\
& \simeq & G_\alpha
\end{eqnarray*}
since $\{\widetilde{U}: U \in \mathcal{T}\}$ forms a basis for the topology of $\widetilde{X}_\mathcal{T}$. This implies $\imin \zeta \circ \zeta_* \simeq \mathrm{id}$. \nopagebreak \begin{flushright}$\Box$\end{flushright}
\section{Examples}\label{Sec examples}
In this section we see some examples of $\mathcal{T}$-sheaves. When $\mathcal{T}$ is the family of semialgebraic open subsets of a locally semialgebraic space we obtain the construction of \cite{De91,DK85}, when $\mathcal{T}$ is the family of relatively compact subanalytic subsets of a real analytic manifold we obtain the subanalytic site of \cite{KS01,Pr1}, in the case of a real vector bundle and conic subanalytic open subsets we obtain the conic subanaytic site, and when $\mathcal{T}$ is the family of open definable subsets of a locally definable space in a given o-minimal structure we obtain in the definable case the o-minimal site of \cite{ejp}. The properties of $\mathcal{T}$-sheaves studied in the previous section apply to these examples. Good references on o-minimality are, for example, the book \cite{VD98} by Van Den Dries and the notes \cite{Co00} by Coste. For semialgebraic geometry relevant to this paper the reader should consult the work by Delfs \cite{De91}, Delfs and Knebusch \cite{DK85} and the book \cite{BCR} by Bochnak, Coste and Roy, for subanalytic geometry we refer to the work \cite{BM88} by Bierstone and Milmann.
\subsection{The semialgebraic site}
Let $R=(R,<, 0,1,+,\cdot )$ be a real closed field. A locally semialgebraic space is a ringed space $(X, \mathcal{O}_X)$ with a sheaf of locally semialgebraic functions which is locally isomorphic to open semialgebraic subspaces $\{(X_i, \mathcal{O}_{X_i})\}_{i\in I}$ with their sheaves of semialgebraic functions. This means that:
\begin{itemize}
\item[(i)]
$X=\bigcup _{i\in I}X_i$;
\item[(ii)]
there are injective maps $\phi _i:X_i\to R^{l_i}$ such that $\phi _i(X_i)$ is a semialgebraic subset of $R^{l_i}$;
\item[(iii)]
for all $j$, $\phi _i(X_i\cap X_j)$ is open in $\phi _i(X_i)$ and the transition maps $\phi _{ij}:\phi _i(X_i\cap X_j)\to \phi _j(X_i\cap X_j):x\mapsto \phi _j(\phi _i^{-1}(x))$ are semialgebaric homeomorphisms.
\end{itemize}
If $I$ is finite, then $X$ is a semialgebraic space. A subset $A$ of $X$ is semialgebraic if and only if there is a finite $I_0\subseteq I$ such that $A\subseteq \bigcup _{i\in I_0}X_i$ and for each $i$, $\phi _i(A\cap X_i)$ is a semialgebraic subset of $\phi _i(X_i)$, and, a subset $B$ of $X$ is locally semialgebraic if and only if for each $i$, $B\cap X_i$ is a semialgebraic subset of $X$. A map between (locally) semialgebraic spaces is (locally) semialgebraic if when it is read through the charts it is (locally) semialgebarci. Thus we have the category of (locally) semialgebraic spaces with (locally) semialgebraic continuous maps.
Let $X$ be a locally semialgebraic space and consider the subfamily of $\mathrm{Op}(X)$ defined by $\mathcal{T}= \{U \in \mathrm{Op}(X): U \,\, \text{is semialgebraic}\}$. The family $\mathcal{T}$ satisfy \eqref{hytau} and the associated site $X_{\mathcal{T}}$ is the semialgebraic site on $X$ of \cite{De91,DK85}. Note also that: (i) the $\mathcal{T}$-subsets of $X$ are exactly the semialgebraic subsets of $X$ (\cite{BCR}); (ii) $\mathcal{T}_{loc}= \{U \in \mathrm{Op}(X): U \,\, \text{is locally semialgebraic}\}$ and (iii) the $\mathcal{T}_{loc}$-subsets of $X$ are exactly the locally semialgebraic subsets of $X$ (\cite{DK85}).\\
For each $F \in \mod(k_{\mathcal{T}})$ there exists a filtrant inductive system $\{F_i\}_{i\in I}$ in $\mathrm{Coh}(\mathcal{T})$ such that $F \simeq \lind i \rho_* F_i$.\\
The subcategory of $\mathcal{T}$-flabby sheaves corresponds to the subcategory of $sa$-flabby sheaves of \cite{De91} and it is injective with respect to $\Gamma(U;\bullet)$, $U \in \mathrm{Op}(X_{\mathcal{T}})$ and $\mathrm{Hom}_{k_{\mathcal{T}}}(G,\bullet)$, $G \in \mathrm{Coh}(\mathcal{T})$. Our results on $\mathcal{T}$-flabby sheaves generalize those for $sa$-flabby sheaves from \cite{De91}.\\
We call in this case the $\mathcal{T}$-spectrum $ \widetilde{X}_{\mathcal{T}}$ of $X$ the semilagebraic spectrum of $X$. The points of $\widetilde{X}_{\mathcal{T}}$ are the ultrafilters $\alpha$ of locally semialgebraic subsets of $X$ such that $U \in \alpha$ for some $U \in \mathrm{Op}(X_{\mathcal{T}})$. This is a locally weakly quasi-compact space with basis of quasi-compact open subsets given by $\{\widetilde{U}:U\in \mathrm{Op}(X_{\mathcal{T}})\}$ and there is an equivalence of categories $\mod(k_{\mathcal{T}}) \simeq \mod(k_{\widetilde{X}_{\mathcal{T}}})$. When $X$ is semialgebraic, then $\widetilde{X}_{\mathcal{T}} =\widetilde{X}$, the semialgebraic spectrum of $X$ from \cite{CR82}, and there is an equivalence of categories $\mod(k_{\mathcal{T}}) \simeq \mod(k_{\widetilde{X}})$ (\cite{De91}).
\subsection{The subanalytic site}
Let $X$ be a real analytic manifold and consider the subfamily of $\mathrm{Op}(X)$ defined by
$\mathcal{T}=\mathrm{Op}^c(X_{sa})=\{U \in \mathrm{Op}(X_{sa}): U\,\, $is subanalytic relatively compact$\}$. The family $\mathcal{T}$ satisfies \eqref{hytau} and the associated site $X_{\mathcal{T}}$ is the subanalytic site $X_{sa}$ of \cite{KS01,Pr1}. In this case $\mathcal{T}_{loc}$-subsets are the subanalytic subsets of $X$.\\
The family $\mathrm{Coh}(\mathcal{T})$ corresponds to the family $\mod_{\mathbb{R}\textrm{-}\mathrm{c}}^c(k_X)$ of $\mathbb{R}$-constructible sheaves with compact support, and for each $F \in \mod(k_{X_{sa}})$ there exists a filtrant inductive system $\{F_i\}_{i\in I}$ in $\mod_{\mathbb{R}\textrm{-}\mathrm{c}}^c(k_X)$ such that $F \simeq \lind i \rho_* F_i$.\\
The subcategory of $\mathcal{T}$-flabby sheaves corresponds to quasi-injective sheaves and it is injective with respect to $\Gamma(U;\bullet)$, $U \in \mathrm{Op}(X_{sa})$ and $\mathrm{Hom}_{k_{X_{sa}}}(G,\bullet)$, $G \in \mod_{\mathbb{R}\textrm{-}\mathrm{c}}(k_X)$. \\
We call in this case the $\mathcal{T}$-spectrum $ \widetilde{X}_{\mathcal{T}}$ of $X$ the subanalytic spectrum of $X$ and denote it by $\widetilde{X}_{sa}$. The points of $\widetilde{X}_{sa}$ are the ultrafilters of subanalytic subsets of $X$ such that $U \in \alpha$ for some $U \in \mathrm{Op}^c(X_{sa})$. Then there is an equivalence of categories $\mod(k_{X_{sa}}) \simeq \mod(k_{\widetilde{X}_{sa}})$.
\
Let $U \in \mathrm{Op}(X_{sa})$ and denote by $U_{X_{sa}}$ the site with the topology induced by $X_{sa}$. This corresponds to the site $X_{\mathcal{T}}$, where
$\mathcal{T}=\mathrm{Op}^c(X_{sa}) \cap U$. In this situation \eqref{hytau} is satisfied.
\subsection{The conic subanalytic site}
Let $X$ be a real analytic manifold endowed with a subanalytic action $\mu$ of $\mathbb{R}^{{\scriptscriptstyle{+}}}$. In other words we have a subanalytic map
$$\mu: X \times \mathbb{R}^{{\scriptscriptstyle{+}}} \to X,$$
which satisfies, for each $t_1,t_2 \in \mathbb{R}^{{\scriptscriptstyle{+}}}$:
$$
\begin{cases}
\mu(x,t_1t_2)=\mu(\mu(x,t_1),t_2), \\
\mu(x,1)=x.
\end{cases}
$$
Denote by $X_{\mathbb{R}^{{\scriptscriptstyle{+}}}}$ the topological space $X$ endowed with the
conic topology,
i.e. $U \in \mathrm{Op}(X_{\mathbb{R}^{{\scriptscriptstyle{+}}}})$ if it is open for the topology of $X$ and invariant by the action of $\mathbb{R}^{{\scriptscriptstyle{+}}}$. We will denote by $\mathrm{Op}^c(X_{\mathbb{R}^{{\scriptscriptstyle{+}}}})$ the subcategory of
$\mathrm{Op}(X_{\mathbb{R}^{{\scriptscriptstyle{+}}}})$ consisting of relatively weakly quasi-compact open
subsets.\\
Consider the subfamily of $\mathrm{Op}(X_{\mathbb{R}^{{\scriptscriptstyle{+}}}})$ defined by
$\mathcal{T}=\mathrm{Op}^c(X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}})=\{U \in \mathrm{Op}^c(X_{\mathbb{R}^{{\scriptscriptstyle{+}}}}): U\,\, \text{is subanalytic}\}$. The family $\mathcal{T}$ satisfies \eqref{hytau} and the associated site $X_{\mathcal{T}}$ is the conic subanalytic site $X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}}$. In this case the $\mathcal{T}_{loc}$-subsets are the conic subanalytic subsets.\\
Set $\mathrm{Coh}(X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}})=\mathrm{Coh}(\mathcal{T})$. For each $F \in \mod(k_{X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}}})$ there exists a filtrant inductive system $\{F_i\}_{i\in I}$ in $\mathrm{Coh}(X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}})$ such that $F \simeq \lind i \rho_* F_i$.\\
The subcategory of $\mathcal{T}$-flabby sheaves is injective with respect to $\Gamma(U;\bullet)$, $U \in \mathrm{Op}(X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}})$ and $\mathrm{Hom}_{k_{X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}}}}(G,\bullet)$, $G \in \mathrm{Coh}(X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}})$.\\
We call in this case the $\mathcal{T}$-spectrum $ \widetilde{X}_{\mathcal{T}}$ of $X$ the conic subanalytic spectrum of $X$ and denote it by $\widetilde{X}_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}}$. The points of $\widetilde{X}_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}}$ are the ultrafilters $\alpha$ of conic subanalytic subsets of $X$ such that $U \in \alpha$ for some $U \in \mathrm{Op}^c(X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}})$. Then there is an equivalence of categories $\mod(k_{X_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}}}) \simeq \mod(k_{\widetilde{X}_{sa,\mathbb{R}^{{\scriptscriptstyle{+}}}}})$.
\subsection{The o-minimal site}
O-minimal structures are a class of ordered structures which are a model theoretic (logic) generalization of interesting classical structures such as:
\begin{itemize}
\item
${\mathbb R}=({\mathbb R},<, 0,1,+,\cdot )$ - the field of real numbers;
\item
${\mathbb R}_{\rm an}=({\mathbb R},<, 0 ,1,+,\cdot, (f)_{f\in {\rm an}})$ - the field of real numbers expanded by restricted globally analytic functions, i.e, functions which are zero outside a compact box and are given by the restriction of a power series converging on a neighborhood of that box (\cite{dd}).
\end{itemize}
More precisely, an ordered structure $${\mathcal M}=(M, <, (c)_{c\in {\mathcal C}}, (f)_{f\in {\mathcal F}}, (R)_{R\in {\mathcal R}})$$
is o-minimal if every definable subset of $M$ in the structure is already definable in the ordered set $(M,<).$ Given an o-minimal structure, we are interested on the geometry of its (locally) definable spaces.
The model theoretic language allows a uniform development of o-minimal geometry in non-standard o-minimal structures. Concrete non-standard o-minimal structures are:
\begin{itemize}
\item
${\mathbb R}((t^{{\mathbb Q}}))=({\mathbb R}((t^{{\mathbb Q}})), <, 0,1, +, \cdot )$ (or any ordered real closed field);
\item
${\mathbb R}((t^{{\mathbb Q}}))_{{\rm an}}=({\mathbb R}((t^{{\mathbb Q}})), <, 0,1, +, \cdot, (f)_{f\in {\rm an}})$
\end{itemize}
where ${\mathbb R}((t^{{\mathbb Q}}))$ is the field of power series with well ordered supports on which every restricted globally analytic function $f\in {\rm an}$ can be interpreted in a canonical way (\cite{dmm}).
There are many important o-minimal expansions
$${\mathcal M}=({\mathbb R},<, 0,1,+,\cdot, (f)_{f\in {\mathcal F}} )$$
of the ordered field of real numbers. For example ${\mathbb R}_{{\rm an} }$, ${\mathbb R}_{{\rm exp }}, $ ${\mathbb R}_{{\rm an},\, {\rm exp}}, $ ${\mathbb R}_{{\rm an} ^*}, $ $ {\mathbb R}_{{\rm an} ^*,\,{\rm exp}}$ see resp., \cite{dd,w,dm94,ds1,ds2}. For each such we have $2^{\kappa }$ many non-isomorphic non standard o-minimal models for each $\kappa >$ cardinality of the language. There is however a non-standard o-minimal structure
$${\mathcal M}=(\bigcup _{n\in {\mathbb N}} {\mathbb R}((t^{\frac{1}{n}})), <, 0,1, +, \cdot, (f_p)_{p\in {\mathbb R}[[\zeta _1, \ldots ,\zeta _n]]})$$
which does not came from a standard one (\cite{LpRob06,HrPet07}).
O-minimal geometry include the geometry of all those (standard) tame analytic structures but it goes beyond and includes also a generalization of PL-geometry: Any ordered vector space over an ordered division ring
$${\mathcal M}=(M, <, 0, +, (\lambda _d)_{d\in D})$$
is an o-minimal structure and its definable sets are the piecewise linear sets with respect to the ordered division ring $D$ (\cite{VD98}).\\
Let ${\mathcal M}=(M,<, (c)_{\in {\mathcal C}}, (f)_{f\in {\mathcal F}}, (R)_{R\in {\mathcal R}} )$ be an arbitrary o-minimal structure. A locally definable space is a space $(X, \mathcal{O}_X)$ with the sheaf of locally definable functions which is locally isomorphic to open definable subspaces $\{(X_i, \mathcal{O}_{X_i})\}_{i\in I}$ with their sheaves of definable functions. This means that:
\begin{itemize}
\item[(i)]
$X=\bigcup _{i\in I}X_i$;
\item[(ii)]
there are injective maps $\phi _i:X_i\to M^{l_i}$ such that $\phi _i(X_i)$ is a definable subset of $M^{l_i}$;
\item[(iii)]
for all $j$, $\phi _i(X_i\cap X_j)$ is open in $\phi _i(X_i)$ and the transition maps $\phi _{ij}:\phi _i(X_i\cap X_j)\to \phi _j(X_i\cap X_j):x\mapsto \phi _j(\phi _i^{-1}(x))$ are definable homeomorphisms.
\end{itemize}
If $I$ is finite, then $X$ is a definable space. A subset $A$ of $X$ is definable if and only if there is a finite $I_0\subseteq I$ such that $A\subseteq \bigcup _{i\in I_0}X_i$ and for each $i\in I_0$, $\phi _i(A\cap X_i)$ is a definable subset of $\phi _i(X_i)$, and, a subset $B$ of $X$ is locally definable if and only if for each $i$, $B\cap X_i$ is a definable subset of $X$. A map between (locally) definable spaces is (locally) definable if when it is read through the charts it is (locally) definable. Thus we have the category of (locally) definable spaces with (locally) definable continuous maps.
Let $X$ be a locally definable space and consider the subfamily of $\mathrm{Op}(X)$ defined by $\mathcal{T}=\mathrm{Op}(X_{\rm def})=\{U \in \mathrm{Op}(X): U\,\, \text{is definable}\}$. The family $\mathcal{T}$ satisfies \eqref{hytau} and the associated site $X_{\mathcal{T}}$ is the o-minimal site $X_{\rm def}$ of \cite{ejp}. Note also that: (i) the $\mathcal{T}$-subsets of $X$ are exactly the definable subsets of $X$ (by the cell decomposition theorem in \cite{VD98}, see \cite{ejp} Proposition 2.1); (ii) $\mathcal{T}_{loc}= \{U \in \mathrm{Op}(X): U \,\, \text{is locally definable}\}$ and (iii) the $\mathcal{T}_{loc}$-subsets of $X$ are exactly the locally definable subsets of $X$.\\
Set $\mathrm{Coh}(X_{\rm def})=\mathrm{Coh}(\mathcal{T})$. For each $F \in \mod(k_{X_{\rm def}})$ there exists a filtrant inductive system $\{F_i\}_{i\in I}$ in $\mathrm{Coh}(X_{\rm def})$ such that $F \simeq \lind i \rho_* F_i$.\\
The subcategory of $\mathcal{T}$-flabby sheaves (or definably flabby sheaves) is injective with respect to $\Gamma(U;\bullet)$, $U \in \mathrm{Op}(X_{\rm def})$ and $\mathrm{Hom}_{k_{X_{\rm def}}}(G,\bullet)$, $G \in \mathrm{Coh}(X_{\rm def})$.\\
We call in this case the $\mathcal{T}$-spectrum $ \widetilde{X}_{\mathcal{T}}$ of $X$ the definable or o-minimal spectrum of $X$ and denote it by $\widetilde{X}_{\rm def}$. The points of $\widetilde{X}_{\rm def}$ are the ultrafilters $\alpha$ of the Boolean algebra of locally definable subsets of $X$ such that $U \in \alpha$ for some $U \in \mathrm{Op}(X_{\rm def})$. This is a locally weakly quasi-compact space with basis of quasi-compact open subsets given by $\{\widetilde{U}:U\in \mathrm{Op}(X_{\rm def})\}$ and there is an equivalence of categories $\mod(k_{X_{\rm def}}) \simeq \mod(k_{\widetilde{X}_{\rm def}})$. When $X$ is definable, then $\widetilde{X}_{\rm def} =\widetilde{X}$, the o-minimal spectrum of $X$ from \cite{Pi88,ejp}, and there is an equivalence of categories $\mod(k_{X_{\rm def}}) \simeq \mod(k_{\widetilde{X}})$ (\cite{ejp}).\\
Finally observe that since:
\begin{itemize}
\item[(i)] an ordered real closed field $R=(R,<, 0,1, +, \cdot )$ is an o-minimal structure and semialgebraic sets in $R$ are exactly the definable sets in $R$, a locally semialgebraic space $X$ is a locally definable space in $R$ and the semialgebraic site on $X$ is the o-minimal site $X_{\rm def}$ on $X$ in $R$;
\item[(ii)] a real analytic manifold $X$ is a locally definable space in the o-minimal structure ${\mathbb R}_{\rm an}=({\mathbb R},<, 0 ,1,+,\cdot, (f)_{f\in {\rm an}})$ and the open relatively compact subanalytic subsets are open definable subsets in ${\mathbb R}_{\rm an}$, the subanalytic site $X_{sa}$ on $X$ is the o-minimal site $X_{\rm def}$ on $X$ in ${\mathbb R}_{\rm an}$.
\end{itemize}
\addcontentsline{toc}{section}{\textbf{References}}
|
\section{Introduction}
Entanglement is considered a fundamental feature and a resource in
the context of quantum computation and information physics
\cite{NieChu}. However entanglement between quantum systems is
easily destroyed when their unavoidable interaction with the
external environment is taken into account \cite{weiss}. Features
and timing of the disentanglement phenomenon depend strongly on the
particular physical system and environment under investigation.
However, even if disentanglement cannot be avoided, quantum
information and computation tasks can still be performed when the
disentanglement time is much longer than the time needed to run the
quantum task. From this point of view it becomes of fundamental
importance to investigate mechanisms and features of disentanglement
in those systems used in the context of quantum computation and
information.
Continuous variable (CV) quantum systems are some of the possible
candidates for quantum protocols \cite{Braunstein}. We consider here
the problem of two independent harmonic oscillators each one leaking
in a bosonic structured reservoir
\cite{PraBec2004,Xiang2008,SerIllPar2004,DoddHall2004,Dodd2004,Hiro2001,MaOliPa}.
This situation has been considered also by the present author in
\cite{Vasile}. However here, as an extension, we study the case of
oscillators with different frequencies, initially in an entangled
TWB state. Our aim is to investigate the entanglement evolution in
the case of two identical high-T Ohmic reservoir with a
Lorentz-Drude cutoff. In this kind of systems, whose dynamics is
determined by a generalized Hu-Paz-Zhang master equation
\cite{HuPazZhang}, all the initial Gaussian states remain Gaussian
during the evolution. For this reason we are entitled to use the
separability criteria for CV systems introduced by Simon
\cite{Simon}, as a marker of entanglement between the oscillators.
The paper is organized as follows. In Section II we introduce the
physical model and its master equation in the weak coupling limit.
In Section III we review the concept of two-mode Gaussian squeezed
state, or Twin Beam state, and provide the solution to the master
equation. In Section IV we introduce the separability criteria for
bipartite continuous variable systems and the Lorentz-Drude spectral
function. Moreover we study the entanglement dynamics as a function
of the initial amount of entanglement and the relative values of the
free oscillator frequencies and the spectral function cut-off
frequency. Finally in Sec. V we give a brief summary of our work.
\section{The master equation}
Our system consists of two non-interacting quantum harmonic
oscillators with frequencies $\omega_1$ and $\omega_2$. Each
oscillator is coupled to its own bosonic reservoir. The total
Hamiltonian can be written as
\begin{eqnarray}\label{Ham}
H&=&\sum_{j}\hbar\omega_j a^{\dag}_j a_j+\sum_{j,k}\hbar\omega_{jk}
b^{\dag}_{jk} b_{jk}+\sum_{j,k}
\gamma_{jk}(a_j+a_j^{\dag})(b_{jk}+b_{jk}^{\dag}),
\end{eqnarray}
where $j=1,2$, $\omega_{1k}$ and $\omega_{2k}$ are the frequencies
of the reservoirs modes, $a_{j}$ ($a_{j}^{\dag}$) and $b_{jk}$
($b_{jk}^{\dag}$) are the annihilation (creation) operators of the
system and reservoirs harmonic oscillators, respectively, and
$\gamma_{jk}$ describe the coupling between the $j$-th oscillator
and the $k$-th mode of its environment. We assume reservoirs with
the same spectral structure and equally coupled to the oscillators.
The dynamics of the two oscillators can be described through the
following non-Markovian local in time master equation
\cite{HuPazZhang}
\begin{eqnarray}\label{HuPaZang}
\dot{\rho}(t)=\sum_j\frac{1}{i\hbar}[H_j^{0},\rho(t)]-\Delta_j(t)
[X_j,[X_j,\rho(t)]]+\Pi_j(t)[X_j,[P_j,\rho(t)]]+\nonumber\\
\frac{i}{2}r_j(t)[X_j^2,\rho(t)]-i\gamma_j(t)[X_j,\{P_j,\rho(t)\}],
\end{eqnarray}
where $\rho(t)$ is the reduced density matrix, $H_j^{0}$ is the free
Hamiltonian of the $j$-th oscillator, and
$X_j=\frac{1}{\sqrt{2}}(a_j+a_j^{\dag}),\quad
P_j=\frac{i}{\sqrt{2}}(a_j^{\dag}-a_j)$ are the quadrature
operators. The interaction with the reservoirs is taken into account
through the time-dependent coefficients of Eq. \eref{HuPaZang}. The
quantities $\Delta_j(t)$ and $\Pi_j(t)$ describe diffusion
processes, $\gamma_j(t)$ is a damping term and $r_j(t)$ renormalizes
the free oscillator frequencies $\omega_j$. For environments in
thermal equilibrium and in the weak coupling limit in ($r_j(t)$
being negligible), they read
\begin{eqnarray}\label{Coeff1}
\Delta_j(t)&=\alpha^2\!\!\int_0^t\!\! ds \!
\int_0^{+\infty}\!\!\!\!\!\!\!\! \!d\omega
J(\omega)[2N(\omega)+1]\cos(\omega s)\cos(\omega_j s), \nonumber\\
\Pi_j(t)&=\alpha^2\!\!\int_0^t\!\! ds \!
\int_0^{+\infty}\!\!\!\!\!\!\!\! \!d\omega
J(\omega)[2N(\omega)+1]\cos(\omega s)\sin(\omega_j s), \nonumber\\
\gamma_j(t)&=\alpha^2\!\!\int_0^t\!\! ds \!
\int_0^{+\infty}\!\!\!\!\!\!\!\! \!d\omega J(\omega)\sin(\omega
s)\sin(\omega_j s).
\end{eqnarray}
\section{Gaussian states and separability condition}
As initial states for our system we consider the class of two-mode
squeezed states (or twin-beams TWB) obtained applying a two-mode
squeezing operator to the vacuum state of the oscillators
\cite{Braunstein}. They belong to the class of Gaussian states
characterized by a Gaussian characteristic function
$\chi_0(\Lambda)=\exp\bigl\{-\frac{1}{2}\Lambda^{T}\boldsymbol \sigma_0\Lambda\bigl\}$,
where $\boldsymbol \sigma_0$ is the covariance matrix and
$\Lambda^T=(x_1,p_1,x_2,p_2)$. For a TWB state we have
\begin{equation}
\boldsymbol \sigma_0 = \left(
\begin{array}{c c}
\boldsymbol A} \newcommand{\bmB}{\boldsymbol B_{0} & \boldsymbol C} \newcommand{\bmR}{\boldsymbol R_0 \\ \boldsymbol C} \newcommand{\bmR}{\boldsymbol R_0^{T} & \boldsymbol A} \newcommand{\bmB}{\boldsymbol B_0
\end{array}
\right),
\end{equation}
with $\mathbf{A}_0=Diag(a,a)$, $\mathbf{C}_0=Diag(c,-c)$,
$a=\cosh(2r)/2$ and $c=\sinh(2r)/2$. The TWB state is thus
determined only by the \emph{squeezing} parameter $r$ which also
determines the initial amount of entanglement. The bigger is $r$,
the larger is the initial entanglement. Using the characteristic
function solution \cite{Intra2003}, it is trivial to verify that the
characteristic function at time $t$ maintains its Gaussian character
$\chi_t(\Lambda)=\exp\bigl\{-\frac{1}{2}\Lambda^{T}\boldsymbol \sigma_t\Lambda\bigl\}$,
where
\begin{equation}\label{CovMatT2}
\boldsymbol \sigma_{t} = \left(
\begin{array}{c c}
\boldsymbol A} \newcommand{\bmB}{\boldsymbol B_{t}^{(1)} & \boldsymbol C} \newcommand{\bmR}{\boldsymbol R_{t} \\ \boldsymbol C} \newcommand{\bmR}{\boldsymbol R_{t}^{T} & \boldsymbol A} \newcommand{\bmB}{\boldsymbol B_{t}^{(2)}
\end{array}
\right),
\end{equation}
with
\begin{equation}
\mathbf{A}_t^{(i)}=\mathbf{A_0}e^{-\Gamma_i}+\left(
\begin{array}{cc}
\Delta_{\Gamma}^{(i)}+\Delta_{co}^{(i)}-\Pi_{si}^{(i)} & -\Delta_{si}^{(i)}+\Pi_{co}^{(i)} \\
-\Delta_{si}^{(i)}+\Pi_{co}^{(i)} & \Delta_{\Gamma}^{(i)}+\Delta_{co}^{(i)}-\Pi_{si}^{(i)} \\
\end{array}
\right),
\end{equation}
\begin{equation}
\mathbf{C}_t=\left(
\begin{array}{cc}
ce^{-(\Gamma_1+\Gamma_2)}\cos[(\omega_1+\omega_2)t] & ce^{-(\Gamma_1+\Gamma_2)}\sin[(\omega_1+\omega_2)t] \\
-ce^{-(\Gamma_1+\Gamma_2)}\sin[(\omega_1+\omega_2)t] & ce^{-(\Gamma_1+\Gamma_2)}\cos[(\omega_1+\omega_2)t] \\
\end{array}
\right),
\end{equation}
with $\Gamma_i(t)=2\int_0^t\gamma(s)ds$. Moreover, because we are
interested in the short time non-Markovian dynamics only, we defined
$\Delta_{\Gamma}^{(i)}(t)\simeq\int_0^t\Delta_i(s)ds$, and the
following \emph{secular coefficients}
$\Delta_{co}^{(i)}(t)\simeq\int_0^t\Delta_i(s)\cos[2\omega_i(t-s)]ds$,
$\Delta_{si}^{(i)}(t)\simeq\int_0^t\Delta_i(s)\sin[2\omega_i(t-s)]ds$,
$\Pi_{co}^{(i)}(t)\simeq\int_0^t\Pi_i(s)\cos[2\omega_i(t-s)]ds$ and
$\Pi_{co}^{(i)}(t)\simeq\int_0^t\Pi_i(s)\sin[2\omega_i(t-s)]ds$.
Details of the calculations can be found in
\cite{Vasile,MaOliPa,Intra2003}. If a suitable environment spectrum
is provided, all previous coefficients can be evaluated analytically
in the high temperature limit. This is the case of the Lorentz-Drude
Ohmic distribution we introduce in the next section. In equations
(6) and (7) we omitted the explicit time dependence for all the
appearing coefficients.
\section{Entanglement dynamics}
In this section we investigate the entanglement dynamics between the
oscillators using the separability criterion of Simon \cite{Simon}.
This criterion is well-suited in the context of two-mode Gaussian
states because it represents a necessary and sufficient condition
for separability and it depends only on the analytic form of the
covariance matrix. In the case of the time-dependent and non
symmetric covariance matrix \eref{CovMatT2}, the Simon criteria is
equivalent to the following algebraical inequality \cite{PraBec2004}
\begin{eqnarray}
S(t)=\det[\mathbf{A}_t^{(1)}\mathbf{A}_t^{(2)}]
+(\frac{1}{4}-|\det\mathbf{C}_t|)^2-Tr[\mathbf{A}_t^{(1)}\mathbf{J}\mathbf{C}_t\mathbf{J}\mathbf{A}_t^{(2)}
\mathbf{J}\mathbf{C}_t^T\mathbf{J}]\nonumber\\
-\frac{\det[\mathbf{A}_t^{(1)}]+\det[\mathbf{A}_t^{(2)}]}{4}\geq 0
\end{eqnarray} with
\begin{equation}
\mathbf{J}=\left(
\begin{array}{cc}
0 & 1 \\
-1 & 0 \\
\end{array}
\right).
\end{equation}
$S(t)$ is called \emph{separability function}. When $S(t)< 0$ the
state is entangled, otherwise it is separable.
We now consider a particular situation with two reservoirs in
thermal equilibrium at high temperature $T$ ($\hbar\omega_c<<k_B T$)
characterized by a Lorentz-Drude \emph{Ohmic} spectral function
\cite{weiss}
\begin{equation}
J(\omega)=\frac{\omega_c^2}{\pi}\frac{\omega}{\omega^2+\omega_c^2}
\end{equation}
where $\omega_c$ is the cut-off frequency of the distribution.
Hitherto we fix the temperature through the condition
$k_BT/\hbar\omega_c=100$ and the coupling constant to the value
$\alpha=0.1$. Moreover we use a dimensionless time variable
$\tau=\omega_c t$ for the evolution of the separabilty. With these
choices we have only two free parameters $x_i=\omega_c/\omega_i$
($i=1,2$), namely \emph{resonance parameters}, describing the
relative positions of the free oscillator frequencies $\omega_i$
with respect to the spectral cut-off frequency $\omega_c$. In
\cite{Vasile} has been observed the existence of two different
resonance parameters regions characterized by different qualitative
and quantitative behavior of entanglement, $x_1=x_2\geq 1$ and
$x_1=x_2<<1$. Exploiting this result we investigate three different
regimes: 1) $x_1,x_2\geq 1$, 2) $x_1,x_2<<1$ and 3) $x_1<<1$,$
x_2\geq 1$.
A second degree of freedom for our analysis is the choice of the
amount of entanglement in the initial TWB state, given by the value
of squeezing parameter $r$. In principle we do not have any
limitation in the choice of the value of $r$, however in real
situations it is not possible to realize two-mode squeezed states
with $r\geq 2$ \cite{Braunstein}. On the other hand when
$r\simeq0.01$ the amount of entanglement is so small that, at least
in the high temperature limit, disentanglement is a very fast
process almost independent from $x_1$ and $x_2$. For these reasons
we restrict our investigation to the intermediate region of
$0.01\leq r\leq1$.
\begin{figure}[!]
\begin{center}
\includegraphics{SepXgrXgr.eps}
\end{center}
\caption{(Colors online) Separability function $S$ vs scaled time
$\tau=\omega_c t$ with $r=1$ and $x_1=1$. The three curves are
relative to different values of $x_2$. The blue solid line
correspond to $x_2=1$, the red dashed line to $x_2=10$ and the black
dotted line to $x_2=100$.}\label{fig:1}
\end{figure}
We start the analysis considering the case $x_1,x_2\geq1$. We show
the entanglement evolution in Fig. \ref{fig:1} where we fixed the
squeezing parameter to the value $r=1$. In this regime the dynamics
is characterized by entanglement sudden death (ESD) appearing
already in the short time non-Markovian region for any value of $r$.
We also fixed the value of the parameter $x_1=1$ and plotted three
curves correspondent to $x_2=1$ (solid blue), $x_2=10$ (dashed red)
and $x_2=100$ (dotted black). We observe that the disentanglement
time $\tau_{dis}$, defined as $S(\tau_{dis})=0$, only slightly
increases as the adimensional frequency \emph{detuning} $\Delta
x=|x_1-x_2|$ increases.
\begin{figure}[!]
\begin{center}
\includegraphics{SepXpiXpi.eps}
\end{center}
\caption{(Colors online) Separability function $S$ vs scaled time
$\tau=\omega_c t$ with $r=0.04$ and $x_1=0.1$. The three curves are
relative to different values of $x_2$. The blue solid line
correspond to $x_2=0.1$, the red dashed line to $x_2=0.2$ and the
black dotted line to $x_2=0.3$.}\label{fig:2}
\end{figure}
The case $x_1,x_2<<1$ is reported in Fig. \ref{fig:2}. As emphasized
in detail in \cite{Vasile}, this regime is characterized by a more
variegate entanglement dynamics, showing oscillations, ESD and
revivals. We report the situation relative to an initially small
amount of entanglement correspondent to $r=0.04$ and $x_1=0.1$. The
three curves correspond to $x_2=0.1$ (solid blue), $x_2=0.2$ (dashed
red) and $x_2=0.3$ (dotted black). In all cases there are
entanglement oscillations and ESD, while an entanglement revival is
also present when $x_2=0.2$. As in the previous regime of Fig.
\ref{fig:1} we notice that disentanglement time is larger when
$\Delta x$ increases with fixed value of $x_1$. However here the
differences in the disentanglement time are much more evident. Also
oscillations are damped out as the detuning increases.
\begin{figure}[!]
\begin{center}
\includegraphics{SepXgrXpi.eps}
\end{center}
\caption{(Colors online) Separability function $S$ vs scaled time
$\tau=\omega_c t$ with $r=0.1$. Blue solid line correspond to the
choice $x_1=x_2=0.1$, red dashed line to $x_1=0.1,x_2=1$, and black
dotted line to $x_1=x_2=1$}\label{fig:3}
\end{figure}
Finally we look at the regime $x_1<<1$ and $x_2\geq1$ whose results
are shown in Fig. \ref{fig:3} with the choice $r=0.1$. The blue
solid line corresponds to $x_1=x_2=0.1$, the dotted black line
describes the $x_1=x_2=1$ case, and the red dashed line represents
the intermediate case of $x_1=0.1$ and $x_2=1$. Here we can observe
two completely different behaviors of $S(\tau)$ relative to the
cases $x_1=x_2=0.1$ and $x_1=x_2=1$. The former case is
characterized by evident entanglement oscillations and a much longer
survival time of entanglement. Instead the latter case shows a fast
ESD without signs of oscillations. The intermediate case of $x_1\neq
x_2$ possesses intermediate features: weak oscillations and a fast
ESD.
The three parameters regimes analyzed display all the possible
features in the entanglement dynamics for our particular system:
entanglement oscillations, ESD and revivals. Thus when the
oscillators frequencies are detuned no new qualitative behavior
emerges, compared to the equal frequency case \cite{Vasile}. On the
other hand we stress the existence of evident and non-negligible
differences. This does not happen in the regime $x_1,x_2>1$, where
the value of disentanglement time increases only slightly as the
oscillator detuning $\Delta x$ increases.
Quantitative differences characterize the opposite case, i.e., the
regime $x_1,x_2<<1$. Even a slight detuning ($\Delta x=0.1$) can
change drastically the value of the disentanglement time and damp
entanglement oscillations. Again this feature is present in Fig.
\ref{fig:3} where these discrepancies are even more pronounced.
From our investigation we can extract two main results. When the
oscillators have different frequencies ($\omega_1\neq\omega_2$), the
entanglement dynamics shows a behavior which is intermediate between
the cases of frequencies both equal to $\omega_1$ and both equal to
$\omega_2$. Moreover the leading role in the disentanglement process
seems to be played by the oscillator with higher resonance parameter
$x$, which forces the entanglement to evolve in a way similar to the
case of both parameters equal to the larger between the two, as
shown in Fig. \ref{fig:3}.
\section{Summary}
In this paper we extended the results obtained in \cite{Vasile} to
the case in which the two oscillators have different free
frequencies, thus when they interact with different parts of the
reservoirs spectra. We considered the particular case of an Ohmic
distribution in the high temperature limit for different initial TWB
states and different values of the free oscillator frequencies.
Quantitative differences in the entanglement dynamics can be
observed in particular when both resonance parameters are in the
regime $x<<1$, where the evolution changes also for a small values
of $\Delta x$. On the contrary the dynamics is not strongly affected
in the opposite regime even for large detuning. Moreover when the
two frequencies lay in opposite parameters regime, the dynamics is
similar to the case $x>>1$, characterized by lack of oscillations
and fast ESD.
\section*{Acknowledgements} The author wish to thank Dr S
Maniscalco, Dr S Olivares and Dr M G A Paris for fruitful
discussions about the present work. Magnus Ernhrooth foundation is
acknowledged for financial support.
\section*{References}
|
\section{Introduction}
The baby Skyrme model was introduced originally as a planar
analogue of the three-dimensional Skyrme model \cite{skyrme}. Also
its target space is simplified accordingly ($S^2$ instead of the
SU(2) target space of the Skyrme model) such that static field
configurations can be classified according to their winding number
in both cases. Like the Skyrme model, also the baby Skyrme model
consists of a quadratic kinetic term (the O(3) nonlinear sigma
model term), a quartic kinetic term (the analogue of the Skyrme
term), and a potential. For the baby Skyrme model, the inclusion
of a potential term is obligatory for the existence of static
finite energy solutions. The specific form of this potential term
is, however, quite arbitrary, and different potentials have been
studied \cite{old} - \cite{other} (for some recent studies see
e.g. \cite{karliner}, \cite{comp-bS}). In addition to its role as
a toy model for the Skyrme model, the baby Skyrme model has also
found some independent applications in condensed matter physics in
the description of the quantum Hall effect \cite{qhe}.
The energy functional of the baby Skyrme model for static
configurations is
\begin{equation} \label{bS}
E = \frac{1}{2} \int d^2 x \left( \partial_i \vec \phi \cdot
\partial_i \vec \phi + \frac{1}{4} (\epsilon_{ij}\partial_i \vec
\phi \times \partial_j \vec \phi )^2 + \mu^2 V(\vec \phi) \right)
\end{equation}
where $\vec \phi $ is a three-component vector field with unit
modulus $|\vec \phi |=1$, and $V$ is the potential. Further, $\mu$
is a positive constant.
Although the baby Skyrme model, due to its simpler structure, may offer a
way for
a better understanding of the solutions in the $(3+1)$ Skyrme model,
it is still a non-integrable, highly complicated, topologically non-trivial,
non-linear field theory. Many properties of baby skyrmions are established
mainly
by numerical simulations, whereas it is much more difficult to gain
analytical understanding. Therefore,
it is important to know whether there exists the possibility for any further
simplification which would keep us in the class of Skyrme-like models and,
nevertheless, allows
for some exact analytical calculations. For example, one may try to identify
which
features of the solutions of the baby Skyrme model are governed by which
part of the model.
Then, neglecting a particular part of the Lagrangian,
one could investigate a simplified model. In case of static solutions only two
simplifications are possible. One may either suppress both the potential and
the quartic, pure Skyrme term, which leads to the scale invariant O(3) sigma
model with its well-known meromorphic solutions. Or one may opt for an energy
functional which is not scale invariant.
In this case, both the
quartic kinetic term and the potential are mandatory for the
evasion of Derrick's theorem, whereas the quadratic term is optional from
this point of view. The latter fact led Gisiger and Paranjape to
consider the model with the quadratic term omitted \cite{GP}, and
with the so-called ``old'' potential
\begin{equation}
V(\vec \phi) = (\hat n - \vec \phi)^2 = 2(1-\hat n \cdot \vec
\phi) \label{old}
\end{equation}
where $\hat n$ is a constant unit vector which selects the vacuum.
The resulting model has a huge amount of symmetry. Indeed, both
the area-preserving diffeomorphisms on base space \cite{GP} and an
abelian subgroup of the area-preserving difeomorphisms on target
space \cite{ab-dif} are symmetries of the static GP (=Gisiger and
Paranjape) model (as the model consists of a potential term and the pure
Skyrme term only, we will also call it the ``pure baby Skyrme model'').
Further, already for a spherically symmetric
ansatz, GP found an infinite number of exact soliton solutions,
which, together with the huge symmetry, points towards
the integrability of this model in $(2+0)$ dimensions \cite{integr}. These
solitons are, in fact,
compact, that is, they differ from the vacuum only in a finite
compact region of the (two-dimensional) base space \cite{rosenau}. A further
explanation of their compact nature can be found in
\cite{comp-bS}, where it is demonstrated that the GP model can, in
fact, be mapped to the signum-Gordon model (the latter model is
known to support compact solutions, see \cite{comp}). In addition,
GP derived a Bogomolny bound in their model, but the solutions
they found do not saturate this bound.
The main aim of the present work is to understand correlations between
properties of soliton solutions in the pure and the full
baby Skyrme models. It occurs that some important topological as well as
geometrical features of baby skyrmions can be captured already
by solutions of the simplified model.
First, we show that the previously proven integrability
of the static version of the pure baby Skyrme models \cite{integr} can be
extended to the $(2+1)$
space-time case. However, now the models are integrable in the sense of
generalized integrability \cite{alvarez}. In particular,
we construct an infinite set of conserved quantities.
In a next step, we generalize the results of GP by studying pure baby
Skyrme models with different potentials. We find that, depending on the
type of potential we choose, there exist soliton solutions with
qualitatively different behavior. In some cases, there exist
compact solitons, whereas in other cases the solitons have an
exponential tail or are even only localized by some inverse powers
of the radius. Further, we find an interesting coincidence between
the behavior of the solitons in the GP like model (with only a
quartic kinetic term), on the one hand, and the existence of
multi-soliton solutions of the full baby Skyrme model, on the
other hand. Multi-solitons in the full model seem to exist only
for potentials such that the GP model has compacton solutions,
whereas they do not exist if the solitons of the corresponding GP
model have an exponential or power-like tail. Moreover, also
the nucleus respectively shell-like structure is preserved when
we go from the pure to the full baby Skyrme model.
We also comment on the issue of Bogomolny bounds.
Specifically, we point out that there exists a second Bogomolny
bound in the GP model, and all exact soliton solutions saturate
this bound. Further, this bound continues to hold for GP like
models with arbitrary potential. This bound was originally found
in \cite{deI-W}, as a contribution to an improved Bogomolny bound
for the full baby Skyrme model. We just add the information that
in GP like models this bound is, in fact, saturated, and solutions
are, therefore, solutions of the corresponding Bogomolny
equations.
Further, this second Bogomolny bound of the
GP model immediately leads to a second, improved Bogomolny bound for the
full baby Skyrme model \cite{deI-W}. Solutions of the full baby
Skyrme model do not saturate this bound for general potentials.
This second bound is, however, much tighter than the more widely
known Bogomolny bound which stems from the quadratic O(3) sigma
model term alone. In addition, the new bound becomes very tight in
the limits of very weak or very strong potential. We explore the
quality (i.e. tightness) of the bound for several parameter
choices. We remark that an equivalent Bogomolny bound may be derived for
generalized Skyrme models in three dimensions \cite{baryons}. There the
pertinent topological charge is the baryon number, and the limiting model
which saturates the Bogomolny bound gives an accurate description of basic
properties of nuclei, like e.g. masses and sizes.
Finally, we comment on time-dependent, non-topological solutions in
the pure baby Skyrme model.
\section{Generalized integrability in (2+1) dimensions}
The model we are going to investigate in this work is defined by
the following Lagrange density
\begin{equation}
L= \frac{1}{2} \left( \frac{1}{2}
(\partial_{\mu} \vec \phi \times \partial_{\nu} \vec \phi )^2
+ \mu^2 \; V(\phi^3), \right)
\end{equation}
where the potential $V$ is any potential depending only on the third component
of the unit vector field.
In applications to the quantum Hall effect, this restricted model corresponds
to the limit of infinitely strong magnetic field.
After using the standard stereographic
projection
$$
\vec{\phi }=\frac{1}{1+|u|^2} \left( u+\bar{u}, -i ( u-\bar{u}),
1 - |u|^2 \right)
$$
we derive equation of motion
$$
\partial_{\mu} \mathcal{K}^{\mu} - \frac{\mu^2}{4} \bar{u} (1+|u|^2)^2 V'=0,
$$
where ($u_\mu \equiv \partial_\mu u$, etc.)
$$
\mathcal{K}^{\mu}=\frac{K^{\mu}}{(1+|u|^2)^2}, \;\;\;
K^{\mu}=(u_{\nu}\bar{u}^{\nu})\bar{u}^{\mu}
-\bar{u}_{\nu}^2u^{\mu}
$$
and the prime denotes differentiation with respect to $u\bar{u}$. The
construction of conserved quantities is performed by a well understood
procedure originally described in \cite{alvarez}. The conserved currents are
chosen as
\begin{equation}
J_{\mu}=\frac{\delta G}{\delta \bar{u}} \mathcal{K}_{\mu}-
\frac{\delta G}{\delta u} \bar{\mathcal{K}}_{\mu},
\;\;\;\; G=G(u\bar{u}),
\end{equation}
where $G$ is an arbitrary function which depends on the modulus of the complex
field.
Obviously, we have infinitely many such currents as these functions are not
restricted
to any particular form. Then, one may easily calculate the divergence
of the currents and
get
$$
\partial_{\mu}J^{\mu}= \frac{\delta^2 G}{\delta \bar{u}^2}
\mathcal{K}_{\mu}\bar{u}^{\mu}+
\frac{\delta^2 G}{\delta \bar{u} \delta u} \mathcal{K}_{\mu}u^{\mu}
+\frac{\delta G}{\delta \bar{u}} \partial^{\mu} \mathcal{K}_{\mu}
-\frac{\delta^2 G}{\delta u^2} \bar{\mathcal{K}}_{\mu}u^{\mu}-
\frac{\delta G}{\delta u \delta \bar{u}} \bar{\mathcal{K}}_{\mu} \bar{u}^{\mu}
-\frac{\delta G}{\delta u} \partial^{\mu} \bar{\mathcal{K}}_{\mu}.
$$
Using the properties of $\mathcal{K}_{\mu}$
$$\mathcal{K}_{\mu}\bar{u}^{\mu}=\bar{\mathcal{K}}_{\mu}u^{\mu}, \;\;\;
\mathcal{K}_{\mu}u^{\mu}= \bar{\mathcal{K}}_{\mu} \bar{u}^{\mu}$$
we find
$$ \partial_{\mu}J^{\mu}=G' ( u\; \partial^{\mu} \mathcal{K}_{\mu} -
\bar{u}\; \partial^{\mu} \bar{\mathcal{K}}_{\mu})=0,$$
where the last step follows from the equation of motion.
\\
In contrast to the pure baby Skyrme model, its full version is not
integrable,
resulting in a rather complicated dynamics of the baby skyrmions.
It has been shown, however,
that the full baby Skyrme model possesses an integrable submodel defined by
an additional constraint which must be imposed in addition to the
equations of
motion, namely the eikonal equation
$(\partial_{\mu} u)^2=0$. This constraint
is quite restrictive and it is not known wether this integrable
submodel
can lead to interesting solutions. On the contrary,
the pure baby Skyrme model is integrable
(in the
specific meaning of the generalized integrability) without any additional
constraint.
\\
One may notice, that the pure baby Skyrme model explores an integrable
limit of the full baby model as was originally considered by
Aratyn et.al. in the
case of the Skyrme-Faddeev-Niemi model \cite{aratyn2}. Indeed, their limit
$m^2 \rightarrow 0$ leads to a model without the quadratic part of the action.
However, as they omitted the potential term, as well, the limit gave a
theory without static soliton solutions, which probably
was the reason why they
did not investigate that model in detail.
\section{Solitons in the pure baby Skyrme model}
\subsection{One-vacuum potentials}
Let us consider the energy functional for the pure baby Skyrme
model
\begin{equation}
E = \frac{1}{2} \int d^2 x \left( \frac{1}{4}
(\epsilon_{ij}\partial_i \vec \phi \times \partial_j \vec \phi )^2
+ \mu^2 V(\phi^3) \right),
\end{equation}
where we specify the potential as
\begin{equation}
V= 4 \left( \frac{1-\phi^3}{2} \right)^s, \label{old s}
\end{equation}
which for $s=1$ gives the old baby Skyrme potential. Here $s >0$. Such a
family of generalized old baby potentials has been analyzed by
Karliner and Hen in the context of the rotational symmetry breaking in the
full baby Skyrme model \cite{karliner}.
\\
After the stereographic projection we get
$$
E = \int d^2 x \left( 2\frac{(\nabla u \nabla \bar{u})^2-(\nabla
u)^2(\nabla \bar{u})^2}{(1+|u|^2)^4}+ 2 \mu^2\left(
\frac{|u|^2}{1+|u|^2} \right)^s \right),
$$
Further, we assume the symmetric ansatz
\begin{equation}
u=e^{i n\varphi} f(r), \label{ansatz}
\end{equation}
which is compatible with the equations of motion, and introduce a
new function
$$
1-g=\frac{1}{1+f^2},
$$
together with a new variable $y=r^2/2$. For topologically
non-trivial solution one has to impose the following boundary
conditions
$$
f(r=0)=\infty, \;\;\; f(r=R)=0, \;\;\; f'(r=R)=0
$$
where $R$ is finite for compactons or $R=\infty$ for standard
solitons (then, the third expression does not lead to a new
condition). In terms of the new function $g$ we get
$$
g(y=0)=1, \;\;\; g(y=Y)=0, \;\;\; g'(y=Y)=0,
$$
Now, the energy takes the simply form
$$
E = 2\pi \int d y \left( 2n^2 g_y^2+ 2 \mu^2 g^s \right).
$$
The finite energy solutions must obey the pertinent first order
Bogomolny equation which is easily obtained from the second order equation
of motion,
$$
n^2 g_y^2 = \mu^2 g^s.
$$
This simple equation supports topological solutions (i.e.,
solutions with the previously specified boundary conditions), which
may be divided into three classes.
\\
{\it (i)} $s \in (0,2)$ {\it - Compactons}
\\
The pertinent solutions are compactons
\begin{equation}
g(y)=\left\{
\begin{array}{cc}
\left( 1 - \frac{y}{y_0} \right)^{\frac{2}{2-s}} & 0 \leq y \leq
y_0
\\
&
\\
0 & y \geq y_0,
\end{array}
\right.
\end{equation}
where the boundary of the compacton is located at
$$
y_0=\frac{2n}{\mu(2-s)}.
$$
The approach to the vacuum becomes more and more rapid as we tend to
$s=2$. Such compact solutions represent rotationally symmetric
multi-solitons configuration which can be understood as a
collection of $n$ $Q=1$ solitons located on top of each other.
However, due to the compacton nature of the solutions one may
easily construct a non-rotational solution with $Q=n$ by a
collection of some solutions with lower topological charges,
provided that they are sufficiently separated and the sum of
charges is equal to $n$.
\\
{\it (ii)} $s=2$ {\it - Exponentially localized solitons}
\\
Now, we get a soliton with the standard exponential tails
\begin{equation}
g(y)=e^{-\frac{\mu y}{n}}.
\end{equation}
Again, higher charge solitons may be interpreted as several $Q=1$ solitons
sitting on top of each other. Further, it is no longer obvious how to
construct non-symmetric solutions and solutions corresponding to
well-separated solitons. We remark that this case has already been
investigated in \cite{Tchr 3}.
\\
{\it (iii)} $s>2$ {\it - Power-like localized solitons}
\\
The third class of solutions is formed by solitons with
power-like tails
\begin{equation}
g(y)= \left( \frac{y_0}{y_0+y}\right)^{\frac{2}{s-2}},
\end{equation}
where
$$
y_0=\frac{2n}{\mu(s-2)}.
$$
Three remarks are appropriate.
\\
Firstly, as one could expect, in the case of solutions of a first order
Bogomolny equation, for all models the total energy is
proportional to the topological charge
$$E= \frac{16\pi}{s+2} \mu |Q|.$$
Secondly, from a geometrical point of view all these solutions share a
common feature. Namely, they are of nucleus type with their energy density
concentrated around the origin. This resembles the baby skyrmions in the
old baby
model, where solitons also possess such a nucleus shape.
\\
Finally, there is a very intriguing coincidence between the compactness
of soliton solutions of the pure model and the existence, for some values of
the parameters, of multi baby-skyrmions in the full model. In fact, for
potentials with $s \in (0,2)$ solitons in the pure model are of the compact
type.
On the other hand, as is known from previous numerical works, full
baby Skyrme models with the same potentials support multi
soliton solutions. Moreover, for baby Skyrme models with potentials with
$s \geq 2$, no
multi-soliton configurations have been found.
It corresponds, on the pure model side,
to non-compact topological solutions
(exponentially or power-like localized).
It seems as if the instability of the multi-skyrmions in the full baby models
has its origin
in the non-compactness of the solitons in the pure models. Equivalently, one
may conjecture that compactons in the pure models are seeds for
multi-skyrmions in the full theory.
\subsection{Two-vacua potentials}
\subsubsection{Generalized new baby potential}
Another type of potentials we are going to analyze is given by the
expression
\begin{equation}
V= 4 \left( \frac{1-(\phi^3)^2}{4} \right)^s,
\end{equation}
with $s >0$. In the limit $s=1$ one can recognize the new baby
potential, a well-known potential with two vacua $\phi^3=\pm 1$. The
pertinent first order equation of motion reads
$$
n^2 g_y^2 = \mu^2 (g(1-g))^s.
$$
For this potential we find two classes of solutions.
\\
{\it (i)} $s \in (0,2)$ {\it - Compactons}
\\
The general solution describes a compact ring. In fact, both vacua
are of the compacton type i.e., a solution tends to them in a
power like manner. More precisely, the solution (not in explicit
form) reads
\begin{equation}
\frac{2}{2-s} _2F_1[1-\frac{s}{2},
\frac{s}{2},2-\frac{s}{2},g]=\frac{\mu(y+y_0)}{n},
\end{equation}
where $_2F_1$ is a hypergeometric function. An example of an
explicit solution can be found for $s=1$
\begin{equation}
g(y)=\left\{
\begin{array}{cc}
\left(\sin \frac{2\mu (y-y_0)}{n} \right)^2 & 0 \leq y \leq y_0 \\
&
\\
0 & y \geq y_0
\end{array}
\right.
\end{equation}
and
$$
y_0=\frac{n \pi}{4 \mu }.
$$
The corresponding energy density is not concentrated at the origin
but has its maximum at $y_{max}=n\pi/8\mu $, forming a ring-like
structure. This solution can be easily extended to a family of
ring-like solutions with the maximum of the energy density located at any
$y \geq y_{max}$. Again, all solutions lead to a linear
energy-charge relation.
\\
Let us underline that the fact that solutions are of the ring type again
resembles the
geometric properties of soliton solutions in the full baby Skyrme model.
Indeed, the baby skyrmions in the new baby Skyrme model have this property.
\\
{\it (ii)} $s \geq 2$ {\it - No solitons}
\\
It is easy to check that there are no topological solitons for
such a value of the parameter. The reason for this is that both
vacua are approached exponentially $(s=2)$ or polynomially
($s>2$). It is of course impossible to join both vacua in such a
way on the semi-infinite segment $y\in [0,\infty)$.
\subsubsection{Generalized new baby potential - non-symmetric case}
\noindent The last family of potentials of this type we want to
analyze has the form
\begin{equation}
V= 4 \left( \frac{1+\phi^3}{2} \right)^a \left( \frac{1-\phi^3}{2}
\right)^b,
\end{equation}
with $a,b > 0$. The properties of the solutions are as follows.
\\
{\it (i) $a \in (0,2)$ and $b \in (0,2)$} {\it - Compactons}
\\
We get compactons as the approach to both vacua is in a power-like
manner.
\\
{\it (ii) $a =2$ and $b \in (0,2)$ or $a \in (0,2)$ and $b =2$}
{\it - Exponentially localized solitons}
\\
We get standard soliton
solutions with exponential tails. As an example, we get
\begin{equation}
g(y)= \frac{2}{1+\cosh \frac{\mu y}{n}}
\end{equation}
for the first case ($a=2$, $b=1$), and
\begin{equation}
g(y)= \left( \tanh \frac{\mu y}{2 n } \right)^2,
\end{equation}
for second case ($a=1$, $b=2$), respectively. Notice that the
second solution possesses the reversed boundary conditions, i.e.,
$g(0)=0$, $g(\infty)=1$.
\\
{\it (iii) $a >2$ and $b \in (0,2)$ or $a \in (0,2)$ and $b > 2$}
{\it - Power-like localized solitons}
\\
We get solitons with power-like tails. As an example one may
consider the choice of parameters $a=4, b=1$. Then, the pertinent
solution reads
\begin{equation}
\sqrt{\frac{1}{g^2} - \frac{1}{g}} + \mbox{ar cosh}
\frac{1}{\sqrt{g}}=y.
\end{equation}
All presented solutions can be extended to shell-like
configurations. As before one gets that the energy is proportional
to the topological charge.
\\
{\it (iii) $a \geq 2$ and $b\geq 2$} {\it - No solitons}
\subsection{Vortex potentials}
A different family of potentials also considered for the baby
Skyrme model is given by the following expression
\begin{equation}
V= 4 \left( (\phi^3)^2 \right)^s \left( \frac{1-\phi^3}{2} \right),
\end{equation}
where $s >0$.
Here the vacuum manifold is completely different from the cases considered
up to now.
It still consists of two disconnected components, but now only one of
the two components is a point (the north pole $\phi^3 =1$ of the target space
$S^2$), whereas the other component of the vacuum is a circle $S^1$: $\phi^3
=0$, $(\phi^1)^2 + (\phi^2)^2 =1$. In our parametrization, this vacuum
corresponds to $g=1/2$.
The presence of this new vacuum allows for field configurations which approach
it in the limit $|\vec x|\to \infty $. Such fields
may be divided into
disjoint topological classes as
$$\vec{n}_{\infty}: \mathbb{R}^2_{r\rightarrow \infty} \simeq \mathbb{S}^1
\longrightarrow U(1)\simeq \mathbb{S}^1.$$
The relevant topological index is the winding number $Q \in
\pi_1(S^1)$, and the corresponding field configurations are
vortices. The important question is, of course, whether such vortex
configurations are possible solutions of the underlying field theory. In this
respect, the full and the pure baby Skyrme model are fundamentally different.
The reason is that the quadratic O(3) sigma model term, which is present only
in the full model, produces
infinite energy for vortex configurations. Indeed, in
polar coordinates the static O(3) term essentially is the sum of a radial
gradient squared and an angular gradient squared. In the limit, the angular
gradient behaves like $r^{-1}\hat e_\varphi \partial_\varphi \exp (in\varphi)
=in r^{-1}\hat e_\varphi$ where $n$ is the vortex number, so for nonzero $n$
the energy density goes like $r^{-2}$ and produces a logarithmic divergency.
As a consequence, there are no vortices with finite energy in the full baby
Skyrme model. In the pure model, on the other hand, only the quartic term is
present in addition to the potential. Further, due to its antisymmetry, {\em
both} gradients have to appear in this term (in fact, both
quadratically). Therefore, the slowly decaying angular gradient is always
multiplied by a rapidly decaying radial gradient, such that the resulting
energy density decays sufficiently fast to have finite energy. In short,
finite energy vortices are not excluded in the pure model, and they turn out
to exist, as we shall see in a moment.
\\
{\it (i)} $s \in (0,1)$ {\it - Vortex compactons}
\\
We get compact vortices. As an example with explicit solutions let
us consider the model with $s=1/2$. Then,
\begin{equation}
g(y)=\left\{
\begin{array}{cc}
\frac{1}{2} \left( \cosh \frac{4\sqrt{2}(y-y_0)}{n} \right)^2 & y
\leq y_0 \\
&
\\
\frac{1}{2} & y > y_0,
\end{array} \right.
\end{equation}
or
\begin{equation}
g(y)=\left\{
\begin{array}{cc}
\frac{1}{2} \left( \sin
\frac{4\sqrt{2}y}{n} \right)^2 & y \leq \tilde{y}_0
\\
&
\\
\frac{1}{2} & y \geq \tilde{y}_0,
\end{array} \right.
\end{equation}
where $$ y_0=\frac{n}{4\sqrt{2}} \mbox{arc cosh} \sqrt{2}, \;\;\;
\tilde{y}_0=\frac{\pi n}{8\sqrt{2}}.$$ The first solution
represents vortex $(g \in (1,1/2))$ whereas the second anti-vortex
$(g \in (0,1/2))$. Quite interestingly, because of the compactness
of the solutions one may construct a configuration carrying a
non-trivial value of the topological charge by gluing in a proper way
together both profile functions (i.e., both vortices). Namely,
\begin{equation}
g(y)=\left\{
\begin{array}{cc}
\frac{1}{2} \left( \cosh \frac{4\sqrt{2}(y-y_<)}{n} \right)^2 & y
\in [0,y_<] \\
&
\\
\frac{1}{2} & y \in [y_<, y_<+y_0]
\\
&
\\
\frac{1}{2} \left( \sin \frac{4\sqrt{2}(y-(y_>+y_0))}{n} \right)^2
& y
\in [y_<+y_0, y_>+y_0] \\
0 & y > y_>+y_0,
\end{array} \right.
\end{equation}
where $$ y_<=\frac{n}{4\sqrt{2}} \mbox{arc cosh} \sqrt{2} $$ $$
y_>=\frac{\pi n}{8\sqrt{2}} (1+ \frac{2}{\pi} \mbox{arc cosh}
\sqrt{2}).
$$ Thus, two vortices with the same winding numbers $n$ form a
baby skyrmion with the topological charge $Q=n$. Therefore, for such
potentials a baby skyrmion seems to dissolve into a pair of $U(1)$
vortices, which now are the true constituent topological objects.
\\
{\it (ii)} $s \geq 1$ {\it - Exponential or polynomial vortices}
\\
A pertinent example is given by $s=1$ case. Then,
\begin{equation}
g(y) = \frac{1}{2} \left(\coth \frac{2(y+y_0)}{n} \right)^2
\end{equation}
or
\begin{equation}
g(y) = \frac{1}{2} \left(\tanh \frac{2y}{n} \right)^2,
\end{equation}
where $y_0$ obeys $$\coth \frac{2y_0}{n} = \sqrt{2}. $$ Of course, now these
vortex solutions cannot be glued together to
form a stable configuration with non-zero
topological charge.
\section{Bogomolny bounds}
In this section we prove that the solutions of the previous section
are indeed
of the Bogomolny type, without performing any dimensional
reduction. Moreover, we derive the pertinent Bogomolny bound and
show that our solutions saturate it. In fact, in an investigation
of the full baby Skyrme model this bound has been originally
reported in \cite{deI-W}.
Let us begin by reviewing briefly the original Bogomolny bound found by GP.
They studied the energy functional
$$
E = \frac{1}{2} \int d^2 x \left( \frac{1}{4}
(\epsilon_{ij}\partial_i \vec \phi \times \partial_j \vec \phi )^2
+ \mu^2 (\hat n - \vec \phi )^2 \right)
$$
where $\hat n$ is a constant unit vector selecting the vacuum.
Further, they derived the following Bogomolny bound
$$
E = \frac{1}{2} \int d^2 x \left( \left( \frac{1}{2} \epsilon_{ij}
\partial_i \vec
\phi \times \partial_j \vec \phi \pm \mu (\hat n - \vec \phi )
\right)^2 \mp \mu (\hat n - \vec \phi )\cdot \epsilon_{ij}
\partial_i \vec \phi \times \partial_j \vec \phi \right)
$$
$$
=\frac{1}{2} \int d^2 x \left( \left( \frac{1}{2} \epsilon_{ij}
\partial_i \vec
\phi \times \partial_j \vec \phi \pm \mu (\hat n - \vec \phi )
\right)^2 \pm \mu \epsilon_{ij} \vec \phi \cdot (\partial_i \vec
\phi \times \partial_j \vec \phi ) \right)
$$
where a total divergence term which gives zero upon integration
has been omitted in the last expression. The energy, therefore,
obeys a Bogomolny bound
$$
E \ge E_B \equiv 4\pi \mu |Q|
$$
were the integer-valued topological charge $Q$ is
$$
Q = \frac{1}{8\pi} \int d^2 x \epsilon_{ij} \vec \phi \cdot
(\partial_i \vec \phi \times \partial_j \vec \phi ) .
$$
The solutions of GP (see also section 3.1), however, do not
saturate this Bogomolny bound. Instead, they satisfy
$$
E_{\rm sol} = \frac{4}{3} E_B
$$
so they are still proportional to the topological charge, just the
coefficient in front is slightly bigger than the coefficient for
the Bogomolny bound. Firstly, let us remark that nontrivial
solutions saturating this bound cannot exist. Indeed, the
corresponding Bogomolny equation is
$$
\frac{1}{2} \epsilon_{ij} \partial_i \vec \phi \times \partial_j
\vec \phi \pm \mu (\hat n - \vec \phi ) =0
$$
and multiplying this by $\times \vec \phi$ one easily concludes
$$
\hat n \times \vec \phi =0
$$
and any solution $\vec \phi$ must be proportional to the trivial
constant vacuum vector $\hat n$.
\\
Secondly, the fact that the explicit solutions still are linear in
the topological charge, just with a slightly bigger prefactor,
already indicates that there might exist another, tighter
Bogomolny bound which is saturated by the energies of
the solutions.
We will demonstrate now that this is indeed the case (the bound itself has
already been found in \cite{deI-W}, as a contribution to an improved bound for
the full baby Skyrme model). For this
purpose, we observe first that the vectorial expression whose
square gives the quartic kinetic term necessarily is proportional
to the vector $\vec \phi$ itself and may therefore be written like
$$
\epsilon_{ij} \partial_i \vec \phi \times \partial_j \vec \phi =
\vec \phi \left( \vec \phi \cdot \epsilon_{ij} \partial_i \vec
\phi \times \partial_j \vec \phi \right) .
$$
Further, we choose the vacuum vector $\hat n = (0,0,1)$ such that the
potential $V=2(1-\phi^3)$ depends only on $\phi^3$. We will, in fact, allow
for general potentials $V=V(\phi^3)$ at the moment, because the Bogomolny
bound can be easily found for them.
Using this, we rewrite the energy functional like
follows,
$$
E = \frac{1}{2} \int d^2 x \left( \frac{1}{4} (\epsilon_{ij}\vec
\phi \cdot (\partial_i \vec \phi \times \partial_j \vec \phi ))^2
+ \mu^2 V( \phi^3 ) \right)
$$
$$
= \frac{1}{2} \int d^2 x \left( \left( \frac{1}{2}
\epsilon_{ij}\vec \phi \cdot (\partial_i \vec \phi \times
\partial_j \vec \phi ) \pm \, \mu \sqrt{V
} \right)^2 \right.
$$
\begin{equation} \label{dIW-bogo}
\left. \mp \, \mu \sqrt{V }
\, \, \, \epsilon_{ij}\vec \phi \cdot (\partial_i \vec \phi \times
\partial_j \vec \phi ) \right) .
\end{equation}
It remains to demonstrate that the last term above is, indeed, a Bogomolny
energy which is bounded by the topological charge.
For this purpose we use the complex field $u$ together with its modulus and
phase $u=fe^{i\Sigma}\equiv \sqrt{F}e^{i\Sigma}$ instead of the unit vector
field $\vec \phi$.
In these terms the topological charge reads
\begin{equation} \label{top-charge}
Q[u] = \frac{-1}{2\pi i} \int d^2 x \frac{\epsilon_{ij} u_i \bar
u_j}{(1+u\bar u)^2} = \frac{1}{2\pi} \int d^2 x
\epsilon_{ij}\frac{ F_i \Sigma_j}{(1+F)^2}.
\end{equation}
For the energy we find from (\ref{dIW-bogo})
$$
E \ge \mp 4\pi \mu \int d^2 x \left[\frac{1}{2\pi} \sqrt{V(F)}
\epsilon_{ij}\frac{F_i \Sigma_j}{(1+F)^2} \right] \equiv 4\pi \mu C_1 |Q|
$$
where the sign has to be chosen accordingly (upper sign for $Q>0$). Further,
$C_1$ is a constant which depends on $V$, and the last equality still
has to be proven. If we replace $V$ by one, then the result is obvious,
because the integrand is just the integrand of the topological charge
(\ref{top-charge}). Equivalently, this expression is just the pullback of the
area form on the target space $S^2$, normalized to one. The base space
integral gives the result $Q$ because the base space sphere is covered $Q$
times while the target space sphere is covered once. But this last result
continues to hold with the factor $\sqrt{V(F)}$ present, up to a constant
$C_1$. Indeed, we just have to find a new target space coordinate $\tilde F$
instead of $F$ such that
\begin{equation} \label{FFtilde}
\frac{\sqrt{V(F)}dF}{(1+F)^2} = C_1 \frac{d\tilde F}{(1+\tilde F)^2}.
\end{equation}
The constant $C_1$, and a second coordinate $C_2$ which is provided by the
integration of Eq. (\ref{FFtilde}), are needed to impose the two conditions
\begin{equation} \label{FFbound}
\tilde F(F=0) =0 \, , \quad \tilde F(F=\infty) =\infty
\end{equation}
which must hold if $\tilde F$ is a good coordinate on the target space $S^2$.
Obviously, $C_1$ depends on the potential. Specifically, for the ``old''
potential
$$
V=2(1-\phi^3) = 4\frac{F}{1+F}
$$
we find
$$
2\sqrt{\frac{F}{1+F}}\frac{dF}{(1+F)^2} = C_1 \frac{d\tilde F}{(1+\tilde F)^2}
$$
$$
\frac{4}{3}\left( \frac{F}{1+F} \right)^\frac{3}{2} = -\frac{C_1}{1+\tilde F}
+ C_2
$$
and, finally, from the boundary conditions (\ref{FFbound}), $C_1 = C_2
=(4/3)$.
The corresponding Bogomolny equations are
$$
\frac{1}{2} \epsilon_{ij}\vec \phi \cdot (\partial_i \vec \phi
\times \partial_j \vec \phi ) \pm \mu \sqrt{V(\phi^3)} =0
$$
or, in terms of $u=fe^{i\Sigma}$, $F=f^2$,
$$
2\epsilon_{ij}[\partial_i (1+F)^{-1} ]\partial_j \Sigma \pm \mu
\sqrt{V(F)} =0.
$$
For $F=F(r)$, $\Sigma = n\varphi$ this just gives the two possible square
roots of the first order equations found in Section 3, as may be checked
easily.
\subsection{Bogomolny bounds in the full baby Skyrme model}
Here we want to discuss the fact that the above Bogomolny bound may be
used immediately to derive a tighter bound also for the full baby
Skyrme model \cite{deI-W}.
The important point here is that the quadratic, O(3)
sigma model part of the full baby Skyrme model, which is absent in
the GP model, has its own, independent Bogomolny bound
$$
E_{O(3)} \ge 4\pi |Q|
$$
in terms of the same winding number. For the full baby Skyrme
model (\ref{bS}) we get, therefore, in the case of the ``old''
potential, the improved Bogomolny bound
$$
E_{bS} = E_{O(3)} + E_{GP} \ge 4\pi |Q|(1+\frac{4}{3}\mu ).
$$
In the full baby Skyrme model, however, solutions no longer
saturate this bound, because they would have to obey the field
equations both of the O(3) model and of the GP model in order to
do so.
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=270,width=0.9 \textwidth]{Energy2.ps}
\caption{Comparison of the energy and Bogomolny bound for $Q=1$
baby skyrmion}\label{rys1}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=270,width=0.9 \textwidth]{energy_all.ps}
\caption{Comparison of the energy and Bogomolny bound for
$Q=2..12$ baby skyrmions}\label{rys2}
\end{center}
\end{figure}
Nevertheless, we expect that the bound should be very tight for
very small and very large $\mu$. The reason for this is as
follows. A simple scaling argument shows that the potential and
the quartic pure Skyrme term always contribute the same amount of
energy to a solution. This implies that in the limit of very small
$\mu$, the energy approaches the pure quadratic O(3) sigma model
energy, whereas in the opposite limit of large $\mu$, the energy
approaches the energy of the GP model. But both the pure O(3)
model energy and the pure GP model energy saturate their
respective Bogomolny bounds. A simple numerical calculation of the
energies confirms this expectation. We make the rotationally
symmetric ansatz $u=f(r)e^{im\varphi}$ and determine $f(r)$ and
the corresponding energy by a simple shooting algorithm. The
deviation between energy and bound reaches its maximum of about
$10\%$ for values of $\mu \sim 1$, and diminishes both for small
and large $\mu$, as expected, see Fig. 1. The case $\mu = \sqrt{0.1}$ and
$Q=1$ corresponds to the numerical solution originally reported in
the first reference of \cite{old}, $E=1.564 \times 4\pi$, in which
case the new bound gives $1.422 \times 4\pi$ which is definitely
better than the old pure O(3) sigma model bound $4\pi$ ($10 \%$
off instead of more than $50 \%$ off).
\\
In Fig. 2 we show the corresponding
results for higher topological charges $Q=2,...,12$.
Observe that although the symmetric ansatz does not necessarily
lead to the energy minima of the full baby Skyrme model, the bound
is again very tight in the large and small $\mu$ limits.
\vspace*{0.2cm}
The generalization to other potentials is obvious. The bound is
always of the type
$$
E_{bS} = E_{O(3)} + E_{GP} \ge 4\pi |Q|(1+C_1 \mu )
$$
where the numerical factor $C_1$ depends on the specific
potential but can be calculated exactly. We remark that the limit of large
$\mu$ has already been studied in \cite{Tchr 3}, although for a different
potential (the square of the ``old'' potential).
\section{Non-existence of non-topological $Q$-balls}
Here we show that there are no topologically trivial $Q$-balls in
the pure baby model with the old potential.
\\
We assume the symmetric ansatz for time-dependent solutions
\begin{equation}
u=e^{i (\omega t + n\phi)} f(r). \label{ansatz t}
\end{equation}
Then, the equation of motion takes the form
$$
f\left[ -\frac{1}{r} \partial_r \left( \frac{rf'f}{(1+f^2)^2}
\left[ \frac{n^2}{r^2}-\omega^2\right] \right) +
\frac{1}{2}\right]=0.
$$
Thus, as always, a solution is given by the vacuum value $f=0$ or
by a solution of
$$
-\frac{1}{r} \partial_r \left( \frac{rf'f}{(1+f^2)^2} \left[
\frac{n^2}{r^2}-\omega^2\right] \right) + \frac{1}{2}=0.
$$
Now we introduce the $g$ function
$$
1-g=\frac{1}{1+f^2}
$$
and get the equation
\begin{equation}
-\frac{1}{r} \partial_r \left( rg' \left[
\frac{n^2}{r^2}-\omega^2\right] \right) + 1=0. \label{eq nontop}
\end{equation}
We are looking for non-topological solutions of the $Q$-ball type.
Hence, the relevant boundary conditions are: for a compacton
starting at $r=0$
$$
g(r=0)=0, \;\;\; g(r=R)=0, \;\;\; g'(r=R)=0
$$
or, for a shell-like compacton
$$
g(r=R_i)=0, \;\;\; g'(r=R_i)=0, \;\;\; i=1,2, \;\;\; R_1 < R_2.
$$
The general solution of the e.o.m. is
$$
g'(r)=\frac{1}{\frac{n^2}{r^2} -\omega^2} \left(\frac{r}{2} + A
\right)
$$
and
$$
g(r)=- \frac{r^2}{4\omega^2} - \frac{n^2}{4\omega^4} \ln
|n^2-\omega^2r^2|-\frac{A}{2\omega^2} \ln |n^2-\omega^2r^2|+B.
$$
First of all we may exclude shell-like compactons, as from
equation for $g'$ we get that
$$
g'(r=R)=0 \;\; \Rightarrow \;\; R^2=-2A.
$$
Hence, there is only one positive $R$. Notice, $A$ must be
negative. Further $g(0)=0$ gives
$$
B=\frac{n^2}{4\omega^4} \ln |n^2 |+\frac{A}{2\omega^2} \ln |n^2|
$$
and
$$
g(r)=- \frac{r^2}{4\omega^2} - \frac{n^2}{4\omega^4} \ln
\left|1-\frac{\omega^2r^2}{n^2}\right|-\frac{A}{2\omega^2} \ln
\left|1-\frac{\omega^2r^2}{n^2}\right| .
$$
The last constant $A$ can be found from $g(R)=0$
$$
0=- \frac{R^2}{4\omega^2} - \frac{n^2}{4\omega^4} \ln
\left|1-\frac{\omega^2R^2}{n^2}\right|-\frac{A}{2\omega^2} \ln
\left|1-\frac{\omega^2R^2}{n^2}\right| .
$$
But $R^2=-2A$, thus (after multiplication by $4\omega^2/n^2$
$$
0= \frac{2A\omega^2}{n^2} - \ln
\left|1+\frac{2A\omega^2}{n^2}\right|-\frac{2A\omega^2}{n^2} \ln
\left|1+\frac{2A\omega^2}{n^2}\right|
$$
or
$$
0=p - (1+p) \ln \left|1+p\right|,
$$
where $p=2A\omega^2/n^2$. This equation has two solutions $p=0$
and $p=p_0 < -1$. But in order to have a non-singular solution
$g$, the radial coordinate cannot be larger than $r<n/w$. So, the
size of the compacton must be also smaller than this critical
value $R<n/w$. As we know $R=\sqrt{-2A}$, where $2A=pn^2/w^2$ if
we use the definition of $p$. Then, $R=\sqrt{-p}n/w < n/w$. Thus
finally, $0<\sqrt{-p}<1$ i.e., $ p \in (-1,0)$, which is in
contradiction with our previous observation that $p_0<-1$.
\vspace*{0.2cm}
The generalization to other potentials is unfortunately difficult.
We are left with a quite complicated
second order nonlinear ODE for $g$. Fortunately, one may prove that for any
old baby potential (\ref{old s}) there are no non-spinning ($n=0$),
non-topological
$Q$-balls. Instead of equation (\ref{eq nontop}), we get
$$
\frac{1}{r} \partial_r \left( rg' \right) + g^{s-1}=0,
$$
where the radial coordinate has been properly re-scaled. Assuming a series
expansion in the vicinity of the origin
$$ g(r)=Ar^a+...$$
we get
$$a=\frac{2}{2-s}, \;\;\;\; a^2+A^{s-2}=0. $$
The second formula leads to a complex or negative value of $A$, which is
unacceptable as
$g$ must be positive. A similar effect occurs if we try an expansion at
any finite $R$.
Then we have $$a(a-1)+A^{s-2}=0. $$
We can get an acceptable $A$ if $0<a<1$. But this gives $s<0$,
which is excluded for obvious reasons.
Hence, no $n=0$, $Q$-balls are possible.
\vspace*{0.2cm}
The non-existence of non-topological $Q$-balls in the pure baby
Skyrme models is a rather new feature which has no analog in the full
baby Skyrme model. On the contrary, in the full theory one expects to find
non-topological
time dependent solutions, compact or non-compact, depending on the
chosen form of the potential \cite{comp-bS}. Moreover, this result is also
unexpected, as topological $Q$-balls have been found by Gisiger and
Paranjape \cite{GP}, with properties similar to $Q$-baby skyrmions.
Apparently, the restricted model leads to non-trivial solutions which
reflect properties of the full model only for field configurations with
a non-trivial topological charge.
\section{Conclusions}
In this paper the pure baby Skyrme model has been analyzed. It was our main
aim to investigate whether solutions of this simplified model may tell us
something about the baby skyrmions, i.e., solitons in the full baby Skyrme
theory. Such a correspondence between solutions of the two models is by no
means
obvious. Due to the huge amount of symmetries in the restricted model,
one could rather
expect a completely different behavior of solutions.
We found, nevertheless, that
many properties of baby skyrmions are sufficiently well described by solutions
of the pure model. Moreover, as the restricted model is integrable
in the $(2+0)$ dimensional case (in $(2+1)$ dimensions the model
has the property of generalized integrability),
we were able to perform all calculations analytically.
\\
The main findings are as follows:
\begin{enumerate}
\item[$\bullet$] Topological (charge) as well as geometrical
(nucleus/shell) features of baby skyr\-mions are captured already by
the soliton solutions of the pure model
\item[$\bullet$] Energies of baby skyrmions are reasonably approximated
by the energies of the solitons of the pure model. The approximation
becomes better and better while $\mu \rightarrow \infty$
\item[$\bullet$] There is a coincidence between the existence
of compact solitons in the pure baby model and the existence
(for some values of the
parameters) of multi-baby-skyrmions in the full baby Skyrme model
\item[$\bullet$] There are no topologically trivial $Q$-ball configurations
\end{enumerate}
The first three results show that the static properties of baby skyrmions
may be qualitatively and quantitatively described by solutions of the
pure model. They indicate that the topological and geometrical properties
are governed by the Skyrme term and the potential, while the quadratic part of
the full baby Skyrme model only quantitatively modifies them.
Especially, the third result can be potentially important
as it gives a chance for an analytical treatment of the issue of the stability
of multi skyrmions. So the purely quartic
model, despite its high symmetry and probable integrability,
shares some properties with the full model, and may
probably be used to approximate the full model in some sense (for approximation
of solitons by compactons in the case of a scalar field theory see \cite{lis}).
\\
The fact that no topologically trivial $Q$-balls exist in the restricted model
leads to a different conclusion. $Q$-balls in the full baby Skyrme model are not
approximated by any solutions of the pure baby model but rather by solutions
of the $CP^1$ model, i.e., another simplified version of the baby Skyrme theory,
which consists of the quadratic part and potential. Of course, such a model
does not allow for static solutions and, therefore, is useless in the
context of the static baby skyrmions.
\\
The identification of simplified models relevant for the study of properties
of topological solitons (the pure baby Skyrme model)
and non-topological solutions
(the $CP^1$ model with potential) is another result of the paper.
\section*{Acknowledgements}
C.A. and J.S.-G. thank MCyT (Spain), FEDER (FPA2005-01963) and
Xunta de Galicia (grant INCITE09.296.035PR and
Conselleria de Educacion) for financial support.
A.W. acknowledges support from the
Ministry of Science and Higher Education of Poland grant N N202
126735 (2008-2010). Further, A.W. thanks Ya. Shnir for discussion.
|
\section{Method}
We have extended this idea to obtain a radical further improvement:
for ${\mathbb Z}^2$ and ${\mathbb Z}^3$ the mean CPU time per attempted pivot, which we
denote $T(N)$, is now only $O(\log N)$ for the range of $N$ studied,
and we have a theoretical argument that the large $N$ behavior is
$O(1)$.
The key observation is that although there are typically $O(N)$ nearest
neighbor contacts for a SAW of length $N$, the
number of contacts between two halves of a SAW is typically $O(1)$, as shown
via renormalization group~\cite{Muller1998SAW} and
Monte Carlo~\cite{Baiesi2001SAW} methods. When we attempt to pivot part
of a SAW, it is guaranteed that each of the two sub-walks remain
self-avoiding, and hence we only need to determine if
the sub-walks intersect. If the resulting walk is self-avoiding,
then we expect, on average, that there will be a constant number
of contacts between the
two sub-walks.
We will now briefly discuss the relevant data structure and
algorithms; full details can be found in \cite{Clisby2009bSAW}.
We implement a binary tree data structure (see
e.g. \cite{Sedgewick1998Books}) which we call a
\emph{SAW-tree}.
The root node of the SAW-tree contains information
about the whole walk, including $R_{\mathrm{e}}^2$,
$R_{\mathrm{g}}^2$, $R_{\mathrm{m}}^2$, and its
{\em minimum bounding box}, which
is the smallest rectangular prism with faces of the form $x_i = c$
which completely contains the walk. The two children of the root node
are valid SAW-trees, and contain bounding box information for the
first and second halves of the SAW, etc., until the leaves of the tree
store individual sites.
The SAW-tree is related to the R-tree~\cite{Guttman1984CS}, a data
structure used in the field of computational geometry, but with additional
information encoding the state of the SAW.
Thus far the SAW-tree has been implemented for ${\mathbb Z}^d$, but
can be straightforwardly adapted to other lattices and the continuum,
as well as other polymer models with short-range interactions.
To guarantee optimal performance, we implement the SAW-tree so that it
is balanced, i.e. so that the
depth never exceeds some fixed constant times $\log N$. We define the
\emph{level} of a node as the number of
generations between a node and the leaves.
Bounding boxes enable us to rapidly determine if two sub-walks
intersect after a pivot attempt: if two
bounding boxes do not intersect, then the sub-walks which they contain
cannot intersect.
If a pivot attempt is successful, then it is necessary to
resolve all intersections between bounding boxes of different nodes in
the tree on opposite sides of the pivot site.
Our implementation
ensures that intersection tests are typically performed between
bounding boxes of nodes which are at the same level.
We argue in \cite{Clisby2009bSAW}
that the nodes at fixed level in the SAW-tree form a renormalized
walk, and the intersections between bounding boxes correspond to
contacts in the original walk. This implies that at each level there
are $O(1)$ intersections, and as the tree has $O(\log N)$ levels this
leads to the conclusion that a successful pivot takes time $O(\log N)$.
Successful pivots occur with probability
$O(N^{-p})$, so overall mean time spent on successful pivots is
$O(N^{-p} \log N)$.
When a pivot attempt is unsuccessful, with high probability the first
intersection occurs near the pivot site. Thus only a small
fraction of the SAW-tree needs to be traversed to find the
intersection, and we argue in \cite{Clisby2009bSAW} that this takes
mean time $O(1)$.
Unsuccessful pivots occur with probability
$O(1)$, and so the overall behavior is $T(N) = O(N^{-p}
\log N + 1) = O(1)$.
In Fig.~\ref{fig:cputimes} we show $T(N)$ for
${\mathbb Z}^2$ and ${\mathbb Z}^3$ from a separate data run, with maximum length $N =
2^{28} - 1 \approx 2.68\e{8}$.
In both cases it is apparent
there is a crossover
due to the shorter latency of cache versus main memory.
In \cite{Clisby2009bSAW} we argue that O(1) behavior may be
reached only for very large $N$, which makes interpretation of
Fig.~\ref{fig:cputimes} difficult.
For ${\mathbb Z}^2$ some curvature is visible,
and the trend appears consistent with $T(N)$ approaching a constant
for sufficiently large $N$. The exponent $p$ is smaller for ${\mathbb Z}^3$ ($p
\approx 0.11$) compared with ${\mathbb Z}^2$ ($p \approx 0.19$); hence,
the approach to a constant is far slower, and in fact
almost no curvature is visible for ${\mathbb Z}^3$.
We believe the numerical evidence provides a strong case that $T(N)$
is at most $O(\log N)$, and is consistent with $T(N) = O(1)$; see
\cite{Clisby2009bSAW} for more details.
$T(N)$ is shown for the various implementations
in
Table \ref{tab:performance}. For SAWs of length $N=10^6$ on the cubic
lattice, the performance gain for our implementation is
approximately 200 when compared with Kennedy's, and over
a thousand
when compared with that of
Madras and Sokal~\footnote{These numbers are intended as
a rough guide, as they are machine and compiler dependent.}. The
dramatic performance gain from the new implementation
not only makes
it possible to obtain large samples of walks with millions of steps, it also makes
the regime of very long walks, of up to $10^9$ steps, accessible to
computer experiments.
\begin{table}[!b]
\caption{$T(N)$, mean time per attempted pivot for $N$-step SAWs.}
\label{tab:performance}
\begin{ruledtabular}
\begin{tabular}{cccc}
Lattice & Madras and Sokal & Kennedy & This Letter \\ \hline
Square & $O(N^{0.81})$ & $O(N^{0.38})$ & $O(1)$\\
Cubic & $O(N^{0.89})$ & $O(N^{0.74})$ & $O(1)$\\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[!ht]
\resizebox{3.375in}{!}{\includegraphics{cputimescombined}}
\caption{$T(N)$ for ${\mathbb Z}^2$ and
${\mathbb Z}^3$. Note that these estimates were obtained in a separate data run
on a different computer from the main experiment, with lengths from
$N = 2^7 - 1$ to $N = 2^{28} - 1$.}
\label{fig:cputimes}
\end{figure}
For SAWs of length $N$, it is expected that the exponential
autocorrelation time is approximately $O(N/f)$~\cite{Madras1988SAW}, where
$f$ is the fraction of
pivot attempts which are successful.
The first $20 N/f$ configurations were discarded, ensuring
that for all practical purposes SAWs were sampled from the uniform
distribution. Batch
estimates of $\langle
R_{\mathrm{e}}^2 \rangle_{33554431}$,
using a batch size of $10^8$, are shown in
Sec.~4 of \cite{Clisby2009bSAW}; in this case
the first 50
batches were discarded,
while initialization bias is visually apparent for (at most) the
first 10 batches.
The computer experiment was performed on a cluster of AMD
Opteron Barcelona 2.3GHz quad core processors,
for a total of 16500
CPU hours. Code was written in C, and compiled with gcc.
$10^{11}$ pivot attempts were made on SAWs of length ranging from 15 to
$3.36\e{7}$, for a grand total of $1.89
\times 10^{13}$ pivot attempts.
Data were collected from every pivot attempt for $f$ and the
Euclidean-invariant moments $R_x^{2k} = \left(R_x^2\right)^k$ with $x \in \{\mathrm{e,g,m}\},
1 \leq k \leq 5$
\footnote{Data from
$R_x^{2k}$ with $2 \leq k \leq 5$ have
not been analyzed here.
These data are inferior for critical exponent
estimates, but may in the future be used to calculate universal
amplitude ratios.}.
The longest walks with
$N=3.36\e{7}$ required 3GB of memory; much longer
walks could conceivably be simulated in the future.
By comparing fits from the whole data set ($N \leq 3.36\e{7}$) with fits
from a reduced
data set ($N \leq 2.1\e{6}$), we confirmed that
data from the longest walks were indeed highly useful in tying down
the various estimates (see Fig.~1 in
\cite{Clisby2009aSAW}). However, the greatest
benefit from the simulation of truly
long walks, of say $10^9$ steps, may be the ability to directly
simulate properties of realistic
systems, such as DNA knotting, rather than determination of
universal parameters.
Monte Carlo estimates of global parameters are collected in Tables
II-V, Sec. 2 of \cite{Clisby2009aSAW}, with confidence intervals
calculated using the standard
binning technique.
For all lengths
studied the integrated autocorrelation time of the Markov chain is much less
than the batch size of $10^8$.
We confirm the accuracy of the confidence interval estimates by studying
the effect of batch size in Sec.~5 of \cite{Clisby2009bSAW}.
\section{Analysis}
We estimated the critical exponents $\nu$ and $\Delta_1 \approx 1/2$ and
associated amplitudes $D_x$ and $b_x$ by fitting the leading
term and
leading correction of
Eq.~(\ref{eq:asymptotic}) via weighted non-linear
regression.
We truncated the data set by requiring $N \geq N_{\mathrm{min}}$,
with $N_{\mathrm{min}}$ a free parameter. We shifted the value of $N$
of Eq.~\ref{eq:asymptotic} by an
amount $\delta N_x$ to obtain smoother
convergence by altering the sub-leading corrections (see
e.g. \cite{Clisby2007SAW}); estimates for $\nu$, $\Delta_1$, $D_x$, and
$b_x$ are unaffected in
the limit $N_{\mathrm{min}} \rightarrow \infty$.
With $\delta N_{\mathrm{e}}= .35$, $\delta N_{\mathrm{g}}=1$,
$\delta N_{\mathrm{m}}=0.4$,
our final model was
\begin{align}
\langle R_x^2 \rangle_N &= D_x (N+\delta N_x)^{2\nu} \left[1 +
\frac{b_x}{(N+\delta N_x)^{\Delta_1}} \right],
\label{eq:model}
\end{align}
Unfortunately we cannot fit the next-to-leading corrections
with exponents $1$, $\Delta_2 \approx 1$,
and $2\Delta_1 \approx 1$ as the differences
between them are far too small to resolve.
For sufficiently large
$N_{\mathrm{min}}$ we found that reduced $\chi^2$ values for all fits
approached 1 from above, indicating Eq.~\ref{eq:model} is asymptotically
correct.
Final estimates of parameters have been made directly from
Figs. \ref{fig:deltafit} and \ref{fig:nufitbias}, combining multiple
sources visually in an
attempt to make estimates robust, and allow the reader to critically
evaluate our final results. We do not distinguish
between \emph{subjective} and
\emph{statistical} errors, as we believe that in this context the
distinction is itself quite
subjective~\footnote{e.g., what value of $N_{\mathrm{min}}$
should be chosen for the statistical error?}.
We provide here some guidance for the interpretation of Figs.
\ref{fig:deltafit} and \ref{fig:nufitbias}, and refer the interested
reader to \cite{Guttmann1989Books} for (much) more information on series
analysis.
\begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{0pt}
\item{We plot estimates against $N_{\mathrm{min}}^{-y}$, where $y$ is chosen such that
$N_{\mathrm{min}}^{-y}$ is
of the same order as the residual error from the fit. The estimates for
$\Delta_1$ and $b_x$ have
$y = 1 - \Delta_1 \approx 0.47$, and $y=1$ for $\nu$ and $D_x$.}
\item{We seek to extrapolate the fits to $N_{\mathrm{min}}=\infty$, or $N_{\mathrm{min}}^{-y} =
0$. Depending upon whether the true value $y_{\mathrm{exact}}$ is less than,
equal to, or greater than $y$, the estimates would approach a
limiting value at $N_{\mathrm{min}}^{-y} = 0$ with infinite, finite, or zero slope
respectively. }
\item{Successive estimates are highly correlated, and so any
trend which lies within the error bars should be disregarded. Only
some of the error bars are plotted in order to reduce
visual noise.}
\item{There are no bounds on the errors of the truncated asymptotic
formulae, and hence the interpretation of the graphs is subjective.
The underlying systematic error is observable dependent, and so
combining estimates from a variety of observables improves
robustness.}
\end{itemize}
In Fig.~\ref{fig:deltafit} we plot estimates of $\Delta_1$ with our
final result plotted at 0; the error bar reflects the scatter
between observables.
In Fig.~\ref{fig:nufitbias} we plot
estimates for $\nu$, biased with the lower and upper limits of our
range for $\Delta_1$, with our final result at 0. Similar plots for
the amplitudes are given in Figs. 5 and 6 of \cite{Clisby2009aSAW}.
We have conservatively chosen the error bar for the final result to
encompass estimates from all observables
$\langle R_x^2 \rangle$.
As the
amplitudes $D_x$ are highly correlated
with estimates for $\Delta_1$, the biasing of $\Delta_1$ greatly
extends the range over which stable fits can be obtained. This is the
reason the biased fits are preferred over the unbiased fits shown in
Fig.~3 of \cite{Clisby2009aSAW}.
\begin{figure}[!b]
\resizebox{3.375in}{!}{\includegraphics{deltafit}}
\caption{Estimates for $\Delta_1$.}
\label{fig:deltafit}
\end{figure}
\begin{figure}[!ht]
\resizebox{3.375in}{!}{\includegraphics{nufitbias}}
\caption{Estimates for $\nu$ from fits with biased $\Delta_1$.
We show
the envelope of
maximum and minimum values of endpoints of error bars for all observables.}
\label{fig:nufitbias}
\end{figure}
\section{Results}
We report our final results in Table~\ref{tab:parameters}, and in
addition we have $D_{\mathrm{m}} = 0.58687(12)$, $b_{\mathrm{e}} =
-0.49(5)$, $b_{\mathrm{g}} = -0.1125(125)$, and $b_{\mathrm{m}} = -0.295(30)$.
If one assumes the hyperscaling relation $d \nu = 2 - \alpha$, then
one also obtains $\alpha = 0.237209(21)$.
The estimates of
$\nu$, $D_{\mathrm{e}}$, and $D_{\mathrm{g}}$ are in accordance with
previous results,
although considerably more accurate.
The estimate for $\Delta_1$ is
more accurate than the previous Monte Carlo
value \cite{Li1995SAW}, but less accurate than the Monte Carlo
renormalization group
estimate of \cite{Belohorec1997SAW}, which relies upon an uncontrolled
although accurate approximation. The claimed accuracy of the
field theory estimates~\cite{Guida1998SAW} for $\Delta_1$ is also
comparable, but as
discussed by Li et al.~\cite{Li1995SAW} these calculations have
underlying systematic errors of uncertain magnitude.
Any desired amplitude ratios can be calculated from the amplitude
estimates. The rational number with smallest denominator within 3
standard deviations of
$\nu=0.587597$ is $161/274$,
suggesting that $\nu$ cannot be
expressed as a rational number with small denominator.
\begin{table}[!ht]
\caption{Comparison of parameter estimates.}
\label{tab:parameters}
\begin{ruledtabular}
\begin{tabular}{lllll}
\multicolumn{1}{c}{Source\footnote{Abbreviations: MC $\equiv$ Monte
Carlo, FT $\equiv$ Field theory, $d=3$ $\equiv$ $d=3$ expansion,
$\epsilon$ bc $\equiv$ $\epsilon$ expansion with boundary
conditions, MCRG $\equiv$ Monte Carlo renormalization group.}} & \multicolumn{1}{c}{$\nu$} & \multicolumn{1}{c}{$\Delta_1$} &
\multicolumn{1}{c}{$D_{\mathrm{e}}$} & \multicolumn{1}{c}{$D_{\mathrm{g}}$}
\\ \hline
This Letter & 0.587597(7) & 0.528(12)& 1.22035(25)& 0.19514(4)\\
\cite{Clisby2007SAW}\footnote{Using Eqs. (74) and (75) with $0.516 \leq \Delta_1
\leq 0.54$.} Series & 0.58774(22) &
&1.2178(54) &\\
\cite{Prellberg2001SAW} MC & 0.5874(2) & & & \\
\cite{MacDonald2000SAW}\footnote{No error
estimates were made in \cite{MacDonald2000SAW}, but estimates for
$\nu$ were in the
range $0.5870 \leq \nu \leq 0.5881$.} Series & 0.58755(55) & & 1.225 &\\
\cite{Guida1998SAW} FT $d=3$& 0.5882(11) & 0.478(10) & & \\
\cite{Guida1998SAW} FT $\epsilon$ bc & 0.5878(11)& 0.486(16) & & \\
\cite{Belohorec1997SAW} MCRG & 0.58756(5) & 0.5310(33) & & \\
\cite{Li1995SAW}\footnote{In addition $b_{\mathrm{e}}=-0.483(39)$, $b_{\mathrm{g}} =
-0.1143(47)$. $D_{\mathrm{e}}$, $D_{\mathrm{g}}$, $b_{\mathrm{e}}$, and $b_{\mathrm{g}}$ estimates were biased
with $\nu=0.5877$,
$\Delta_1=0.56$; the confidence intervals were not
intended to be taken seriously.} MC & 0.5877(6) & 0.56(3) & 1.21667(50) & 0.19455(7) \\
\end{tabular}
\end{ruledtabular}
\end{table}
We would like to stress that, due to the neglect of sub-leading terms,
there are underlying
systematic errors in our estimates
which \emph{are not} and
\emph{cannot be} controlled. We have the luxury of high quality data
from long walks, and have attempted to be conservative with
our claimed errors, but acknowledge there is a risk that the
(subjective) confidence intervals may not be sufficiently large.
\section{Conclusion}
In summary, an efficient version of the pivot
algorithm for SAWs has been implemented and used to calculate $\nu$;
the algorithms
developed promise to be widely useful in the Monte Carlo
simulation of SAWs and related models of polymers.
\begin{acknowledgments}
I thank I.~G.~Enting, A.~J.~Guttmann, G.~Slade, A.~Sokal, and an
anonymous referee for useful
comments on the manuscript.
Computations were performed using the
resources of VPAC.
Financial support from the Australian Research Council is gratefully
acknowledged.
\end{acknowledgments}
\FloatBarrier
|
\section{\label{sec:intro}Introduction}
Launch of LHC on the rated capacity assumes not only the answer
on the question on existence of Higgs boson, the proof of
supersymmetry display and a new insight on origin of Dark Matter
\cite{LHC},
but permits one to reconsider the problems of an
unique description of variety of elementary particles and all
known interactions. In this relation, the development of
higher-spin (HS) field theory in view of its close relation to
superstring theory on constant curvature spaces, which operates
with an infinite set of massive and massless bosonic and fermionic
HS fields subject to multi-row Young tableaux (YT)
$Y(s_1,...,s_k)$, $k \geq 1$ (see for a review, \cite{reviews})
seems by actual one. The paper considers the last results of
constructing
Lagrangian formulations (LFs) for free integer and half-integer HS fields on
$AdS_d$-space with $Y(s_1,s_2)$ in Fronsdal metric-like formalism
within BFV-BRST approach \cite{BFV} as a starting point for an
interacting HS field theory in the framework of conventional
Quantum Field Theory, and in part based on the results presented
in \cite{adsfermBKR,flatfermmix, 0812.2329, Sakharov}.
This method of Lorentz-covariant constructing LF for HS fields,
developed originally in a way that applies to Hamiltonian
quantization of gauge theories with a given LF, consists in a
solution of the \textit{problem inverse} to that of the method
\cite{BFV} (as in the case of string field theory \cite{SFT} and
in the early papers on HS fields \cite{Ouvry}) in the sense of
constructing a classical gauge LF with respect to a nilpotent
BFV--BRST operator $Q$.
In detail, the solution of \emph{inverse problem} includes 4
steps:
\begin{itemize}\vspace{-0.5ex}
\item the realization of initial irrep conditions
of AdS group, that extract the fields with a definite mass $m$ and
generalized spin $\mathbf{s}=(s_1,...,s_k)$ \cite{Metsaev} as
operator mixed-class constraints $o_I$ in a special Fock space
$\mathcal{H}$; \vspace{-1ex}
\item the additive conversion
(following to \cite{conversion}) of algebra $o_I$ into one of
$O_I$: $O_I=o_I+o'_I$, $[o_I,o'_J\}=0$, determined on wider Fock
space, $\mathcal{H}\bigotimes \mathcal{H}'$ with only first-class
constraints $O_\alpha \subset O_I$; \vspace{-1ex}
\item the construction of the
Hermitian nilpotent BFV-BRST operator $Q'$ for non-linear superalgebra of
converted operators $O_I$ which contains the BFV-BRST operator $Q$ for only
subsystem of $O_\alpha$;
\vspace{-1ex}
\item the finding of Lagrangian $\mathcal{L}$ for given HS field
through corresponding scalar product $\langle \ | \
\rangle$ like $\mathcal{L} \sim \langle \chi |Q |\chi \rangle$,
to be invariant with respect to
gauge transformations $\delta|\chi \rangle = Q
|\Lambda \rangle$ with $|\chi \rangle$ containing initial HS
field.
\end{itemize}
\vspace{-1.0ex} As compared to application of above algorithm for
bosonic \cite{flatbos} and fermionic \cite{flatfermmix, flatferm}
HS fields on $\mathbf{R}^{1,d-1}$ with standard resolution of the
2nd and 3rd steps due to the same Lie
(super)algebra structure for $o_I, o'_I, O_I$: $[o_I,
o_J\} = f_{IJ}^K o_K$, their resolution already for
totally-symmetric HS fields on $AdS_d$ space \cite{adsfermBKR,
0206027} are not so easy. It is revealed on stages of Verma module
(VM) construction for $o'_I$ and its non-polynomial(!) oscillator
realization in $\mathcal{H}'$ because of $AdS$-radius
$r^{-\frac{1}{2}}$ ($r = R/d(d-1)$ for scalar curvature $R$)
\cite{adsfermBKR, 0206027}. In turn, a construction of BFV-BRST
operator $Q'$ does not have the Lie-algebra form, ${Q'} =
\mathcal{C}^I {O}_I + \textstyle\frac{1}{2}
\mathcal{C}^I\mathcal{C}^{J}f^{K}_{JI}\mathcal{P}_K$ for
(super)algebra of
$O_I$ in transiting to $AdS$-space.
The main goals of the paper are to apply the above strategy to
construct LFs for bosonic and fermionic HS fields on
$AdS_d$-spaces subject to $Y(s_1,s_2)$.
\section{Bosonic fields in AdS spaces}
A massive integer spin $\mathbf{s}=(s_1,s_2)$, ($s_1 \geq s_2$),
representation of the AdS group in an $AdS_d$ space is realized in
a space of mixed-symmetry tensors,
\begin{equation}\label{Young k2}
\Phi_{(\mu)_{s_1},(\nu)_{s_2}} \hspace{-0.2em}\equiv
\hspace{-0.2em}
\Phi_{\mu_1\ldots\mu_{s_1},\nu_1\ldots\nu_{s_2}}(x)
\hspace{-0.3em}\longleftrightarrow \hspace{-0.3em}
\begin{array}{|c|c|c c c|c|c|c|c|c| c|}\hlin
\!\mu_1 \!&\! \mu_2\! & \cdot \ & \cdot \ & \cdot \ & \cdot\ & \cdot\ & \cdot\ &
\cdot\ &\!\! \mu_{s_1}\!\! \\
\hlin
\! \nu_1\! &\! \nu_2\! & \cdot\
& \cdot\ & \cdot & \cdot & \cdot & \!\!\nu_{s_2}\!\! \\
\cline{1-8
\end{array}\ ,
\end{equation}
subject to the Klein-Gordon (\ref{Eq-0b}), divergentless,
traceless and mixed-symmetry equations (\ref{Eq-1b}) [for $\beta =
(2;3) \Longleftrightarrow (s_1>s_2; s_1 = s_2)$]:
\begin{eqnarray}
\label{Eq-0b} &&\bigl[\nabla^2 +r[(s_1-\beta-1+ d)(s_1-\beta) -
s_1- s_2]+m^2 \bigr]\Phi_{(\mu)_{s_1},\ (\nu)_{s_2}} =0,\\
&&\bigl(\nabla^{\mu_1}, \ \nabla^{\nu_1}, \ g^{\mu_1\mu_2} ,\
g^{\nu_1\nu_2},\ g^{\mu_1\nu_1} \bigr)\Phi_{(\mu)_{s_1},\
(\nu)_{s_2}} = \Phi_{\{(\mu)_{s_1},\nu_1\}\nu_2...\nu_{s_2}}=0.
\label{Eq-1b}
\end{eqnarray}
To obtain HS symmetry algebra (of $o_I$) for a
description of all integer HS fields, we introduce a Fock
space $\mathcal{H}$, generated by 2 pairs of creation $a^i_\mu(x)$
and annihilation $a^{j+}_\mu(x)$ operators, $i,j =1,2, \mu,\nu
=0,1...,d-1$: $[a^i_\mu, a_\nu^{j+}]=-g_{\mu\nu}\delta_{ij}$, and
a set of constraints for an arbitrary string-like vector
$|\Phi\rangle \in \mathcal{H}$,
\begin{eqnarray}
\label{PhysState} \hspace{-2ex}&& \hspace{-2ex} |\Phi\rangle =
\textstyle\sum_{s_1=0}^{\infty}\sum_{s_2=0}^{s_1}\Phi_{(\mu)_{s_1},(\nu)_{s_2}}(x)\,
a^{+\mu_1}_1\ldots\,a^{+\mu_{s_1}}_1a^{+\nu_1}_2\ldots\,a^{+\nu_{s_2}}_2|0\rangle,\\
\label{l0} \hspace{-2ex}&& \hspace{-3ex} {\tilde{l}}_0|\Phi\rangle
= \bigl(l_0+ \tilde{m}^2_b + r \bigl((g_0^1-2\beta-2)g_0^1 -
g_0^2 \bigr)\bigr)|\Phi\rangle=0 , \ l_0 = [D^2 -
r\textstyle\frac{d(d-6)}{4}],\\
\label{lilijt} \hspace{-2ex} && \hspace{-2ex} \bigl({l}^i, l^{ij},
t \bigr)|\Phi\rangle = \bigl(-i a^i_\mu D^\mu,
\textstyle\frac{1}{2}a^{i}_\mu a^{j\mu}, a^{1+}_\mu a^{2\mu}\bigr)
|\Phi\rangle=0,\ i\leq j,
\end{eqnarray}
with number particles operators, $g_0^i =
-\frac{1}{2}\{a^{i+}_\mu, a^{\mu{}i}\}$, central charge
$\tilde{m}^2_b = {m}^2$ + $r\beta(\beta+1)$, operator $D_\mu =
\partial_\mu-\omega_\mu^{ab}(x)\bigl(\sum_{i}a_{i{}a}^+a_{i{}b}
\bigr)$, $a^{(+)\mu}_i(x)=e^\mu_a(x)a^{(+)a}_i$ with vielbein
$e^\mu_a$, spin connection $\omega_\mu^{ab}$, tangent indices
$a,b$. Operator $D_\mu$ is equivalent in its action in
$\mathcal{H}$ to the covariant derivative $\nabla_{\mu}$ [with
d'Alambertian $D^2 = (D_a + \omega^{b}{}_{ba} )D^a$]. The set of 7
primary constraints (\ref{l0}), (\ref{lilijt}) with $\{o_\alpha\}$
= $\bigl\{{\tilde{l}}_0, {l}^i, l^{ij}, t \bigr\}$ are equivalent
to Eqs. (\ref{Eq-0b}), (\ref{Eq-1b}) for all spins.
For Hermiticity of BFV-BRST operator (reality Lagrangian
$\mathcal{L}$) the algebra with $o_\alpha$ must be enlarged by
adding the operators $\bigl(l^+_i, l^+_{ij}, t^+ \bigr)$,
resulting the HS symmetry algebra in AdS${}_d$ space with
$Y(s_1,s_2)$, denoted as $A(Y(2), AdS_d))$. The maximal Lie
subalgebra of operators $l_{ij}, t, g_0^i, l^+_{ij}, t^+$ is
isomorphic to $sp(4)$ whereas the only nontrivial quadratic
commutators in $A(Y(2), AdS_d))$ are due to operators with
$D_{\mu}$: $l_i, \tilde{l}_0, l^+_i$. For the aim of LF
construction it is enough to have a simpler, (so called
\textit{modified}) algebra $A_{mod}(Y(2), AdS_d))$, with operator
$l_0$ (\ref{l0}) instead of ${\tilde{l}}_0$, so that
AdS-mass term, $\tilde{m}^2_b + r
\bigl((g_0^1-2\beta-2)g_0^1 -
g_0^2 \bigr)$, will be restored later within conversion
and properly construction of LF.
Algebra $A_{mod}(Y(2), AdS_d))$
contains 1 first-class constraint $l_0$, 4 differential $l_i,
l_i^+ $, 8 algebraic $ t, t^+, l_{ij}, l_{ij}^+$ second-class
constraints $\theta_{\mathbf{a}}$, operators $g_0^i$, composing an
invertible matrix: $\|[\theta_{\mathbf{a}},
\theta_{\mathbf{b}}\}\| = \|\Delta_{\mathbf{ab}}(g_0^i)\| +
(o_{I})$, and satisfies the non-linear relations (additional to
ones for $sp(4)$) given by Table~\ref{table}.
\hspace{-1ex}\begin{table}[t] {
\begin{center}
\begin{tabular}{||c||c|c|c|c|c|c|c||c||}\hline\hline
$\hspace{-0.2em}[\; \downarrow, \rightarrow \}$\hspace{-0.7em}&
$t$ & $t^+$ & $l_0$ &
$l^i$ &$l^{i{}+}$ & $l^{ij}$
&$l^{ij{}+}$ &$g^i_0$ \\
\hline\hline $l_0$
& $0$ & $0$ & $0$
& \hspace{-0.3em}
$\hspace{-0.2em}-r{\mathcal{K}}^{bi+}_1$\hspace{-0.5em} & \hspace{-0.3em}
$r{\mathcal{K}}^{bi}_1$\hspace{-0.3em}
& $0$ & $0$ & $0$ \\
\hline $l^k$
& \hspace{-0.5em}$- l^2\delta^{k1}$ \hspace{-0.5em} & \hspace{-0.5em}$
-l^1\delta^{k2}$ \hspace{-0.9em}&\hspace{-0.9em}
$r{\mathcal{K}}^{bk+}_1$\hspace{-1.1em}
& \hspace{-0.3em}${W}^{ki}_b$ \hspace{-0.3em} & \hspace{-0.3em}
${X}^{ki}_b$\hspace{-0.3em}
& $0$ & \hspace{-0.5em}$- \textstyle\frac{1}{2}l^{\{i+}\delta^{j\}k}$
\hspace{-0.9em} & \hspace{-0.4em}
$l^i\delta^{ik}$\hspace{-0.4em} \\
\hline $l^{k+}$ & \hspace{-0.5em}$l^{1+}
\delta^{k2}$\hspace{-0.7em} & \hspace{-0.7em}$l^{2+}\delta^{k1}$ \hspace{-1.0em} &
\hspace{-0.5em}$-r{\mathcal{K}}^{bk}_1$ \hspace{-1.1em}&\hspace{-0.3em}
$-{X}^{ik}_b$\hspace{-0.3em}
&\hspace{-0.5em} $- {W}^{ki+}_b$\hspace{-0.5em}
&\hspace{-0.7em} $ \textstyle\frac{1}{2}l^{\{i}\delta^{j\}k}$\hspace{-0.7em} & $0$ &\hspace{-0.7em} $-l^{i+}\delta^{ik}$\hspace{-0.5em} \\
\hline\hline
\end{tabular}
\end{center}} \vspace{-2ex}\caption{The non-linear part of algebra $A_{mod}(Y(2),
AdS_d)$.}\label{table}
\end{table}
In the Table~\ref{table}, the quantities ${\mathcal{K}}^{bk}_1, {W}^{ki}_b,
X^{ki}_b$ are quadratic in $o_I$ (see \cite{0812.2329} for
details)
\begin{eqnarray}
{} {W}^{ki}_b & = & 2r\varepsilon^{ki}\left[(g_0^2-g_0^1)l^{12} -
t l^{11} + t^+ l^{22}\right], \quad
\varepsilon^{ki}=-\varepsilon^{ik}, \varepsilon^{12}=1,
\label{lilj}\\
{\mathcal{K}}^{bk}_1& = &\Bigl(4\textstyle\sum_{i}l^{ki+}l^i
+l^{k+}(2g_0^k-1) -2l^{2+}t \delta^{k1} - 2l^{1+}t^+
\delta^{k2}\Bigr) \label{l'0li+} ,
\\
{} {X}^{ik}_b
& =
&\bigl\{{{l}}_{0}+ r\bigl( K^{0i}_0 +
\mathcal{K}^{12}_0 \bigr)\bigr\}\delta^{ij} + r\bigl\{
\bigl[4\textstyle\sum\nolimits_{j} l^{ 1j+}l^{ j2} +
\textstyle(\sum_{j}g_0^{ j}
-{2}
)t\bigr]\delta^{k1}\delta^{i2}\nonumber\\
&& + r\bigl\{ \bigl[4\textstyle\sum\nolimits_{j} l^{ j2+}l^{ 1j} +
t^+\textstyle(\sum_{j}g_0^{ j}
-{2}
)\bigr]\delta^{k2}\delta^{i1}, \label{lilj+b}
\end{eqnarray} with $K^{0i}_0, i=1,2$, $\mathcal{K}^{12}_0$ composing
a Casimir operator $\mathcal{K}$ for $sp(4)$ algebra
\begin{equation}\label{Casimirs}
\mathcal{K}_0 =
\textstyle\sum_{i}K_0^{0i} +
2\mathcal{K}_0^{12} =
\sum_{i}\bigl((g_0^i)^2-2g_0^i -4l^{ii+}l^{ii}\bigr)+2 \bigl(t^{ +}t - g_0^{ 2} -
4l^{ 12 +}
l^{ 12}\bigr).
\end{equation}
Algebra $A(Y(2), AdS_d)$ generalizes its analogs both for
mixed-symmetry HS fields on flat spaces and for
totally-symmetric HS fields on AdS spaces \cite{flatbos,
0206027}.
\section{Verma module, Fock space realization}
The procedure of additive conversion for non-linear algebra
$A(Y(2), AdS_d)$ of $o_I$ into algebra $A_c(Y(2), AdS_d)$ of
converted constraints $O_I$, $O_I=o_I+o'_I$, $[o_I, o'_J\}$ acting
in Hilbert space $\mathcal{H}\bigotimes \mathcal{H}'$ with only
first-class constraints $O_\alpha$ \cite{0206027, adsfermBKR} may
be realized in two ways resulting either to unconstrained or to
constrained LF. For the former case, it means the conversion of
the total set of second-class constraints $\{\theta_\mathbf{a}\}$
for the latter the conversion of only differential and part of
algebraic constraints: $l_i, l_i^+, t, t^+$ having restricting the
algebra $\mathcal{A}(Y(2),AdS_d)$ to the reduced non-linear
algebra $\mathcal{A}_r(Y(2),AdS_d)$ = $\{l_0, g_0^i, l_i, l_i^+,
t, t^+\}$ with off-shell traceless conditions on the fields and
gauge parameters of final LF. For the aims of QFT an unconstrained
LF is more preferable.
To find additional parts $o'_I$ we, first, determine the
multiplication law for the algebra $\mathcal{A}'(Y(2),AdS_d)$,
which for compactly written multiplication law for $o_I$ given by
Table~\ref{table} reads (see Refs.\cite{adsfermBKR, 0812.2329,
0206027} for details)
\begin{eqnarray}
[\,o_I',o_J'\} = f_{IJ}^Ko_K'- f_{IJ}^{KM}o_M'o_K'
\quad \mathrm{if} \quad [{o}_I,{o}_J\} =
f_{IJ}^K{o}_K+f_{IJ}^{KM}{o}_K{o}_M. \label{addal}
\end{eqnarray}
Second, following generalization of Poincare--Birkhoff--Witt
theorem, we construct VM, based on Cartan-like decomposition
enlarged from one for $sp(4)$
\begin{equation}\label{Cartandecomp}
\mathcal{A}'(Y(2),AdS_d) = \{l^{\prime +}_{ij},
t^{\prime+}, l^{\prime +}_i\} \oplus \{g_0^{\prime i}, l_0'\}
\oplus \{l^{\prime }_{ij}, t', l^{\prime }_i\} \equiv
\mathcal{E}^-\oplus H \oplus\mathcal{E}^+.
\end{equation}
Note, that in contrast to the case of Lie (super)algebra and
totally-symmetric HS fields on AdS space \cite{flatfermmix,
adsfermBKR, 0206027}, the negative root vectors $l^{\prime +}_1,
t^{\prime+}, l^{\prime +}_2$ are not commuted, making the
arbitrary vector $|\vec{N}\rangle_V = \left| {n}_{11},{n}_{12}, {n}_{22},n_1, n, n_2
\rangle_V \right.$
\begin{equation}\label{VM}
|\vec{N}\rangle_V \equiv \textstyle\bigl(l^{\prime
+}_{11}\bigr){}^{n_{11}}\bigl(l^{\prime +}_{12}\bigr){}^{
n_{12}}\bigl(l^{\prime +}_{22}\bigr){}^{ n_{22}}
\Bigl(\frac{l^{\prime +}_1}{m_1}\Bigr){}^{ n_1}\bigl(t^{\prime
+}\bigr){}^{n} \Bigl(\frac{l^{\prime
+}_{2}}{m_2}\Bigr){}^{n_{2}}|0\rangle_V, \
\mathcal{E}^+|0\rangle_V=0
\end{equation}
from VM, (for highest weight vector
$|0\rangle_V$, $n_{ij}, n_i, n \in \mathbf{N}_0$, and arbitrary
constants $m_i$ with dimension of mass) by not proper one for
$t^{\prime+}, l^{\prime
+}_2$!
That \emph{nontrivial entanglement} are resolved within iterative
procedure, so that the VM for algebra $\mathcal{A}'(Y(2),AdS_d)$
may be constructed. Omitting tedious technical details (see
Ref.\cite{Sakharov}), note that by the crucial difference here is
the presence of so-called \emph{basic block} operator $\hat{t}'$
enlarging Lie part ${t}'_{L}$ in ${t}'={t}'_{L} + \hat{t}'$:
$\hat{t}'_{\vert r=0} =0$,
from which the result of the action of
all $o'_I$ on vector $|\vec{N}\rangle_V$ is found. The realization
of the VM in formal power series (due to $r$) in degrees of
creation and annihilation operators $(B,B^+) = b_i, b_i^+, b_{ij},
b_{ij}^+, b, b^+$ in $\mathcal{H}'$ whose number coincides to
ones of 2nd-class constraints is solved too. As a result, the
Cartan generators have the boundary conditions: $(g_0^{\prime i},
l_0')=(h^i, m_0^2) + O(B,B^+)$, with constants $h^i, m_0^2$
permitting then under special choices to restore the correct
equations of motion within LF. The explicit form of basic block
$\hat{t}'(B,B^+)$ (in ${t}'={t}'_{L} + \hat{t}'$) reads,
{\small\begin{eqnarray}
&& \hspace{-1.0em}{t}'_{L} =\left(h^1-h^2 - b^+_2b_2 - b^+b\right)b - b_{11}^+b_{12}
-2b_{12}^+b_{22}\,,
\label{t'Lf} \\
&& \hspace{-1.0em}\hat{t}' = \sum_{k=0}\Biggl[\sum_{{}^1m=0}\hspace{-0.2em}
\sum_{{}^1l=0} ... \sum_{{}^km=0}\hspace{-0.2em}\sum_{{}^kl=0}
(-1)^k\hspace{-0.2em}
\left(\hspace{-0.2em}\frac{-2r}{m_2^2}\hspace{-0.2em}\right)^{\hspace{-0.2em}\sum_{i=1}^k({}^im+{}^il)+k}
\hspace{-0.2em}\prod_{i=1}^k
\frac{1}{(2{}^im+1)!}\frac{1}{(2{}^il+1)!}\nonumber
\\
\hspace{-1.0em} && \hspace{1.0em}\times \bigl(b_{22}^+\bigr)^{\sum_{i=1}^k({}^im+{}^il)+k}\Biggl\{
b_2^+bb_2 -\frac{m_1}{m_2}\sum_{m=0}\left(\frac{-2r}{m_2^2}\right)^m
\frac{(b_{22}^+)^m}{(2m+1)!}b_1^+
b_2^{2m+1}\nonumber \\
\hspace{-1.0em} &&\hspace{1.0em} +\sum_{m=1}\left(\frac{-2r}{m_2^2}\right)^m
(b_{22}^+)^{m-1}\Bigl[b_{12}^+
\Bigl\{
\frac{h^2-h^1+2b^+b}{(2m)!}+\frac{b^+_2b_2}{(2m+1)!}\Bigr\}
\nonumber \\
\hspace{-1.0em} && \hspace{1.0em}- \frac{b_{11}^+b^+}{(2m)!} - \frac{b_{22}^+}{(2m)!}(h^2-h^1+b^+b)b\Bigr]b^{2m}_2 \nonumber \\
\hspace{-1.0em} && \hspace{1.0em} - \sum_{m=0}\sum_{l=0}\left(\frac{-2r}{m_2^2}\right)^{m+l+1}
\frac{1}{(2m+1)!}(b_{22}^+)^{m+l}\Bigl[ b_{12}^+
\Bigl\{\frac{h^2-h^1+2b^+b+b_2^+b_2}{(2l+1)!}\nonumber\\
\hspace{-1.0em} && \hspace{1.0em} -
\frac{b_2^+b_2}{(2l+2)!}
\Bigr\}
-\frac{b_{11}^+b^+}{(2l+1)!} -\frac{b_{22}^+}{(2l+1)!}(h^2-h^1+b_2^+b_2+b^+b)b \nonumber\\
\hspace{-1.0em} && \hspace{1.0em}+
\frac{m_1}{m_2}\frac{b_{22}^+}{(2l+2)!}b_1^+b_2
\Bigr] b_2^{2(m+l+1)}
\Biggr\}(b_2)^{2(\sum_{i=1}^k({}^im+{}^il)+k)}\Biggr],
\end{eqnarray}}
whereas the rest operators
$o'_I(B,B^+)$ may be found in \cite{Sakharov}.
Therefore, the solution of the second problem on construction of
the VM for algebra $A'(Y(2), AdS_d)$ and its oscillator
realization is found.
\section{BRST operator for non-linear algebra}
The system of ${O}_I$ forming non-linear algebra
$\mathcal{A}_{c}(Y(2), AdS_d)$ with multiplication law following
from Eq. (\ref{addal}):
\begin{equation}\label{multAc} [O_I, O_J\} =
F_{IJ}^K(O)O_K,\quad F_{IJ}^K(O) = f_{IJ}^K
-\bigl(f_{IJ}^{MK}+
f_{IJ}^{KM}\bigr)o_M' + f_{IJ}^{MK}{O}_M
\end{equation} and
Table~\ref{table} now has no the form of closed algebra,
because of presence nontrivial Jacobi identities
for 6 triples $(L_1, L_2, L_0)$, $(L_1^+, L_2^+, L_0)$, $(L_i,
L_j^+, L_0)$. Indeed, there exists a set of third order structural
functions in terminology of Ref. \cite{BFV} resolving those
identities (see \cite{0812.2329}).
The construction of a BFV-BRST operator ${Q}'$ for $\mathcal{A}_{c}(Y(2),
AdS_d)$, in case of the Weyl ordering for quadratic
combinations of ${O}_I$ in the r.h.s. of $[{O}_I, {O}_J\}$ and for
the $(\mathcal{C}\mathcal{P})$-ordering for the ghost coordinates
$\mathcal{C}^I$: $\eta_0$,
$\eta^i_G$ $\eta_i^+$, $\eta_i$, $\eta_{ij}^+$, $\eta_{ij}$,
$\eta$, $\eta^+$, and their conjugated momenta $\mathcal{P}_I$:
${\cal{}P}_0$, ${\cal{}P}^i_G$ , ${\cal{}P}_i$, ${\cal{}P}_i^+$,
${\cal{}P}_{ij}$, ${\cal{}P}_{ij}^+$, $\mathcal{P}^+$,
$\mathcal{P}$, is a more complicated task than for
totally-symmetric HS fields on AdS${}_d$ \cite{adsfermBKR,
0206027}.
The nilpotent operator ${Q}'$ has the terms in third degree in
ghosts $C^I$ \cite{0812.2329},
\begin{eqnarray}\label{explQ'}
&&\hspace{-1.5em} {Q'}
= Q'_1 + Q'_2 + \Bigl[r^2\eta_0\eta_i\eta_j\varepsilon^{ij}\Bigl\{
\textstyle \frac{1}{2}\sum_k\Bigl(G^k_0[\mathcal{P}
\mathcal{P}^{+}_{22}
- \mathcal{P}^+
\mathcal{P}^{+}_{11}+ {i}
\mathcal{P}^{+}_{12}\sum_l(-1)^l\mathcal{P}_G^l]
\nonumber\\
&& \textstyle -
{i}({L}^{+}_{11}\mathcal{P}^+-{L}^{+}_{22}\mathcal{P})
\mathcal{P}_G^k + 4L^{kk}\mathcal{P}^{+}_{k2}
\mathcal{P}^{+}_{1k}\Bigr)
- {L}^{+}_{12}\mathcal{P}_G^1\mathcal{P}_G^2 +2
L^{12}\mathcal{P}^{+}_{22}\mathcal{P}^{+}_{11}\Bigr\}
\nonumber\\
&& \hspace{-0.7em}\textstyle + r^2\eta_0\eta^+_i\eta_j\Bigl\{
\Bigl[\sum_{k}(-1)^k\frac{i}{2}G^k_0 \sum_l\mathcal{P}_G^l +
2(L^{+}_{22}\mathcal{P}^{22}-L^{11}\mathcal{P}^{+}_{11})\Bigr]
\mathcal{P}\delta^{1j} \delta^{2i} \nonumber\\
&& \hspace{-0.7em}+
\varepsilon^{\{1j}\delta^{2\}i}\Bigl(\imath\textstyle\sum_k\Bigl[
\frac{1}{2}T\mathcal{P}^+ - 2 L^{12}\mathcal{P}^{+}_{12}
(-1)^k\Bigr]\mathcal{P}_G^k +2(
L^{12}\mathcal{P}-T\mathcal{P}^{12})\mathcal{P}^{+}_{22}
\nonumber\\
&& \hspace{-0.7em} +
2(T^+\mathcal{P}^{12}-L^{12}\mathcal{P}^+)\mathcal{P}^{+}_{11}
\Bigr) -T\Bigl[\mathcal{P}_G^1\mathcal{P}_G^2\delta^{2i}
\delta^{1j}+2\mathcal{P}^{11}\mathcal{P}^{+}_{22}\delta^{1i}\delta^{2j}\Bigr]
- \textstyle
2\sum_{k}(-1)^k\nonumber\\
&& \hspace{-0.7em}
\times\Bigl[(G^k_0\mathcal{P}^{11}+\imath{L}^{11}\mathcal{P}_G^k)
\delta^{1i}\delta^{2j}- (G^k_0\mathcal{P}^{22} +
\imath{L}^{22}\mathcal{P}_G^k) \delta^{2i}\delta^{1j} \Bigr]
\mathcal{P}^{+}_{12}
\Bigr\} + h. c.\Bigr], \end{eqnarray}
with the standard form for linear $Q_1'$ and quadratic $Q_2'$
terms in ghosts $\mathcal{C}^I$. The Hermiticity of operator $Q'$
is defined by the rule: $
Q^{\prime +}K = K Q'$, {for} operator $K = \hat{1} \otimes K' \otimes
\hat{1}_{gh}$, with non-degenerate operator $K'$ providing the Hermiticity of
the additional parts $o'_I$ in $\mathcal{H}'$.
\section{Unconstrained Lagrangian formulation}
BFV-BRST operator $Q$ for 1st-class constraints
$O_\alpha = \{L_0,L^i,L^{ij},T\}$ is extracted from $Q'$
(\ref{explQ'}) by collecting the terms with ghosts $\eta^i_{G}$
being by BRST-invariant
enlarged number particle operators $\sigma^i + h^i = G_0^i + ghosts$ and
inessential operators $\mathcal{B}^i$ for further derivation of an LF,
\begin{eqnarray}
{Q}' = {Q} + \eta^i_G(\sigma^i+h^i) + \mathcal{B}^i
\mathcal{P}^i_G. \label{decomposQ'}
\end{eqnarray}
The same is applied to a physical vector $|\chi\rangle \in
\mathcal{H}_{tot}=\mathcal{H}\bigotimes
\mathcal{H}'\bigotimes\mathcal{H}_{gh}$,
$|\chi\rangle$=$|\Phi\rangle + |\Phi_A\rangle$,
$|\Phi_A\rangle_{\vert \mathcal{H}}$ = $0$, with $|\Phi\rangle$
given in (\ref{PhysState}). From commutativity, $[Q,
\sigma^k\}=0$ and choice of a representation for Hilbert space (as
in SFT \cite{SFT})) it follows the spectral problem from the
equation ${Q}'|\chi\rangle = 0$ \cite{0812.2329},
\begin{equation}
\label{Qchi} {Q}|\chi\rangle = 0, \quad
(\sigma^i+h^i)|\chi\rangle=0, \quad {gh}(|\chi\rangle)=0,
\end{equation}
thus determining the spectrum of spin values and proper
eigenvectors,
\begin{eqnarray}
h^i(s_i)&=& \textstyle -\bigl( s_i + \frac{d - 5 }{2}
-2\delta^{i2} \bigr)\,, \label{hchin1n2} \quad
|\chi\rangle_{(s_1,s_2)},
\end{eqnarray}
As differed to the second equation in (\ref{Qchi}) the first
equation is valid only in the subspace of $\mathcal{H}_{tot}$ with
the zero ghost number.
After substitution: $h^i \to h^i(s_i)$ operator $Q_{(s_1,s_2)}$,
is nilpotent on each subspace $H_{tot{}(s_1,s_2)}$ whose vectors
satisfy to the Eqs.(\ref{Qchi}) for (\ref{hchin1n2}). Hence, the
equations of motion (one to one correspond to
Eqs.(\ref{Eq-0b}),(\ref{Eq-1b})),~a~sequ\-ence of reducible gauge
transformations and Lagrangian action have the form
\begin{eqnarray}
\hspace{-1em}&& Q_{(s_1,s_2)}|\chi^0\rangle_{(s_1,s_2)}=0, \
\delta|\chi^{l} \rangle_{(s_1,s_2)}
=Q_{(s_1,s_2)}|\chi^{1+1}\rangle_{(s_1,s_2)},
\ l = 0,...,6,\label{LEoM} \\
\hspace{-1em}&& {\cal S}_{(s_1,s_2)} = \int d \eta_0 \;
{}_{(s_1,s_2)}\langle \chi^0 |K_{(s_1,s_2)} Q_{(s_1,s_2)}| \chi^0
\rangle_{(s_1,s_2)},\texttt{ for } |\chi^0\rangle\equiv
|\chi\rangle. \label{S}
\end{eqnarray}
The corresponding LF for bosonic field with spin $\mathbf{s}$
subject to $Y(s_1,s_2)$ is a reducible gauge theory of maximally
$L = 6$-th stage of reducibility.
\section{Fermionic fields in AdS spaces} The
construction of LF for half-integer HS fields subject to
$Y(s_1,s_2)$, $s_i=n_i+\frac{1}{2}$ is more complicated due to
presence of fermionic constraints, which follows from respective
irrep conditions on spin-tensor $\Phi_{(\mu)_{n_1},\ (\nu)_{n_2}}$
(with suppressed spinor index $A$) being by Dirac, gamma-traceless
and mixed-symmetry equations
\begin{eqnarray}
\Bigl(\bigl[i\gamma^{\mu}\nabla_{\mu}
-r^\frac{1}{2}(n_1 + \textstyle\frac{d}{2}-\beta)-m
\bigr],\ \gamma^{\mu_1} ,\ \gamma^{\nu_1}
\Bigr)\Phi_{(\mu)_{n_1},\ (\nu)_{n_2}} \hspace{-0.5em}=
\Phi_{\{(\mu)_{n_1},\nu_1\}\nu_2...\nu_{n_2}}=0.\nonumber
\end{eqnarray}
The only peculiarities are as follows. Instead of case of HS
symmetry algebra, we need to consider the case of respective
superalgebra, and therefore the construction of VM and BFV-BRST
operator is more non-trivial, but exists (see for details
\cite{adsfermBKR, Sakharov}). Because of presence in the total
Hilbert space $\mathcal{H}_{tot}$ the bosonic ghosts of any power
the stage of reducibility grows with spin $(s_1,s_2)$. At last,
for the case of fermionic HS fields we must obtain Lagrangian
which is linear in derivatives $\nabla_{\mu}$. But prescription
which follows from bosonic case leads to second-order Lagrangian.
However, this problem are effectively overcome by means of partial
gauge fixing and solving of part of equations of motion. Doing so,
the LF for mixed-symmetry fermionic HS fields may be constructed
\cite{flatfermmix, Sakharov}.
\section{Summary}
We have briefly considered the method of constructing the LF for
free massive mixed-symmetry HS fields on AdS${}_d$ space in the
framework of BFV-BRST approach. To do so, we have constructed new
auxiliary representation for non-linear algebra which serve for
conversion procedure of initial HS symmetry algebra. Then, we have
sketched a details of systematic way to find BFV-BRST operator for
non-linear operator algebra and presented a proper construction of
gauge LF basically for bosonic HS fields. The
Eqs.(\ref{LEoM}),({\ref{S}}) present the basic results of the
paper being the first step to interacting theory, following in
part to the research \cite{BuchbinderTsulaia}.
\paragraph{Acknowledgements} The author
thanks the organizers of the Bo\-go\-lyu\-bov'09 Conference (RAS -
JINR, Moscow - Dubna, Russia) for support and hospitality.
|
\section{Introduction}
One of the key-science goals of the SKA is to detect most of the
neutral hydrogen (HI; 1420.4~MHz in the rest frame) content of
galaxies out to cosmological redshifts of z $\sim$ 1 (for further
details see e.g. {\it Science with the Square Kilometre Array}
eds. Carilli and Rawlings or {\it Cosmology, Galaxy Formation and
Astroparticle Physics on the Pathway to the SKA} eds. Kl\"ockner,
Rawlings, Jarvis and Taylor). The technical parameters that determine
the performance of the SKA have been identified previously during the
SKADS program and are defined in the Benchmark scenario
(\cite{bench}). At the moment the SKA design includes three distinct
telescope technologies in order to cover the required frequency range,
i.e. between a few hundreds of MHz to 10$-$25~GHz. The parameter
space one needs to cover to assess the performance of the
low-frequency sparse dipole array, the mid-frequency aperature array
(AA), or the high-frequency dish array is enormous, and this is
impossible to accomplish via a single telescope simulation. To get a
basic understanding of the array parameters and their influence on the
quality and sensitivity of the final images, one can parameterise each
of the different antenna types in terms of a dish equivalent. For a
dish array, two basic parameters are the system equivalent flux
density (SEFD) and the image sensitivity ($\Delta$I). Respectively,
these are given by
\begin{equation}
{\rm SEFD} = \frac{{\rm T}_{\rm sys} 2 {\rm k}}{\eta_a \, {\rm A}},
\end{equation}
\noindent where T$_{\rm sys}$ is the system temperature [K], k is the Boltzmann constant, A is
the collecting area [m$^2$], and $\eta_a$ is the aperture efficiency, and (for a naturally weighted image) the image sensitivity can be calculated via:
\begin{equation}
{\rm \Delta I} = \frac{\rm SEFD}{\sqrt{{\rm N (N-1) {\rm N_{\rm Stokes}} t} \Delta \nu }},
\label{eq:sens}
\end{equation}
\noindent where N is the number of Antennas, N$_{\rm Stokes}$ is the number of Stokes parameters,
t integration time [s], and $\Delta \nu$ is the bandwidth [Hz] (\cite{ww1999}; a sensitivity
calculator can be found at
www-astro.physics.ox.ac.uk/$\sim$hrk/ARRAY$_{-}$EXPOSURE.html).
In order to increase the image sensitivity both equations dictate that
the system temperature and the effective area are crucial to the
telescope performance. However in the case of continuum emission that
is spectrally well-behaved (spectral index of around zero), the image
sensitivity can be increased by trading the SEFD for increased
bandwidth. For spectral line observations, where the bandwidth is
tailored to the width of the expected signal, this cannot be done.
Bandpass stability also plays a crucial role in the image quality
in this regime.
In this article we describe an end-to-end (e2e) HI simulation plan
that is focused on exploring a more manageable parameter space defined
by the system temperature, effective area, and the spatial
configuration of the array.
\section{Simulation}
Any aperture synthesis telescope acts as a spatial frequency filter
and an analytic determination of the final image quality is in most
cases impossible. Therefore the purpose of any interferometry
simulation is to define a practical sensitivity limit with respect to
the theoretical estimates. Generally one would like to have a full
array simulation that simulates the complete signal path from the
astronomical source up to and including the receiver electronics. Such
an approach is very complicated, computationally expensive and for the
purposes of a full SKA simulation completely impractical. The e2e
simulation can be broken down into individual components each of which
is treated as a standalone simulation. If one is interested in the
electromagnetic properties (e.g.~the directional gain and its
stability) of individual dish designs one needs to take into account
the telescope structure and make use of a full electromagnetic
simulation (e.g. \cite{holler}). Similar simulations are needed if one
is interested in the performance of an aperature array and its primary
beam pattern (e.g. OSKAR; www.oerc.ox.ac.uk/research/oskar). So far
little has been done to simulated the electronic path from the
receiver to the correlator, but that is certainly in the scope of the
PrepSKA (www.jb.man.ac.uk/prepska) program. Finally, after the
correlation stage the data quality and array performance can be
evaluated by the simulating the response of individual baselines, or
by simulating and assessing the final image.
Radio astronomy software packages which are currently available and
offer simulation functionality (e.g. AIPS, CASA, MeqTrees) generally
focus on generating a model visibility set given a model sky and a set
of parameters which describe the observation. Such functionality has
emerged naturally, due to the self-calibration process relying on the
generation of model visibilities, however the sophistication of the
simulations offered by packages has generally evolved beyond this
fundamental role, particularly in the case of MeqTrees.
\begin{figure}
\centering
\includegraphics[bb=181 30 680 560,width=8.5cm,clip]{HRK_FIG1.PS}
\caption{Overview of a full simulation path.
}
\label{FIG:simflow}
\vspace{-0.5cm}
\end{figure}
An overview of the developed e2e simulation path is shown in
Fig.~1. The grey box shows the steps we take to generate a model
sky. In our case this is based on the SKADS Simulated Skies (S$^3$),
which are a series of databases describing the properties of a range
of simulated astrophysical objects. The databases themselves are
discussed further in the next section. Specific subsets of these
databases can be retrieved and processed, and finally converted into
either a 2-D image or a 3-D datacube (the third axis being frequency)
with optional Gaussian noise\footnote{A suite of Python-based routines
with user-friendly GUIs, collectively known as the S3-Tools allows
access, manipulation and imaging of the SKADS Simulated Skies.
These tools have been developed by F. Levrier and a general overview
can be found in this volume (Levrier, 2009 in these
proceedings).}. These represent idealised radio skies which can then
be fed into a telescope simulation package. The array simulator is
represented in Fig.~1 by the black box, which will take the model sky
image and generate a model data set based on this. The final step is
the analysis of this end product. In the case of the e2e HI simulation
the analysis step involves running both the `idealised' and `observed'
HI datacubes through a source finder algorithm and comparing the two
resulting catalogues. This comparison can then be used to benchmark a
specific telescope design or observing strategy. In addition, the
ability of the S$^3$-tools to produce maps with Gaussian noise allows
to test the source finding algorithm and provide measures for
completeness studies.
\subsection{The sky simulation}
Here we present a short summary of the sky simulations that can be
used within the array simulator. Further details on the simulations
together with a webform which can be used to query the databases is
available on the Oxford S$^3$
webpage\footnote{s-cubed.physics.ox.ac.uk}. The full suite of
simulations is presented using two distinct products. Properties of
individual extragalactic objects and of Galactic pulsars are stored in
databases (SEX, SAX, PUL), whereas morphologically complex structures
such as the Global Sky Model (GSM) and the signals of the Epoch of
Reionization (EOR) are available as images.
The GSM is the radio foreground of our Galaxy which has
been modeled by the radio and (sub)- millimeter emission between
10~MHz to 100~GHz. In addition to the diffuse Galactic emission it also includes
emission from individual point sources e.g. supernova remnants
(\cite{gsm}; space.mit.edu/home/angelica/gsm).
The EOR images display the HI line signal of the Intergalactic medium
(IGM) during the Epoch of Reionization. This simulation covers the
redshift range between z = 5.6 and z = 23.6. In addition to the
ionization field, the effect of inhomogeneous heating of the IGM by
X-rays and the Lyman-$\alpha$ radiation field are taken into
account. The simulations have been produced in a cubic simulation box
with a side length of s$_{box}$ = 100 Mpc/h and a particle mass resolution
of 3$\times$10$^6$/h solar masses (Santos et al. 2008).
The semi-empirical simulation of extragalactic sources (SEX) describes
the radio continuum emission in a sky area of 20$\times$20 deg$^2$
out to a cosmological redshift of z = 20. As the name suggests, the sources were drawn from
observed (or extrapolated) luminosity functions and grafted onto an
underlying dark matter density field with biases which reflect their
measured large-scale clustering. This approach puts an emphasis on
modelling the large-scale cosmological distribution of radio sources
rather than the internal structure of individual galaxies. \\
\noindent Five source types of radio sources have been included in the simulation:
\begin{itemize}
\item[-]Radio-quiet AGN [1 core; 36,132,566 sources]
\item[-]Radio-loud AGN of the FRI class [1 core + 2 lobes; 23,853,132 sources]
\item[-]Radio-loud AGN of the FRII class [1 core + 2 lobes + 2 hot-spots; 2,345 sources]
\item[-]Quiescent star-forming galaxies [1 disk; 207,814,522 sources]
\item[-]Starbursting galaxies [1 disk; 7,267,382 sources]
\end{itemize}
\noindent For each of the source types, the database provides the
radio fluxes at observer frequencies at 151 MHz, 610 MHz, 1.4 GHz,
4.86 GHz and 18 GHz, down to flux density limits of
10~nJy. Intermediate frequencies can be determined by using the
S3-tools. In addition to the continuum emission this simulation
provides a rough estimate of the HI mass of the starbursting and
star-forming galaxies (\cite{wilman}). A secound version of these
simulations extending the simulated properties into the far-infrared
(principally for comparison with data from the Herschel satellite) has
now been completed (\cite{wilman2}).
The semi-analytic simulation (SAX) provides the properties of neutral
atomic (HI) and molecular (H$_2$) hydrogen in galaxies and associated
radio and sub-millimeter emission lines, i.e. the HI-line and various
CO transition lines. This simulation relies on the Millennium
simulation of cosmic structure (\cite{springel}), which reliably recovers comoving
length scales from 10 kpc to several hundred Mpc and galaxies with
cold hydrogen masses (HI+H2) above 10$^8$~M$_{\odot}$.
There are two versions of the SAX database, reflecting the different
versions of the Millennium simulation: the full Millennium simulation
(s$_{box}$ = 500/h Mpc; $\sim$ 685 Mpc) and the smaller test version,
called the Milli-Millennium simulation (s$_{box}$ = 62.5/h Mpc; $\sim$
85.6 Mpc) where s$_{box}$ defines the diameter of the simulated box.
Both of the SAX databases have been produced by constructing a mock
observing cone from the corresponding simulation box. The opening
angle or the field of view (FoV) of these simulations therefore depends
on the values of the maximum redshift one requires (z$_{max}$), e.g. for a
redshift of 1 the FoV of the simulation would be 12$\times$12
deg$^2$. More information about the FoV of the simulation can be
obtained via the simulation page. Currently, radio continuum data is
not available, although efforts are being made to add this information
(Obreschkow et al. 2009a; Obreschkow et al. 2009b; Obreschkow et
al. 2009c)
S3-Tools is used to build mock radio maps or cubes in which the
radio emission of the extragalactic radio sources could be combined with
the diffuse radio emission of the GSM or EOR.
\subsection{Array simulation}
The developed array simulator\footnote{A copy of this simulator can be
downloaded via www-astro.physics.ox.ac.uk/$\sim$hrk} is based on ``classical'' AIPS and
ParselTongue (\cite{aips,PT2006}), the Python interface to AIPS. The
core function of the array simulator makes use of the AIPS task {\tt
UVCON}. The basic input of this task is a list of antenna locations,
together with properties of each antenna, such as diameter, system
temperatures and aperture efficiencies. Additional inputs are the
total time of observation, the integration time per visibility and the
input sky model which also defines the pointing position on the
sky. The output is a standard UV-FITS data file in which the
visibilities correspond to the input model with added Gaussian noise
appropriate for the specified antenna characteristics. Dirty or
deconvolved images can be produced by invoking the task {\tt IMAGR}.
If the input model consists of a cube instead of a 2-D image then two
different simulation paths can be used to generate the ``observed
sky''. If the output map is to be a continuum image formed at the
central observing frequency, the visibilities are produced per
frequency step and are finally merged into a single visibility set. If
instead the output should be a 3-D datacube, the simulation produces a
unique visibility set at each frequency step, each of which is then
imaged independently. Each plane is finally combined into a cube by
using the task {\tt MCUBE}. For cases where the duration of the
observation equals or exceeds that necessary for a full UV coverage, the
visibilities are generated for a full synthesis and the noise scaled
down accordingly. This cuts down on both processing overheads and the size
of the resulting UV data files.
This style of simulation is well suited to the investigation of the
completeness of surveys as well as to the understanding of the imaging
capabilities when observing individual galaxies. In particular one can
investigate the quality of snapshot observations for different array
layouts. There are, however, several AIPS--based limitations which
have a direct influence on the questions we can ask, and more
complicated simulations are needed to address these. Currently the
number of array elements which AIPS can handle is limited to 255
(E. Greisen 2009, private communication), which is sufficient for most of
the current arrays but is not enough for a full simulation of, for
example, the LOFAR array at the level of individual dipole
elements. One can circumvent this to some extent by combining several
elements into a single station. This is relevant since LOFAR does
indeed perform correlations between the beamformed data from each
station, and not between individual dipoles. This is a model which the
SKA is likely to follow in order to minimize the data stream.
Two properties of a real interferometric array which are absent from
this simulation framework are primary beam effects and bandpass
variability. These are two factors which significantly affect the
dynamic range and image quality of a real observation, particularly
the former in the case of an aperture array.
Despite the various limitations we feel that AIPS is likely to be the
most tested software package and our approach provides a basic
simulation pipeline which can provide a reliable check for more
complicated simulation efforts (e.g. within MeqTrees or CASA).
In the redshift range around z~=~1 the SKA's mapping speed has the
potential to be revolutionized by mid-frequency aperture arrays (AA).
As mentioned before, the simulation software is limited to 255
antennas and it is necessary to combine the aperture arrays
into individual stations. The following equation shows how to
calculate the total area of an aperture array, assuming Hertzian
dipoles, and therefore the equivalent dish diameter of an AA station:
\begin{equation}
\rm A_{\rm AA} = N_{\rm tile} N_{\rm dipole} \frac{3}{8 \pi} (\frac{c}{\nu})^2 N_{\rm stokes},
\end{equation}
\noindent where N$_{\rm tile}$ is the number of tiles and N$_{\rm
dipole}$ is the number of dipoles per tile. Assuming a square
kilometre of collecting area the diameter of a SKA AA station would be
56~m. Note that an aperture efficiency of 80\% and a total of 255
stations has been assumed for a S$^3$ default SKA realization
(S$^3_{\rm real}$). This value only corresponds to observations
towards the zenith, and the effective station area varies with
elevation. However simulations to investigate the performance of such
a simplified AA are still valuable because the layout of the array
will directly affect the synthesized beam and therefore the imaging
capability. Figure~3 shows a cut through the dirty beam pattern
showing a high sidelobe pattern which can be minimized by varying the
station distribution within the array configuration (e.g. AntConfig
can be used for this purpose; www.kat.ac.za/public/wiki/AntConfig).
The current design of the aperture array has two layouts that have the
same core, but differ in the outer regions (R. Bolton 2009, private
communication). The core has a radius of 2.5~km and has 165 randomly
placed stations which are separated at least by 96~m.
\begin{itemize}
\item{The ``concentrated'' layout has 72 stations beyond the core
placed in 5 spiral arms out to 10 km (radius). Beyond this 13
stations are placed on the same spiral arms out to 180~km (S$^3_{\rm real}$C).}
\item{The ``not-concentrated'' layout has 85 stations beyond the core
logarithmically placed in 5 spiral arms out to 180~km (S$^3_{\rm real}$NC).}
\end{itemize}
\noindent Figure~2 shows the UV coverage of a test simulation of 1
hour duration. For displaying purposes each visibility has an
integration time of 10~minutes. The high density of visibilities and
the filled central core will provide high sensitivity and image
fidelity when observing diffuse HI emission. However deconvolution of
such structure will be difficult because of the dirty beam pattern
shown in Fig.~3. The broad sidelobe pattern will add ambiguities
during attempts to recover diffuse, extended HI emission.
\begin{figure}
\centering
\includegraphics[width=8cm,clip]{HRK_FIG3.PS}
\caption{The UV coverage of the ``concentrated'' aperture-array
layout (S$^3_{\rm real}$C). The simulation is based on 1~hr integration with an
integration time of 10 minutes per visibility. Note that the 10
minutes integration per visibility is for displaying purposes
only.}
\label{FIG:uvcover}
\end{figure}
\begin{figure}
\centering
\includegraphics[bb=40 140 565 642,width=8cm,clip]{HRK_FIG4.PS}
\caption{Cut through the synthesized beam (S$^3_{\rm real}$C). Due to the high density
of stations in the core area, the synthesizes beam profile
plateaus around 30 arcsec and shows strong sidelobes at a level of
$\sim$10\% . }
\label{FIG:beamcut}
\end{figure}
A final limitation of the simulation pipeline arises at the imaging stage.
The images that can be handled by AIPS are limited to 8096 pixels per
dimension. For the current SKA layout the maximum baseline length of
360~km would provide sub-arcsec angular resolution at frequencies
higher than 200~MHz. Such resolution limits the field-of-view that
can be simulated, and producing a simulation covering several square
degrees requires the sky to be divided into sub-patches. For example,
the spatial resolution of the array configuration at 700~MHz is
0.3~arcsec. If we Nyquist-sample the sky then these images
need to have a pixel resolution of about 0.1~arcsec per pixel. Taking
the AIPS limitation into account, such images would cover a sky area of
0.67$\times$0.67 deg$^2$ and thus to simulate a 4~deg$^2$ field 9
sub-images are necessary.
\subsection{The HI e2e simulation}
\begin{figure}
\centering
\includegraphics[bb=30 180 540 580,width=8.5cm,clip]{HRK_FIG2.PS}
\caption{Flux density of HI masses with a fixed line width of
164~kms$^{-1}$ versus redshift. The three lines show the
entire HI mass range of the HIPASS survey. Note for detection
arguments one needs to assume a channel resolution similar to the
HI line width, otherwise the sensitivity is reduced via
Eq.~\ref{eq:sens}.}
\label{FIG:sensi}
\end{figure}
The anticipated HI simulation will make use of the SAX simulated
sky. These simulations do not include a physical model of the
associated continuum emission. However, to evaluate the influence of
continuum emission to the HI simulation a mock continuum
component may be added to the line emission (using the task {\tt
UVMOD}). The HI emission of each galaxy has been pasted into the
model cube by using S3-tools, selecting the ``Oxford'' HI templates in the map making tool. The
following equation is given to provide a general understanding of the detectability of HI.
The basic relationship between HI mass and HI
line flux density is defined by
\begin{equation}
M_{\rm HI} = \frac{2.36 \times 10^5}{(z + 1)} {\rm D^2_L} \frac{\pi}{\sqrt{\rm 2 ln(2)}} {\rm S_p} \Delta {\rm V} ,
\end{equation}
\noindent where D$_{\rm L}$ is the luminosity distance [Mpc], assuming
a Gaussian line profile where $S_{\rm p}$ is the peak flux density
[Jy] and $\Delta {\rm V}$ is the full line width measured at
half-maximum [FWHM; kms$^{-1}$]. To investigate how much of the HI
mass function one can trace with the SKA, the expected HI flux density
is shown in Fig.~4. The HI flux density has been calculated by using
average values over the HIPASS sample for the HI line width (FWHM; 164
kms$^{-1}$) and for the HI masses (range between 2$\cdot$10$^{7}$ ---
6$\cdot$10$^{10}$~M$_{\odot}$) (\cite{zwaan1,zwaan2}). Note that in Fig.~4
to detect the HI mass at specific redshift one assumes that the entire
HI signal is confined within one channel. For a freely chosen 1-sigma
rms noise of 1~$\mu$Jy one could detect M$^*$$_{\rm HI}$ galaxies at
the 3-sigma level out to a redshift of 1 (Note that a rms noise of
1~$\mu$Jy would correspond to an integration time of 36 hours, assuming
a channel width of 250~kHz and a T$_{\rm sys}$ of 50~K.). This is the
anticipated aim of the SKA and the HI simulations need to be able to
investigate such a cosmic volume. Furthermore, if there is little
evolution in the HI mass function one could expect to detect the most
massive HI galaxies up to redshifts of about 4.
The FoV of the Millenium simulation box corresponds to an area of
116$\times$116~deg$^2$ at redshift 0.1, which shrinks to
5.4$\times$5.4~deg$^2$ at redshift 4. At low redshifts a HI
simulation would help to investigate potential systematic errors in a
measurement of the faint end of the HI mass function. Such a
simulation would require around 30276 individual simulation runs,
based on the field-of-view limitations per-facet, as discussed above.
This is a computationally very expensive exercise, and is only
worthwhile once the SKA array configuration has been finalised. For
investigating and optimising array configurations it makes more sense
to perform smaller scale simulations which still allow us to quantify
the performance of the array.
\begin{table}
\caption[]{Simulation parameters at various redshifts. Bandwidth corresponding to a fixed redshift interval of $\Delta z = 0.1$. The fraction of the HIPASS volume to the co-moving volume of 1 square degree and $\Delta z = 0.1$ (e.g. at redshift 1 such volume would correspond to 0.0008~Gpc$^3$ deg$^{-2}$). The FoV of the SAX sky simulations (values are obtained from s-cubed.physics.ox.ac.uk). Expected HI sources per square degree
for a flux limit of 3~$\mu$Jy (assuming a rms of 1~$\mu$Jy). For example to investigate the HIPASS volume at redshift 1 a 4$\times$4 deg$^{2}$ field must be simulated and need to be split into 36 individual sub-array-simulations to handle the AIPS image limitation.}
\label{tab:1}
$$
\begin{array}{ccccc}
\hline
\noalign{\smallskip}
{\rm redshift} & {\rm bandwidth} & \frac{\rm HIPASS vol}{\rm co-volume} & {\rm SAX sky} & {\rm SAX sources}\\
& {\rm [MHz]} & {\rm [deg^{2}]} & {\rm [deg^2]} & {\rm [deg^{-2}]}\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
0.5 & 59 & 37 & 21.4 \times 21.4 & 15136\\
1 & 34 & 16 & 12.0 \times 12.0 & 7136\\
1.5 & 22 & 12 & 9 \times 9 & 1897 \\
2 & 15 & 11 & 7.5 \times 7.5 & 816 \\
\noalign{\smallskip}
\hline
\end{array}
$$
\vspace{-0.2cm}
\end{table}
We propose such a HI simulation here, which will partly match the
co-moving volume of the HIPASS dataset. This requires a smaller number
of individual simulation runs, and it is thus possible to re-run the
simulation several times.
HIPASS is a blind HI radio survey of the sky at declinations
southwards of 25 degrees. However to calculate the co-moving volume we
use the initially presented catalogue covering the hole southern
hemisphere. The survey covers a sky area of 21314~deg$^2$ , a redshift
range of z = 0.001--0.042 (corresponding to a co-moving volume of
0.013 Gpc$^3$), and has a sensitivity limit approximately
corresponding to a peak flux density of 0.05~Jy (see \cite{zwaan1} for
a full description of the sample completeness). This survey yields a
total number of 4315 sources, all of which are subsequently identified
with optical galaxies (\cite{doyle}).
Using the SAX simulation to mimic the HIPASS survey no
extrapolation of the sky simulation is needed because the simulated
volume is large enough to cover the entire HIPASS volume. In fact this
can been shown by comparing the HIPASS co-moving distance of 171~Mpc
(z=0.042 and using the cosmological parameter of the Millenium
simulation) which is smaller than the radius of the simulation box,
i.e. 500/2~Mpc/h $\sim$ 342~Mpc.
Using the online query of the SAX simulation and defining the redshift
and the sensitivity specifications\footnote{SAX query input: select
count(*), from galaxies$_{-}$line where zapparent between 0.001 and 0.042
and hiintflux*hilumpeak$>$0.05 } of the HIPASS catalogue one obtain
4545 sources. This means that the SAX simulation predicts that the
HIPASS catalog contains 4545 sources, which matches the observed
number of 4315 sources within 5~\%. This difference can be partially
attributed to the fact that HIPASS yields a continuously varying
completeness function rather than a strict peak flux limit. Assuming
no flux limits the SAX simulation predicts 772120 source within
the HIPASS volume, i.e. 180-times more sources than picked up by HIPASS.
However Table~\ref{tab:1} shows the expected number counts assuming a
minimum peak flux density of 3~$\mu$Jy. The simulation still contain
enough source to address e.g. the study of the faint end of the HI
mass function and to make statistical significant predictions out to
high cosmological distances.
\begin{figure}
\centering
\includegraphics[bb=70 195 526 605,width=8.7cm,clip]{HRK_FIG5.PS}
\caption{Example of an HI input sky. The cube has 11 spectral
channels covering a frequency range of 11$\times$62.5~kHz. For
illustration purposes the image shows the channel averaged line
emission (using {\tt SQASH}) and in addition the averaged image has been
convolved with a 0.3~arcsec Gaussian beam (using {\tt CONVL}).}
\label{FIG:mom0}
\end{figure}
Figure~5 shows the HI intensity map of an input cube. No
continuum emission has been added and the individual galaxies are
unresolved. The corresponding ``observed sky'' is shown in Fig.~6. In
order to analyse these images the automated source finder,
Duchamp\footnote{www.atnf.csiro.au/people/Matthew.Whiting/Duchamp}
will be used to generate ``observed'' source catalogues for comparison
to the input catalogues. An important part of this analysis will be to
produce idealised input images with purely Gaussian noise of an
equivalent level to cross--check the reliability and completeness of
the source finding software in the presence of the image artifacts
introduced by an interferometer.
\begin{figure}
\centering
\includegraphics[bb=30 150 540 645,width=8cm,clip]{HRK_FIG6.PS}
\caption{Example of a simulated HI sky. The rms of this image
including sources is of the order of 8~$\mu$Jy. The cleaned image
(using 200 clean components) displays the channel averaged line
emission.\vspace{0cm}}
\label{FIG:mom0sim}
\end{figure}
\section{Conclusions}
A full e2e simulation path has been developed. The array simulator
makes use of images or datacubes based on the S$^3$ catalogues to
simulate an observation. In the simulations no errors due to
calibration or telescope hardware have been introduced, but a
simplistic treatment of gain and phase errors could be included at
later stages. The analysis of the simulated data sets (either images
or cubes) will make use of automated source finder software. In
addition to this, a more sophisticated analysis of the simulated data
is possible because the simulator produces visibility data as well as
images. For example one could investigate the sensitivity of an array
to diffuse emission by using the UV-gap ($\frac{\Delta{\rm U}}{\rm
U}$) analysis techniques (\cite{lalduu}) or by analysing the
statistics of Fourier phases (\cite{francois}).
The proposed HI simulations match the requirements for studying the
capability of the SKA aperture array when imaging HI structures in
nearby galaxies. Such simulations will also investigate the impact of
different telescope designs on the proposed SKA blind HI surveys. The
input HI sky will have an equivalent co-moving volume to that of the
HIPASS survey, and this relatively small volume makes it feasible to
re-run the simulations with different parameters within a reasonable
time. With this setup we are able to analyse the ``observed skies''
and study the influence that the antenna layout has on a blind HI
galaxy survey by investigating following points:
\begin{itemize}
\item[-] completeness (peak flux density)
\item[-] positional accuracy
\item[-] redshift determination
\item[-] necessity of subtracting continuum emission and our ability to do so
\end{itemize}
\begin{acknowledgements}
HRK would like to thank Rosie Bolton for the two array configuration files.
\end{acknowledgements}
|
\section{Introduction}\hspace{5mm}
In 1991, the two-parameter family of $p,q$-deformed quantum
oscillators has been introduced \cite{Cha-Ja}. Soon after, in
\cite{Ar_Fib} the family was named "Fibonacci oscillators", due to
the basic property of the family encoded in the three-term linear
recurrence relation \cite{Ar_Fib}
\begin{equation}
E_{n+1}=\lambda E_n+\rho E_{n-1}, \hspace{12mm} n\geq 1,
\hspace{12mm} \lambda, \rho \ \epsilon \ \mathbb{R} ,\label{1}
\end{equation}
for any three consecutive values of energy levels from the spectrum
of respective deformed quantum oscillators, with $\lambda$ and
$\rho$ definite coefficients that depend on specific model of
deformed oscillator. The usual Fibonacci numbers form the sequence
generated by equation like (\ref{1}) in which $\lambda=\rho=1$ and
the first two members of the sequence are fixed as 1 and 1; the
property states that each number in the sequence results from the
sum of two preceding ones.
Following \cite{Ar_Fib}, in our paper we call the Fibonacci
oscillators those oscillators whose any three energies $E_{n-1}$,
$E_n$, $E_{n+1}$ satisfy the Fibonacci relation (FR) given by
(\ref{1}).
They form the Fibonacci class of oscillators.
There exist a number of known oscillators that are the Fibonacci ones.
For instance, usual harmonic oscillator certainly is the Fibonacci
oscillator in the sense of eq. (\ref{1}) with $\lambda=2$ and
$\rho=-1$.
As already mentioned, the family of $p,\!q$-oscillators which
contains well-known models of one-parameter $q$-oscillators, belongs to the
Fibonacci class.
However, the question naturally arises about possible existence of other models of
deformed oscillators which satisfy (\ref{1}), i.e., possess the
Fibonacci property (FP).
Recently, in conjunction with generalized Heisenberg algebras \cite{SCRM}
possible extensions of the FP were studied, either in the direction
of nonlinearization of the relation, or towards so-called $k$-step
extension (in particular, Tribonacci sequence), see \cite{Schork}.
Also, in ref. \cite{GR-5} some classes of non-Fibonacci (namely
so-called $k$-bonacci) oscillators have been explored.
Our goal in this paper is two-fold.
First, we study from the viewpoint of possessing the FP some
multi-parameter extended
families of deformed oscillators, including 3- and 4-parametric ones from
\cite{Chung,Borzov, Mizrahi} as well as their 5-parameter extended
model given in \cite{Burban}.
We prove that all these multi-parameter deformations belong to
the class of Fibonacci oscillators. On the other hand, we examine
certain as yet not well studied deformation called $\mu$-oscillator,
which appeared in \cite{Jan}, and present the proof that this model
lives outside the Fibonacci class (that is, it is non-Fibonacci one).
This conclusion served for us as motivation to explore possible
(still linear) extensions{\footnote{For comparison see e.g. ref. \cite{Wei}
for an interesting example of nonlinear three-term relation giving the values
of (non-equidistant) area spectrum of 5-dim Gauss-Bonnet black hole.}}
of the FR and to find among them most adequate and natural form of
generalization which is suitable for the $\mu$-oscillator.
In other words, our second goal is to provide a proper generalization
of the FP that is adequate for both the $\mu$-oscillator and its several
multi-parameter extensions. The employed particular generalization
of the FP can be termed "quasi-Fibonacci"\ property.
Correspondingly, the $\mu$-oscillator \cite{Jan} and its appropriate
extensions belong to the class of quasi-Fibonacci (QF) oscillators.
In some sense, this can also be viewed as the deformation of quantum
oscillator whose energy values form a "locally-Fibonacci"\ sequence
of values.
\section{Deformed oscillators of the Fibonacci class}\hspace{5mm}
In this section we examine the energy spectra of the models of
multi-parameter deformed oscillators from \cite{Chung,Borzov,
Mizrahi,Burban}, from the viewpoint of possessing the FP.
But first it will be useful to remind main facts about some well-known
models of deformed oscillators.
General approach for treatment of deformed oscillators
involves the notion \cite{Man,Mel,Bona} of the structure function
$f(N)$ or $\varphi(N)$ that serves to define the deformation.
In terms of the structure function $\varphi(N)$ given by the
equalities
\begin{equation}
\varphi(N)=a^{\dagger}a\ , \hspace{8mm} \varphi(N+1)=aa^{\dagger},
\label{2}
\end{equation}
the most general commutation relation can be written in the form
\[\varphi(N+1)-F(N)\varphi(N)=G(N)\ ,
\]
with $F(N)$ and $G(N)$ real functions. Every special choice of
$F(N)$ and $G(N)$ provides, through solving the latter equation
\cite{Mel}, the particular explicit form of the structure function
$\varphi(N)$ and thus the corresponding model of deformed
oscillator. With fixed $\varphi(N)$, the commutation relations for
the generating elements $a$, $a^{\dagger}$ and $N$ of the concrete
model of deformed oscillator are given as
\begin{equation}
[N,a^{\dagger}]=a^{\dagger}, \hspace{8mm} [N,a]=-a\ , \label{3}
\end{equation}
\begin{equation}
aa^{\dagger}-a^{\dagger}a=\varphi(N+1)-\varphi(N). \label{4}
\end{equation}
That means we use the form of basic commutation relation such that
$F(N)=1$ and $G(N)=\varphi(N+1)-\varphi(N)$. In this case, fixation
of the structure function $\varphi(N)$ completely specifies the
(deformed oscillator) model.
\subsection{Well-known Fibonacci oscillators}\hspace{5mm}
Let us recall some most popular models of deformed oscillators which
involve one or two deformation parameters and belong to the
Fibonacci class of oscillators.
$\bullet$ {\it {Arik-Coon (AC) model} } \cite{AC}
\vspace{2mm} \noindent This most early known model is given by the
commutation relations (\ref{3}) and
\begin{equation}
aa^{\dagger}-qa^{\dagger}a=1. \label{5}
\end{equation}
In what follows, we use for all the models the same definition of
the Hamiltonian as for the usual harmonic oscillator:
\begin{equation}
H=\frac{1}{2}(aa^{\dagger}+a^{\dagger}a)\ \label{6}
\end{equation}
where $\hbar \omega=1$ is meant.
Using the appropriately modified (deformed) version of the Fock
space \cite{Mel,Bona} wherein $H|n\rangle=E_n|n\rangle$, and
$\varphi_{AC}(N)|n\rangle=\varphi_{AC}(n)|n\rangle$, the energy spectrum of AC model is
\begin{equation}
E_n=\frac{1}{2}\Bigl(\varphi_{AC}(n+1)+\varphi_{AC}(n)\Bigr)=\frac{1}{2}\Bigl([n+1]_{AC}+[n]_{AC}\Bigr)=
\frac{1}{2}\biggl( \frac{q^{n+1}-1}{q-1}+\frac{q^n-1}{q-1}\biggr)
\label{7}
\end{equation}
where the definition of the $AC$-type $q$-bracket (for $X$ either a
number or an operator) and the structure function is given as
\begin{equation}
[X]_{AC}=\frac{q^X-1}{q-1} \hspace{8mm} \Leftrightarrow
\hspace{8mm} \varphi_{AC}(n)=\frac{q^n-1}{q-1}=[n]_{AC} \label{8}
\end{equation}
so that the latter satisfies (\ref{5}).
One can easily check that the AC-oscillator energies (\ref{7})
satisfy the FR (\ref{1}) if $\lambda=1+q$ and $\rho=-q$.
\vspace{2mm}
$\bullet$ {\it {Biedenharn-Macfarlane (BM) model} } \cite{Bied,Mcf}
\vspace{2mm} The model is determined by the commutation relations
(\ref{3}) and the relations
\begin{equation}
aa^{\dagger}-qa^{\dagger}a=q^{-N} , \hspace{12mm}
aa^{\dagger}-q^{-1}a^{\dagger}a=q^N. \label{9}
\end{equation}
The corresponding structure function which solves (\ref{9}) is
\begin{equation}
\varphi_{BM}(N)=\frac{q^N-q^{-N}}{q-q^{-1}}\equiv [N]_{BM}\ ,
\label{10}
\end{equation}
and the energy spectrum reads
\begin{equation}
E_n=\frac{1}{2}\Bigl(\varphi_{BM}(n+1)+\varphi_{BM}(n)\Bigr)=
\frac{1}{2}\biggl(
\frac{q^{n+1}-q^{-(n+1)}}{q-q^{-1}}+\frac{q^n-q^{-n}}{q-q^{-1}}\biggr).
\label{11}
\end{equation}
The BM oscillator satisfies the FR (\ref{1}) with $\lambda=q+q^{-1}$
and $\rho=-1$.
\vspace{2mm}
$\bullet$ {\it {$p,\!q$-deformed oscillator}} \cite{Cha-Ja}
\vspace{2mm} This model is determined by the commutation relations
(\ref{3}) and the relations
\begin{equation}
aa^{\dagger}-q a^{\dagger}a=p^{N} , \hspace{12mm}
\hspace{8mm} aa^{\dagger}-p\,a^{\dagger}a=q^N , \label{12}
\end{equation}
possessing the $q\leftrightarrow p$ symmetry.
The relevant structure function satisfying (\ref{12}) is
\begin{equation}
\varphi_{p,q}(N)=\frac{q^N-p^N}{q-p}\equiv[N]_{p,q}\ , \label{13}
\end{equation}
and hence the energy spectrum reads
\begin{equation}
E_n=
\frac{1}{2}\biggl(\frac{q^{n+1}-p^{n+1}}{q-p}+\frac{q^n-p^{n}}{q-p}\biggr).
\label{14}
\end{equation}
The model of $p,q$-deformed oscillator is popular enough: it
finds a number of interesting applications, see e.g.
\cite{AdG,SIGMA,Crn,Algin,Algin2,Ben,UFZh,Plethora}.
From it, by setting $p=1$ the AC-model is retrieved.
At $p=q^{-1}$ we recover the BM-type oscillator.
Finally, the peculiar case of $p=q$ corresponds to the so-called
Tamm-Dancoff (TD) deformed oscillator \cite{Odaka,Jagan}.
This model, though not so popular as the former two, nevertheless
possesses some interesting properties.
More details concerning the TD-type oscillator model including some of its
properties can be found in the paper \cite{GR-1}.
Note that the TD oscillator arises also as a particular
one-dimensional case in the $SU(d)$ covariant $d$-dimensional so-called
$q$-deformed Newton oscillator \cite{AAW}.
With (\ref{14}), the two-parameter $p,q$-model of deformed oscillators
satisfies the FR
(\ref{1}) for $\lambda=q+p$ and $\rho=-qp$.
Accordingly, the AC-oscillator is the Fibonacci oscillator with
$\lambda=1+q$ and $\rho=-q$, whereas the energy levels of the
BM-oscillator satisfy (\ref{1}) if we take $\lambda=q+q^{-1}$ and $\rho=-1$.
Finally, to confirm that the TD-model of deformed oscillator
does satisfy (\ref{1}), we put $\lambda=2q$ and $\rho=-q^2$.
\vspace{5mm}
\subsection{A family of multi-parameter deformed oscillators}\hspace{5mm}
In this subsection we focus on those deformed multi-parameter
oscillators which possess more then two deformation parameters in
their defining relations and thus generalize the $q$- and
$p,\!q$-deformed oscillators. It is of special interest to examine
the properties of such deformed oscillators in dependence on the
number of parameters.
Our main goal is to examine whether these models of oscillators satisfy
the FR (\ref{1}) and in case they don't, to search for a way of
appropriately modifying the FR.
As the example of multi-parameter oscillators consider the
$(p,q,\alpha,\beta,l)$-deformed oscillator family formulated in
\cite{Burban} and given by the defining relations
\[aa^{\dagger}-q^{l}a^{\dagger}a=p^{-\alpha N-\beta} ,
\hspace{8mm} aa^{\dagger}-p^{-l}a^{\dagger}a=q^{\alpha N+\beta}.
\]
Here $p,q,\alpha,\beta,l$ are the parameters of deformation.
Creation and annihilation operators act upon the Fock space basis
state $|n\rangle$ by the formulas
\begin{equation}
aa^{\dagger}|n\rangle=\frac{q^{\alpha n+\beta+l}-p^{-(\alpha
n+\beta+l)}}{q^l-p^{-l}}|n\rangle\ ,
\hspace{5mm}
a^{\dagger}a|n\rangle=\frac{q^{\alpha n+\beta}-p^{-(\alpha
n+\beta)}}{q^l-p^{-l}}|n\rangle. \label{16}
\end{equation}
With the Hamiltonian as in (\ref{6}) the eigenvalues which form the
energy spectrum are
\begin{equation}
E_n=\frac12\biggl(\frac{q^{\alpha n +\beta +l}-p^{-(\alpha n +\beta
+l)}}{q^l-p^{-l}}+ \frac{q^{\alpha n +\beta }-p^{-(\alpha n
+\beta)}}{q^l-p^{-l}}\biggr). \label{17}
\end{equation}
\noindent
Unlike the models mentioned above, for this 5-parameter model
the formula for $E_n$ cannot be written as
$E_n=\frac{1}{2}(\varphi_{n+1}+\varphi_n)$, but rather as
$E_n=\frac{1}{2}(\varphi_n+\varphi_{n+\frac{l}{\alpha}})$; only in
the 4-parameter case wherein $l=\alpha$ the presentation
$E_n=\frac{1}{2}(\varphi_{n}+\varphi_{n+1})$ is retrieved.
The models of deformed oscillators examined in Subsection 2.1
obviously follow from the $(p,q,\alpha,\beta,l)$-deformed oscillator
as particular cases .
Moreover, the 5-parameter family reduces at $l=1$ to the 4-parameter
deformed oscillator treated in Ref. \cite{Borzov,Mizrahi} which in its turn
goes over into the three-parameter deformed quantum oscillators
first formulated in \cite{Chung}.
\vspace{2mm}
\subsection{Fibonacci property of multi-parameter
oscillator models}\hspace{5mm}
The constants $\lambda$ and
$\rho$ in (\ref{1}) for the AC, BM and $p,q$-models of deformed
oscillators given in Subsection 2.1 can be deduced easily by solving
the corresponding systems of equations. The principal feature of
those models is that the coefficients $\lambda$, $\rho$ are indeed
constant, i.e., they do not depend on the eigenvalue $n$ of the
(particle number) operator $N$.
Suppose that (\ref{1}) is true for the extended 3-, 4- and
5-parameter oscillators. Then the following statement is true.
{\bf Proposition 1.} {\it The five-parameter deformed oscillator
whose energies $E_n$, $n\geq 0$, are given by (\ref{17}), satisfies,
with certain constants $\lambda$, $\rho$ each of the following two
relations}
\begin{equation}
\begin{cases} E_{n+1}=\lambda E_n+\rho E_{n-1}\ , \cr
E_{n+2}=\lambda E_{n+1}+\rho E_{n}\ , {\label{18}}
\end{cases}
\end{equation}
{\it and thus belongs to the class of Fibonacci oscillators}.
The proof proceeds through solving for $\lambda$ and $\rho$ the
system (\ref{18}) with account of (\ref{17}). Recall, these
$\lambda$ and $\rho$ should not depend on $n$. As follows from
(\ref{18}),
\begin{equation}
\lambda=\frac{E_{n+2}E_{n-1}-E_{n+1}E_n}{E_{n+1}E_{n-1}-E^2_n}\
,\hspace{5mm}
\rho=\frac{E^2_{n+1}-E_{n+2}E_n}{E_{n+1}E_{n-1}-E^2_n}\ . \label{19}
\end{equation}
Inserting from (\ref{17}) the expression for $E_j$ , $j=n-1, n, n+1,
n+2$, we obtain\
$ \lambda=B/C$\ where
$B=B(n;p,q,\alpha,\beta,l),\ C=C(n;p,q,\alpha,\beta,l)$, and
\[
B=\left(\frac{q^{\alpha(n+2)+\beta+l}-p^{-\alpha(n+2)-\beta-l} +
q^{\alpha(n+2)+\beta}-
p^{-\alpha(n+2)-\beta}}{q^l-p^{-l}}\right)\times
\]
\[\hspace{5mm} \times\left(\frac{q^{\alpha(n-1)+\beta+l}-
p^{-\alpha(n-1)-\beta-l}+
q^{\alpha(n-1)+\beta}-p^{-\alpha(n-1)-\beta}}{q^l-p^{-l}}\right)-
\]
\[ \hspace{5mm}
-\left(\frac{q^{\alpha(n+1)+\beta+l}-
p^{-\alpha(n+1)-\beta-l}+q^{\alpha(n+1)+\beta}-
p^{-\alpha(n+1)-\beta}}{q^l-p^{-l}}\right)\times
\]
\[
\hspace{5mm} \times\left(\frac{q^{\alpha n+\beta+l}-p^{-\alpha
n-\beta-l}+q^{\alpha n+\beta}-p^{-\alpha
n-\beta}}{q^l-p^{-l}}\right)\ ,
\]
\[
C=\left(\frac{q^{\alpha(n+1)+\beta+l}-p^{-\alpha(n+1)-\beta-l}+q^{\alpha(n+1)+\beta}-
p^{-\alpha(n+1)-\beta}}{q^l-p^{-l}}\right)\times
\]
\[ \hspace{5mm}
\times\left(\frac{q^{\alpha(n-1)+\beta+l}-p^{-\alpha(n-1)-\beta-l}+
q^{\alpha(n-1)+\beta}-p^{-\alpha(n-1)-\beta}}{q^l-p^{-l}}\right)-
\]
\[
\hspace{6mm} -\left(\frac{q^{\alpha n+\beta+l}- p^{-\alpha
n-\beta-l}+q^{\alpha n+\beta}-p^{-\alpha
n-\beta}}{q^l-p^{-l}}\right)^2.
\]
Calculation shows remarkable cancelation of $n$-dependence: we get
$B/C=B'/C'$ where
\[
B'=-q^{3\alpha}p^{2\alpha+l}-q^{3\alpha}p^{2\alpha}-q^{3\alpha+l}p^{2\alpha+l}-
q^{3\alpha+l}p^{2\alpha}-p^{-\alpha}-p^{-\alpha+l}-q^lp^{-\alpha}-
\]
\[
\hspace{7mm} -q^lp^{-\alpha+l}+q^\alpha
p^l+q^\alpha+q^{\alpha+l}p^l+q^{\alpha+l}+
q^{2\alpha}p^{\alpha+l}+q^{2\alpha}p^\alpha+
q^{2\alpha+l}p^{\alpha+l}+q^{2\alpha+l}p^\alpha\ ,
\]
\[
C'=-p^l-1-q^lp^l-q^l+2q^\alpha p^{\alpha+l}+ 2q^\alpha
p^\alpha+2q^{\alpha+l}p^{\alpha+l}+
\]
\[\hspace{7mm}
\hspace{-1mm}+2q^{\alpha+l}p^\alpha-q^{2\alpha}p^{2\alpha+l}-
q^{2\alpha}p^{2\alpha}-q^{2\alpha+l}p^{2\alpha+l}-
q^{2\alpha+l}p^{2\alpha}\ .
\]
Moreover, it is easy to check that $(q^\alpha+p^{-\alpha})C'=B'$.
Hence finally
\begin{equation}
\lambda=q^\alpha+p^{-\alpha}\ . \label{20}
\end{equation}
The same procedure applied to $\rho$ yields $\rho=D/E$ where
\[
D=q^{3\alpha}p^\alpha+q^{3\alpha+l}p^\alpha+ q^\alpha
p^{-\alpha}+q^\alpha p^{-\alpha+l}+
q^{\alpha+l}p^{-\alpha}+q^{\alpha+l}p^{-\alpha+l}-\]
\[
\hspace{5mm}
-2q^{2\alpha}p^l-2q^{2\alpha}-2q^{2\alpha+l}p^l-2q^{2\alpha+l}+
q^{3\alpha}p^{\alpha+l}+q^{3\alpha+l}p^{\alpha+l}\ ,
\]
\[
E=-p^l-1-q^lp^l-q^l+2q^\alpha p^{\alpha+l}+ 2q^\alpha
p^\alpha+2q^{\alpha+l}p^{\alpha+l}+
\]
\[
\hspace{6mm} +2q^{\alpha+l}p^\alpha-q^{2\alpha}p^{2\alpha+l}-
q^{2\alpha}p^{2\alpha}-q^{2\alpha+l}p^{2\alpha+l}-q^{2\alpha+l}p^{2\alpha}\
,
\]
and it is easily checked that $D=(-q^\alpha p^{-\alpha}) E$. As
result,
\begin{equation}
\rho=-q^\alpha p^{-\alpha}\ . \label{21}
\end{equation}
So we conclude: thus found $\lambda$ and $\rho$ in (\ref{20}) and
(\ref{21}) do solve (\ref{18}), and do not depend on $n$.
Hence we have proven the Proposition.
\section{A case for generalizing the Fibonacci property}\hspace{5mm}
In this section we will examine, from the
viewpoint of (non)validity of FP, the $\mu$-deformed oscillator
formulated in \cite{Jan}. We will prove that the FP is not valid in
that case. Then our next goal is to find generalization of the FP
appropriate for the $\mu$-oscillator.
For proper extension, we adopt the so-called "quasi-Fibonacci"\ (QF)
property encapsulated in the QF relation satisfied by members of the
QF sequence. But, first let us recall necessary setup of the
$\mu$-deformed oscillator.
\subsection{Nonlinear $\mu$-deformed oscillator} \hspace{5mm}
The $\mu$-oscillator \cite{Jan} is defined in terms of the unital
algebra whose generating elements $a$, $a^{\dagger}$ and $N$
satisfy
\[
[N,a]=-a\ , \hspace{10mm} [N,a^{\dagger}]=a^{\dagger} \ ,
\hspace{10mm} [N,aa^{\dagger}]=[N,a^{\dagger}a]=0\ ,
\]
\[aa^{\dagger}-a^{\dagger}a=\varphi_{\mu}(N+1)-\varphi_{\mu}(N)\ ,\]
and its structure function is given as
\begin{equation}
a^{\dagger}a=\varphi_{\mu}(N)\equiv\frac{N}{1+\mu N}, \hspace{10mm}
aa^{\dagger}=\varphi_{\mu}(N+1)\equiv\frac{N+1}{1+\mu(N+1)}
\label{22}
\end{equation}
with $\mu$ a deformation parameter.
That is, the basic commutation relations for the $\mu$-oscillator
are eq. (\ref{3}) and
\begin{equation}
[a,a^{\dagger}]=\frac{N+1}{1+\mu (N+1)}-\frac{N}{1+\mu N}\ .
\label{23}
\end{equation}
In the $\mu$-deformed version of Fock space, with normalized ground
state $|0\rangle$ such that
\begin{equation}
a|0\rangle=0, \hspace{5mm} N|0\rangle=0, \hspace{5mm}
\varphi_{\mu}(0)=0, \label{24}
\end{equation}
we have the infinite set of basis states
\begin{equation}
|n\rangle=\frac{(a^{\dagger})^n}{\sqrt{\varphi_{\mu}(n)!}}|0\rangle\
, \hspace{8mm} \langle n|m \rangle=\delta_{mn}, \hspace{8mm}
n,m=0,1,2,..., \label{25}
\end{equation}
$\varphi_{\mu}(n)!=\varphi_{\mu}(n)\varphi_{\mu}(n-1)...\varphi_{\mu}(1)$
on which
\[ N|n\rangle=n|n\rangle, \hspace{8mm} \varphi_{\mu}(N)|n\rangle=\varphi_{\mu}(n)|n\rangle\ , \]
\[\langle n-1|a|n \rangle=\langle
n|a^{\dagger}|n-1\rangle=\sqrt{\varphi_{\mu}(n)}=\Bigl(\frac{n}{1+\mu
n}\Bigr)^{\frac{1}{2}}.\]
Since
\begin{equation}
\varphi_{\mu}(n)=\frac{n}{1+\mu n}\ , \label{26}
\end{equation}
the energy spectrum of this model reads
\begin{equation}
E_n=\frac12\Bigl(\varphi_{\mu}(n+1)+\varphi_{\mu}(n)\Bigr)=\frac{1}{2}\biggl(
\frac{n+1}{1+\mu(n+1)}+\frac{n}{1+\mu n}\biggr) \label{27}
\end{equation}
where the deformation parameter $\mu$ is assumed to satisfy $\mu\geq
0$. Setting $\mu=0$ recovers the known formulae for the standard
quantum harmonic oscillator.
\subsection{Non-Fibonacci nature of the $\mu$-oscillator}
\hspace{5mm}
Suppose the $\mu$-oscillator, with $\varphi(N)$ given in (\ref{22}),
obeys the FP stating that
\begin{equation}
E_{n+1}-\lambda E_n-\rho E_{n-1}=0 \label{28}
\end{equation}
with {\it constant} $\lambda$ and $\rho$. With account of (\ref{27})
this rewrites as
\begin{equation}
\frac{n+2}{1+\mu(n+2)}+\frac{n+1}{1+\mu(n+1)}- \lambda \biggl(
\frac{n+1}{1+\mu(n+1)}+\frac{n}{1+\mu n}\biggr)- \rho
\biggl(\frac{n}{1+\mu n}+\frac{n-1}{1+\mu (n-1)}\biggr)=0\ .
\label{29}
\end{equation}
Then the following statement is true.
\vspace{2mm}
{\bf Proposition 2.} {\it The $\mu$-oscillator is not the
Fibonacci one, i.e., (\ref{29}) fails for it.}
\vspace{1mm}
To prove, we take into account the equality, stemming from (\ref{29}),
of the two polynomials one of which is zero.
This leads to the following set of equations:
\vspace{3mm} $n^4$:\ \ \ \ $-1+\lambda+\rho=0$\ ;
\vspace{3mm} $n^3$:\ \ \ \ $\lambda(3+2\mu)+\rho(3+2\mu)-3-2\mu=0$\
;
\vspace{3mm} $n^2$:\ \ \ \
$\lambda(6+9\mu-2\mu^2)+\rho(6+7\mu-2\mu^2)- 6-11\mu+2\mu^2=0$\ ;
\vspace{3mm} $n^1$:\ \ \ \
$\lambda(\mu^{2}(\mu^{-1}-1)+(\mu^{-1}+2)(2\mu^{-2}+\mu^{-1}-2))+
\rho(-\mu^{2}(\mu^{-1}+1)+$
\vspace{3mm}\hspace{6mm} \ \ \
$+(\mu^{-1}+2)(2\mu^{-2}-\mu^{-1}-2))-2-10\mu+\mu^{2}+4\mu^3=0$\ ;
\vspace{3mm} $n^0$:\ \ \ \
$2\mu^2-2+\mu(1+2\mu)[(\lambda-1)(\mu^{-1}-1)+\rho(\mu^{-1}+1)]=0$\
. \vspace{2mm}
\noindent We have to solve this set for $\lambda$ and $\rho$.
The first two equations yield $\lambda+\rho=1$.
However, insertion this in the rest of equations leads us to
inconsistency, as there is no solution satisfying the whole system,
for arbitrary $\mu$. So we conclude: the $\mu$-oscillator with
$\varphi(n)$ and $E_n$ from (\ref{26})-(\ref{27}) {\em does not
possess the Fibonacci property} (\ref{28}), with constant $\lambda$
and $\rho$.
Therefore, we have to find a modification of the Fibbonacci property
which would be adequate for the $\mu$-oscillator and also for some
models that extend it.
\begin{center}
\it{Possible ways of generalizing the FP}
\end{center}
Among possible modifications, we could try an extended version of
the usual Fibonacci relation (\ref{1}) obtained merely by adding
extra terms. Say we could consider the $k$-term extended (so-called
$k$-bonacci) relation of the form (see e.g. \cite{Schork,GR-5}):
\[
E_{n}=\alpha_1 E_{n-1}+\alpha_2E_{n-2} +\alpha_3
E_{n-3}+...+\alpha_{k} E_{n-k}\ .
\]
For the $\mu$-oscillator, let us check the first extended case of
$k=3$ (so-called "Tribonacci relation"). It is proved that the
$\mu$-oscillator does not satisfy the Tribonacci relation. The proof
goes in complete analogy with the above case of $k=2$, see
Proposition 2: by deriving the corresponding system of (now six)
equations and then trying to solve that system. As result, we are
led to the inconsistency of the system and thus to the conclusion
that the $\mu$-oscillator does not satisfy the Tribonacci relation.
Moreover, this negative result for the $\mu$-oscillator extends to
all other cases $k\geq 3$ of $k$-bonacci relation.
Next, the most radical possibility would be to search for some
nonlinear generalization of (\ref{1}), say in the form
$E_{n+1}=F(E_n, E_{n-1})$ with $F(x, y)$ an appropriate function
\cite{SCRM}. We prefer however to {\it preserve both the linearity
and the three-term form} of relation. But some price should be payed
for such a choice, and the price is nothing but the loss of constant
nature of $\lambda$ and $\rho$. In other words, these coefficients
inevitably should be $n$-dependent.
\subsection{Quasi-Fibonacci property of $\mu$-deformed oscillator}\hspace{5mm}
We will call "quasi-Fibonacci"\ oscillators those deformed
oscillators the energy spectrum of which satisfy the extended or
quasi-Fibonacci (QF) relation involving
$\lambda=\lambda(n)\equiv\lambda_n$ and $\rho=\rho(n)\equiv\rho_n$:
\begin{equation}
E_{n+1}=\lambda_n E_n+\rho_n E_{n-1}. \label{30}
\end{equation}
We will show that the $\mu$-deformed oscillator belongs to the set
of QF oscillator models. That is, by passing from the Fibonacci
property to the QF one, we will prove that it is the QF property
which is adequate to the energy spectrum of $\mu$-oscillator.
There exist three different ways of finding $\lambda_n$ and $\rho_n$
needed to prove the QF property.
\subsubsection{ First way of finding
$\lambda_n$, $\rho_n$ (by the "splitting" ansatz)}\hspace{5mm}
To start with, let us recall that the energy spectrum $E_n$ of the
deformed oscillator consists of two terms given by the structure
function $\varphi(n)$:
\begin{equation}
E_n=\frac{1}{2}\bigl(\varphi(n)+\varphi(n+1)\bigr)\ . \label{31}
\end{equation}
From (\ref{30}) and (\ref{31}), using the "splitting", we have the
following system of equations:
\begin{equation}
\begin{cases} \varphi(n+1)=\lambda_n \varphi(n)+\rho_n \varphi(n-1) , \cr
\varphi(n+2)=\lambda_n \varphi(n+1)+\rho_n \varphi(n) . \label{32}
\end{cases}
\end{equation}
Solving the latter system, for $\lambda_n$ and $\rho_n$ we have
\begin{equation}
\lambda_n=\frac{\varphi(n+1)-\rho_n\,\varphi(n-1)}{\varphi(n)}\ ,
\hspace{5mm}
\rho_n=\frac{\varphi(n+2)\varphi(n)-\varphi^2(n+1)}{\varphi^2(n)-
\varphi(n+1)\varphi(n-1)}\ . \label{33}
\end{equation}
From (\ref{33}), using the expression (\ref{26}) for $\varphi(n)$ of
the $\mu$-oscillator, we finally obtain
\begin{equation}
\lambda_n=2\frac{(1+\mu(1+2n))}{(1+2\mu n)}\cdot\frac{(1+\mu
n)}{1+\mu(n+2)}\ , \label{34}
\end{equation}
\begin{equation}
\rho_n=-\frac{(1+2\mu(n+1))}{(1+2\mu n)}\cdot \frac{(1+\mu
n)(1+\mu(n-1))}{(1+\mu(n+2))(1+\mu(n+1))} \ . \label{-35}
\end{equation}
It is clear that, for consistency, from (\ref{34}) and (\ref{-35})
in the limit $\mu\rightarrow 0$ we should recover the {\it
constants}:
\begin{equation}
\lim_{\mu\rightarrow 0}\lambda_n=2\ , \hspace{5mm}
\lim_{\mu\rightarrow 0}\rho_n=-1\ , \label{-36}
\end{equation}
i.e., the values $\lambda$, $\rho$ of the usual harmonic oscillator.
Obviously, that is true. Hence we conclude that the $\mu$-oscillator
satisfies (\ref{30}) with (\ref{27}) and (\ref{34}), (\ref{-35}),
and thus is quasi-Fibonacci one.
\subsubsection{Second way to find $\lambda_n$ and
$\rho_n$ (by the substitution ansatz)}\hspace{5mm}
Recall that most general form of defining commutation
relation looks as \cite{Mel}
\begin{equation}
\varphi(N+1)-F(N)\varphi(N)=G(N)\ . \label{-37}
\end{equation}
By treating this as the recursion relation and fixing the initial
conditions
\begin{equation}
\varphi(0)=0, \hspace{5mm} \varphi(1)=G(0)=\varepsilon\
,\hspace{5mm} \varepsilon \ \epsilon \ \Re, \label{38}
\end{equation}
the structure function $\varphi(n)$ can be obtained as
\begin{equation}
\varphi(n)=F(n-1)!\Biggl(\varepsilon+\sum^{n-1}_{j=1}\frac{G(j)}{F(j)!}\Biggr)\
,\label{39}
\end{equation}
where $F(j)!=F(j)F(j-1)...F(1)$ and $F(0)!=1$ (at $\varepsilon=0$
the corresponding formula has been given in \cite{Mel}).
For our aim, we proceed in analogy with (\ref{-37})-(\ref{39}). So
we use the ansatz
\begin{equation}
\rho_n=\lambda_{n-1} \hspace{5mm} {\rm or} \hspace{5mm}
\rho_{n+1}=\lambda_{n}. \label{40}
\end{equation}
With account of this, the QF relation (\ref{30}) rewrites in the
form
\begin{equation}
E_{n+2}=\lambda_{n+1}E_{n+1}+\lambda_nE_n\ \label{41}
\end{equation}
or, for more explicit analogy with (\ref{-37}), in the form
\begin{equation}
\lambda_{n+1}-\Bigl(-\frac{E_n}{E_{n+1}}\Bigr)\lambda_n=\frac{E_{n+2}}{E_{n+1}}\
,
\hspace{5mm} n\geq 1 \ . \label{42}
\end{equation}
Fixing the initial conditions as
\begin{equation}
\lambda_0=\lambda(0)=0\ , \hspace{5mm}
\lambda_1=\lambda(1)=\frac{E_2}{E_1} \label{43}
\end{equation}
we derive for $\lambda_n$ the expression
\begin{equation}
\lambda_n=\Bigl(-\frac{E_{n-1}}{E_n}\Bigr)!
\sum^{n-1}_{j=0}\frac{E_{j+2}/E_{j+1}}{(-E_j/E_{j+1})!}\ .
\label{44}
\end{equation}
The coefficient $\rho_n$ then follows from (\ref{40}). In another
way, we may use the relation (\ref{42}) to deduce
$\lambda_n$ recursively. Recall that the lowest values of
$\lambda_n$ besides (\ref{43}), are:
\[
\lambda_2=-1+\frac{E_3}{E_2}\ , \hspace{28mm}
\lambda_3=-1+\frac{E_4+E_2}{E_3}\ ,
\]
\[
\lambda_4=-1+\frac{E_5+E_3-E_2}{E_4}\ , \hspace{8mm}
\lambda_5=-1+\frac{E_6+E_4-E_3+E_2}{E_5}\ ,
\]
and the formula for generic $\lambda_n$ (which is equivalent to
(\ref{44}), is
\begin{equation}
\lambda_n=-1+\frac{E_{n+1}+E_{n-1}+\sum^{n-2}_{j=2}(-1)^{n-k+1}E_j}{E_n}
= \frac{1}{E_n} \sum^{n+1}_{j=2}(-1)^{n-j+1}E_j\ . \label{45}
\end{equation}
Let us remark that instead of the initial condition
$\lambda(0)=0$ in eq. (\ref{43}), we can set more general one,
namely
\begin{equation}
\lambda_0=\lambda(0)=c\ , \hspace{5mm} c=E_1/E_0. \label{46}
\end{equation}
Then it follows from (\ref{41}) that
\[
\lambda_1=\frac{E_2-cE_0}{E_1}, \hspace{5mm}
\lambda_2=\frac{E_3-E_2+cE_0}{E_2}\ , \hspace{5mm}
\lambda_3=\frac{E_4-E_3+E_2-cE_0}{E_3}\ ,
\]
and by applying induction, we arrive at the desired result
\begin{equation}
\lambda_n=\frac{\sum_{k=2}^{n+1}(-1)^{n-k+1}E_k+(-1)^ncE_0}{E_n}\ .
\label{47}
\end{equation}
Of course, putting $c=0$ reproduces (\ref{45}). Using $c=E_1/E_0$ in
(\ref{46}) the formula (\ref{47}) takes the form
\begin{equation}
\lambda_n=(-1)^{n+1}\frac{1}{{E_n}}\sum_{k=1}^{n+1}(-1)^{k}E_k
\label{48}
\end{equation}
which after substitution of (\ref{27}) yields the desired result:
\begin{equation}
\lambda_n=\left\{\frac{(-1)^n}{1+\mu}+\frac{n+2}{1+\mu(n+2)}\right\}
\frac{(1+\mu n)(1+\mu(n+1))}{n(1+\mu(n+1))+(n+1)(1+\mu n)}\ ,
\label{49}
\end{equation}
\begin{equation}
\rho_n=\left\{\frac{(-1)^{n-1}}{1+\mu}+\frac{n+1}{1+\mu(n+1)}\right\}
\frac{(1+\mu(n-1))(1+\mu n)}{(n-1)(1+\mu n)+n(1+\mu(n-1))}\ .
\label{50}
\end{equation}
The solution (\ref{49})-(\ref{50}) of eq. (\ref{30}) obviously
differs from the solution (\ref{34})-(\ref{-35}) as the ansatz
applied in subsection 3.3.2 is essentially different from the one
used in 3.3.1.
\vspace{5mm} \subsubsection{Third (most general) approach for
deriving $\lambda_n$ and $\rho_n$}\hspace{5mm}
We first remark that the formulae derived in the preceding
two subsections do not constitute general result. The goal of
this subsection is just to find $\lambda_n$ and $\rho_n$ from the
QF relation (\ref{30}) within most general approach.
If we take instead of (\ref{30}) the system
\begin{equation}
\begin{cases} E_{n+1}=\lambda_n E_n+\rho_n E_{n-1} , \cr
E_{n+2}=\lambda_{n+1} E_{n+1}+\rho_{n+1} E_{n} , \cr
\end{cases} \label{51}
\end{equation}
(as 2 equations with 4 unknowns) we cannot calculate
$\lambda$ and $\rho$ as before, see (\ref{18})-(\ref{19}).
To find explicitly $\lambda_n$ and $\rho_n$ satisfying
(\ref{30}) along with (\ref{27}), that is, the relation
\begin{equation}
\frac{n+2}{1+\mu(n+2)}=(\lambda_n-1) \frac{n+1}{1+\mu(n+1)}
+\left(\lambda_n + \rho_n \right) \frac{n}{1+\mu n} + \rho_n
\frac{n-1}{1+\mu(n-1)} \ , \label{52}
\end{equation}
we now follow more general scheme:
we replace (\ref{52}) with the pair of equations
\begin{equation}
\frac{2n+2}{1+\mu(n+2)}=\bigl(\lambda_n-1\bigr)\frac{n+1}{1+\mu(n+1)}+
\frac{K(n;\mu)(n+1)}{1+\mu(n+1)}\ ,\label{53}
\end{equation}
\begin{equation}
\frac{-n}{1+\mu(n+2)}=\bigl(\lambda_n+\rho_n\bigr)\frac{n}{1+\mu
n}+\rho_n\frac{n-1}{1+\mu(n-1)}-\frac{K(n;\mu)(n+1)}{1+\mu(n+1)} \ .
\label{54}
\end{equation}
Note that, the unspecified function $K=K(n;\mu)$ introduced here,
due to its arbitrariness, encapsulates the non-uniqueness of
splitting (\ref{52}) into two equations and thus provides
equivalence of (\ref{52}) with the pair (\ref{53})-(\ref{54}).
Indeed, adding (\ref{53}) and (\ref{54}) yields (\ref{52}).
Or, isolating the last term (with $K(n;\mu)$) in
(\ref{53}) and the identical term in (\ref{54}), we get the
equivalence: (\ref{52}) $\leftrightarrow$ (\ref{53})\&(\ref{54}).
Now, (\ref{53})-(\ref{54}) are easily solved for $\lambda_n$ and
$\rho_n$ that yields the desired general result
\begin{equation}\hspace{-8mm}
\lambda_n=1-K(n;\mu)+2\frac{1+\mu (n+1)}{1+\mu(n+2)} \ ,
\label{55}
\end{equation}
\begin{equation}
\rho_n=\frac{1+\mu(n\!-\!1)}{1+2(n\!-\!1)(1+\mu n)} \
\!\left\{K(n;\mu)\,\frac{2(n+1)(1+\mu n)\!-\!1}
{1+\mu (n+1)}-\frac{4n(1+\mu(n+1))}{1+\mu (n+2)} \right\}.
\label{56}
\end{equation}
{\bf Remark 1}. Various choices of the function $K(n;\mu)$ are of
interest. Say, by properly specified choice $K=K^{(1)}$
(respectively
$K=K^{(2)}$), we can reproduce from
(\ref{55})-(\ref{56}) the expressions (\ref{34}), (\ref{-35}) in
subsection 3.3.1 (respectively the formulae (\ref{49}), (\ref{50})
in subsection 3.3.2). It is easy to verify that the corresponding
choices of $K(n;\mu)$ are:
\begin{equation}
K^{(1)}=\frac{1+3\mu n(1+2\mu)+2\mu(\mu n^2+1)}{(1+2\mu
n)(1+\mu(n+2))}\ , \label{57}
\end{equation}
\[
K^{(2)}=\frac{\mu}{1+\mu(n+2)}+
\frac{(-1)^{n+1}(1+\mu n)(1+\mu(n+1))} {(1+\mu)\bigl\{n\,(1+\mu(n+1))+(n+1)(1+\mu
n)\bigr\}}+\
\]
\begin{equation}
\hspace{10mm}+\frac{(n+1)\bigl\{ 3\mu n+(5n+1)(1+\mu n)\bigr\}}
{(1+\mu(n+2))\bigl\{n\,(1+\mu(n+1))+(n+1)(1+\mu n)\bigr\}}\ .
\label{58}
\end{equation}
From viewing the quasi-Fibonacci relation as "locally-Fibonacci"
property, see end of Introduction, it follows that the (auxiliary)
function $K(n;\mu)$, introduced in order to account the
non-uniqueness of replacing (\ref{52}) by the pair
(\ref{53})$\&$(\ref{54}) plays the role somewhat resembling gauge
freedom in gauge theories: the quantities $\lambda_n$ and $\rho_n$
depend on $K(n;\mu)$ while the energies $E_n$ do not, see (\ref{27})
and (\ref{51})-(\ref{52}).
{\bf Remark 2}. As mentioned above, the usual harmonic oscillator
is Fibonacci oscillator for which $\lambda=2$, $\rho=-1$.
Its structure function and energy spectrum obviously stem from
(\ref{26}), (\ref{27})
if $\mu\rightarrow 0$.
However, sending $\mu\rightarrow 0$ in the formulas (\ref{55})-(\ref{56})
gives
\begin{equation}
\lambda_n=3-K(n;0), \hspace{18mm}
\rho_n=\frac{K(n;0)(2n+1)-4n}{2n-1} \ , \label{59}
\end{equation}
i.e., there is a residual dependence on $n$ if $K(n;0)\ne 1$.
This looks as a kind of surprise: due to $n$-dependent $\lambda_n$,
$\rho_n$ the usual oscillator {\it can also be treated as
quasi-Fibonacci oscillator} (satisfying (\ref{30})).
Only if $K(n;0)=1$ is put, we get $\lambda=2$,
$\rho=-1$, and the usual oscillator becomes genuine Fibonacci oscillator.
\vspace{2mm}
Concerning the last phrase it is worth to add the following.
>From (\ref{57}) and (\ref{58}) at $\mu\rightarrow0$ we have:
$K^{(1)}|_{\mu=0}=1$, but $K^{(2)}|_{\mu=0}\neq1$. This explains the
fact that $\lambda_n$ and $\rho_n$ from (\ref{34})-(\ref{-35}) do
lead at $\mu\rightarrow0$ to the values $\lambda=2$ and $\rho=-1$ of
the ordinary oscillator, while those in (\ref{49})-(\ref{50}) do not
(the reason is the very ansatz $\rho_n=\lambda_{n-1}$ in
(\ref{40})).
{\bf Remark 3.} The extension based on replacing the constant
$\lambda$, $\rho$ with $\lambda_n=\lambda(n)$ and $\rho_n=\rho(n)$
is of principal value for the QF property. However, for the true QF
property to be valid it is in fact enough
that only one of the coefficients depends on $n$, the other being constant.
For instance, consider the case: $\lambda_n=\lambda(n)$ and $\rho=-1$
for the considered $\mu$-oscillator. Imposing $\rho=-1$ in
(\ref{56}) yields the specified $K(n;\mu)$-function:
\[ K(n;\mu)_{\rho=-1} = \frac{(1+\mu(n+1))\{2\mu^2n^2(3n+1)+
4\mu
n(3n+2\mu)+\mu(n-2)+6n-1\}}{(1+\mu(n-1))(1+\mu(n+2))\{2(n+1)(1+\mu
n)-1\}}
\]
With this particular $K$-function, we find the $\lambda_n$ related
to $\rho=-1$:
\begin{equation}
\lambda(n)_{\rho=-1}=1-\frac{(1+\mu(n+1))\{10\mu n^2(2+\mu
n)+(1-4\mu)(3\mu n+1)+12n\}}{(1+\mu(n-1))(1+\mu(n+2))\{2(n+1)(1+\mu
n)-1\}}\ . \label{60}
\end{equation}
It is worth to mention that similar form of the QF relation, with
$\rho=-1$ and certain $\lambda=\lambda(n)$,
appears in \cite{d-Negro}, where the members of the QF sequence
therein have the physical sense of (the Fourier transform of) the
oscillation amplitude of dipole with number $n$ in the non-periodic
chain of dipoles.
Likewise, we may impose $\lambda=2$ and then find from
(\ref{55})-(\ref{56}) the relevant $K$-function and the
corresponding $\rho=\rho_n$:
\[\hspace{-8mm} K(n;\mu)_{\lambda=2}= 2\frac{(1+\mu(n+1))}{(1+\mu(n+2))}-1\ ,
\]
\begin{equation}
\rho(n)_{\lambda=2}=-\frac{1+\mu(n-1)}{1+2(n+1)(1+\mu
n)}\cdot\frac{n(2+5\mu)+4\mu
n(n+\mu)+2\mu^2n^2(n+3)-1}{(1+\mu(n+1))(1+\mu(n+2))}\ . \label{61}
\end{equation}
\vspace{3mm} {\bf Remark 4.} Let $K$ in (\ref{55}), (\ref{56}) be
equal to zero. Then $\lambda_n=\frac{P_1}{Q_1}$, where $P_1$ and
$Q_1$ are linear in $n$, while $\rho_n=\frac{P_3}{Q_3}$ where $P_3$
and $Q_3$ are cubic in $n$. At $K=1$, however, $\lambda_n$ is
again "linear/linear"\ expression in $n$, but $\rho_n$ is the
"quadratic/quadratic"\ one. Let us compare the structure of the
coefficients $\lambda_n$ and $\rho_n$ which were obtained in
different ways, see subsections 3.3.1-3.3.3. While $\lambda_n$,
$\rho_n$ in (\ref{34}), (\ref{-35}) are of the form:
$\lambda_n=\frac{P_2^{(1)}}{Q_2^{(1)}}$ and
$\rho_n=\frac{\widetilde{P}_3^{(1)}}{\widetilde{Q}_3^{(1)}}$, the
structure seen in (\ref{49}), (\ref{50}) looks as
$\lambda_n=\frac{P_3^{(2)}}{Q_3^{(2)}}$ and
$\rho_n=\frac{\widetilde{P}_3^{(2)}}{\widetilde{Q}_3^{(2)}}$.
Likewise, an appropriate choice of the function $K(n;\mu)$ in the
general expressions (\ref{55}), (\ref{56}), yields another
particular case for $\lambda_n$, $\rho_n$ such that they also both
take the form $\frac{P_3^{(3)}}{Q_3^{(3)}}$. Indeed, the choice
\[
K=\frac{\bigl(1+\mu(n+1)\bigr)\bigl(1+2(n-1)(1+\mu
n)\bigr)}{\bigl(1+\mu(n+2)\bigr)\bigl(-1+2(n+1)(1+\mu n)\bigr)}
\]
again leads to the "cubic/cubic"\ expressions for both $\lambda_n$
and $\rho_n$:
\[
\lambda_n=1+\frac{1+\mu(n+1)}{1+\mu(n+2)}\cdot\frac{3+2(1+\mu
n)(3n+1)}{1-2(n+1)(1+\mu n)},
\]
\[
\rho_n=-\frac{1+\mu(n-1)}{1+2(n-1)(1+\mu
n)}\cdot\frac{1+2n(1+\mu(n+3))}{1+\mu(n+2)}.
\]
\vskip5mm
To end this section, we find it instructive to tabulate
$\lambda_n$ and $\rho_n$ from (\ref{55})-(\ref{56}) for low values of $n$, together with
$E_n$, at fixed $K(n;\mu)=1$, see Table 1.
\begin{center}
\noindent{\footnotesize{\bf%
Table 1. Explicit form of \boldmath{$\lambda_n$} and
\boldmath{$\rho_n$} at low values of \boldmath{$n$}.
}}
\end{center}
\vspace{5mm} \vskip1mm
\tabcolsep20.1pt
\noindent\begin{tabular}{c|c|l|l}
\hline%
\hline%
\rule{0pt}{10pt}
$n$ & $\lambda_n$ & $\rho_n$ & $E_n$ \\
\hline%
\hline%
\rule{0pt}{2pt} \vspace{-3mm}
{} & {} & {} & {}\\
$0$ & \hspace{-2mm}$\lambda_0=\frac{2+2\mu}{1+2\mu}$ & \hspace{-3mm} $\rho_0=\frac{\mu-1}{1+\mu}$ &
\hspace{-2mm} $E_0=\frac{1}{2(1+\mu)}$ \\
\vspace{-4mm}
{} & {} & {} & {}\\
\hline%
\vspace{-3mm}
{} & {} & {} & {}\\
$1$ & \hspace{-2mm}$\lambda_1= \frac{2+4\mu}{1+3\mu}$ & \hspace{-2mm}$\rho_1\!=\frac{3+4\mu}{1+2\mu}-\frac{4(1+2\mu)}{1+3\mu}$ &
\hspace{-2mm} $E_1=\frac{1}{2(1+\mu)}+\frac{1}{1+2\mu}$ \\
\vspace{-4mm}
{} & {} & {} & {}\\
\hline%
\vspace{-3mm}
{} & {} & {} & {} \\
$2$ & \hspace{-2mm}$\lambda_2 = \frac{2+6\mu}{1+4\mu}$ & \hspace{-3mm}
$\rho_2\!=\!\frac{1+\mu}{3+4\mu}\!\left(\!\frac{5+12\mu}{1+3\mu}\!-\!\frac{8(1+3\mu)}{1+4\mu}\!\right)$ \hspace{-3mm} &
\hspace{-2mm} $E_2=\frac{1}{1+2\mu}+\frac{3}{2(1+3\mu)}$ \\
\vspace{-4mm}
{} & {} & {} & {}\\
\hlin
\vspace{-3mm}
{} & {} & {} & {}\\
$3$ & \hspace{-2mm}$\lambda_3= \frac{2+8\mu}{1+5\mu} $ & \hspace{-3mm}
$\rho_3\!=\!\frac{1+2\mu}{5+12\mu}\!\left(\!\frac{7+24\mu}{1+4\mu}\!-\!\frac{12(1+4\mu)}{1+5\mu}\!\right)$ \hspace{-3mm}&
\hspace{-2mm} $E_3=\frac{3}{2(1+3\mu)}+\frac{2}{1+4\mu}$ \\
\vspace{-4mm}
{} & {} & {} & {}\\
\hline%
\vspace{-3mm}
{} & {} & {} & {}\\
$4$ & \hspace{-2mm}$\lambda_4= \frac{2+10\mu}{1+6\mu} $ & \hspace{-3mm}
$\rho_4\!=\!\frac{1+3\mu}{7+24\mu}\!\left(\!\frac{9+40\mu}{1+5\mu}\!-\!\frac{16(1+5\mu)}{1+6\mu}\!\right)$ \hspace{-3mm}&
\hspace{-2mm} $E_4=\frac{2}{1+4\mu}+\frac{5}{2(1+5\mu)}$ \\
\vspace{-3mm}
{} & {} & {} & {}\\
\hline%
\vspace{-3mm}
{} & {} & {} & {}\\
\vspace{-3mm}
$5$ & \hspace{-2mm}$\lambda_5=\frac{2+12\mu}{1+7\mu} $ & \hspace{-3mm}
$\rho_5\!=\!\frac{1+4\mu}{9+40\mu}\!\left(\!\frac{11+60\mu}{1+6\mu}\!-\!\frac{20(1+6\mu)}{1+7\mu} \!\right)$ \hspace{-3mm}&
\hspace{-2mm} $E_5=\frac{5}{2(1+5\mu)}+\frac{3}{1+6\mu}$ \\
{} & {} & {}& {} \\
\hline%
\end{tabular}
\section{Mixed cases of deformed oscillators}\hspace{5mm}
It is possible to construct new models of deformed oscillators which
are quasi-Fibonacci extensions of a particular Fibonacci oscillator
of Section 2, by combining the latter with the $\mu$-oscillator
which plays the role of the basic quasi-Fibonacci "building block".
Clearly, the procedure brings in some additional deformation
parameters. On this way we naturally obtain new models with two
deformation parameters $\mu$ and $q$: say, the mixed $\mu\!-\!AC$
case, and also
the mixed $\mu\!-\!BM$ and $\mu\!-\!TD$ cases. However, since the
models of $q$-deformed oscillators are contained as particular cases
in the $\!p,q$-family, see sections 2.1-2.3, it is useful to start
with the three-parameter, mixed $(\mu;p,q)$-family of models.
\begin{center}
\vspace{3mm} {\it The three-parameter "mixed"\ family of
$(\mu;p,q)$-deformed oscillators} \vspace{1mm}
\end{center}
\hspace{5mm}This family of deformed oscillator models arises due to
combining the $p,\!q$-oscillator with the $\mu$-oscillator and is
given by the structure function
\begin{equation}
\varphi_{\mu,p,q}(n)=\frac{[n]_{p,q}}{1+\mu n} \ , \hspace{12mm}
[n]_{p,q}\equiv\frac{p^n-q^n}{p-q}. \label{62}
\end{equation}
Of course, the choice (\ref{62}) of "mixed" deformation is not
unique: for the "mixed" structure function one could also use, say,
$\psi_{\mu,p,q}(n)=\frac{[n]_{p,q}}{1+\mu[n]_{p,q}}$ or
$\chi_{\mu,p,q}(n)=[\frac{n}{1+\mu n}]_{p,q}$. Our choice however
(i) is simpler from the viewpoint of further use in applications
like that in \cite{mu-bose} and (ii) better correlates with the
ideas of \cite{Jan}.
So we expect that this $(\mu;p,q)$-oscillator, with the energy
spectrum
\[
E_n=\frac{1}{2}\Bigl(\varphi_{\mu,p,q}(n+1)+\varphi_{\mu,p,q}(n)\Bigr),
\] obeys the QF relation.
It is clear that the $(\mu;p,q)$-deformed family cannot satisfy pure
Fibonacci relation, because of its $\mu$-component as carrier of the
QF property; only if $\mu=0$ it reduces to the pure
$(p,q)$-oscillator for which the status of Fibonacci oscillator is
recovered.
So, consider the $(\mu;p,q)$-oscillator as
quasi-Fibonacci one.
The QF property of it
will be certified if the explicit expressions for $\lambda_n$ and
$\rho_n$ are found such that the QF relation (\ref{30}) holds true:
\[
\frac{[2]_{p,q}[n+1]_{p,q}-pq[n]_{p,q}}{1+\mu(n+2)}+
\frac{[n+1]_{p,q}}{1+\mu(n+1)}= \nonumber
\]
\begin{equation}
=\lambda_n\biggl(
\frac{[n+1]_{p,q}}{1+\mu(n+1)}+\frac{[n]_{p,q}}{1+\mu n}\biggr)+
\rho_n\biggl( \frac{[n]_{p,q}}{1+\mu n}+
\frac{[n-1]_{p,q}}{1+\mu(n-1)}\biggr) \ . \label{63}
\end{equation}
Note that in the LHS the identity:
\begin{equation}
[n+2]_{p,q}=[2]_{p,q}[n+1]_{p,q}-pq[n]_{p,q} \label{64}
\end{equation}
has been used. Instead of (\ref{63}) we may {\it equivalently} exploit the
following two relations:
\[
\bigl(\lambda_n-1\bigr)\frac{[n+1]_{p,q}}{1+\mu(n+1)}=
\frac{[2]_{p,q}
[n+1]_{p,q}}{1+\mu(n+2)}-K\frac{[n+1]_{p,q}}{1+\mu(n+1)}\ ,
\]
\[
\bigl(\lambda_n+\rho_n\bigr)\frac{[n]_{p,q}}{1+\mu
n}+\rho_n\frac{[n-1]_{p,q}}{1+\mu(n-1)}=
\frac{-pq[n]_{p,q}}{1+\mu(n+2)}+K\frac{[n+1]_{p,q}}{1+\mu(n+1)}\ .
\]
Let us emphasize that the arbitrary function $K=K(n;\mu,p,q)$
involves both the variable $n$ and the three deformation parameters.
Like before, it is the unspecified function $K(n;\mu,p,q)$ which
guarantees the equivalence with (\ref{63}).
From the pair of equations, using $[2]_{p,q}=p+q$ as implied by
(\ref{62}), the desired solution does follow, namely
\begin{equation}
\lambda_n\equiv
\lambda_n(\mu,p,q)=1-K+[2]_{p,q}\frac{1+\mu(n+1)}{1+\mu(n+2)}\
, \label{65}
\hspace{5mm}
\end{equation}
\[
\rho_n\equiv\rho_n(\mu,p,q)=\frac{1+\mu(n-1)}{[n]_{p,q}(1+\mu(n-1))+
[n-1]_{p,q}(1+\mu n)}\ \times
\]
\begin{equation}
\times \Biggl\{K\biggl([n]_{p,q}+\frac{[n+1]_{p,q}(1+\mu
n)}{1+\mu(n+1)}\biggr)-[n]_{p,q}\biggl(1+\frac{[2]_{p,q}(1+\mu(n+1))+qp(1+\mu
n)}{1+\mu(n+2)}\biggr)\Biggr\} \label{66}
\end{equation}
This constitutes our general result for the $(\mu;p,q)$-deformed
oscillators. From this, by fixing $K(n;\mu,p,q)$ we can generate for
$\lambda_n$ and $\rho_n$ various special expressions.
Let us consider some special cases of (\ref{65}), (\ref{66}), with
fewer deformation parameters.
1) If $p=q=1$ and $\mu\neq0$, it is easily checked that one recovers
the expressions (\ref{55}), (\ref{56}) in Sec. 3.3 for $\lambda(n)$
and $\rho(n)$ of the pure $\mu$-oscillator, the latter being the
carrier of QF property.
2) If we put $\mu=0$ in (\ref{65}), (\ref{66}), we have:
\[\lim_{\mu\to 0} \lambda_n(\mu,p,q) = \lambda(p,q)=\lambda=1-K(n;p,q)+[2]_{p,q}\ ,
\]
\[\lim_{\mu\to 0} \rho_n(\mu,p,q) = \rho(p,q)=\rho=
\frac{K(n;p,q)([n+1]+[n])-[n](1+pq+[2]_{p,q})}{[n]_{p,q}+[n-1]_{p,q}}\
.
\]
In the latter expressions $\lambda$ and $\rho$ still depend on $n$.
However, at $K=1$ we recover $\lambda=[2]_{p,q}$ and $\rho=-pq$ of
the pure $(p,q)$-oscillator in Sec.2. The cancelation of the
$n$-dependence in $\lambda$, see $K(n;p,q)$, and in $\rho$,
follows by applying the identity (\ref{64}).
3) Let $p=q^{-1}$ and $q\rightarrow1$. The corresponding values of
the coefficients then read
\[
\lambda_n=1-K(n;1,1)+\frac{2(1+\mu(n+1))}{1+\mu(n+2)}\ ,
\]
\[
\rho_n=\frac{1+\mu(n-1)}{1+2(n-1)(1+\mu
n)}\!\left(\!K(n;1,1)\frac{2(n+1)(1+\mu n)\!-\!1}{1+\mu(n+1)}\!-\!
\frac{4n(1+\mu(n+1))}{1+\mu(n+2)}\right)
\]
and if moreover $K=1$, we have from the letter:
\[\lambda_n|_{K=1}\!=\!2\frac{1+\mu(n+1)}{1+\mu(n+2)},
\]
\[
\rho_n|_{K=1}\!=\!\frac{1+\mu(n\!-\!1)}{1+2(n\!-\!1)(1+\mu
n)}\left(\!\!1\!-
\!2n+\mu\!\left(\!\frac{4n}{1+\mu(n+2)}\!-\!\frac{n+1}{1+\mu(n+1)}\right)\right)\
.
\]
4) Putting $\mu=0$, $p=q=1$ in (\ref{65})-(\ref{66}),
for $\lambda_n$ and $\rho_n$ yields
\[\lambda_n|_{\mu=0}\!=\!3-K, \hspace{5mm} \rho_n|_{\mu=0}\!=\!\frac{K(2n+1)-4n}{2n-1}\ .\]
see Remark 2 for more details.
As mentioned at the beginning of this section, since we have at our
disposal the explicit formulae (\ref{65}), (\ref{66}) for the
3-parameter $(\mu;p,q)$-family, we immediately get the corresponding
results for the three distinguished two-parameter cases. Let us
illustrate this (recall (8), (10), and $[n]_{TD}\equiv nq^{n-1}$ for
the AC, BM and TD cases correspondingly):
\vspace{2mm} $\bullet$ let $p\to 1$.
\ \ \ In this case, $ \varphi_{\mu,AC}(n)=\frac{[n]_{AC}}{1+\mu n}$
\hspace{12mm} (mixed $\mu$-AC case);
\vspace{2mm} $\bullet$ put $p=q^{-1}.$ In this case
$\varphi_{\mu,BM}(n)=\frac{[n]_{BM}}{1+\mu n}$ \hspace{11mm} (mixed
$\mu$-BM case);
\vspace{2mm} $\bullet$ put $p=q$. \ \ \ In this case
$\varphi_{\mu,TD}(n)=\frac{[n]_{TD}}{1+\mu n}$ \hspace{13mm} (mixed
$\mu$-TD case).
\vspace{1mm} \noindent
We end with the comment on the hybrid case of six
parameters: the $(\mu;p,q,\alpha,\beta,l)$-case.
\vspace{2mm}
{\bf Remark 5.}
All the above treatment carried out for the three-parameter
$(\mu;p,q)$-family of deformed oscillators can be easily extended to
the case of six-parameter family which combines the 5-parametric
case (\ref{17}) and the formula (\ref{27}) of the $\mu$-oscillator
case.
Also, one can put $\l=\alpha$ in order to have
unambiguous definition of the structure function, consistent with
the relation $E_n=\frac12\bigr(\varphi(n)+\varphi(n+1)\bigr)$.
\subsection*{Conclusions and outlook}\hspace{5mm}
In this paper we dealt with two essentially different classes of
deformed oscillators: the Fibonacci class and the (much more rich)
quasi-Fibonacci class. As our first result, we have proven that
besides the well-studied 2-parameter family of $p,q$-deformed
oscillators known \cite{Ar_Fib} as Fibonacci oscillators, there
exists more general, with three additional parameters, the
$(p,q,\alpha,\beta,l)$-family of deformed oscillators which does
also belong to the Fibonacci class. What is rather unexpected, the
coefficients $\lambda$ and $\rho$ for the 5-parametric family of
deformed oscillators depend on $p$, $q$ (as in the $p,\!q$-deformed
case) and else only on one parameter $\alpha$, from the remaining
three parameters.
On the other hand, according to our second main result reflected in
Proposition 2, the $\mu$-oscillator \cite{Jan} is not in Fibonacci
class, but belongs to the class of QF oscillators whose basic
feature is that the sequence of values of energy levels obeys the QF
relation, being as well linear and three-term (or two-step) one, but
involving the \underline{non-constant} coefficients $\lambda_n$ and
$\rho_n$ which depend besides the deformation parameter $\mu$ also
on the number $n$ of the energy level $E_n$.
The peculiar feature of QF oscillators is that the
$n$-dependent coefficients $\lambda_n$ and $\rho_n$ are obtained
non-uniquely, as it was demonstrated with the three different
methods of solving the eq.(\ref{30}). While the first two ways lead
to partial solutions, the third method is general one due to the
(arbitrary) function $K(n;\mu)$ being involved.
We have shown that the $\mu$-oscillator is not the unique one
possessing the QF property. Indeed, it can be utilized as a basic
ingredient for constructing other families of QF deformed
oscillators which have additional deformation parameter(s).
An example of the three-parameter family of $(\mu;p,q)$-deformed
oscillators obtainable through combining the $p,q$-oscillator with the
$\mu$-oscillator has been considered.
From that, a particular two-parameter families
(the $\mu,q$-oscillator as the Arik-Coon $\&$ the $\mu$-type hybrid,
the BM $\&$ the $\mu$-type hybrid, and the TD $\&$ the $\mu$-type
hybrid) of deformed oscillators naturally follow.
We believe our results provide the basis for
constructing in a regular manner numerous new models of multi-parameter
deformed nonlinear oscillators belonging to different extensions of
the Fibonacci class, in particular, those possessing the quasi-Fibonacci property.
It is also worth to note that interesting {\it classes of
quasi-Fibonacci oscillators} have been explored in our recent paper \cite{GR-5}.
The polynomially deformed non-Fibonacci oscillators treated therein are rather amusing
as they can be viewed in three different ways: 1) as quasi-Fibonacci
oscillators; 2) as the oscillators obeying inhomogeneous Fibonacci
relation; 3) as $k$-bonacci oscillators. Unlike the classes just
mentioned, the $\mu$-oscillator studied in the present paper along
with its direct extensions admit only the quasi-Fibonacci type (way)
of description.
For each Fibonacci oscillator there exist diverse quasi-Fibonacci
extensions, see e.g. the note below (\ref{62}) and Remark 5.
This fact, all that is said in the preceding paragraph, and the results of Sec.3, Sec.4
clearly demonstrate that the class of QF models based on the
novel concept of quasi-Fibonacci oscillators is really rich and
worth of detailed study. Although exploring of physical applications of
quasi-Fibonacci oscillators is at the very beginning, we may quote
the recent work \cite{mu-bose} on the $\mu$-Bose gas model employing
$\mu$-oscillators.
Let us finally note that in view of existing deformed fermionic
Fibonacci oscillators \cite{AAA} it is worth to study, along
the lines of the present paper, possible models of fermionic
quasi-Fibonacci oscillators, as yet another class.
\section*{Acknowledgement}
The authors are thankful to I.M. Burban for interesting discussions,
and to the referees for valuable remarks and suggestions which led
to improved presentation. This research was partially supported by
the Grant 29.1/028 of the State Foundation of Fundamental Research
of Ukraine and by the Special Program of the Division of Physics and
Astronomy of the NAS of Ukraine.
|
\section{Introduction}
\label{intro}
Since the early work of \citet{elme77}, it is known that during the supersonic expansion of an HII region,
a dense layer of material can be collected between the ionization and the shock fronts.
This layer can be fragmented into massive condensations that may then collapse to form new massive stars and/or clusters.
Several observational studies have inferred that this process, known as ``collect and collapse'', triggers massive
star formation (see e.g., \citealt{poma09,zav07}, and references therein).
As \citet{deha05} point out, only the presence of either a dense
molecular shell surrounding the ionized gas of an HII region or massive fragments regularly
spaced along the ionization front, can prove that we are dealing with the collect and collapse process.
Using Galactic Legacy Infrared Mid-Plane Survey Extraordinaire (GLIMPSE)
data, \citet{church06,church07} cataloged almost 600 infrared (IR) dust
bubbles, full or partial rings bordered by a photodissociation region (PDR) and detected mainly at 8 $\mu$m,
which usually enclose ionized gas and hot dust observed at 24 $\mu$m. Most of these bubbles are HII regions
and several have morphologies that are indicative of triggered star formation.
On the other hand, \citet{cyga08} also using GLIMPSE data, identified more than 300 extended 4.5 $\mu$m sources,
so-called ``extended green objects (EGOs)'', applying the common coding of the [4.5] band as green in three-color
composite Infrared Array Camera images. According to the authors, an EGO is probably a massive young stellar
object (MYSO) driving outflows. The extended emission in the 4.5 $\mu$m band is produced by
H$_{2}$ ($\nu =$ 0--0, S(9,10,11)) lines and CO ($\nu = 1-0$) band heads, which are excited by the shock of
the outflows propagating in the interstellar medium (ISM). The majority of EGOs are associated with infrared dark
clouds (IRDCs) and some lie over the border of IR dust bubbles.
From the catalog of \citet{church06}, we select the IR dust bubble N65,
which harbors the EGO G35.03+0.35 in one of its borders (indicative of active star formation),
to study its environment.
N65 is a complete (i.e., closed ring) IR dust bubble centered
on $\alpha_{2000} =$ 18$^{\mathrm h}$ 54$^{\mathrm m}$ 02$^{\mathrm s}$, $\delta_{2000} = +$01$^\circ$ 59$^\prime$ 30$^{\prime\prime}$~($l = 35$\fdg$000$, $b = +0$\fdg$332$)
with a radius of about 2$^\prime$.6 \citep{church06}. Close to N65, over one of its borders,
centered on $\alpha_{2000} =$ 18$^{\mathrm h}$ 54$^{\mathrm m}$ 04.2$^{\mathrm s}$, $\delta_{2000} =
+$02$^\circ$ 01$^\prime$ 33.9$^{\prime\prime}$~($l = 35$\fdg$035$, $b = +0$\fdg$338$) lies the IR source IRAS 18515+0157, where
several molecular lines tracers of star formation were detected. Maser emission from H$_{2}$O and OH was detected by
\citet{forster89}, as well as methanol \citep{caswell95}. \citet{bronf96} and \citet{jiji99} also detected
CS and NH$_{3}$, respectively.
The detection of formaldehyde at v$_{LSR} \sim 57$ km s$^{-1}$~was used by \citet{watson03}
to determine that this star-forming region lies at the kinematic
distance of 10 kpc, the farthest possible given the distance ambiguity of the first Galactic quadrant.
On the other hand, close to this region, at $l = 35$\fdg$015$, $b = +0$\fdg$356$ and also across the border
of N65, \citet{anderson09} cataloged the UCHII region G35.02+0.35 with a velocity of $\sim 57.2$ km s$^{-1}$,
which they locate at the near distance of $\sim 3.6$ kpc.
In this work, we present a molecular and IR study of the environment surrounding
the IR dust bubble N65 to explore its surrounding ISM and search for signatures of
star formation. We describe the data used in Sect. \ref{data}, the results and the
discussion of them are presented in Sect. \ref{resultdisc} and Sect. \ref{sum} summarizes the
results.
\section{Data}
\label{data}
We analyzed data extracted from five large-scale surveys: the
Two Micron All Sky Survey (2MASS)\footnote{2MASS 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.},
Galactic Legacy Infrared Mid-Plane Survey Extraordinaire (GLIMPSE),
MIPSGAL, GRS\footnote{Galactic Ring Survey \citep{jackson06}}, and the SCUBA Legacy Catalogue.
GLIMPSE is a mid-infrared survey of the inner
Galaxy performed with the {\it Spitzer Space Telescope}. We used the mosaicked images from
GLIMPSE and the GLIMPSE Point-Source Catalog (GPSC) in the {\it Spitzer}-IRAC (3.6, 4.5, 5.8, and 8 $\mu$m).
IRAC has an angular resolution of between 1\farcs5 and 1\farcs9 (see \citealt{fazio04} and \citealt{werner04}).
MIPSGAL is a survey of the same region as GLIMPSE, using the MIPS instrument (24 and 70 $\mu$m) on {\it Spitzer}.
The MIPSGAL resolution is 6$^{\prime\prime}$~at 24 $\mu$m.
The GRS is being performed by the Boston University and the
Five College Radio Astronomy Observatory (FCRAO). The survey maps the galactic Ring in the $^{13}$CO J=1--0 line
with an angular and spectral resolution of 46$^{\prime\prime}$~and 0.2 km s$^{-1}$, respectively (see \citealt{jackson06}).
The observations were performed in both position-switching and on-the-fly mapping modes, achieving an
angular sampling of 22$^{\prime\prime}$. The SCUBA Legacy Catalogues\footnote{http://www2.cadc-ccda.hia-iha.nrc-cnrc.gc.ca/community/scubalegacy/} are two sets of continuum maps and catalogs using
data at 450 and 850 $\mu$m obtained with the Submillimetre Common User Bolometer Array (SCUBA) with
angular resolutions of 19$^{\prime\prime}$~and 11$^{\prime\prime}$, respectively \citep{difranc08}.
\section{Results and discussion}
\label{resultdisc}
Figure \ref{iracmips} ({\it left}) shows a {\it Spitzer}-IRAC three color image of N65. The three IR bands are 3.6 $\mu$m
(in blue), 4.5 $\mu$m (in green), and 8 $\mu$m (in red). The location of the EGO G35.03+0.35
\citep{cyga08} is indicated.
Figure \ref{iracmips} ({\it right}) displays a composite two-color image towards N65. Red and green represent
the {\it Spitzer}-IRAC emission at 8 $\mu$m and the {\it Spitzer}-MIPSGAL emission at 24 $\mu$m, respectively.
Both figures clearly show the PDR visible in the 8 $\mu$m emission, which originates mainly
in the polycyclic aromatic hydrocarbons (PAHs). The PAH emission delineates the HII region
boundaries because these large molecules are destroyed inside the ionized region, but are excited in the
PDR by the radiation leaking from the HII region \citep{poma09}.
The 24 $\mu$m emission, displayed in green in Fig. \ref{iracmips} ({\it right}), corresponds to hot dust,
distributed mainly towards the eastern border of N65 that coincides with the central position of IRAS 18515+0157.
\begin{figure}[h]
\includegraphics[width=14cm]{Fig01.ps}
\caption{Mid-IR emission of the IR dust bubble N65. {\it Left}: {\it Spitzer}-IRAC three-color image
(3.6 $\mu$m $=$ blue, 4.5 $\mu$m $=$ green, and 8 $\mu$m $=$ red). The location of the EGO G35.03+0.35 is indicated.
{\it Right}: red is the {\it Spitzer}-IRAC 8 $\mu$m emission and green is the {\it Spitzer}-MIPSGAL emission at 24 $\mu$m.}
\label{iracmips}
\end{figure}
From Fig. \ref{iracmips}, a probably smaller
IR dust bubble can be discerned in the southwest border of N65, centered on $l = 34$\fdg$96$, $b = +0$\fdg$31$,
which is displayed in Fig. \ref{smallbub}.
The emission at 8 $\mu$m (Fig. \ref{smallbub} {\it left}) shows an almost circular PDR,
which encloses a region of about 18$^{\prime\prime}$~in diameter.
At the center of the bubble, a source is visible with emission
in the 3.5, 4.5, 8, and 24 $\mu$m bands, which is probably the exciting star of this
structure surrounded by hot dust emitting at 24 $\mu$m
(green in Fig. \ref{smallbub} {\it right}).
\begin{figure}[h]
\includegraphics[width=13cm]{Fig02.ps}
\caption{Mid-IR emission of the possible smaller IR dust bubble that lies in the southwest border of N65.
{\it Left}: {\it Spitzer}-IRAC three-color image (3.6 $\mu$m $=$ blue, 4.5 $\mu$m $=$ green and 8 $\mu$m $=$ red).
{\it Right}: red is the {\it Spitzer}-IRAC 8 $\mu$m emission and green is the {\it Spitzer}-MIPSGAL
emission at 24 $\mu$m.}
\label{smallbub}
\end{figure}
\subsection{Molecular and dust analysis}
\label{molec}
We inspected the molecular gas around N65 from the GRS data in the whole velocity range and found an interesting
feature around v $\sim$ 50 km s$^{-1}$.
Figure \ref{molecgas} shows the integrated velocity maps of the $^{13}$CO J=1-0 emission every 1.05 km s$^{-1}$~between
47 and 56 km s$^{-1}$. The circle shows the position and size of N65.
A molecular cloud with an arc-like shape open to the south is evident that encloses
the IR bubble over one of its borders. The molecular gas exhibits several clumps along the PDR.
The distribution and morphology of this material suggests that the
collect and collapse process might be occurring.
\begin{figure}[h]
\includegraphics[width=13cm]{Fig03.ps}
\caption{Integrated velocity maps of the $^{13}$CO J=1-0 emission every 1.05 km s$^{-1}$. The grey scale is in K km s$^{-1}$.
The contour levels are 2, 4 and 6 K km s$^{-1}$. The circle represents the position and size of N65.}
\label{molecgas}
\end{figure}
In Fig. \ref{molec+scuba}, we compare the $^{13}$CO emission integrated within the interval 47 and 55 km s$^{-1}$~with
the continuum dust emission at 850 $\mu$m extracted from the SCUBA (in yellow contours).
In this figure, it is evident that the dust emission coincides with the brightest molecular clump centered
on $l \sim 35\fdg02$, $b \sim 0\fdg35$ and with the position of the EGO G35.03+0.35.
The presence of the EGO together with all the previous results mentioned in Sect. \ref{intro} provide clear
evidence that this region might be active in star formation.
If we assume that the molecular gas in the velocity range considered before is physically associated
with N65, we can adopt $\sim 50$ km s$^{-1}$~as the systemic velocity of the infrared dust bubble.
According to the galactic rotation model of \citet{fich89} (with $R_{\odot} = 8.5$ kpc and $v_{\odot} = 220$ km s$^{-1}$),
we obtain kinematic
distances of either 3.5 or 10.5 kpc. Considering that the detection rate for IR dust bubbles in
GLIMPSE peaks at the distance of 4.2 kpc within a horizon of 8 kpc \citep{church06},
the bubbles at the near kinematic distance are more likely to be detected than those at the far distance.
Thus, we assume a distance of 3.5 kpc for N65, in coincidence with the UCHII region G35.02+0.35 cataloged
by \citet{anderson09}.
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{Fig04.ps}
\caption{$^{13}$CO emission integrated between 47 and 55 km s$^{-1}$, the contours levels are 19, 23, and 33 K km s$^{-1}$. The yellow
contours correspond to the 850 $\mu$m continuum emission obtained from SCUBA and the levels are 0.3, 0.6, 0.9, and 2.3
Jy beam$^{-1}$. The beams of each emission are included in the bottom left corner.}
\label{molec+scuba}
\end{figure}
Using the $^{13}$CO J=1--0 line and assuming local thermodynamic equilibrium (LTE), we estimate the H$_{2}$ column density
toward the brightest molecular clump shown in Fig. \ref{molec+scuba}. We use
$${\rm N(^{13}CO)} = 2.42 \times 10^{14} \frac{T_{\rm ex} \int{\tau_{13} dv}}{1 - exp(-5.29/T_{\rm ex})},$$
to obtain the $^{13}$CO column density, where $\tau_{13}$ is the optical depth of the line and
following \citet{anderson09} we considered $T_{\rm ex} = 20$ K.
We assume that the $^{13}$CO emission is optically thin and use the relation N(H$_{2}$)/N($^{13}$CO)$ \sim 5 \times 10^5$
(e.g., \citealt{simon01}) to estimate a column density of N(H$_{2}$) $\sim 1.6 \times 10^{22}$ cm$^{-2}$.
From this value, we estimate the mass of the molecular clump to be $\sim 2 \times 10^{3}$ M$_\odot$.
This value was obtained from
$$ {\rm M} = \mu~m_{{\rm H}} \sum{\left[ D^{2}~\Omega~{\rm N(H_{2})} \right] }, $$
where $\Omega$ is the solid angle subtended by the $^{13}$CO J=1--0 beam size, $m_{\rm H}$ is the hydrogen mass,
$\mu$ is the mean molecular weight assumed to be 2.8 by taking into account a relative helium abundance
of 25 \%, and $D$ is the distance. Summation was performed over all the observed positions within the
33 K km s$^{-1}$~ contour level (see Fig. \ref{molec+scuba}). Finally from this mass value, we derive a density of
$n \sim 10^{4}$ cm$^{-3}$.
Using the 850 $\mu$m continuum emission, we estimate the dust mass from the relation of \citet{tej06}:
$$ M_{dust} = 1.88 \times 10^{-4} \left(\frac{1200}{\nu}\right)^{3+\beta} S_{\nu} \left(e^{0.048 \nu/T_{d}} - 1\right) D^{2},$$
where $S_{\nu}$ is the flux density at the frequency $\nu$. We use
$S_{353GHz} = 19.42$ Jy as obtained by \citet{difranc08} for this region. We assume that
$T_{d}$ the dust temperature is 20 K, $\beta$ the dust emissivity index is 2.6 for the assumed
dust temperature in this region according to \citet{hill06}, and $D$ the distance to N65 is 3.5 kpc.
This equation assumes a standard opacity of $k_{1200GHz} = 0.1$ cm$^{2}$ g$^{-1}$. We obtain a dust mass
of $M_{dust} \sim 60$ M$_\odot$. The dust-to-gas mass ratio is then $\sim 0.03$, which is 3 times higher than the
canonical Galactic ratio.
\subsection{IR photometry}
\label{photom}
We analyze the distribution of IR point sources in the surroundings of N65 to search
for signs of star formation and try to identify the ionizing-star candidates.
Figure \ref{ccdiagr} shows the [5.8]-[8.0] versus [3.6]-[4.5]
color-color (CC) diagram for the sources extracted from the GLIMPSE Point Source Catalog in the {\it Spitzer}-IRAC bands
within a circle of 8$^\prime$~in radius centered on
$l = 35$\fdg$00$, $b = +0$\fdg$32$. The size of this region completely covers the extension
of N65 and the surrounding molecular clumps, where stars may be forming.
We only consider sources with detection in the four bands.
The regions in the figure indicate the stellar evolutionary stages
based on the criteria of \citet{allen04}: class I sources are protostars with circumstellar envelopes,
class II are disk-dominated objects, and class III are main sequence and giant stars.
We search for ionizing star candidates among class III sources, while class I and II sources
are chosen to be the YSO candidates. In the following sections, we report the results of these analyses.
\begin{figure}[h]
\centering
\includegraphics[angle=-90,totalheight=0.35\textheight,viewport=0 0 550 700,clip]{Fig05.ps}
\caption{\ GLIMPSE color-color diagram [5.8]-[8.0] versus [3.6]-[4.5] for sources within a circle
of 8$^\prime$~in radius centered at N65. We only consider sources with detection in the four {\it Spitzer}-IRAC bands.
The regions indicate the stellar evolutionary stage as defined by \citet{allen04}.
Class I sources are protostar with circumstellar envelopes;
class II are disk dominated objects; and class III are main sequence and giant stars. }
\label{ccdiagr}
\end{figure}
\subsection{Identifying the exciting star(s) of N65}
\label{ioniz}
We attempt identify or at least propose the possible exciting stars(s) of N65.
According to \citet{church06}, IR dust bubbles may be produced by O- and/or B-type stars.
Given that we do not find any cataloged massive star toward N65, we look for the exciting star(s) among
the class III sources because they include main sequence stars.
The ionizing star(s) are not always located close to the geometrical
center of the bubble because of their proper motion. However, we expect to find them inside the
PDR as IR dust bubbles are young objects and the ionizing star(s) may not have enough time
to overcome the shell. To perform near IR photometry using the 2MASS {\it JHK} bands, we select
class III objects within the N65 PDR that have detections in the aforementioned near IR bands.
Figure \ref{cc} shows the ({\it H-K}) versus ({\it J-H}) color-color (CC) diagram of the 22
sources found using this criteria. Figure \ref{cm} shows the {\it K} versus ({\it H-K})
color-magnitude (CM) diagram. Finally, in Fig. \ref{excitstar} we show the location of these sources.
To identify possible exciting star(s) candidates among the 22 sources found,
we select objects that meet both of the following criteria simultaneously:
(i) according to their position in the CC diagram (Fig. \ref{cc}), the sources
may be main sequence stars and (ii) according to their position in the CM diagram (Fig. \ref{cm}), the sources
may be O- or B-type stars. These sources are indicated by red numbers in Figs. \ref{cc} and \ref{cm} and
red crosses in Fig. \ref{excitstar}.
\begin{figure}[h]
\centering
\includegraphics[width=8cm,angle=-90]{Fig06.ps}
\caption{({\it H-K}) versus ({\it J-H}) color-color (CC) diagram of the class III objects within the N65 PDR.
The two solid curves represent the location of the main
sequence (thin line) and the giant stars (thicker line) derived from
\citet{bessell88}. The parallel dashed lines are reddening
vectors. We have assumed the interstellar reddening law of \citet{rieke85}
($A_J/A_V$=0.282; $A_H/A_V$=0.175 and $A_K/A_V$=0.112). The sources with red numbers are the ones selected
to be the candidates for exciting stars of N65.}
\label{cc}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=8cm,angle=-90]{Fig07.ps}
\caption{({\it H}) versus ({\it H-K}) color-magnitude (CM) diagram of the class III objects within the N65
PDR. The position of the zero-age-main-sequence stars with A$_{v} = 0$ mag at a distance of 3.5 kpc are
indicated by black stars. The reddening curve for an O3 star is shown with a black line. The sources with red
numbers represent are the ones selected to be candidate exciting stars of N65.}
\label{cm}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=8cm]{Fig08.ps}
\caption{Possible exciting star(s) over the 8 $\mu$m emission of N65. The red crosses are the sources selected,
according to the CM and the CC diagrams, to be candidate exciting stars.}
\label{excitstar}
\end{figure}
As an additional constraint of the ionizing star(s) candidate, we derive the spectral energy distribution
(SED) of the selected sources by fitting the fluxes in
the IRAC and 2MASS bands with a Kurucz photospheric model using the tool developed by \citet{robit07}
and available online\footnote{http://caravan.astro.wisc.edu/protostars/}.
The fitting tool requires a distance range and a range of visual extinction.
We assume a distance of 3.5 kpc and derive the extinction for each source from the 2MASS
CC diagram {\it (H-K)} versus {\it (J-H)} presented in Fig. \ref{cc}.
Once the SED is obtained, we check whether the fitted effective
temperature is consistent with the temperature of an O- and B-type star.
Our analysis is only qualitative because of the uncertainty in the determination of
the spectral type from the CM diagram and because the fitting photosphere model extends only to 50,000 K.
The fitted effective temperatures of stars 5, 10, and
16 obtained from the SED is $\sim$ 6,000 K and for star 20 is $\sim$ 20,000 K.
These values are too low for a O- or early B-type star. For the remaining sources 8, 9, 12, 15, 17, 18, 20, and 21
the fitted temperature is $\sim$ 50,000 K, which is in closer agreement with the effective temperature
expected for an O-type star (see \citealt{scha97}). Thus, based on their position inside the bubble,
we suggest that sources 8, 12, and 15 are the most reliable ionizing-star candidates of N65.
\subsection{Star formation around N65}
\label{yso}
Figure \ref{pointsrc} shows the distribution of both class I (yellow crosses) and class II (magenta crosses) point
sources around N65. We note the presence of a group of these sources right upon the
molecular clumps that surround the bubble above its PDR. As molecular clouds are star birth places, we expect
some of these sources to be YSOs whose formation could have been triggered by the expanding
IR dust bubble.
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{Fig09.ps}
\caption{\ Two-color image towards N65: $^{13}$CO emission integrated between 47 and 55 km s$^{-1}$ (in green with white contours) and
8 $\mu$m emission (in red). We show the distribution of class I (yellow crosses) and class II (magenta crosses)
sources around N65 derived from the CC diagram of Fig. \ref{ccdiagr}. We have labeled the sources located
upon the molecular cloud and close to the PDR. }
\label{pointsrc}
\end{figure}
We perform a fitting of the fluxes of YSO candidates in the IRAC and 2MASS bands,
to derive their spectral energy distribution (SED) and constrain
their evolutionary stage. We limit our study to the sources superimposed on the molecular cloud around N65
(see Fig. \ref{pointsrc}). The SED was obtained using the tool developed by
\citet{robit07} available online\footnote{http://caravan.astro.wisc.edu/protostars/}.
This tool consists of a grid of precomputed radiative transfer models with a fast $\chi^{2}$
~minimization algorithm. We select the models that satisfied the condition
\begin{equation}
\chi^{2} - \chi^{2}_{min} < 2N,
\label{chi2}
\end{equation}
where $\chi^{2}_{min}$~is the minimum value of the $\chi^{2}$~among all models,
and {\it N} is the number of input data fluxes (fluxes specified as upper limit do not contribute
to {\it N}). Hereafter, we refer to models satisfying Eq. (\ref{chi2}) as ``selected models".
To relate the SED to the evolutionary stage of the YSO, \citet{robit06}
defined three different stages as a complementary classification of Allen's classes. The stage
classification is based on the values of the central source mass $M_\star$, the disk mass $M_{disk}$,
the envelope mass $M_{env}$, and the envelope accretion rate $\dot{M}_{env}$ of the YSO.
Stage I YSOs are those that have $\dot{M}_{env}/M_\star>10^{-6}~yr^{-1}$, i.e., protostars with
large accretion envelopes;
stage II are those with $\dot{M}_{disk}/M_\star>10^{-6}~yr^{-1}$ and $\dot{M}_{env}/M_\star<10^{-6}~yr^{-1}$, i.e.,
young objects with prominent disks; and stage III are those with $\dot{M}_{disk}/M_\star<10^{-6}~yr^{-1}$ and
$\dot{M}_{env}/M_\star<10^{-6}~yr^{-1}$, i.e., evolved sources where the flux is dominated by the central source.
The stage classification is based on the physical properties of the YSO,
rather than on the spectral index derived from the slope of
its SED. However, as pointed by \citet{robit06}, the stage classification is equivalent to
the class classification for a great number of models. Therefore, an easy way of
identifying young objects is looking for sources that can be fitted only by selected models with $\dot{M}_{env}>0$.
Sources that are fitted by models with zero and non-zero values of $\dot{M}_{env}$ may be
more evolved YSOs whose evolutionary stage is not constrained by the data (see \citealt{poulton08}).
In Table \ref{avelines}, we report the main results of the fitting output for class I and II sources
projected onto the molecular cloud. The YSO identification (Col. 1)
is taken from Fig. \ref{pointsrc}. In Col. 2 and 3, we report the $\chi^{2}$~per data point of the best-fit model and
the number on models satisfying Eq. (\ref{chi2}), respectively. The remaining columns
report the physical parameters of the source, specifying the range of values of the selected models:
central source mass, disk mass, envelope mass, and envelope accretion rate, respectively.
These results permit us to confirm the presence of young objects in the vicinity of the IR dust bubble N65.
\begin{table}[h]
\caption{Parameters of class I and II YSO candidates derived from the SED fitting.}
\label{avelines}
\centering
\begin{tabular}{ccccccc}
\hline\hline
Source& $\chi^{2}_{min}/{\it N}$ &{\it n} &M$_{\star}$ &M$_{disk}$ &M$_{env}$ &$\dot{M}_{env}$ \\
& & & (M$_{\odot}$) & (M$_{\odot}$) & (M$_{\odot}$) & (M$_{\odot}/yr$) \\
\hline
1 & 0.13& 484& $1 - 11$ & $8.3\times{10^{-8}} - 0.4$ & $6.4\times{10^{-9}} - 1800$ & $0 - 0.003$\\
2 & 2.2& 41 & $5 - 8$ & $7.1\times{10^{-4}} - 0.7$ & $1.1\times{10^{-8}} - 1200$ & $0 - 0.002$\\
3 & 0.1& 84 & $1 - 11$& $1.1\times{10^{-6}} - 0.71$ & $1.7\times{10^{-8}} - 140$ & $0 - 0.001$\\
4 & 0.2 & 273 & $0.2 - 11$ & $5.5\times{10^{-5}} - 0.6$ & $3.8\times{10^{-9}} - 360$ & $0 - 0.003$\\
5 & 5 & 4 & 9 & 0 & 28 & $3.8\times{10^{-4}}$ \\
6 & 2.2 & 2 & 4.6 & 0.014 & 37 & $2.1\times{10^{-4}}$ \\
7 & 10 & 9 & $4 - 5$ & $7.4\times{10^{-5}} - 0.02$ & $0.1 - 4$ & $5\times{10^{-6}} - 9\times{10^{-5}} $\\
8 & 1.3 & 1 & 9.6 & 0.75 & 130 & $2.1\times{10^{-4}}$\\
9 & 5 & 61 & $2 - 7$& $1.6\times{10^{-7}} - 0.13$ & $1.8\times{10^{-7}} - 17$ & $0 - 5.4\times{10^{-4}}$\\
10& 4 & 9 & $5$ & $4.5\times{10^{-5}} - 0.001 $ & $0.09 - 1.5$ & $9.5\times{10^{-7}} - 1.4\times{10^{-5}}$ \\
11& 3.3 & 10 & $2 - 4$ & $6.5\times{10^{-4}} - 0.03$ & $0.02 - 16$ & $4\times{10^{-6}} - 5\times{10^{-4}}$ \\
12& 11 & 2 & $3 -4$ & $0.016 - 0.02$ & $2 - 8 $ & $ 3.5\times{10^{-6}} - 2\times{10^{-5}}$ \\
13& 2.4 & 13 & $1 - 7$ & $1.1\times{10^{-6}} - 0.3 $ & $0.006 - 43$ & $3.5\times{10^{-8}} - 1.1\times{10^{-4}}$ \\
14& 1.4 & 298 & $1 - 11$ & $3\times{10^{-6}} - 0.7$ & $5\times{10^{-9}} - 140$ & $ 0 - 0.002$\\
15& 1 & 3 & $8 - 9$ & $0.03 - 0.2 $ & $70 - 140$ & $1.4\times{10^{-8}} - 2\times{10^{-4}} $ \\
16& 0.03 & 430 & $1 - 8$ & $3\times{10^{-8}} - 0.4 $ & $1\times{10^{-8}} - 23$ & $0 - 4\times{10^{-4}}$\\
17& 3 & 2 & 4.6 & 0.014 & 37 & $2.1\times{10^{-4}}$ \\
18& 0.8 & 13 & $2 - 9$ & $6.7\times{10^{-4}} - 0.7$ & $0.1 - 140$ & $3\times{10^{-6}} - 4\times{10^{-4}} $ \\
19& 0.4 & 85& $2 - 21$ & $0 - 1.1$ & $0.4 - 1300$ & $4.1\times{10^{-5}} - 0.002$\\
20& 2& 2 & $9 - 10$ & $0.3 - 0.8$ & $65 - 130$ & $2\times{10^{-4}}$ \\
21&0.003& 1652& $0.7 - 15$ & $6\times{10^{-7}} - 0.9 $ & $4.5\times{10^{-9}} - 560$ & $0 - 0.001$ \\
22& 1.3& 50 & $3 - 15$& $9\times{10^{-9}} - 0.2$ & $1.4\times{10^{-8}} - 200$ & $ 0 - 6\times{10^{-4}}$ \\
\hline
\end{tabular}
\end{table}
We identify 12 sources that are only fitted by models with $\dot{M}_{env}>0$; they are sources 5, 6, 7, 10,
11, 12, 13, 15, 17, 18, 19, and 20. These are presumably young objects embedded in prominent envelopes.
In every case, the selected models are prevalently stage I with a few stage II.
The only exception is source 13, which is fitted by both stage I and stage III models.
Sources 5, 15, and 20 appear to be the youngest YSOs. They may all be massive central sources (between 8 and 10 M$_\odot$)
surrounded by envelopes of several solar masses. In the case of 15 and 20, their ages are estimated to be
$\sim{10^{4}}$ years. Regarding the source 5, its position is coincident with the ultra-compact HII region G35.02+0.35 and
its age is about $10^{3}$ years. This is a really young massive source and a clear sign that star formation is
still occurring in the vicinity of N65.
The selected models (i.e., models satisfying Eq. \ref{chi2}) for sources 7, 10, 11, and 12 spread
over a similar range of values among them. These YSOs contain
a central source of mass less than 5 M$_\odot$~and ages between $10^{4}$ and $10^{5}$ years.
For source 19, we obtain a large number of selected models (85) because for this source we only fitted the four
IRAC fluxes. However, all of them are stage I so this source is probably a young object. The large spread in the
range of parameters does not allow us to constraint its mass and age.
Projected onto the molecular cloud, we also identify sources with zero and non-zero values
for $\dot{M}_{env}$: sources 1, 2, 3, 4, 9, 14, 16, 21, and 22. The evolutionary stage of these sources
cannot be assured without doubt because the parameters of the selected models are spread over a wide range.
We should include fluxes measured at other wavelengths to reduce the number of models satisfying
the condition defined in Eq. (\ref{chi2}). However, we can identify sources that are in their early stages
of evolution (embedded or disk dominated) because the selected models are either stage I or II.
This is the case for sources 1, 2, 4, 14, and 16.
For the remaining sources, 3, 9, 16, and 22, the selected models are distributed between stages I, II, and III so their
evolutionary stage cannot be established without controversy.
We note that source 21 is located at the center of the smaller IR bubble
introduced in Sect. \ref{resultdisc} that lies right upon the PDR of N65 (see Fig. \ref{smallbub}).
We fitted the SED of the source and obtained more than
1600 models satisfying Eq. (\ref{chi2}) and, consequently, a wide range of physical parameters for the YSO,
so we cannot establish its evolutionary stage.
However, its location in the center of the smaller IR dust bubbles suggests that
this source may be a second generation massive star blowing a stellar
wind strong enough to evacuate a bubble around itself. More observations are needed to investigate this scenario.
In conclusion, the SED study confirms the existence of several YSOs possibly embedded in the molecular clumps that
surround N65, which is indicative of the collect and collapse process occurring in this region.
\subsection{SED analysis of the EGO G35.03+0.35}
\label{ego}
We derive the SED for the EGO G35.05+0.35 using the fluxes in the {\it Spitzer}-IRAC 4.5, 5.8, and 8.0 $\mu$m bands
(the GPSC reports no detection in the 3.6 $\mu$m band and the source is absent in the 2MASS catalog). We also use
fluxes in the far IR of the SCUBA bands of 450 and 850 $\mu$m and in the band of
1200 $\mu$m measured by the SIMBA bolometer at SEST \citep{hill06}. Both the SCUBA and SIMBA datasets
have a lower angular resolution than that of the GPSC and the measured fluxes may include
contributions from other sources near the EGO. For this reason, we perform a fitting of the SED by considering
two different cases: a$)$ assuming that the far IR data represent the flux emitted only by the EGO, the SED
of the best-fit model being shown in Fig. \ref{sedEGO} ({\it left}); and b$)$ using the
far IR data as an upper limit (i.e., assuming that the emission originates in the EGO and, probably,
other sources around it), the SED of the best-fit model in this case being shown in Fig. \ref{sedEGO} ({\it right}).
In Table \ref{tablaEGO}, we report the physical parameters of the selected models,
considering both SCUBA and SIMBA data as flux emitted only by the EGO (case {\it a}) and upper limits (case {\it b}).
The column designation is the same as in Table \ref{avelines}.
\begin{figure}[h]
\centering
\includegraphics[height=5.5cm]{Fig10.ps}
\includegraphics[height=5.5cm]{Fig11.ps}
\caption{\ Best-fit SED model for the EGO G35.05+0.35. The points in blue are the measured fluxes at
the {\it Spitzer}-IRAC bands of 4.5, 5.8, and 8.0 $\mu$m, the SCUBA bands of 450 and 850 $\mu$m, and the band of
1,200 $\mu$m measured with the SIMBA bolometer at SEST.
{\it Left}: SCUBA and SIMBA data are assumed to represent flux emitted only by the EGO.
The total flux (black line) is completely dominated by the
flux from the envelope as would be expected in a young stellar object at its earlier stages of evolution.
{\it Right}: SCUBA and SIMBA data are taken to be an upper limit.
The total flux at the longer wavelengths is dominated by the
envelope, while at the shorter wavelengths the main contributions originates in the disk (red line)
and the central source (blue line). }
\label{sedEGO}
\end{figure}
\begin{table}[h]
\caption{Physical parameters of EGO G35.05+0.35 derived from the SED fitting for the selected models satisfying
Eq. (\ref{chi2}). }
\label{tablaEGO}
\centering
\begin{tabular}{ccccccc}
\hline\hline
{\it Case} & $\chi^{2}_{min}/{\it N}$ &{\it n} &M$_{\star}$ &M$_{disk}$ &M$_{env}$ &$\dot{M}_{env}$ \\
& & & (M$_{\odot}$) & (M$_{\odot}$) & (M$_{\odot}$) & (M$_{\odot}/yr$) \\
\hline
a & 30 & 2 & $27 - 33$ & $0 - 0.3$ & $1900 - 2500$ & $ 0.004 - 0.006$\\
b & 3.3 & 16 & $6 - 10$ & $0.005 - 0.8$ & $0.7- 1100$ & $ 2\times{10^{-5}} - 0.002 $\\
\hline
\end{tabular}
\end{table}
According to Fig. \ref{sedEGO} ({\it left}), if we consider fluxes at the longer wavelengths
to be emitted only by the EGO, the total flux of the best-fit model is completely dominated by the envelope.
As pointed out by \citet{robit06}, the envelope of
sources with high accretion rates becomes optically thick. In our case of $\dot{M}_{env}\sim10^{-3}$~M$_{\odot}/yr$,
radiation emitted
by both the central source and the disk is probably absorbed and re-emitted by the envelope of gas and dust.
As can be seen in Table \ref{tablaEGO}, the physical parameters of the EGO indicate that it is a massive
central source surrounded by a massive envelope.
According to Fig. \ref{sedEGO} ({\it right}), if we take SCUBA and SIMBA measurements to be upper limits, the total flux
of the best-fit model is still dominated by the envelope but only at the longer wavelengths. In the near and
mid IR, the main contribution to the total flux comes from the disk.
The accretion rate values that we derived for the envelope are
lower than those derived in the previous case (see Table \ref{tablaEGO}).
Thus, the envelope is expected to be more transparent to radiation coming from the inner
regions (disk and central source).
Both cases show that the fitted SED of the EGO is indicative of a massive central source
surrounded by a massive envelope. The age of the YSO is estimated in between $10^{4}$~and $10^{5}$~years.
From these results, we suggest that the EGO G35.05+0.35 is a candidate massive young stellar
object at the earlier stages of evolution with
outflowing activity, in agreement with that proposed by \citet{cyga08}.
The presence of this object is a clear sign that star
formation is occurring in the vicinity of N65.
\section{Summary}
\label{sum}
We have conducted a study of the surroundings of the IR dust bubble N65 using large-scale surveys
and archival data in the radio and infrared bands.
The main results can be summarized as follows:
\begin{enumerate}
\item We discovered the presence of a molecular cloud with an arc-like shape enclosing N65
over its east, north and west borders and centered on a velocity of $\sim 50$ km s$^{-1}$.
The cloud appears to be fragmented into smaller molecular clumps that are seen right upon the PDR.
Based on the distribution and morphology of the molecular gas between 47 and 55 km s$^{-1}$, we
suggest that the collect and collapse process may be occurring.
\item The EGO G35.05+0.35 lies in the most intense molecular clump, which belongs to the
discovered molecular shell. From the $^{13}$CO J=1--0 emission, we estimated its molecular mass and density to be
$\sim 2 \times 10^{3}$ M$_\odot$~and $\sim 10^{4}$ cm$^{-3}$, respectively. We found that the
dust continuum emission, traced by the 850 $\mu$m band, also peaks in this region. We
estimated a dust-to-gas ratio of $\sim 0.003$ for this clump.
\item From the analysis of the infrared photometry of point sources around N65,
we identified three O- or early B-type stars as the most probable candidates responsible for the creation of N65.
In addition, we found many YSOs distributed right upon
the molecular clumps around N65. The analysis of the SED of these sources confirms that many of them
are in the embedded or disk-dominated stage. This suggests that the gas collected into the molecular
clumps may have collapsed to form a second generation of stars, showing another galactic case
of the collect and collapse process.
\item The SED of EGO G35.05+0.35 as derived by fitting the fluxes at the mid and far IR, confirms
that it is a massive young stellar object.
\end{enumerate}
\section*{Acknowledgments}
A.P. is a doctoral fellow of CONICET, Argentina.
S.P. and E.G. are members of the {\sl Carrera del investigador cient\'\i fico} of CONICET, Argentina.
This work was partially supported by the CONICET grant PIP 112-200801-02166, UBACYT A023 and ANPCYT
PICT-2007-00902.
\bibliographystyle{aa}
|
\section{Introduction}
It is well known that cosmic strings are produced in the early Universe
at the phase transition associated with spontaneous symmetry breaking in
the Grand-Unified-Theories (GUTs)~\cite{Vilenkin}. Although cosmic strings
formed before inflation are diluted away, some GUTs predict series
of phase transitions where symmetry such as $U(1)_{B-L}$ ($B$ and $L$
denotes baryon and lepton number) breaks at low energy and produces
strings after inflation.
Cosmic strings are also produced at the end of brane inflation
in the framework of the superstring theory~\cite{Sarangi:2002yt,Dvali:2003zj}.
Once produced, they
survive until now and can leave observable signatures. Thus, the cosmic
strings provide us with an opportunity to probe unified theories in
particle physics which cannot be tested in terrestrial experiments.
Various cosmological and astrophysical signals of cosmic strings have
been intensively studied for decades. Especially, many authors have
studied gravitational waves (GW) emitted from the cosmic string network,
in particular, GWs from cosmic string loops. Cosmic string loops
oscillate by their tension and emit low frequency GWs corresponding to
their size~\cite{Caldwell:1991jj,DePies:2007bm}. Moreover, string loops
generically have cusps and kinks, and these structures cause high
frequency modes of GWs, i.e. GW bursts~%
\cite{Damour:2000wa,Damour:2001bk,Damour:2004kw,Siemens:2006yp}.
On the other hand, GWs from infinite strings (long strings which lie
across the Hubble horizon) have attracted much less interest than those
from loops. This is because there are only a few infinite strings
in one Hubble horizon, so GWs from them are much weaker than those from
loops whose number density in the Hubble volume is much larger.\footnote
For cosmic strings whose reconnection probability is small, such as
cosmic superstrings, many infinite strings can exist in a Hubble
horizon. For such kinds of cosmic strings, however, the number of loops
in a horizon is accordingly large, hence the situation that the loop
contribution is stronger than infinite strings remains unchanged.}
Loops can emit GWs with frequency $\omega \gtrsim (\alpha t)^{-1}$,
where $\alpha$ is the parameter which represents the typical loop size
normalized by the horizon scale. Therefor, unless $\alpha$ is much less
than $1$, GWs from loops dominate in the wide range of frequency bands
detected by GW detection experiments, e.g. pulsar timing or ground
based or space-borne GW detectors.
However, there are several reasons to consider GWs of infinite strings.
First, recent papers suggest that $\alpha$
is much less than previously thought~\cite{Siemens:2002dj,Polchinski:2006ee},
and hence loops cannot emit low frequency GWs which can be detected
by pulsar timing experiments.
If this is true,
GWs from infinite strings can give dominant contribution to the GW
background at those low frequencies, as we will show in this paper.
Second, infinite strings can emit GWs with wavelength of horizon scale
because the characteristic scale of infinite strings is as long as the size
of the horizon. This may be detected by future and on-going CMB surveys.
Loops cannot emit GWs with such long wavelength.\footnote
Other recent simulations imply $\alpha \sim 0.1$, which is much bigger
than $G\mu$. In this case loops can emit GWs whose wavelength is
comparable to or somewhat shorter than the horizon scale~\cite{Ringeval:2005kr}.}%
Thus, we are lead to consider GWs from infinite strings.
A structure on infinite strings which is responsible for GW emission is a kink.\footnote
Generically, infinite strings do not have any cusp since its existence depends
on the boundary condition of the string, i.e., the periodic condition. }%
Kinks are produced when infinite strings reconnect and kinks on the
infinite strings can produce GW bursts. The kinks are sharp when they
first appear, but are gradually smoothened as the Universe expands.
Moreover, when an infinite string self-intercommutes
and produces a loop, some kinks immigrate from the infinite string to the loop,
and hence the number of kinks on the infinite string decreases.
Recently, Copeland and Kibble derived the distribution function of
kinks on an infinite string taking into account these effects~\cite{Copeland:2009dk}.
However, they did not study the GWs from kinks.
Therefore, in this paper, we investigate GWs from
kinks using the kink distribution obtained in~\cite{Copeland:2009dk}
and discuss the detectability of them by future GW experiments,
especially by pulsar timing, such as Square Kilometer Array (SKA)
and space-based detectors such as DECIGO and BBO.
This paper is organized as follows.
In section 2, we briefly review a part of the basis of cosmic strings.
In section 3, we derive the distribution function of kinks on infinite strings according to
~\cite{Copeland:2009dk}.
In section 4, we derive the formula of the energy radiated from kinks per unit time.
In section 5, we calculate the spectrum of stochastic GW background originating from kinks, using formula derived in section 4.
Section 6 is devoted to summary and discussion.
\section{Dynamics of cosmic strings}
The dynamics of a cosmic string, whose width can be neglected, is described
by the Nambu-Goto action,
\begin{equation}
S=-\mu \int d^2 \zeta \sqrt{-\det(\gamma_{ab})}. \label{NGaction}
\end{equation}
where $\zeta^a \ (a=0,1)$ are coordinates on the world sheet of the cosmic
string, $\gamma_{ab} = g_{\mu \nu} x^{\mu}_{, a} x^{\nu}_{,b}$
($x^{\mu}_{,a}=\frac{\partial x^{\mu}}{\partial \zeta^a}$ ) is the induced
metric on the world sheet, and $\mu$ is the tension of the string.
The energy-momentum tensor is
\begin{equation}
T^{\mu \nu}(x)=\mu \int d^2\zeta \sqrt{-\det(\gamma_{ab})}
\gamma^{ab}x^{\mu}_{,a}x^{\nu}_{,b} \delta^{4}(x-X(\zeta)),
\label{EMtensor}
\end{equation}
where $X=X(\zeta)$ is embedding of the world sheet on the background
metric.
If the background space-time is Minkowski one, we can select the coordinate
system $(\zeta^0,\zeta^1)=(\tau, \sigma)$ which satisfies the gauge conditions
\begin{equation}
\tau = t \ (\mathrm{physical \ time}),
~~~~~~\ x_{,\tau} \cdot x_{,\sigma}=0, \
~~~~~~x^2_{,\tau} + x^2_{,\sigma} =0.
\end{equation}
The time scale of a GW burst is much shorter than the Hubble expansion,
hence we consider an individual burst event on the Minkowskian background.
The general solution of the equation of motion derived from the action
(\ref{NGaction}) is
\begin{equation}
x^{\mu} = \frac{1}{2} (a^{\mu}(u) + b^{\mu}(v)), \
\mathbf{a}^{\prime 2}(u) = \mathbf{b}^{\prime 2}(v)=1
\end{equation}
where $u=\sigma+t, v=\sigma-t$. We call $a(u)$ $(b(v))$ the
left (right)-moving mode.
Then, Eq.~(\ref{EMtensor}) can be rewritten in terms of $a(u)$ and $b(v)$,
\begin{equation}
T^{\mu \nu}(k) = \frac{\mu}{4}(I^{\mu}_+(k)I^{\nu}_-(k)
+ I^{\mu}_-(k)I^{\nu}_+(k)),
\label{EMtensor2}
\end{equation}
\begin{equation}
I_+^{\mu}(k)=\int du a^{\prime\mu}(u) e^{ik\cdot a(u)/2}, \
I_-^{\mu}(k) = \int dv b^{\prime\mu}(v) e^{ik\cdot b(v)/2}, \label{I}
\end{equation}
where $T^{\mu \nu}(k)$ is the Fourier transform of the $T^{\mu \nu}(x)$, i.e.
$T^{\mu \nu}(k) = \int d^4x T^{\mu \nu}(x)e^{ik\cdot x}$.
\section{Distribution function of kinks}
Kinks can be defined as discontinuities of $\mathbf{a}^{\prime}$ or
$\mathbf{b}^{\prime}$.
They are produced when two infinite strings collide and reconnect
because $\mathbf{a}^{\prime}$ and $\mathbf{b}^{\prime}$ on the new infinite
string are created by connecting $\mathbf{a}^{\prime}$s or
$\mathbf{b}^{\prime}$s on two different strings.
Let us suppose that $\mathbf{a}^{\prime}$ jumps from $\mathbf{a}^{\prime}_-$ to $\mathbf{a}^{\prime}_+$
at a kink. Then the ``sharpness'' of a kink is defined by
\begin{equation}
\psi = \frac{1}{2}(1-\mathbf{a}^{\prime}_+ \cdot \mathbf{a}^{\prime}_-).
\end{equation}
Thus the norm of the difference between $\mathbf{a}^{\prime}_-$ and $\mathbf{a}^{\prime}_+$ is
$|\Delta \mathbf{a}^{\prime} | = 2\sqrt{\psi}$.
The production rate of kinks is given by~\cite{Copeland:2009dk}
\begin{equation}
\dot{N}_{\rm production} = \frac{\bar{\Delta}V}{\gamma^4t^4} g(\psi) ,
\end{equation}
where $N(t,\psi)d\psi$ denotes the number of kinks
with sharpness between $\psi$ and $\psi+d\psi$ in the volume $V$,
$\bar{\Delta}$ and $\gamma$ are constants related to string networks,
whose values are~\cite{Copeland:2009dk},
\begin{equation}
\bar{\Delta}_r \simeq 0.20, \ \bar{\Delta}_m \simeq 0.21, \
\gamma_r\simeq0.3, \ \gamma_m\simeq0.55,
\end{equation}
Here the subscript $r$($m$) denotes the value in the
radiation(matter)-dominated era.
\begin{equation}
g(\psi)=\frac{35}{256}\sqrt{\psi}(15-6\psi -\psi^2) \label{g}
\end{equation}
and we set $g(\psi)=0$ for $\psi<0, 1<\psi$.
The correlation length of the cosmic strings $\xi$ is given by
$\xi \simeq \gamma t$.
Produced kinks are blunted by the expansion of the Universe.
The blunting rate of the kink with the sharpness $\psi$ is given
by~\cite{Copeland:2009dk}
\begin{equation}
\frac{\dot{\psi}}{\psi}\bigg|_{\rm stretch} = -2\zeta t^{-1},
\end{equation}
where $\zeta$ is a constant which, in the radiation(matter)-dominated era,
is given by $\zeta_r \simeq 0.09~ (\zeta_m \simeq 0.2)$.
On the other hand, the number of kinks on an infinite string decreases
when it self-intercommutes since some kinks are taken away by the loop
produced. The decrease rate of kinks due to this effect is given
by~\cite{Copeland:2009dk}
\begin{equation}
\frac{\dot{N}}{N}\bigg|_{\rm to~loop}=-\frac{\eta}{\gamma t},
\end{equation}
where $\eta$ is constant which, in the radiation(matter)-dominated era,
is given by $\eta_r \simeq 0.18 ~(\eta_m \simeq 0.1$).
Taking into account these effects, the evolution of the kink number $N$
obeys the following equation,
\begin{equation}
\dot N = \frac{\bar\Delta V}{\gamma^4 t^4}g(\psi) +
\frac{2\zeta}{t}\frac{\partial}{\partial\psi}(\psi N) -
\frac{\eta}{\gamma t}N. \label{Neq}
\end{equation}
\begin{table}[t]
\begin{center}
\begin{tabular}{ | c | c | c | c | }
\hline
~ & definition & radiation dom. & matter dom.
\\ \hline
$\gamma$ & - & 0.3 & 0.55 \\
$\zeta$ & - & 0.09 & 0.2 \\
$\beta$ & - & 1.1 & 1.2 \\
$A$ & $2\zeta-\beta$ & $-0.92$ & $-0.8$ \\
$B$ & $4\zeta-\beta$ & $-0.74$ & $-0.4$ \\
$C$ & $(\beta-8\zeta)/[3(\beta-2\zeta)]$ & 0.14 & $-0.17$ \\
\hline
\end{tabular}
\caption{
Various constants appearing in the calculation of kink distribution and
the spectrum of GWs from kinks.
}
\label{table:constants}
\end{center}
\end{table}
We have to solve this equation under an appropriate initial condition.
The initial condition to be imposed depends on how strings emerge.
We consider two typical scenarios.
As a first scenario, let us suppose that
cosmic strings form when spontaneous symmetry breaking (SSB) occurs
in the radiation dominated Universe.
After the formation, cosmic strings interact with particles in thermal bath,
which acts as friction on the string.
At first this friction effect overcomes the Hubble expansion.
(This epoch is called friction-dominated era.)
The important observation is that the friction effect smoothens strings
and washes out the small scale structure on them~\cite{Garriga:1993gj}.
The temperature at which friction domination ends is given by~\cite{Vilenkin:1991zk}
\begin{equation}
T_c \sim G\mu M_{\rm pl}, \label{Tefd}
\end{equation}
where $G$ is the Newton constant and $M_{\rm pl}$ is the Planck scale.
If the temperature at the string formation is lower than this critical
value,
the friction effect can be neglected from the formation epoch to the present
day, and kinks emerge at the same time as appearance of strings.
As a second scenario, let us suppose that cosmic strings are generated
at the end of inflation, like strings made by condensation of waterfall
fields in supersymmetric hybrid
inflation~\cite{Dvali:1994ms,Binetruy:1996xj,Jeannerot:1997is},
and cosmic superstrings left after annihilation of the D-brane and
anti D-brane in brane inflation~\cite{Sarangi:2002yt,Dvali:2003zj}.
In this case, cosmic strings form in the inflaton-oscillation dominated era,
which resembles the matter-dominated era.
Although strings feel frictions from dilute plasma existing before the
reheating completes, the effect is negligible as long as the reheating
temperature after inflation $T_{r}$ is lower than $\sim G\mu M_{\rm pl}$.
In this case, kinks begin to be formed right after the formation of strings.
Otherwise, the temperature at which kinks begin to appear is given by
Eq.~(\ref{Tefd}).
In anyway, even if such kinks survive friction, they become extremely
dense in the later period, so that they do not affect the observable
part of the GW spectrum, as we will see later.
On the other hand, if the reheating temperature is lower than $T_c$,
the kinks from the first matter era definitely continue to exist,
and if the reheating temperature is extremely low, GWs from these kinks
can be observed.
Therefore, we concentrate on the situation where the reheating temperature
is low enough so that the kinks generated in the first matter era survive
without experiencing the friction domination.
We denote the time when kinks starts to be formed by $t_*$.
Then we can get the solution, but its precise form is very complicated.
Here we assume that the matter-dominated epoch follows after inflation, and
the reheating completes at $t=t_r$ after which the radiation dominated era
begins.
If we focus on only the dominant term and neglect $\mathcal{O}(1)$ numerical
factor, we get
\begin{equation}
\frac{dN}{d\psi}(t,\psi) \sim
\begin{cases}
\psi^{-\beta_m/2\zeta_m} t^{-1}
& {\rm for~~}\psi> \left( \frac{t_*}{t} \right)^{2\zeta_m} \\
\left( \frac{t}{t_*} \right)^{\beta_m+\zeta_m}\psi^{1/2} t^{-1}
& {\rm for~~}\psi< \left( \frac{t_*}{t} \right)^{2\zeta_m}
\end{cases}
\label{bunpueasymat1}
\end{equation}
in the first matter era,
\begin{equation}
\frac{dN}{d\psi}(t,\psi) \sim
\begin{cases}
\psi^{-\beta_r/2\zeta_r} t^{-1}
& {\rm for~~}\psi> \psi^{\rm (RD)}_1 (t) \\
\left( \frac{t}{t_r} \right)^{\beta_r-\beta_m\zeta_r/\zeta_m}
\psi^{-\beta_m/2\zeta_m} t^{-1}
& {\rm for~~}\psi^{\rm (RD)}_2 (t)
<\psi< \psi^{\rm (RD)}_1 (t) \\
\psi^{1/2} \left( \frac{t}{t_r}\right)^{\beta_r+\zeta_r}
\left( \frac{t_r}{t_*} \right)^{\beta_m+\zeta_m}t^{-1}
& {\rm for~~}\psi < \psi^{\rm (RD)}_2 (t)
\end{cases}
\label{bunpueasyrad}
\end{equation}
in the radiation era where
\begin{eqnarray}
&\psi^{\rm (RD)}_1 (t)&= \left( \frac{t_r}{t} \right)^{2\zeta_r},\\
&\psi^{\rm (RD)}_2 (t)&=
\left(\frac{t_r}{t}\right)^{2\zeta_r} \left(\frac{t_*}{t_r}\right)^{2\zeta_m},
\end{eqnarray}
and
\begin{align}
\frac{dN}{d\psi}(t,\psi) \sim
&
\begin{cases}
\psi^{-\beta_m/2\zeta_m} t^{-1}
& {\rm for~~}\psi > \psi^{\rm (MD)}_1(t)\\
\left( \frac{t}{t_{eq}} \right)^{\beta_m-\beta_r\zeta_m/\zeta_r}
\psi^{-\beta_r/2\zeta_r} t^{-1}
&{\rm for~~} \psi^{\rm (MD)}_2(t)
<\psi< \psi^{\rm (MD)}_1(t) \\
\left( \frac{t_{eq}}{t_r} \right)^{\beta_r-\beta_m\zeta_r/\zeta_m}
\psi^{-\beta_m/2\zeta_m} t^{-1}
&{\rm for~~} \psi^{\rm (MD)}_3(t) <\psi < \psi^{\rm (MD)}_2(t) \\
\psi^{1/2} \left(\frac{t}{t_{eq}}\right)^{\beta_m+\zeta_m}
\left( \frac{t_{eq}}{t_r}\right)^{\beta_r+\zeta_r}
\left( \frac{t_r}{t_*} \right)^{\beta_m+\zeta_m}t^{-1}
&{\rm for~~}\psi < \psi^{\rm (MD)}_3(t)
\end{cases} \label{bunpueasymat2}
\end{align}
in the second matter era,
where
\begin{eqnarray}
&\psi^{\rm (MD)}_1 (t)&= \left( \frac{t_{eq}}{t} \right)^{2\zeta_m},
\label{psiMD1} \\
&\psi^{\rm (MD)}_2 (t)&=
\left(\frac{t_{eq}}{t}\right)^{2\zeta_m}
\left(\frac{t_r}{t_{eq}}\right)^{2\zeta_r}, \\
&\psi^{\rm (MD)}_3 (t)&=
\left(\frac{t_{eq}}{t}\right)^{2\zeta_m}\left(\frac{t_r}{t_{eq}}\right)^{2\zeta_r}
\left(\frac{t_*}{t_r}\right)^{2\zeta_m}.
\label{psiMD3}
\end{eqnarray}
Here $\beta$ is the constant related to the string network ($\beta_r
\simeq 1.1, \ \beta_m \simeq 1.2$), and $t_{eq}$ denotes the matter-radiation equality epoch.
We have converted $N$, distribution in the volume $V$, to
$dN/d\psi= N(t,\psi)/(V/\xi^2)$, which is the distribution per unit length.
The derivation of the above expression of $dN/d\psi$ is described in Appendix~\ref{sec:app0}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1\linewidth]{kinkdistribution.eps}
\caption{The distribution function of kinks on infinite strings produced
at the end of the inflation in the second matter era.
$\psi_1^{\rm (MD)}$, $\psi_2^{\rm (MD)}$, $\psi_3^{\rm (MD)}$ are given by
Eqs.~(\ref{psiMD1})-(\ref{psiMD3}). }
\label{fig:kinkdist}
\end{center}
\end{figure}
Let us consider the physical meaning of this distribution function.
It is not difficult to consider how the number of kinks in the horizon
changes as time goes on.
When a kink is born, its sharpness ranges from 0 to 1,
but the typical value is $\mathcal O(0.1)$.
Therefore, the kink distribution has a peak at $\psi\sim
\mathcal O(0.1)$ at the very early stage.
Then kinks are made blunt by the cosmic expansion.
Thus newly produced kinks are sharp (large $\psi$) and old ones are blunt
(small $\psi$).
Therefore, the peak value of $\psi$ is much smaller than 0.1 at the
late stage.
This peak consists of the oldest kinks.
Fig.~\ref{fig:kinkdist} roughly sketches the shape of the distribution
function in the second matter era.
It should be noted that the above distribution is derived
without considering gravitational backreaction.
Since the most abundant kinks are extremely blunt,
they might be influenced by the gravitational backreaction and disappear.
However, it is difficult and beyond the scope of this paper to take backreaction
into account.
We will comment on this issue in Appendix~\ref{sec:appB}.
\section{THE SPECTRUM OF GRAVITATIONAL WAVES FROM KINKS}
Given the energy-momentum tensor of the source,
one can calculate the energy of the GW in the direction of $\hat{k}$
with frequency $\omega$ as~\cite{CandG}
\begin{eqnarray}
\frac{dE}{d\Omega}(\hat{k}) & =
& 2G\int^{\infty}_{0}d\omega \omega^2
\left( T^{\mu \nu *}(k)T_{\mu \nu}(k)
- \frac{1}{2}|T^{\mu}_{\mu}(k)|^2 \right) \nonumber \\
& = &2G \Lambda_{ij.lm}(\hat{k})\int^{\infty}_0 d\omega \omega^2
T^{ij*}(k)T^{lm}(k)
\end{eqnarray}
\begin{equation}
\Lambda_{ij,lm}(\hat{k}) \equiv
\delta_{il} \delta_{jm} - 2\hat{k}_j\hat{k}_m\delta_{il} +
\frac{1}{2}\hat{k}_i\hat{k}_j\hat{k}_l\hat{k}_m-\frac{1}{2}\delta_{ij}\delta_{lm}
+ \frac{1}{2}\delta_{ij}\hat{k}_l\hat{k}_m
+\frac{1}{2}\delta_{lm}\hat{k}_i\hat{k}_j
\end{equation}
Thus by substituting the energy-momentum tensor (\ref{EMtensor2}),
we find the energy of the GW after computing integral $I_{\pm}$
in Eq.~(\ref{I}).
It is given by
\begin{align}
\frac{dE}{d\omega d\Omega} (k)= \frac{G\mu^2\omega^2}{4}
\bigg(& |\vec{I}_+|^2|\vec{I}_-|^2 + |\vec{I}_+ \cdot \vec{I}_-|^2 - |\vec{I}_+\cdot\vec{I}_-^*|^2
\nonumber \\
&-|\vec{I}_+|^2|\hat{k} \cdot \vec{I}_-|^2-|\vec{I}_-|^2|\hat{k}\cdot\vec{I}_+|^2
+ |\hat{k}\cdot\vec{I}_+|^2|\hat{k}\cdot\vec{I}_-|^2 \nonumber \\
&- (\vec{I}_+\cdot\vec{I}_-^*)(\hat{k}\cdot\vec{I}_+^*)(\hat{k}\cdot\vec{I}_-)
- (\vec{I}_+^*\cdot\vec{I}_-)(\hat{k}\cdot\vec{I}_+)(\hat{k}\cdot\vec{I}_-^*)
\nonumber \\
&+ (\vec{I}_+\cdot\vec{I}_-)(\hat{k}\cdot\vec{I}_+^*)(\hat{k}\cdot\vec{I}_-^*)
+(\vec{I}_+^*\cdot\vec{I}_-^*)(\hat{k}\cdot\vec{I}_+)(\hat{k}\cdot\vec{I}_-) \bigg)
\label{Egeneral}
\end{align}
where $\vec{I}_{\pm} \equiv (I^1_{\pm}(k),I^2_{\pm}(k),I^3_{\pm}(k))$.
The method to calculate $I_{\pm}$ is described
in~\cite{Damour:2001bk,Binetruy:2009vt}.
In the limit $\omega \rightarrow \infty$, $I_{\pm}$ exponentially reduces to $0$,
unless at least one of the following conditions on the integrand is met.
One is the existence of discontinuities of $a^{\prime i}(u)$ (or $b^{\prime i}(v)$).
The contribution of a discontinuity of $\mathbf{a}^{\prime}$ to $I_+(k)$ is
\begin{equation}
I^i_+(k) \simeq -\frac{2}{i\omega}
\left(\frac{a^{\prime i}_+}{1-\hat{k}\cdot\mathbf{a}^{\prime}_+}
- \frac{a^{\prime i}_-}{1-\hat{k}\cdot\mathbf{a}^{\prime}_-} \right)
e^{i\omega(u_*-\hat{k}\cdot\mathbf{a}_*)/2}
\label{Idisc}
\end{equation}
where $u_*$ is the position of the discontinuity, $\mathbf{a}_*=\mathbf{a}(u_*)$
and we assume $\mathbf{a}^{\prime}$ jumps from $\mathbf{a}^{\prime}_-$ to $\mathbf{a}^{\prime}_+$ at $u=u_*$.
The region of length $\sim\omega^{-1}$ around $u=u_*$ contributes to this value.
We find $I^i_+(k) \sim \psi/\omega$, where $\psi$ denotes
sharpness of the kink.
The other condition is the existence of stationary points of the phase of the
integrand, i.e.
$\omega (u-\hat{k}\cdot\mathbf{a}(u))/2$ (or $\omega (-v-\hat{k}\cdot\mathbf{b}(v))/2$).
This condition is expressed as
\begin{equation}
1-\hat{k}\cdot\mathbf{a}^{\prime}(u_s)=0 \ (\mathrm{or} \ -1-\hat{k}\cdot\mathbf{b}^{\prime}(v_s)=0) \label{sta-con}
\end{equation}
at the point $u=u_s$($v=v_s$). The contribution of the stationary point of the phase to $I_+(k)$ is
\begin{equation}
I^i_+(k) \simeq \frac{1}{\omega^{2/3}} a^{\prime \prime}_i(u_s)
e^{i\omega(u_s-\hat{k}\cdot\mathbf{a}(u_s))/2}
\left( \frac{12}{|\hat{k}\cdot\mathbf{a}^{\prime \prime \prime}(u_s)|}\right)^{2/3}
\frac{i}{\sqrt{3}}\Gamma(2/3). \label{Ista}
\end{equation}
For any value of $\mathbf{a}^{\prime} \ (\mathbf{b}^{\prime})$, there is one direction $\hat{k}$ that
satisfies (\ref{sta-con}), i.e. $\hat{k} = \mathbf{a}^{\prime} \ (-\mathbf{b}^{\prime})$.
Therefore, every point of $u$ can contribute to $I^i_+(k)$ for one direction $\hat{k}$.
One can find the energy of the GW burst from ONE kink by picking up the contribution
from the discontinuity for one of $I_{\pm}$ (say, $I_+$) and
from the stationary point for the other (say, $I_-$).
The energy emitted when the kink is located at the world sheet coordinate
$(u,v)=(u_*,v_s)$ is evaluated by substituting (\ref{Idisc}) into $I_+$ and
(\ref{Ista}) into $I_-$ in (\ref {Egeneral}).
Then we obtain
\begin{align}
\frac{dE}{d\Omega d\omega}(\omega,-\omega \mathbf{b}^{\prime}_s)=&
\frac{\Gamma ^2\left(\frac{2}{3}\right)}{3}G\mu^2 \omega^{-\frac{4}{3}}
\left(\frac{12}{|\mathbf{b}^{\prime}_s\cdot\mathbf{b}^{\prime \prime \prime}_s|}
\right)^{\frac{4}{3}} \nonumber \\
&\times
\left(\frac{1}{(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{+} )^2}-
\frac{2 \mathbf{a}^{\prime}_+ \cdot\mathbf{a}^{\prime}_-}
{(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{+} )
(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{-} )}
+\frac{1}{(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{-} )^2}
\right)\mathbf{b}^{\prime \prime 2}_s,
\end{align}
where the subscript $s$ represents the value at $v=v_s$.
The energy is radiated at every moment toward the direction of $-\mathbf{b}^{\prime}$ from
the kink.
This is the formula which represents energy emitted in a short period in a small
solid angle.
The total energy emitted in the short period $\Delta t$ is found by multiplying
$\Delta \Omega$,
which is the solid angle that the GW sweeps in this short period.
The extent of the radiation has the solid angle
$\sim ((\omega / |\mathbf{b}^{\prime \prime}_s|)^{-1/3})^2$~\cite{Damour:2001bk}.
The variation of the direction of the GW is roughly estimated by
$|\Delta \mathbf{b}^{\prime}| \sim |\mathbf{b}^{\prime \prime} \Delta t|$.
As a result, we obtain
\begin{equation}
\Delta \Omega \sim (\omega / |\mathbf{b}^{\prime \prime}_s|)^{-1/3}
|\mathbf{b}^{\prime \prime}_s|\Delta t
\end{equation}
and the energy emitted per unit time is
\begin{equation}
\begin{split}
\frac{dP}{d\omega} \sim
&\frac{\Gamma ^2(2/3)}{3}G\mu^2 \omega^{-5/3}
\left(\frac{12}{|\mathbf{b}^{\prime}_s\cdot\mathbf{b}^{\prime \prime \prime}_s|}
\right)^{4/3} \\
&\times
\left(\frac{1}{(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{+} )^2}-
\frac{2 \mathbf{a}^{\prime}_+ \cdot\mathbf{a}^{\prime}_-}
{(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{+} )
(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{-} )}
+\frac{1}{(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_{-} )^2}
\right)|\mathbf{b}^{\prime \prime }_s|^{10/3},
\label{dPdomega}
\end{split}
\end{equation}
The terms in the large parenthesis in the 2nd line of Eq.~(\ref{dPdomega})
can be estimated by taking average
over the angle between the left- and right-moving mode as
$\langle \frac{1}{(1+\mathbf{b}^{\prime}_s \cdot \mathbf{a}^{\prime}_{\pm})^2} \rangle \sim
\langle \frac{1}{(1+\mathbf{b}^{\prime}_s \cdot \mathbf{a}^{\prime}_+)(1+\mathbf{b}^{\prime}_s\cdot\mathbf{a}^{\prime}_-)}
\rangle \sim 1$.
The magnitudes of $\mathbf{b}^{\prime \prime}_s$ and $\mathbf{b}^{\prime
\prime \prime}_s$ should be $\xi^{-1} (\sim t^{-1})$ and $\xi^{-2} (\sim t^{-2})$.\footnote{
Although $\mathbf{b}^{\prime}$ has small and dense discontinuities which represent kinks,
$v_s$ is
generally not a kink point. So the rough shape of $\mathbf{b}_s^{\prime}$ is determined by
the global
appearance of the string network and hence the length scale of its
variation is roughly
the curvature radius of the network.
}
After making these substitutions and neglecting $\mathcal{O}(1)$ numerical
factors, we find
\begin{equation}
\frac{dP}{d\omega} \bigg|_{\mathrm{one \ kink}}
\sim 10 G \mu^2 \psi \omega^{-5/3} t^{-2/3}.
\end{equation}
So far we have evaluated the GW spectrum from one kink.
However, there exist many kinks on an infinite string and the
final observable GW spectrum is made from sum of contribution from these kinks.
Therefore, $I_+$ picks contributions of many kinks. Formally,
\begin{equation}
I^i_+(k) = \sum_m I^i_{+,m}(k),
\end{equation}
where an integer $m$ labels each kink.
Thus Eq.~(\ref{Egeneral}) has cross terms of the contributions from
different kinks, e.g., $\vec{I}_{+,m} \cdot \vec{I}^*_{+,n}$.
However, such cross terms must vanish since a GW burst from a kink is a local
phenomenon which relates only the region around the kink.
In fact, the structure of the string around a specific point arises as a result
of nonlinear evolution of the string network, and hence the values of
$\mathbf{b}_s^{\prime \prime}$, $\mathbf{b}_s^{\prime \prime \prime}$ and so on,
are stochastic.
We show that ensemble averages of cross terms vanish as expected, i.e.
$\langle I^i_{+,m}I^{j*}_{+,n}\rangle=0$ in Appendix~\ref{sec:appA},
where it is also shown that the kinks that dominantly contribute to
the power of GWs with frequencies $\sim \omega$ are ones which satisfy
\begin{equation}
\left(\psi \frac{dN}{d\psi} \right)^{-1} \sim \omega^{-1}.
\label{kinkcon}
\end{equation}
In other words, if the interval of kinks with sharpness $\sim \psi$ is similar
to the period of the GW under consideration, these kinks make dominant
contribution to the GW.
In the first matter era, using Eq.~(\ref{bunpueasymat1}),
Eq.~(\ref{kinkcon})
simplifies to\footnote
Strictly speaking, Eq.~(\ref{kinkcon}) has two solution for $\psi$,
but it is sufficient to take larger one.
\begin{equation}
\psi \sim (\omega t)^{2\zeta_m/A_m}
\end{equation}
for $\omega < (t_*/t)^{A_m}t^{-1}$.
(Here, we set $A \equiv 2\zeta-\beta$. $A_r \simeq -0.92, \ A_m\simeq -0.8$.)
For $\omega > (t_*/t)^{A_m}t^{-1}$,
$\left(\psi \frac{dN}{d\psi} \right)^{-1} > \omega^{-1}$ is satisfied by
an arbitrary value of $\psi$.
In Appendix~\ref{sec:appA} it is shown that for $\omega > (t_*/t)^{A_m}t^{-1}$
the main contribution to $\omega \frac{dP}{d\omega}$ comes from the kinks
corresponding to the peak of the distribution, i.e. $\psi \sim (t_*/t)^{2\zeta_m}$.
If we denote the value of sharpness of kinks which make dominant contribution
to $\omega \frac{dP}{d\omega}$ as $\psi_{\rm max}(\omega,t)$, it is given by
\begin{equation}
\psi_{\rm max}(\omega,t) \sim
\begin{cases}
(\omega t)^{2\zeta_m/A_m}
& {\rm for~~}\omega < (t_*/t)^{A_m}t^{-1} \\
\left(\frac{t_*}{t}\right)^{2\zeta_m}
& {\rm for~~}\omega > (t_*/t)^{A_m}t^{-1}
\end{cases} \label{psimaxmat1}
\end{equation}
in the first matter era. In the radiation era, $\psi_{\rm max}(\omega,t)$ is
found in a similar way as
\begin{equation}
\psi_{\rm max}(\omega,t) \sim
\begin{cases}
(\omega t)^{2\zeta_r/A_r}
& {\rm for~~}\omega<\omega^{\rm (RD)}_1 (t)\\
\left( \frac{t_r}{t} \right)^{2D/A_m}(\omega t)^{2\zeta_m/A_m}
& {\rm for~~}\omega^{\rm (RD)}_1 (t) < \omega < \omega^{\rm (RD)}_2 (t) \\
\left(\frac{t_r}{t}\right)^{2\zeta_r} \left(\frac{t_*}{t_r}\right)^{2\zeta_m}
& {\rm for~~}\omega >\omega^{\rm (RD)}_2(t),
\end{cases}. \label{psimaxrad}
\end{equation}
where
\begin{eqnarray}
&\omega^{\rm (RD)}_1 (t)&= \left( \frac{t_r}{t}\right)^{A_r}t^{-1},\\
&\omega^{\rm (RD)}_2 (t)&= \left(\frac{t_r}{t} \right)^{A_r} \left(\frac{t_*}{t_r} \right)^{A_m}t^{-1}.
\end{eqnarray}
In the second matter era, $\psi_{\rm max}$ is estimated as
\begin{equation}
\psi_{\rm max}(\omega,t) \sim
\begin{cases}
(\omega t)^{2\zeta_m/A_m}
&{\rm for~~}\omega<\omega^{\rm (MD)}_1(t) \\
\left( \frac{t_{eq}}{t} \right)^{-2D/A_r}(\omega t)^{2 \zeta_r/A_r}
&{\rm for~~}\omega^{\rm (MD)}_1(t) < \omega < \omega^{\rm (MD)}_2(t) \\
\left( \frac{t_r}{t_{eq}} \right)^{2D/A_m}(\omega t)^{2\zeta_m/A_m}
&{\rm for~~}\omega^{\rm (MD)}_2(t)< \omega < \omega^{\rm (MD)}_3(t) \\
\left(\frac{t_{eq}}{t}\right)^{2\zeta_m}
\left(\frac{t_r}{t_{eq}}\right)^{2\zeta_r}\left(\frac{t_*}{t_r}\right)^{2\zeta_m}
&{\rm for~~}\omega > \omega^{\rm (MD)}_3(t),
\end{cases}. \label{psimaxmat2}
\end{equation}
where
\begin{eqnarray}
&\omega^{\rm (MD)}_1 (t)&= \left( \frac{t_{eq}}{t}\right)^{A_m}t^{-1},\\
&\omega^{\rm (MD)}_2 (t)&=
\left(\frac{t_{eq}}{t} \right)^{A_m} \left(\frac{t_r}{t_{eq}} \right)^{A_r}t^{-1}, \\
& \omega^{\rm (MD)}_3 (t)&=
\left(\frac{t_{eq}}{t} \right)^{A_m} \left(\frac{t_r}{t_{eq}} \right)^{A_r}
\left(\frac{t_*}{t_r}\right)^{A_m}t^{-1}.
\end{eqnarray}
Here, we set $D\equiv \beta_r\zeta_m-\beta_m\zeta_r \simeq 0.11$.
As a result, assuming that the powers of GW from different kinks are roughly
same as far as their sharpnesses are in the same order, (in other words,
assuming that the quantities concerned with kinks, such as
$\mathbf{b}^{\prime \prime}_s$, except their sharpness, are roughly same)
we can estimate the total power of GWs with frequencies $\sim \omega$ from
all of the kinks in a horizon as,
\begin{equation}
\omega \frac{dP}{d\omega} \bigg|_{\rm tot} \sim
\omega \frac{dP}{d\omega} \bigg|_{\mathrm{one \ kink}} (\omega,\psi_{\rm max}(\omega,t))
\times \psi \frac{dN}{d\psi} \bigg|_{\psi = \psi_{\rm max}(\omega,t)} \times t.
\label{omedPdome}
\end{equation}
The first factor denotes the power of GWs from one kink.
The second factor denotes the number of kinks which satisfy
$\psi \sim \psi_{\rm max}(\omega,t)$ per unit length.
The third factor is length of an infinite string in a horizon.
Then we finally obtain
\begin{equation}
\omega \frac{dP}{d\omega} \bigg|_{\rm tot} \sim
\begin{cases}
10G\mu^2(\omega t)^{C_m}
& {\rm for~~}t^{-1}<\omega < (t_*/t)^{A_m}t^{-1}\\
10G\mu^2\left(\frac{t_*}{t} \right)^{B_m}(\omega t)^{-2/3} &
{\rm for~~}\omega > (t_*/t)^{A_m}t^{-1}
\end{cases} \label{powtotmat1}
\end{equation}
in the first matter era,
\begin{equation}
\omega \frac{dP}{d\omega} \bigg|_{\rm tot} \sim
\begin{cases}
10G\mu^2(\omega t)^{C_r}
&{\rm for~~}t^{-1}<\omega<\omega^{\rm (RD)}_1(t) \\
10G\mu^2 \left(\frac{t_r}{t} \right)^{2D/A_m}(\omega t)^{C_m}
& {\rm for~~} \omega^{\rm (RD)}_1(t) <
\omega < \omega^{\rm (RD)}_2(t) \\
10G \mu^2 \left( \frac{t_*}{t_r}\right)^{B_m}
\left( \frac{t_r}{t}\right)^{B_r}(\omega t)^{-2/3}
&{\rm for~~}\omega > \omega^{\rm (RD)}_2(t)
\end{cases}
\label{powtotrad}
\end{equation}
in the radiation era, and
\begin{align}
\omega \frac{dP}{d\omega} \bigg|_{\rm tot} \sim
\begin{cases}
10G\mu^2(\omega t)^{C_m}
&{\rm for~~}t^{-1}<\omega<\omega^{\rm (MD)}_1(t) \\
10G\mu^2 \left(\frac{t_{eq}}{t} \right)^{-2D/A_r}(\omega t)^{C_r}
&{\rm for~~} \omega^{\rm (MD)}_1(t)< \omega < \omega^{\rm (MD)}_2(t)\\
10G\mu^2 \left(\frac{t_{eq}}{t_r} \right)^{-2D/A_m}(\omega t)^{C_m}
&{\rm for~~} \omega^{\rm (MD)}_2(t) < \omega
< \omega^{\rm (MD)}_3(t) \\
10G \mu^2 \left( \frac{t_*}{t_r}\right)^{B_m}
\left( \frac{t_r}{t_{eq}}\right)^{B_r}
\left(\frac{t_{eq}}{t}\right)^{B_m}(\omega t)^{-2/3}
&{\rm for~~} \omega >\omega^{\rm (MD)}_3(t)
\end{cases} \label{powtotmat2}
\end{align}
in the second matter era, where
$B\equiv 4\zeta-\beta~~(B_r\simeq -0.74, B_m\simeq -0.4)$ and
$C\equiv (\beta-8\zeta)/3(\beta-2\zeta)~~(C_r\simeq 0.14,C_m\simeq-0.17$).
A robust lower bound on the frequency of GWs is set to be $t^{-1}$
corresponding to the horizon scale.
\section{THE STOCHASTIC BACKGROUND OF GRAVITATIONAL WAVES FROM KINKS}
We have found Eqs.~(\ref{powtotmat1}),(\ref{powtotrad}) and
(\ref{powtotmat2}) as the total energy radiated per unit time
in a horizon from kinks on an infinite string.
Now we are in a position to calculate the density parameter of GWs defined by
\begin{equation}
\Omega_{\rm gw}(\omega) \equiv
\frac{\omega}{\rho_c} \frac{d\rho}{d\omega}(\omega),
\end{equation}
where $\rho_c$ denotes the critical energy density of the present
Universe.
Noting that the energy density of GWs decreases as $a^{-4}$
and the frequency redshifts as $a^{-1}$, we get
\begin{equation}
\Omega_{\rm gw}(\omega) \sim
\frac{1}{\rho_c} \int^{t_0}_{t_*} dt \frac{1}{t^3} \left(\omega^{\prime}
\frac{dP}{d\omega^{\prime}}\right) \bigg|_{\omega^{\prime}
= \omega \times a_0/a(t)}
\left(\frac{a(t)}{a_0}\right)^4, \label{forOmega}
\end{equation}
where $a(t)$ is the scale factor and $a_0$ represents its present value.
Using Eqs.~(\ref{powtotmat1}),(\ref{powtotrad}) and (\ref{powtotmat2}),
we get the spectrum as
\begin{align}
&\Omega_{\rm gw} (\omega)\sim \nonumber \\
&
\begin{cases}
60\pi(G\mu)^2(\omega t_0)^{C_m}
\qquad \quad \ \quad \qquad \qquad
&{\rm for~} t_0^{-1}<\omega< \omega_1 \\
60\pi(G\mu)^2\left(\frac{\Omega_r}{\Omega_m}\right)^{-3D/A_r}
(\omega t_0)^{C_r}
\qquad
&{\rm for~} \omega_1 < \omega < \omega_2 \\
60\pi(G\mu)^2\left(\frac{\Omega_r}{\Omega_m}\right)^{-4D/A_m}
\left(\frac{T_0}{M_{\rm pl}}\right)^{4D/A_m}
\left(\frac{T_r}{M_{\rm pl}}\right)^{-4D/A_m}
(\omega t_0)^{C_m}
&{\rm for~} \omega_2 < \omega < \omega_3 \\
60\pi(G\mu)^2(\omega t_0)^{-2/3}
\left( \frac{\Omega_r}{\Omega_m}\right)^{3B_m/2-4B_r}
\left(\frac{T_0}{M_{\rm pl}}\right)^{2B_r}
\left(\frac{T_r}{M_{\rm pl}}\right)^{2B_m-2B_r}
\left(\frac{H_*}{M_{\rm pl}}\right)^{-B_m}
&{\rm for~}\omega > \omega_3
\end{cases}, \label{Omegagw}
\end{align}
where
\begin{eqnarray}
&\omega_1 &= \omega^{\rm (MD)}_1(t_0) =
\left( \frac{\Omega_r}{\Omega_m}\right)^{3A_m/2}t^{-1}_0,
\label{omegaMD1} \\
&\omega_2 &= \omega^{\rm (MD)}_2(t_0) =
\left( \frac{\Omega_r}{\Omega_m}\right)^{3A_m/2-2A_r}
\left(\frac{T_0}{M_{\rm pl}}\right)^{2A_r}
\left(\frac{T_r}{M_{\rm pl}}\right)^{-2A_r}t^{-1}_0, \\
& \omega_3 &= \omega^{\rm (MD)}_3(t_0) =
\left( \frac{\Omega_r}{\Omega_m}\right)^{3A_m/2-2A_r}
\left(\frac{T_0}{M_{\rm pl}}\right)^{2A_r}
\left(\frac{T_r}{M_{\rm pl}}\right)^{2A_m-2A_r}
\left(\frac{H_*}{M_{\rm pl}}\right)^{-A_m}t^{-1}_0
\label{omegaMD3}
\end{eqnarray}
where $t_0$ and $T_0$ denote the present age and temperature of the Universe,
$T_r$ is the reheating temperature
and $H_*$ is the Hubble parameter at the end of inflation.\footnote{
Here we neglected the effect of cosmological constant.
The GW spectrum will be slightly modified if the
cosmological constant is taken into account.
Detailed estimation of this effect is beyond the scope of this paper
since it needs a simulation of cosmic string network evolution in the
cosmological constant dominated Universe.
}
This formula has complicated exponents.
Using $\beta$ and $\zeta$ with values given above,
$\Omega_m/\Omega_r \simeq 5.5\times10^3$ and
$T_0/M_{\rm pl}\simeq 9.6\times 10^{-32}$,
Eq.~(\ref{Omegagw}) simplifies to
\begin{align}
\Omega_{\rm gw}(\omega) \sim
&\begin{cases}
10^2(G\mu)^2(\omega t_0)^{-0.17}
& {\rm for~~}t_0^{-1} < \omega < \omega_1\\
10(G\mu)^2(\omega t_0)^{0.14}
& {\rm for~~}\omega_1 < \omega
< \omega_2 \\
10^{18}(G\mu)^2\left(\frac{T_r}{M_{\rm pl}}\right)^{0.56}(\omega t_0)^{-0.17}
& {\rm for~~}\omega_2 <\omega
< \omega_3 \\
10^{40}(G\mu)^2\left(\frac{T_r}{M_{\rm pl}}\right)^{0.68}
\left(\frac{H_*}{M_{\rm pl}}\right)^{0.4} (\omega t_0)^{-2/3}
& {\rm for~~} \omega > \omega_3 ,
\end{cases}.
\end{align}
where
\begin{eqnarray}
&\omega_1 & \sim 10^4t_0^{-1},\\
&\omega_2 & \sim
10^{-3}\left(\frac{T_r}{T_0}\right)^{1.8}t_0^{-1}, \\
& \omega_3 & \sim
10^{54}\left(\frac{T_r}{M_{\rm pl}}\right)^{0.24}
\left(\frac{H_*}{M_{\rm pl}}\right)^{0.8}t_0^{-1}
\end{eqnarray}
\begin{figure}[t]
\begin{center}
\includegraphics[width=18cm]{spectrum.eps}
\caption{A schematic picture of
gravitational wave spectrum from kinks (\ref{Omegagw}).
$\omega_1$, $\omega_2$, $\omega_3$
are given by Eqs.~(\ref{omegaMD1})-(\ref{omegaMD3}).
}
\label{fig:spectrum}
\end{center}
\end{figure}
We find that the integral in Eq.~(\ref{forOmega}) is dominated by the
contribution from the period near the present for the whole range
of $\omega$.
In other words, almost all of the present energy of GWs from
the kinks on infinite strings comes from those radiated
around the present epoch.
This does not mean that the kinks produced at present make
dominant contribution to the energy of GWs.
We see above that for given frequency $\omega$,
the kinks which dominantly contribute to $\omega \frac{dP}{d\omega}$ are
determined by Eqs.~(\ref{psimaxmat1}),(\ref{psimaxrad}) or
(\ref{psimaxmat2}).
Accordingly high frequency modes arise from dense, blunt and old kinks,
and low frequency modes arise from thin, sharp and new ones.
The first line of Eq.~(\ref{Omegagw}) corresponds to GWs from new kinks which
were born after the matter-radiation equality, the second line corresponds
to GWs from old kinks which were born in the radiation era,
the third line corresponds to GWs from older kinks produced between the end
of the inflation and the start of the radiation era and
the last line corresponds to GWs which came from the most abundant
and oldest kinks.
Figure~\ref{fig:spectrum} sketches the shape of the spectrum.
The spectrum has three inflection points at
$\omega_1, \omega_2$ and $\omega_3$.
This is due to change of the type of kinks which mainly contribute to
$\Omega_{\rm gw}$.
Positions of these inflection points depend on the reheating
temperature $T_r$.
If $T_r$ takes the value around its lower bound, say,
$10$~MeV~\cite{Kawasaki:1999na}, the second inflection point falls in the
observable region, as we will see.
Even if $T_r$ is so low, the natural value of $H_*$ makes the third inflection
far above the observable region.
If the reheating ends immediately and $H_*$ is as small as possible,
the region between the second and third inflection is so short that
the third inflection enters the observable region.
This is a crude estimation, and in order to derive the realistic spectrum
we should take into account subtlety described in~\cite{Damour:2001bk,Damour:2004kw}
where the authors claimed that GWs from kinks are burst-like and hence
GW bursts with rare event rate (``isolated'' GWs ) should not be counted as
constituent of the stochastic GW background.
We should calculate the stochastic GW background spectrum using the following
formulae~\cite{Damour:2001bk,Damour:2004kw}
\begin{equation}
\Omega_{\rm gw}(f) \sim \frac{3\pi^2}{2}(f t_0)^2h_{\rm conf}^2(f)
\label{OmegagwDV}
\end{equation}
\begin{equation}
h_{\rm conf}^2(f)=\int \frac{dz}{z}\theta(n(f,z)-1) n(f,z)h^2(f,z)
\label{hconf}
\end{equation}
\begin{equation}
n(f,z)=\frac{1}{f}\frac{d\dot{N}}{d\ln z} \label{n}
\end{equation}
\begin{equation}
d\dot{N}\sim \frac{1}{4}\theta_m(f,z)(1+z)^{-1}
\psi_{\rm max}(\omega_z,t)\tilde{N}(\psi_{\rm max}(\omega_z,t),z)
t^{-1}(z)dV(z),
\label{dN}
\end{equation}
\begin{equation}
h(f,z)=\frac{G\mu [\psi_{\rm max}(\omega_z,t)]^{1/2}t(z)}{[(1+z)ft(z)]^{2/3}}
\frac{1+z}{t_0z}\theta(1-\theta_m(f,z)), \label{h}
\end{equation}
\begin{equation}
\theta_m(f,z) = [(1+z)f t(z)]^{-1/3}, \label{thetam}
\end{equation}
\begin{equation}
dV=
\begin{cases}
54\pi t_0^3 \sqrt{\frac{\Omega_m}{\Omega_r}\frac{T_0}{T_r}}(1+z)^{-9/2}dz
& (\mathrm{1st \ matter \ era})\\
72\pi t_0^3\sqrt{\frac{\Omega_m}{\Omega_r}}(1+z)^{-5}dz
& (\mathrm{radiation \ era})\\
54\pi t_0^3((1+z)^{1/2}-1)^2(1+z)^{-11/2}dz
& (\mathrm{2nd \ matter \ era})
\end{cases}
\label{dV}
\end{equation}
\begin{equation}
t(z)=
\begin{cases}
\sqrt{\frac{\Omega_m}{\Omega_r}\frac{T_0}{T_r}}(1+z)^{-3/2}t_0
& (\mathrm{1st \ matter \ era})\\
\sqrt{\frac{\Omega_m}{\Omega_r}}(1+z)^{-2}t_0 & (\mathrm{radiation \ era}) \\
(1+z)^{-3/2}t_0 & (\mathrm{2nd \ matter \ era})
\end{cases}
\label{tz}
\end{equation}
where $f=\omega/2\pi$ and $\omega_z = \omega (1+z)$.
$dV$ means the proper spatial volume between redshifts $z$ and $z+dz$.
$t(z)$ represents the cosmic time at the redshift $z$.
$n(\omega,z)$ represents the number of GW bursts at redshift $\sim z$
with frequencies $\omega$ superposed in a period
of $\sim \omega^{-1}$.
$\tilde{N}(\psi,z)dVd\psi$ is the number of kinks with sharpness
$\psi \sim \psi +d\psi$
in the volume $dV$ at redshift $z$, so
$\tilde{N}(\psi,z)t^2(z) \sim \frac{dN}{d\psi}(t(z),\psi)$.
Isolated GW bursts are excluded from the calculation by inserting the step function
in the integral in Eq.~(\ref{hconf}).
$h(f,z)$ is the logarithmic Fourier component of the waveform of the GW burst
from one kink located at redshift $z$ and can contribute
to GW with frequency $f$.
The results of calculation are shown in Figure~\ref{Omegazuhigh} and
Figure~\ref{Omegazulow}.
We take $G\mu\sim10^{-7}$, close to the current upper bound from CMB
observation~\cite{Wyman:2005tu}, and show the spectrum with frequency
$\omega$ from the band of CMB experiments to that of ground-based GW detectors.
These figures include both our crude estimate (\ref{Omegagw}) and the
improved one given by Eqs.~(\ref{OmegagwDV})-(\ref{tz}).
We see that the latter is much smaller than the former.
This is because the spectrum is dominated by GWs emitted recently,
and recent GW bursts have a more tendency to be isolated.
The fact that the difference between the two estimates becomes larger
in higher frequency band might disagree with intuition,
since the kinks corresponding to high frequency GWs are more abundant.
However, the higher frequencies of GW are, the smaller the possibility
that they overlap, because the period of oscillation becomes shorter and the
extent of the GW beam becomes narrower.
As a result, higher frequency GWs are more likely isolated in time.
\begin{figure}[t]
\begin{center}
\includegraphics[width=15cm]{omegagwhighTini.eps}
\caption{$\Omega_{\rm gw}$ in the case where strings emerge
at the phase transition in the radiation era
for $G\mu=10^{-7}$ and $T_*\sim 10^{12}$~GeV.
The upper line represents the estimate using Eq.~(\ref{Omegagw}), i.e.
including ``rare bursts'',
and the lower line represents the estimate using Eqs.~(\ref{OmegagwDV})-
(\ref{thetam}),
i.e. excluding ``rare bursts''. Sensitivity curves of various experiments are shown.
That of DECIGO is derived from~\cite{Seto:2001qf}.
That of BBO correlated is derived from~\cite{Buonanno:2004tp}.
Others are derived from~\cite{Smith:2005mm}.
\label{Omegazuhigh}}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=15cm]{omegagwlowTini1.eps}
\caption{$\Omega_{\rm gw}$ in the case where strings emerge at the end of inflation
for $G\mu=10^{-7}$, $T_r\sim 10$~MeV.
The upper line represents the estimate including ``rare bursts'',
and the lower line represents the estimate excluding ``rare bursts''.
\label{Omegazulow}}
\end{center}
\end{figure}
In Figure~\ref{Omegazuhigh} we assume that cosmic strings were born by SSB
at GUT scale and first kinks appeared at the temperature $10^{12}~$GeV (\ref{Tefd}).
In Figure~\ref{Omegazulow} we assume that strings were produced at the
end of the inflation with extremely low reheating temperature ($T_r=10$~MeV).
This assumption makes the second inflection point of the spectrum visible
in the observable frequency band.
We also assume the inflation energy scale is sufficiently high so that third
inflection point on the spectrum is far from the observable region.
If we could observe the second inflection, we can deduce the reheating temperature.
Note that the spectrum depends on $G\mu$ via the overall factor $(G\mu)^2$.
Therefore, when we vary the value on $G\mu$, the spectrum only moves upward or
downward, and its shape (e.g., the position of bending) does not change.
This situation is different from the case of GWs from cosmic string loops.
In the case of cosmic string loops,
since different values of $G\mu$ give different values of lifetime of loops,
the resultant spectral shape is different~\cite{Caldwell:1991jj,DePies:2007bm}.
Let us discuss detectability of this GW background.
We have to see whether GWs from kinks exceeds not only thresholds of various
experiments but also GWs from other source.
The spectrum of GWs from loops was discussed in~\cite{Damour:2000wa,Damour:2001bk,Damour:2004kw,Siemens:2006yp,Caldwell:1991jj,DePies:2007bm},
and the contribution of loops to GW background is much larger
than that of kinks if they coexist in some frequency band.
However, a loop cannot emit GWs with frequencies smaller than the inverse of
its size.
Thus there is a cut on the low frequency side of the spectrum
of GWs from loops
corresponding to the inverse of the loop size $\sim (\alpha t)^{-1}$.
If $\alpha \sim G\mu$, the spectrum of GWs from loops begins to appear
at $\omega \gtrsim 10^{-12}$~Hz,
and this covers the frequency band where both pulsar timing arrays
and GW detectors have good sensitivity.
However, $\alpha$ is one of the most unknown parameters in
the cosmic string model.
According to some recent simulations~\cite{Ringeval:2005kr},
$\alpha$ may be much greater and the broader region may be covered by loops' GW.
On the other hand, some recent studies~\cite{Siemens:2002dj,Polchinski:2006ee}
show the possibility that $\alpha$ is extremely small,
say, $\alpha\sim (G\mu)^n$ with $n\gtrsim 1$.
In such a case, GWs from loops dominate only very high-frequency region and
GWs from kinks may be observable at low-frequency region.
For example, if $\alpha \lesssim 10^{-9}$,
the band of SKA~\cite{Kramer:2004rwa} can be used for detection of GWs from kinks,
and for $G\mu \sim 10^{-7}$, $\Omega_{\rm gw}$ reaches the sensitivity of SKA.
In a more extreme case $\alpha \lesssim 10^{-13}$,
the sensitivity band of space-borne detectors,
BBO~\cite{Crowder:2005nr} and DECIGO~\cite{Seto:2001qf} are open for detection of
kink-induced GWs,
and $\Omega_{\rm gw}$ exceeds the sensitivity of correlated analysis of BBO
for $G\mu \sim 10^{-7}$,
and that of ultimate-DECIGO for $G\mu \gtrsim 10^{-9}$.\footnote
Note that stochastic GWs from astrophysical sources such as white-dwarf binaries
make a dominant contribution for $\omega \lesssim 1~$Hz~\cite{Farmer:2003pa}
and the observable frequency range is somewhat limited.
In even more extreme case, $\alpha \lesssim 10^{-16}$,
it may be possible to detect the inflection point in Figure 4
for $G\mu \gtrsim 10^{-10}$ and determine the reheating temperature as
$T_r \sim 10$~MeV.
Moreover, GWs from kinks may be detected through CMB observations.
As opposed to GWs from loops, kinks can emit GWs with wavelength
comparable to the horizon scale.
These GWs induce B-mode polarizations, which is a target of on-going and
future CMB surveys.
The spectrum of GWs from kinks is quite different from inflationary GWs
and hence its effect on CMB is also expected to be distinguished from that
of inflationary origin. We will study this issue elsewhere.
\section{Conclusions} \label{sec:conc}
In this paper, we have considered gravitational waves emitted by kinks on
infinite cosmic strings.
We have calculated the spectrum of the stochastic background of such
gravitational waves and discussed their detectability by pulsar timing
experiments and space-borne detectors.
It is found that if the size of cosmic string loops is much smaller than
that of Hubble horizon, some frequency bands are open for detection of GWs
originating from kinks.
It can be detected by pulsar timing experiments for $G\mu \gtrsim 10^{-7}$,
and by space-borne gravitational wave detectors for much smaller $G\mu$,
although the latter may be hidden by the loop contribution unless
the typical loop size is extremely small.
If it is detected, it will provide information on the physics of the early
Universe, such as phase transition and inflation models.
Moreover, the spectrum shape depends on the thermal history of the Universe,
and hence GWs from cosmic strings can be used as a direct probe into the early
evolution of the Universe.
Notice that the inflationary GWs also carry information on the thermal history of the Universe
~\cite{Seto:2003kc,Boyle:2005se,Nakayama:2008ip}.
Although the inflationary GWs are completely hindered by GWs from cosmic strings
if the value of $G\mu$ is sizable and GWs from kinks come to dominate in
low-frequency region,
GWs from cosmic strings also have rich information on the physics of the early Universe.
|
\section*{Introduction}\label{sec:1}
The study of the dynamics of discrete groups of automorphisms of
$\mathbb P_{\mathbb C}^2$ is in its childhood, and one of the first basic problems is
understanding how the various possible notions of ``limit set"
relate among themselves. This is the main topic of this article.
If $\Gamma \subset PSL(2,\mathbb{C})$ is a classical Kleinian group, then
its limit set $\Lambda$ is defined as the set of accumulation points
of the $\Gamma$ orbits in $\mathbb P_{\mathbb C}^1 \cong S^2$. Well known arguments
show that $\Lambda$ is independent of the choice of orbit, whenever
$\Gamma$ is non-elementary, and its complement $\Omega$ in $\mathbb P_{\mathbb C}^1$ is
the largest discontinuity domain. Moreover, $\Omega$ coincides also
with the region of equicontinuity, admits a complete metric and
volume which are invariant under the action of $\Gamma$, facts which
are essential for defining the Patterson-Sullivan measure for
Kleinian groups (see for instance \cite{Pat, Su}). When we go look
at discrete subgroups of $PSL(3, \mathbb C)$ acting on $\mathbb P_{\mathbb C}^2$, there
are examples where the action on the complement of the set of
accumulation points of the orbits is neither discontinuous nor
equicontinuous. There are also examples (see \cite{Ku}) where there
is a largest discontinuity set but it does not agree with the
equicontinuity set. Hence it seems that there are different notions
of ``limit set" for discrete subgroups of $PSL(3,\mathbb{C})$ acting on
$\Bbb{P}^2_\Bbb{C}$.
For discrete actions in general topological spaces there is a notion
of limit set due to Kulkarni, see \cite{Ku}, which has the important
property of granting that the action on its complement is
discontinuous and in the case of conformal automorphisms of the
$n$-sphere $\Bbb{S}^n$ agrees with the usual notion of limit set. In
\cite{SV1, SV2} there are some interesting families of projective
groups and their limit set in Kulkarni's sense. Later in the
articles \cite {cano, Na1, Na2}, the authors provide sufficient
conditions, for particular subgroups $\Gamma$ of $PSL(3, \mathbb{C})$, to
grant that the equicontinuity set and Kulkarni's discontinuity
region, $\Omega(\Gamma)$, agree. It was also shown that, in such
particular case, $\Omega(\Gamma)$ is the largest discontinuity set.
One of the goals of this paper is the following:
\begin{theorem} \label{main1}
Let $\Gamma\subset PSL(3,\mathbb{C})$ be a discrete group, then the
equicontinuity set of $\Gamma$ is a discontinuity region. Moreover,
if $ U$ is an open $\Gamma$ invariant subset with at least three
lines in general position lying on its complement, then $U$ is
contained in the equicontinuity set of $\Gamma$. Furthermore, if
Kulkarni's limit set $\Lambda(\Gamma)$ contains at least three lines
in general position, then Kulkarni's discontinuity region
$\Omega(\Gamma)= \mathbb P_{\mathbb C}^2 \setminus \Lambda(\Gamma)$ is equal to the
equicontinuity set of the group $\Gamma$.
\end{theorem}
In section \ref{psprojective},we split the Theorem \ref{main1} in
three parts: Theorem \ref{gamactsdisconeq}, Theorem \ref{dissubeq}
and Theorem \ref{eqeqdis}. The proofs of these theorems rely on
{\it pseudo-projective} maps, which provide a compactification of
the non-compact Lie group $PSL(3,\mathbb{C})$.
Kulkarni's limit set $\Lambda(\Gamma)$, for a subgroup $\Gamma$ of
$PSL(3, \mathbb{C})$ acting on $\mathbb P_{\mathbb C}^2$, may seem a complicated object at
first sight. However, when $\Gamma =\langle \alpha \rangle$ is a
cyclic group, generated by the element $\alpha \in PSL(3, \mathbb{C})$, then
the limit set of this cyclic group, denoted $\Lambda(\alpha)$, can
be found (see \cite{Na2}) and it is one of the following (according
to the Jordan canonical form of $\alpha$):
\begin{itemize}
\item The empty set,
\item all of $\mathbb P_{\mathbb C}^2$,
\item one single complex line,
\item one complex line and one point,
\item two distinct complex lines.
\end{itemize}
In section \ref{lambdaunlin}, we define the set
$C(\Gamma)=\overline{\cup _{\gamma \in \Gamma} \Lambda(\gamma)}$ and
we prove the following:
\begin{theorem}\label{main2}
Let $\Gamma\subset PSL(3,\Bbb{C})$ be a discrete group, if the
number of complex lines in general position in $\Lambda(\Gamma)$ and
$C(\Gamma)$ is at least three, then
\[
\Lambda(\Gamma)=C(\Gamma)=\overline{\bigcup_{\gamma\in \Gamma}
\Lambda(\gamma)},
\]
and $\Lambda(\Gamma)$ is the union of complex lines. Moreover the
discontinuity region according to Kulkarni, $\Omega(\Gamma)$, agrees
with the equicontinuity set. Furthermore, when $\Gamma$ acts on
$\mathbb P_{\mathbb C}^2$ without global fixed points, then $\Omega(\Gamma)$ is the
largest open set where $\Gamma$ acts properly and discontinuously.
\end{theorem}
If $\Gamma \subset PU(2,1)$ is a discrete subgroup, then $\Gamma$
acts on the complex hyperbolic plane $\mathbb H ^2_\mathbb{C}$ by isometries. In
this case, the limit set according to Chen-Greenberg of $\Gamma$ is
obtained as the set of accumulation points, in $\partial \mathbb H ^2 _\mathbb{C}$,
of the $\Gamma$-orbit of any point in $\mathbb H ^2 _\mathbb{C}$. Also, it can be
obtained as the closure of the set of fixed points of non-elliptic
elements in $\Gamma$. When we consider the action of $\Gamma$ in
$\mathbb P_{\mathbb C}^2$, the Kulkarni's limit set $\Lambda(\Gamma)$ is given in the
following way (see \cite{Na1}): Let $F_{\Gamma}$ be the set of
points $p \in
\partial \mathbb H^2 _\mathbb{C}$ such that there exists a non-elliptic element
$\gamma \in \Gamma$ that satisfies $\gamma p=p$. In what follows,
$(\mathbb P_{\mathbb C}^2)^*$, denotes the space of complex lines in $\mathbb P_{\mathbb C}^2$. If
$\mathcal{E}(\Gamma)$ denotes the subset of $(\mathbb P_{\mathbb C}^2)^*$ consisting of
those complex lines tangent to $\partial \mathbb H ^2 _\mathbb{C}$ at points of
$F_{\Gamma}$, then
$$\Lambda(\Gamma) = \bigcup _{l \in \overline{\mathcal{E}(\Gamma)}} l=
\overline{\bigcup_{l \in \mathcal{E}(\Gamma)} l}.$$
Moreover, when $l\in \overline{\mathcal{E}(\Gamma)}$, the closure of
the $\Gamma$-orbit of the complex line $l$ in $(\mathbb P_{\mathbb C}^2)^*$ is equal
to $\overline{\mathcal{E}(\Gamma)}$. In other words, the action of
$\Gamma$ in the set of lines $\overline{\mathcal{E}(\Gamma)}$ is
minimal.
Let $\Gamma \subset PU(2,1)$ be a discrete subgroup such that
$\Lambda(\Gamma)=\mathbb P_{\mathbb C}^2 \setminus \mathbb H ^2 _\mathbb{C}$, then there exist
complex lines contained in $\Lambda(\Gamma)$ not belonging to
$\overline{\mathcal{E}(\Gamma)}$, so that $\mathcal{E}(\Gamma)$
gives us a canonical choice of lines to describe the limit set
$\Lambda(\Gamma)$ as a union of lines. In section \ref{lambdaunlin},
we prove the following generalization of the main theorems in
\cite{Na1, BN}:
\begin{theorem} \label{main3}
Let $\Gamma \subset PSL(3, \mathbb{C})$ be an infinite discrete subgroup,
without fixed points nor invariant lines. Let $\mathcal{E}(\Gamma)$
be the subset of $(\mathbb P_{\mathbb C}^2 )^*$ consisting of all the complex lines $l$
for which there exists an element $\gamma \in \Gamma$ such that $ l
\subset \Lambda(\gamma)$.
\begin{enumerate}
\item[a)] $Eq(\Gamma)= \Omega (\Gamma)$, is the maximal open set on which
$\Gamma$ acts properly and discontinuously. Moreover, if
$\mathcal{E}(\Gamma)$ contains more than three complex lines then
every connected component of $\Omega(\Gamma)$ is complete Kobayashi
hyperbolic (compare with \cite{BN}).
\item[b)] The set
$$\Lambda(\Gamma)= \overline{\bigcup _{ l \in \mathcal{E}(\Gamma)} l}
=\bigcup _{ l \in \overline{\mathcal{E}(\Gamma)}} l =
\overline{\bigcup _{\gamma \in \Gamma} \Lambda(\gamma)}$$ is
path-connected.
\item[c)]If $\mathcal{E}(\Gamma)$ contains more than three lines,
then $\overline{\mathcal{E}(\Gamma)} \subset (\mathbb P_{\mathbb C} ^2)^*$ is a perfect
set. Also, it is the minimal closed $\Gamma$-invariant subset of
$(\mathbb P_{\mathbb C}^2 )^*$.
\end{enumerate}
\end{theorem}
\section{Preliminaries and Notation.}
\subsection{Projective Geometry}
We recall that the complex projective plane
$\mathbb{P}^2_{\mathbb{C}}$ is defined as the quotient space
$$(\mathbb{C}^{3}\setminus \{(0,0,0)\})/\mathbb{C}^*,$$
where $\mathbb{C}^*= \mathbb{C} \setminus \{0\}$ acts on $\mathbb{C}^3\setminus \{(0,0,0)\}$ by the usual scalar
multiplication. This is a complex $2$-dimensional Riemannian
manifold which is compact, connected and is naturally equipped with
the Fubini-Study metric. Let $[\mbox{ }]:\mathbb{C}^{3}\setminus
\{(0,0,0)\}\rightarrow \mathbb{P}^{2}_{\mathbb{C}}$ be the quotient
map. If $\beta=\{e_1,e_2,e_3\}$ is the standard basis of
$\mathbb{C}^3$, we write $[e_j]=e_j$ and if $w=(w_1,w_2,w_3)\in
\mathbb{C}^3\setminus \{(0,0,0)\}$ then we write
$[w]=[w_1:w_2:w_3]$. Also, the set $l\subset
\mathbb{P}^2_{\mathbb{C}}$ is said to be a complex line if
$[l]^{-1}\cup \{(0,0,0)\}$ is a complex linear subspace of dimension
$2$. Given two distinct points $p,q\in \mathbb{P}^2_{\mathbb{C}}$,
there is a unique complex line passing through $p$ and $q$, such
line is denoted by $\Flecha{p,q}$.
Any 2-dimensional complex vector subspace of $\mathbb{C} ^3$ can be
expressed as the set of points $(z_1, z_2, z_3)$ satisfying an
equation of the form $A z_1+ B z_2 + C z_3 =0$. Conversely, any
element $(A, B, C) \in \mathbb{C}^3 \setminus \{(0,0,0)\}$ defines, up to a
non-zero scalar multiple, the 2-dimensional complex vector subspace
of $\mathbb{C} ^3$ consisting of the points $(z_1, z_2, z_3)$ such that $A
z_1+ B z_2 + C z_3 =0$. In this way, the space of complex lines in
$\mathbb P_{\mathbb C}^2$, denoted by $(\mathbb P_{\mathbb C}^2)^*$, can be identified with $\mathbb P_{\mathbb C} ^2$.
Consider the action of $\mathbb{Z}_{3}$ (the cubic roots of the
unity) on $SL(3,\mathbb{C})$ given by the usual scalar
multiplication, then
$PSL(3,\mathbb{C})=SL(3,\mathbb{C})/\mathbb{Z}_{3}$ is a Lie group
whose elements are called projective transformations. Let $[\mbox{
}]:SL(3,\mathbb{C})\rightarrow PSL(3,\mathbb{C})$ be the quotient
map, we say that $\tilde\gamma \in GL(3,\mathbb{C})$ is a lift of
$\gamma \in PSL(3,\mathbb{C})$ if there is a cubic root $\tau$ of
$(Det(\tilde{\gamma}))^{-1}$ such that $[\tau
\widetilde\gamma]=\gamma$, also, we will use the notation
$(\gamma_{ij})$ to denote elements in $SL(3,\Bbb{C})$. One can show
that $ PSL(3, \mathbb{C})$ is a Lie group that acts transitively,
effectively and by biholomorphisms on $\mathbb{P}^2_{\mathbb{C}}$ by
$[\gamma]([w])=[\gamma(w)]$, where $w\in \mathbb{C}^3\setminus
\{(0,0,0)\}$ and $\gamma\in SL(3,\mathbb{C})$. Also it is well
known that projective transformations take complex lines into
complex lines. Moreover, if the complex line $l \in (\mathbb P_{\mathbb C}^2)^*$ is
represented by $[A: B: C]$ and $\gamma =[\tilde{\gamma}] \in PSL(3,
\mathbb{C})$, then $\gamma (l) \in (\mathbb P_{\mathbb C}^2)$ is represented by $[(A, B, C)
\tilde{\gamma}^{-1}]$.
\subsection{Kulkarni's Limit Set for Subgroups of $PSL(3,\mathbb{C})$}
The Kulkarni's limit set is defined for actions of groups on very
general topological spaces (see \cite{Ku}), but we restrict our
attention to subgroups of $PSL(3,\mathbb{C})$ acting on $\mathbb P_{\mathbb C}^2$.
Let $\Gamma\subset PSL(3,\mathbb{C})$ be a subgroup:
\begin{enumerate}
\item $L_0(\Gamma)$ is defined as the closure of the points in
$\mathbb{P}^2_{\mathbb{C}}$ with infinite isotropy group.
\item $L_1(\Gamma)$ is the closure of the set of cluster points of
the $\Gamma$-orbit of the point $z$, where $z$ runs over
$\mathbb{P}^2_{\mathbb{C}}\setminus L_0(\Gamma)$.
Recall that $q$ is a cluster point for the family of sets
$\{\gamma (K)\}_{ \gamma \in \Gamma}$, where $K \ne \varnothing$ is
a subset of $\mathbb{P}^2_{\mathbb{C}}$, if there is a sequence
$(k_m)_{m\in\mathbb{N}}\subset K$ and a sequence of distinct
elements $(\gamma_m)_{m\in\mathbb{N}}\subset \Gamma$ such that
$\gamma_m(k_m)\xymatrix{ \ar[r]_{m \rightarrow \infty}&} q$.
\item $L_2(\Gamma)$ is the closure of cluster points of the
family of compact sets $\{\gamma (K)\}_{ \gamma \in \Gamma}$, where
$K$ runs over all the compact subsets of
$\mathbb{P}^2_{\mathbb{C}}\setminus (L_0(\Gamma) \cup L_1(\Gamma))$.
\item The \textit{Limit Set in the sense of Kulkarni} for $\Gamma$ is
defined as: $$\Lambda (\Gamma) = L_0(\Gamma) \cup L_1(\Gamma) \cup
L_2(\Gamma).$$
\item The \textit{Discontinuity Region in the
sense of Kulkarni} of $\Gamma$ is defined as:
$$\Omega(\Gamma) = \mathbb{P}^2_{\mathbb{C}}\setminus
\Lambda(\Gamma).$$
\end{enumerate}
We say that $\Gamma$ is a \textit{Complex Kleinian Group} if
$\Omega(\Gamma)\neq \emptyset$, see \cite{SV1}. The following
proposition is obtained from \cite{Ku}.
\begin{proposition}
Let $\Gamma\subset PSL(3,\Bbb{C})$ be a complex Kleinian group.
Then:
\begin{enumerate}
\item \label{i:pk1} $\Gamma$ is discrete and countable.
\item\label{i:pk2}
$\Lambda(\Gamma),\,L_0(\Gamma),\,L_1(\Gamma),\,L_2(\Gamma)$ are
closed $\Gamma$-invariant sets.
\item \label{i:pk3} $\Gamma$ acts
discontinuously on $\Omega(\Gamma)$.
\end{enumerate}
\end{proposition}
As the reader can notice, the computation of the limit set
$\Lambda(\Gamma)$ can be very complicated. The following lemma (see
\cite{Na2}) helps us in the task of the computation of
$L_2(\Gamma)$.
\begin{lemma} \label{cuasiminimalidad}
Let $\Gamma$ be a subgroup of $PSL(3, \mathbb C)$. If $C \subset
\mathbb P_{\mathbb C}^2 $ is a closed set such that for every compact subset $K \subset
\mathbb P_{\mathbb C}^2 \setminus C$, the cluster points of the family of compact
sets $\{ \gamma(K) \} _{\gamma \in \Gamma}$ are contained in
$L_0(\Gamma) \cup L_1(\Gamma)$, then $L_2(\Gamma) \subseteq C$.
\end{lemma}
In what follows, the limit set of the cyclic group $\langle \gamma
\rangle$ will be denoted $\Lambda(\gamma)$.
The non-trivial elements of $PSL(3,\mathbb{C})$ can be classified as
elliptic, parabolic or loxodromic (see \cite{Na2}):
The \emph{elliptic} elements in $PSL(3,\mathbb{C})$ are those elements
$\gamma$ that have a lift to $SL(3, \mathbb{C})$ whose Jordan canonical form
is
\begin{displaymath}
\left(
\begin{array}{ccc}
e^{i\theta _1} & 0 & 0 \\
0 & e^{i \theta _2} & 0 \\
0 & 0 & e^{i \theta _3}
\end{array}
\right).
\end{displaymath}
The limit set $\Lambda(\gamma)$ for $\gamma$ elliptic is
$\varnothing$ or all of $\mathbb P_{\mathbb C}^2$ according to whether the order of
$\gamma$ is finite or infinite. It is the appropriated moment to
remark that those subgroups of $PSL(3, \mathbb{C})$ containing an elliptic
element of infinite order cannot be discrete.
The \emph{parabolic} elements in $PSL(3, \mathbb{C})$ are those elements
$\gamma$ such that the limit set $\Lambda(\gamma)$ is equal to one
single complex line. If $\gamma$ is parabolic then it has a lift to
$SL(3,\mathbb{C})$ whose Jordan canonical form is one of the following
matrices:
\begin{displaymath}
\left(
\begin{array}{ccc}
1 & 1 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array}
\right) , \left(
\begin{array}{ccc}
1 & 1 & 0 \\
0 & 1 & 1 \\
0 & 0 & 1
\end{array}
\right) , \left(
\begin{array}{ccc}
e^{2\pi i t} & 1 & 0 \\
0 & e^{2\pi i t} & 0 \\
0 & 0 & e^{-4\pi i t}
\end{array}
\right) , e^{2 \pi i t} \ne 1
\end{displaymath}
In the first case $\Lambda(\gamma)$ is the complex line consisting
of all the fixed points of $\gamma$, in the second case
$\Lambda(\gamma)$ is the unique $\gamma$-invariant complex line. In
the last case $\Lambda(\gamma)$ is the complex line determined by
the two fixed points of $\gamma$.
There are four kinds of \emph{loxodromic} elements in $PSL(3,\mathbb{C})$:
\begin{itemize}
\item The \emph{complex homotheties} are those elements $\gamma \in PSL(3, \mathbb{C})$ that
have a lift to $SL(3,\mathbb{C})$ whose Jordan canonical form is
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda & 0 & 0 \\
0 & \lambda &0 \\
0 & 0 & \lambda ^{-2}
\end{array}
\right) , \quad |\lambda| \ne 1,
\end{displaymath}
and its limit set $\Lambda(\gamma)$ is the set of fixed points of
$\gamma$, consisting of a complex line and a point.
\item The \emph{screws} are those elements $\gamma \in PSL(3, \mathbb{C})$
that have a lift to $SL(3, \mathbb{C})$ whose Jordan canonical form is
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda & 0 & 0 \\
0 & \mu &0 \\
0 & 0 & (\lambda \mu) ^{-1}
\end{array}
\right) , \quad \lambda \ne \mu, \, |\lambda| =|\mu| \ne 1,
\end{displaymath}
and its limit set $\Lambda(\gamma)$ consists of the complex line,
$l$, on which $\gamma$ acts as an elliptic transformation of $PSL(2,
\mathbb{C})$ and the fixed point of $\gamma$ not lying in $l$.
\item The \emph{loxoparabolic} elements $\gamma \in PSL(3, \mathbb{C})$ have
a lift to $SL(3, \mathbb{C})$ whose Jordan canonical form is
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda & 1 & 0 \\
0 & \lambda &0 \\
0 & 0 & \lambda ^{-2}
\end{array}
\right) , \quad |\lambda| \ne 1,
\end{displaymath}
and the limit set $\Lambda(\gamma)$ consists of two
$\gamma$-invariant complex lines. The element $\gamma$ acts on one
of these lines as a parabolic element of $PSL(2, \mathbb{C})$ and on the
other as a loxodromic element of $PSL(2, \mathbb{C})$.
\item The \emph{strongly loxodromic} elements $\gamma \in PSL(3, \mathbb{C})$
have a lift to $SL(3, \mathbb{C})$ whose Jordan canonical form is
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda _1 & 0 & 0 \\
0 & \lambda _2 &0 \\
0 & 0 & \lambda _3
\end{array}
\right) , \quad |\lambda_1| < |\lambda_2|< |\lambda_3|.
\end{displaymath}
This kind of transformation has three fixed points, one of them is
attracting, other is repelling and the last one is a saddle. The
limit set $\Lambda(\gamma)$ is equal to the union of the complex
line determined by the attracting and saddle points and the complex
line determined by the saddle and repelling points.
\end{itemize}
\begin{theorem}\label{polloteo} \cite{cano}
Let $\Gamma \subset PSL(3, \mathbb{C})$ be a discrete and infinite group
then
\begin{enumerate}
\item there exists $\gamma _0 \in \Gamma$ such that $\gamma _0$ has infinite
order,
\item If $\Gamma$ acts properly and discontinuously on $U \subset \mathbb P_{\mathbb C}^2$, then
at least one of the complex lines in $\Lambda(\gamma _0)$, is
contained in $\mathbb P_{\mathbb C}^2 \setminus U$. In particular, there exists a
complex line $l_0$ such that $l_0 \subset \Lambda(\gamma _0)$ and
$l_0 \subset \Lambda(\Gamma)$.
\end{enumerate}
\end{theorem}
\section{ Pseudo-Projective Maps and Equicontinuity. }
\label{psprojective}
We denote by $M_{3 \times 3}(\mathbb{C})$ the space of all $3 \times 3$
matrices with complex entries equipped with the standard topology.
The quotient space
\[ (M_{3 \times 3}(\mathbb{C}) \setminus \{\mathbf{0}\})/\mathbb{C}^* \]
is called the space of \emph{pseudo-projective maps of} $\mathbb P_{\mathbb C}^2$ and
it is naturally identified with the projective complex space $\mathbb P_{\mathbb C}^8$.
Since $GL(3,\mathbb{C})$ is an open, dense, $\mathbb{C}^*$-invariant set of $M_{3
\times 3}(\mathbb{C}) \setminus \{\mathbf{0}\}$, we obtain that the space of
pseudo-projective maps of $\mathbb P_{\mathbb C}^2$ is a compactification of $PSL(3,
\mathbb{C})$. As in the case of projective maps, if $\mathbf{s} \in M_{3
\times 3}(\mathbb{C}) \setminus \{\mathbf{0}\}$, then $[\mathbf{s}]$ denotes
the equivalence class of the matrix $\mathbf{s}$ in the space of
pseudo-projective maps of $\mathbb P_{\mathbb C}^2$. Also, we say that $\mathbf{s}\in
M_{3 \times 3}(\mathbb{C})\setminus \{\mathbf{0}\}$ is a lift of the
pseudo-projective map $S$, whenever $[\mathbf{s}]=S$.
Let $S$ be an element in $( M_{3 \times 3}(\mathbb{C})\setminus
\{\mathbf{0}\})/ \mathbb{C} ^*$ and $\mathbf{s}$ a lift to $M_{3 \times
3}(\mathbb{C})\setminus \{\mathbf{0}\}$ of $S$. The matrix $\mathbf{s}$
induces a non-zero linear transformation $s:\mathbb
{C}^{3}\rightarrow \mathbb {C}^{3}$, which is not necessarily
invertible. Let $Ker(s) \subsetneq \mathbb{C} ^3$ be its kernel and let
$Ker(S)$ denote its projectivization to $\mathbb P_{\mathbb C}^2$, taking into account
that $Ker(S):= \varnothing$ whenever $Ker(s)=\{(0,0,0)\}$.
Then $S$ can be considered as a map
$$S:\mathbb {P}^{2}_\mathbb {C}\setminus Ker(S) \rightarrow \mathbb {P}^{2}_\mathbb {C}$$
$$ S([v])=[ s(v)];$$
this is well defined because $v\notin Ker(s)$. Moreover, the
commutative diagram below implies that $S$ is a holomorphic map:
$$
\xymatrix{
\mathbb {C}^{3}\setminus Ker ( s) \ar[rr]^{s} \ar[d]_{[\, \, \,
\,]}& & \mathbb {C}^{3}\setminus \{(0,0,0)\}\ar[d]^{[\, \, \,
\,]}\\
\mathbb {P}^{2}_\mathbb {C}\setminus Ker(M) \ar[rr]_{S} && \mathbb
{P}^{2}_\mathbb {C} } \;
$$
The \emph{image} of $S$, denoted $Im(S)$, is defined as the subset
of $\mathbb P_{\mathbb C}^2$ given by
$$Im(S):=[s(\mathbb {C}^{3})\setminus\{(0,0,0)\}] \, ,$$
and we have that $$dim_\mathbb {C}(Ker(S))+dim_\mathbb {C}(
Im(S))=1\,.$$
\begin{definition} \label{defeq}
The {\it equicontinuity set} for a family $\mathcal{F}$ of
endomorphisms of $\mathbb {P}^2_\mathbb {C}$, denoted
$Eq(\mathcal{F})$, is defined to be the set of points $z\in
\mathbb{P}^2_\mathbb{C}$ for which there is an open neighborhood $U$
of $z$ such that $\{ f\vert_U : f \in \mathcal{F}\}$ is a normal
family (where normal family means that every sequence of distinct
elements has a subsequence which converges uniformly on compact
sets).
\end{definition}
It is not hard to see that $Eq(\mathcal F)$ is an open set and, in
the particular case when the family $\mathcal F$ consists of the
elements of a group $\Gamma \subset PSL(3, \mathbb{C})$, the equicontinuity
set $Eq(\Gamma)$ is $\Gamma$-invariant. The following lemma helps us
to relate the equicontinuity set of a discrete group $\Gamma \subset
PSL(3,\mathbb{C})$ to the set of pseudo-projective maps obtained as limits
of the elements of $\Gamma$ .
\begin{lemma}\label{delasucesion}
Let $(\gamma _n)$ be a sequence of elements in $PSL(3,
\mathbb{C})$, then there exists a subsequence, still denoted
$(\gamma_n)$, and a pseudo-projective map $S$ such that:
\begin{itemize}
\item[1)] The sequence $(\gamma_n)$ converges uniformly to $S$ on compact subsets of
$\mathbb{P}^2_{\mathbb{C}}\setminus Ker(S)$.
\item[2)] If $Im(S)$ is a complex line, then there exists a
pseudo-projective map $T$ such that $\gamma ^{-1}_n \underset{n \to
\infty}{\longrightarrow} T$ uniformly on compact subsets of
$\mathbb{P}^2_{\mathbb{C}}\setminus Ker(T)$. Moreover, $Im(S)=
Ker(T)$ and $Ker(S)=Im(T)$.
\item[3)] If $Ker(S)$ is a complex line, then
there exists a pseudo-projective map $T$ such that $\gamma ^{-1}_n
\underset{n \to \infty}{\longrightarrow} T$ uniformly on compact
subsets of $\mathbb{P}^2_{\mathbb{C}}\setminus Ker(T)$ and $Im(S)
\subset Ker(T)$. Moreover, if $l$ is a complex line not passing
through $Im(S)$ then the sequence of complex lines $\gamma_n
^{-1}(l)$ goes to the complex line $Ker(S)$ as $n \to \infty$.
\end{itemize}
\end{lemma}
\begin{proof}
1) For every $n \in \mathbb N$, let $\mathbf g _n =(g_{ij}^{(n)})$
be a lift to $GL(3,\mathbb{C})$ of $\gamma_n$. The unitary ball in
$Mat_{3\times 3}(\mathbb{C})=\mathbb{C} ^9$ with respect to the norm $\| \mathbf g
\| _{\infty} = \underset{1 \leqslant i,j \leqslant
3}{\max}|g_{ij}|$, is a compact set. It follows that the sequence
$\tilde{\mathbf{g}}_n = \frac{\mathbf{g}_n}{\|\mathbf{g}_n
\|_{\infty}}$ has a subsequence, still denoted
$\tilde{\mathbf{g}}_n$, such that $\tilde{\mathbf{g}}_n \to
\mathbf{s} \in Mat _{3 \times 3}( \mathbb{C})$, as $n \to \infty$. We remark
that $\mathbf{s} \ne \mathbf{0}$. Now, if $K$ is a compact subset of
$\mathbb P_{\mathbb C}^2 \setminus Ker(S)$, we consider the sets $C = \{v \in \mathbb{C}
^3\setminus \{(0,0,0)\} : [v] \in K \}$ and $\tilde{K}=\{ v/\| v \|
\in \mathbb{C} ^3: v \in C \}$. We notice that $\tilde{K}$ is a compact
subset of $\mathbb{C} ^3\setminus \{(0,0,0) \}$ and $[\tilde{K}]=K$. Since
$\tilde{\mathbf{g}}_n \underset{n \to \infty}{\longrightarrow}
\mathbf{s}$ uniformly on $\tilde{K}$ we obtain that
$\gamma_n=[\tilde{\mathbf{g}}_n] \underset{n \to
\infty}{\longrightarrow} [\mathbf{s}]=S$ uniformly on $K$.
2) For every $n \in \mathbb N$, let $\mathbf g _n =(g_{ij}^{(n)})$
be a lift to $GL(3, \mathbb{C})$ of $\gamma_n$. We can assume that $\mathbf
g _n \to \mathbf{s}$ as $n \to \infty$, where $\mathbf s$ is a lift
to $Mat _{3 \times 3}( \mathbb{C})$ of the pseudo-projective map $S$. The
cofactor matrix of $\mathbf{g}_n$, denoted $cof(\mathbf{g}_n)$ is a
lift to $GL(3,\mathbb{C})$ of $\gamma _n ^{-1}$. Since the computation of
the cofactor matrix is a continuous function of the matrix entries,
we have that $cof(\mathbf{g}_n) \to cof(\mathbf s)$ as $n \to
\infty$. It follows that $\gamma_n ^{-1}$ converges to the
pseudo-projective map $[cof(\mathbf s)]:=T$ uniformly on compact
subsets of $\mathbb P_{\mathbb C}^2 \setminus Ker(T)$.
The matrix $\mathbf{s}$ may have the following Jordan canonical
forms.
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda _1 & 0 & 0 \\
0 & \lambda _2 & 0 \\
0 & 0 & 0
\end{array}
\right) , \left(
\begin{array}{ccc}
\lambda _3 & 1 & 0 \\
0 & \lambda _3 & 0 \\
0 & 0 & 0
\end{array}
\right) , \lambda _j \ne 0 \textrm{ for } j=1,2,3.
\end{displaymath}
In both cases the matrix $cof(\mathbf{s})$ is equal, up to
conjugation, to
\begin{displaymath}
\left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & \mu
\end{array}
\right) , \quad \mu \ne 0,
\end{displaymath}
and the proof of 2) follows from a straightforward computation.
3) By part 1) we can assume that $\gamma _n ^{-1}$ converges to a
pseudo-projective map $T$ uniformly on compact subsets of $\mathbb P_{\mathbb C}^2
\setminus Ker(T)$. We proceed by contradiction and we assume that
$Im(S)$ is not contained in $Ker(T)$. By hypothesis, $Ker(S)$ is a
complex line, so that $Im(S)$ consists of one single point denoted
by $p$, and $p \notin Ker(T)$. Now, if $x \in \mathbb P_{\mathbb C}^2 \setminus Ker(S)$
then $S(x)=p$, it follows that $\gamma _n(x) \underset{n \to
\infty}{\longrightarrow}S(x)=p$. Since $p \notin Ker(T)$, we can
assume that the compact set $\{ \gamma _n(x): n \in \mathbb{N} \}
\cup \{ p\}$ is contained in $\mathbb P_{\mathbb C}^2\setminus Ker(T)$. Hence $x=
\gamma _n ^{-1}(\gamma _n (x)) \underset{n \to
\infty}{\longrightarrow} T(p)$, which contradicts the fact that $T$
is a function defined on $\mathbb P_{\mathbb C}^2 \setminus Ker(T)$. Therefore, $Im(S)
\subset Ker(T)$.
For every $n \in \mathbb N$, let $\mathbf g _n =(g_{ij}^{(n)})$ be a
lift to $GL(3,\mathbb{C})$ of $\gamma_n$. We can assume that $\mathbf g _n
\to \mathbf{s} \in Mat_{3 \times 3}$ as $n \to \infty$, and
$[\mathbf{s}]=S$. Without loss of generality we can assume that
\begin{displaymath}
\mathbf s = \left(
\begin{array}{ccc}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & 1
\end{array}
\right) .
\end{displaymath}
In this case $Ker(S)=\Flecha{e_1, e_2} $ is precisely the set of
points in $\mathbb P_{\mathbb C}^2$ that satisfy the equation $0z_1+0z_2+1 z_3=0$, then
the line $Ker(S)$ can be identified with $[0:0:1]$. Now, if $l$ is a
complex line not passing through $Im(S)=[e_3]$ then $l$ can be
identified with $[A:B:C]$, $C \ne 0$.
It follows that the complex line $\gamma_n ^{-1}(l)$ is identified
with
$$[(A, B, C) (\mathbf g _n ^{-1})^{-1}]=[(A, B, C)\mathbf g _n ],$$
which implies that the sequence of complex lines $\gamma_n ^{-1}(l)$
converges to the complex line identified with
$$[(A, B, C) \mathbf g] = [0: 0: C]=[0:0:1],$$
and this line is $Ker(S)$.
\end{proof}
\begin{lemma}\label{delasuc2}
Let $( \gamma_n)_{n\in \Bbb{N}}\subset PSL(3,\mathbb {C})$ be a
sequence that converges uniformly to the pseudo-projective map $S$
on compact subsets of $\mathbb{P}^2_{\mathbb{C}}\setminus Ker(S)$.
Then
$$Eq\{ \gamma_n:n\in \mathbb {N}\}=\mathbb {P}^2_\mathbb {C}\setminus
Ker(S).$$
\end{lemma}
\begin{proof}
The inclusion
$\mathbb P_{\mathbb C}^2 \setminus Ker(S)\subset Eq\{\gamma_n : n\in\mathbb{N} \}$ is
obtained from part 1) of Lemma \ref{delasucesion}.
Now, we split the proof in two cases according to whether $Ker(S)$
is a point or a complex line.
If $Ker(S)$ is a point, denoted $k$, then $Im(S)$ is a complex line,
and given two distinct points $p$ and $q$ in $Im(S)$, there exist
points $p'$ and $q'$ arbitrarily close to $k$ such that $S(p')=p$
and $S(q')=q$ implying that $S$ cannot be extended to a continuous
function in a neighborhood of $k$, which implies $k \notin
Eq\{\gamma _n : n \in \mathbb N \}$. Therefore $Eq\{\gamma _n : n
\in \mathbb N \} \subset \mathbb P_{\mathbb C}^2\setminus Ker(S)$.
Now, let us assume that $Ker(S)$ is a line, and let $p$ be a point
in $Ker(S)\setminus {Im (S)}$, we must prove that $p\notin
Eq\{\gamma_n : n\in\mathbb{N} \}$. We choose a complex line $l_1$
different from $Ker(S)$ and passing through $p$. Since $(\mathbb P_{\mathbb C}^2)^*$ is
a compact space, we can assume that there exists a complex line
$l_2$ such that $\gamma_n(l_1)$ converges to $l_2$ as $n \to
\infty$. Given that $\gamma _n \underset{n \to
\infty}{\longrightarrow} S$ and $Im(S)$ is a point denoted $q$, we
have that $q$ lies in $l_2$.
If $q'$ is any point in $l_2\setminus \{q\}$ then there exists a
sequence of points $x_n$ in $l_1$ such that $ \gamma_n(x_n) \to q'$
as $n \to \infty$. Since $l_1$ is compact, we can assume that there
exists a point $x\in l_1$ such that $x_n \to x$ as $n\to\infty$. We
notice that $x \in Ker(S)$ because otherwise we can assume that
$\overline{\{x_n : n\in \mathbb{N}\}}$ is a compact subset of
$\mathbb{P}^2_{\mathbb{C}}\setminus Ker(S)$, then $\gamma_n(x_n) \to
q$ as $n \to \infty$, so that $q = q'$, which is absurd. Hence $x=
p$, then $x_n \to p$ and $\gamma_n(x_n)\to q' \ne q$, which implies
that $p\notin Eq\{\gamma_n : n\in\mathbb{N} \}$. It follows that
$Ker(S)\setminus Im(S) \subset Eq\{\gamma_n : n\in\mathbb{N} \}$,
since the equicontinuity set is an open set, we have that $Ker(S)
\subset \mathbb P_{\mathbb C}^2 \setminus Eq\{\gamma_n : n\in\mathbb{N} \}$. Therefore,
$\mathbb P_{\mathbb C}^2 \setminus Ker(S)= Eq\{\gamma_n : n\in\mathbb{N} \}$.
\end{proof}
\begin{theorem} \label{gamactsdisconeq}
Let $\Gamma \subset PSL(3, \mathbb{C})$ be a discrete group then $\Gamma$
acts properly and discontinuously on the open $\Gamma$-invariant set
$Eq(\Gamma)$.
\end{theorem}
\begin{proof}
By contradiction, let $K$ be a compact subset of $Eq(\Gamma)$ such
that there exists a sequence $(\gamma _n)$ of distinct elements of
$\Gamma$ that satisfy $\gamma_n (K) \cap K \ne \varnothing$. For
every $n \in \mathbb{N}$ there exists $k_n \in K$ such that $\gamma
_n (k_n) \in K$. Without loss of generality we can assume that $k_n
\underset{n \to \infty}{\longrightarrow} k \in K$ and $\gamma
_n(k_n) \underset{n \to \infty}{\longrightarrow} \kappa \in K$.
Without loss of generality, we can assume, by lemma
\ref{delasucesion} 1), that there exists a pseudo-projective map $S$
such that $\gamma _n \underset{n \to \infty}{\longrightarrow} S$
uniformly on compact subsets of $\mathbb P_{\mathbb C}^2 \setminus Ker(S)$. Since $K
\subset Eq(\Gamma)$, the lemma \ref{delasuc2} implies that $K
\subset \mathbb P_{\mathbb C}^2\setminus Ker(S)$. As\'i que $\gamma_n(k_n) \underset{n
\to \infty}{\longrightarrow} S(k) $ then $S(k)=\kappa \in K$ which
means that $\kappa \in Im(S)$.
The Lemma \ref{delasucesion} implies that there is a pseudo
projective map $T$ such that $Im(S) \subset Ker(T)$, and the Lemma
\ref{delasuc2} implies that $Ker(T) \subset \mathbb P_{\mathbb C}^2 \setminus
Eq(\Gamma)$. Hence $\kappa \in Im(S) \subset Ker(T) \subset \mathbb P_{\mathbb C}^2
\setminus Eq(\Gamma)$, which contradicts that $K \subset
Eq(\Gamma)$. Therefore $\Gamma$ acts properly and discontinuously on
$Eq(\Gamma)$.
\end{proof}
\begin{theorem}\label{dissubeq}
If $\Gamma \subset PSL(3, \mathbb{C})$ is a discrete group,
and $U$ is a $\Gamma$-invariant open subset of $\mathbb P_{\mathbb C}^2$
such that $\mathbb P_{\mathbb C}^2 \setminus U$ contains at least three complex lines
in general position, then $$U \subset Eq(\Gamma).$$
\end{theorem}
\begin{proof}
It suffices to prove that $Ker(S) \subset \mathbb P_{\mathbb C}^2 \setminus U ,$
whenever $S$ is a pseudo-projective map which is the limit of a
sequence $(\gamma _n)_{n \in \mathbb N} \subset \Gamma$.
If $Ker(S)$ is a point, denoted $p$, then $Im(S)$ is a complex line.
Moreover, there exists a pseudo-projective map $T$ such that $\gamma
_n ^{-1}$ goes to $T$, as $n \to \infty$, uniformly on compact
subset of $\mathbb P_{\mathbb C}^2 \setminus Ker(T)$, such that $Ker(S)= Im(T)$ and
$Im(S)=Ker(T)$. There is a complex line $l_1 \ne Ker(T)$ such that
$l_1 \subset \mathbb P_{\mathbb C}^2 \setminus U$. It follows that $\gamma _n ^{-1}(x)$
goes to $p$ as $n \to \infty$ for every point $x \in l_1 \setminus
Ker(S)$, which implies that $p \in \mathbb P_{\mathbb C}^2 \setminus U$ (because $U$ is
a $\Gamma$-invariant open set).
If $Ker(S)$ is a complex line, then there exists a complex line
$l_1$ not passing through $Im(S)$ and contained in $\mathbb P_{\mathbb C}^2 \setminus
U$, because $\mathbb P_{\mathbb C}^2 \setminus U$ contains at least three lines in
general position. By the Lemma \ref{delasucesion} part iii), $\gamma
_n ^{-1} (l_1)$ goes to $Ker(S)$, as $n \to \infty$, which implies
that $Ker(S) \subset \mathbb P_{\mathbb C}^2 \setminus U$.
\end{proof}
\begin{theorem}\label{eqeqdis}
Let $\Gamma \subset PSL(3, \mathbb{C})$ be a discrete subgroup such that
$\Lambda(\Gamma)$ contains at least three lines in general position,
then $$\Omega(\Gamma)= Eq(\Gamma).$$
\end{theorem}
\begin{proof}
By Theorem \ref{gamactsdisconeq} we have that $L_0(\Gamma) \cap
Eq(\Gamma)= \varnothing = L_1(\Gamma) \cap Eq(\Gamma)$. Now we prove
that $L_2(\Gamma) \cap Eq(\Gamma)= \varnothing$. Let us denote by
$C$ the complement of $Eq(\Gamma)$ in $\mathbb P_{\mathbb C}^2$, and let $K$ be a
compact subset of $\mathbb P_{\mathbb C}^2 \setminus C = Eq(\Gamma)$. By Lemma
\ref{cuasiminimalidad}, it suffices to prove that the accumulation
points of the orbit of $K$ are contained in $L_0 (\Gamma) \cup
L_1(\Gamma)$. Let $x$ be one of such accumulation points, then there
exists a sequence, $k_n$, of points in $Eq(\Gamma)$ such that $k_n
\to k \in Eq(\Gamma)$ as $n \to \infty$, and there is a sequence
$\gamma _n$ of different elements in $\Gamma$, such that $\gamma _n
(k_n) \to x$, as $n \to \infty$. By the Lemma \ref{delasucesion} we
can assume that there exists a pseudo-projective map $S$ such that
$\gamma _n$ goes to $S$, as $n \to \infty$, uniformly on compact
subsets of $\mathbb P_{\mathbb C}^2 \setminus Ker(S)$. We have two cases according to
whether $Ker(S)$ is a point or a complex line.
If $Ker(S)$ is a complex line, then the Lemma \ref{delasucesion}
iii) implies that $\{k_n : n \in \mathbb N\} \cup \{k\} \subset
Eq(\Gamma) \subset \mathbb P_{\mathbb C}^2 \setminus Ker(S)$. Thus, $\gamma_n(k_n) \to
x$ implies that $\gamma _n(k) \to S(k)=x$, and given that $k \in
Eq(\Gamma) \subset \mathbb P_{\mathbb C}^2 \setminus L_0 (\Gamma)$, we deduce that $x
\in L_1(\Gamma)$.
If $Ker(S)$ is a point then $(\{k_n\} \cup \{k\}) \cap Ker(S)=
\varnothing$ because $\{k_n\} \cup \{k\} \subset Eq(\Gamma)$, and
the proof follows as in the former case.
Therefore $L_2(\Gamma) \cap Eq(\Gamma)= \emptyset$, then $Eq(\Gamma)
\subset \Omega(\Gamma)$. The proof of the reverse inclusion follows
easily from the Theorem \ref{dissubeq}.
\end{proof}
\section{$\Lambda(\Gamma)$ is union of complex lines}
\label{lambdaunlin} In this section $\Gamma \subset PSL(3, \mathbb{C})$ is a
Complex Kleinian Group, and $C(\Gamma)$ denotes the set
$$C(\Gamma) = \overline{ \bigcup _{\gamma \in \Gamma} \Lambda(\gamma)}$$
\begin{lemma}\label{CsubLambda}
If $\Lambda(\Gamma)$ has at least three complex lines in general
position, then $\Lambda(\gamma) \subset \Lambda(\Gamma)$ for every
$\gamma \in \Gamma$. In particular, $C(\Gamma)\subset
\Lambda(\Gamma)$.
\end{lemma}
\begin{proof}
Let $\gamma$ be an element in $\Gamma$.
There are three cases depending on whether $\gamma$ is elliptic,
parabolic or loxodromic (\cite{Na2})
(1) If $\gamma$ elliptic then $\Lambda(\gamma)=\varnothing$ and the
result follows immediately.
(2) If $\gamma$ es parabolic, then we have three subcases:
\quad (i) We can assume, without loss of generality, that $\gamma$
has a lift given by the matrix.
\begin{displaymath}
\left(
\begin{array}{ccc}
1 & 1 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array}
\right) ,
\end{displaymath}
then $\Lambda(\gamma)=L_0(\gamma) = \overleftrightarrow{e_1, e_2}
\subset \L_0(\Gamma) \subset \Lambda(\Gamma)$.
\quad (ii) We can assume, without loss of generality, that $\gamma$
has a lift $\tilde{\gamma}$ which has the form
\begin{displaymath}
\left(
\begin{array}{ccc}
1 & 1 & 0 \\
0 & 1 & 1 \\
0 & 0 & 1
\end{array}
\right) .
\end{displaymath}
It follows that, for any $n \in \mathbb N$,
\begin{displaymath}
\tilde{\gamma} ^n =\left(
\begin{array}{ccc}
1 & n & \frac{n(n-1)}{2} \\
0 & 1 & n \\
0 & 0 & 1
\end{array}
\right) .
\end{displaymath}
If $\Lambda(\gamma)= \Flecha{e_1 , e_2}$ is not contained in
$\Lambda(\Gamma)$, then there exists $z \in \mathbb{C}$ such that $[z:1:0]
\in \Omega(\Gamma)$. Let $w$ be any complex number, then the
sequence $a_n(w):= [z:1: \frac{2(w-n)}{n(n-1)}]$, $n \in \mathbb N$
tends to $[z:1:0]$ as $n \to \infty$. Thus, for $N>0$ large enough
($N$ depends on $w$), we have that $(a_n(w)) _{n \geqslant N}
\subset \Omega(\Gamma)$ for all $n >N$. We note that
$$\gamma ^n (a_n)=[z+ w : \frac{2w-(n+1) }{n-1} : \frac{2(w-n)}{n(n-1)}]
\underset{n \to \infty}{\longrightarrow} [-z -w: 1:0].$$ Hence, $[-z
-w: 1:0] \in \Lambda(\Gamma)$, and taking $w=-2z$ we obtain a
contradiction. Therefore, $\Lambda(\gamma) \subset \Lambda(\Gamma)$.
\quad (iii) We can assume, without loss of generality, that $\gamma$
has a lift $\tilde{\gamma}$ of the form:
\begin{displaymath}
\tilde{\gamma} = \left(
\begin{array}{ccc}
\lambda & 1 & 0 \\
0 & \lambda & 0 \\
0 & 0 & \lambda ^{-2}
\end{array}
\right) , \quad |\lambda|=1.
\end{displaymath}
We proceed by contradiction. Let's assume that $\Lambda(\gamma)=
\Flecha{e_1, e_3}$ is not contained in $\Lambda(\Gamma)$, then there
exists $z \in \mathbb{C}$ such that $[z:0:1] \in \Omega(\Gamma)$. We note
that $a_n:= [z:1/n^2 : 1] \underset{n \to \infty}{\longrightarrow}
[z:0:1]$, so that for $N>0$ large enough, we have that $(a_n) _{n
\geqslant N} \subset \Omega(\Gamma)$ for all $n
>N$. The sequence,
$$\gamma ^n (a_n)=[z \lambda ^n+ \frac{n \lambda ^{n-1}}{n^2} : \frac{\lambda ^n}{n^{2}} : \lambda ^{-2n}]
=[z \lambda ^{3n}+ \frac{\lambda ^{3n-1}}{n}: \frac{\lambda
^{3n}}{n^2}:1], \quad n >N,$$ has cluster point $[z:0:1]$ which
implies that $[z:0:1] \in \Lambda(\Gamma)$, a contradiction.
(3) If $\gamma$ is loxodromic, we have four subcases:
\quad (i) If $\gamma$ has a lift $\tilde{\gamma}$ whose normal
Jordan form is:
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda _1 & 0 & 0 \\
0 & \lambda _2 & 0 \\
0 & 0 & \lambda _3
\end{array}
\right) , \quad |\lambda _1|<|\lambda _2|<|\lambda _3|,
\end{displaymath}
then we can assume that $\tilde \gamma$ is such matrix. It follows
that $\Lambda(\gamma)= \Flecha{e_1, e_2} \cup \Flecha{e_2, e_3}$,
and the lambda-lemma implies that
$$\Flecha{e_1,e_2} \subset \Lambda(\Gamma) \, \textrm{ \'o } \,
\Flecha{e_2, e_3} \subset \Lambda(\Gamma).$$ Let's assume, without
loss of generality, that $ \Flecha{e_1,e_2} \subset
\Lambda(\Gamma)$, then there exists $l \subset \Lambda(\Gamma)$ such
that $l$ does not pass through the point $[1:0:0]$ (because
$\Lambda(\Gamma)$ contains at least three complex lines in general
position). We note that the sequence of complex lines $\gamma ^n
(l)$ goes to the complex line $\Flecha{e_2, e_3}$ as $n \to \infty$,
therefore $\Flecha{e_2, e_3}\subset \Lambda(\Gamma)$.
\quad (ii) If $\gamma$ has a lift $\tilde{\gamma}$ whose Jordan
normal form is
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda _1 & 0 & 0 \\
0 & \lambda _1 & 0 \\
0 & 0 & \lambda _1 ^{-2}
\end{array}
\right) , \quad |\lambda _1| \ne 1,
\end{displaymath}
then we can assume that $\tilde \gamma$ is such matrix. In this case
$\Lambda (\gamma)=L_0(\gamma) \subset L_0(\Gamma) \subset
\Lambda(\Gamma)$.
\quad (iii) If $\gamma$ has a lift $\tilde{\gamma}$ whose Jordan
normal form is
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda _1 & 0 & 0 \\
0 & \lambda _2 & 0 \\
0 & 0 & \lambda _3
\end{array}
\right) , \quad |\lambda _1| = |\lambda _2| < |\lambda _3|,
\end{displaymath}
then we can assume that $\tilde \gamma$ is such matrix, so that
$$\Lambda(\gamma)= \Flecha{e_1, e_2} \cup \{[0:0:1]\}.$$
Let $l$ be a complex line contained in $\Lambda(\Gamma)$ not passing
through $[0:0:1]$ then $\gamma ^{-n} (l)$ is a sequence of complex
lines that tends to the complex line $\Flecha{e_1, e_2}$ as $n$
tends to $\infty$, which implies that $\Flecha{e_1, e_2} \subset
\Lambda(\Gamma)$. Moreover, $\gamma ^n(z) \underset{n \to
\infty}{\longrightarrow} [0:0:1]$ whenever $z \in P^2 _\mathbb{C} \setminus
\Flecha{e_1, e_2}$. Therefore $[0:0:1] \in \Lambda(\Gamma)$.
\quad (iv) If $\gamma$ has a lift $\tilde{\gamma}$ whose Jordan
normal form is
\begin{displaymath}
\left(
\begin{array}{ccc}
\lambda & 1 & 0 \\
0 & \lambda & 0 \\
0 & 0 & \lambda ^{-2}
\end{array}
\right) , \quad |\lambda | < 1,
\end{displaymath}
then we can assume that $\tilde \gamma$ is such matrix, in this
case, $\Lambda(\gamma)=\Flecha{e_1, e_2} \cup \Flecha{e_1,e_3}$. We
have that
\begin{displaymath}
\tilde{\gamma} ^n= \left(
\begin{array}{ccc}
\lambda ^n & n \lambda ^{n-1} & 0 \\
0 & \lambda ^n & 0 \\
0 & 0 & \lambda ^{-2n}
\end{array}
\right) , n \in \mathbb N .
\end{displaymath}
Let's assume that $\Flecha{e_1, e_2}$ is not contained in
$\Lambda(\Gamma)$, then there exists $z \in \mathbb{C}$ such that $[z:1:0]
\in \Omega(\Gamma)$. If $w$ is a non-zero complex number, then the
sequence $[wz:w: n\lambda ^{3n-1}]$ tends to $[z:1:0]$ as $n \to
\infty$. Moreover $$ \gamma ^n([z:1: \frac{n \lambda ^{3n-1}}{w}])
\underset{n \to \infty}{\longrightarrow} [w:0:1],
$$
which implies that the complex line $\Flecha{e_1, e_3}$ is contained
in $\Lambda(\Gamma)$.
Now, assuming $\Flecha{e_1, e_2} \subset \Lambda(\Gamma)$, let $l$
be a complex line different from $\Flecha{e_1, e_2}$, which is
contained in $\Lambda(\Gamma)$ and does not pass through $[0:0:1]$,
then $l$ has an equation of the form
$$Az_1 +Bz_2 +Cz_3 =0, \quad |A|^2+|B|^2 \ne 0, C \ne 0.$$
It is not hard to check that $\gamma ^n (l)$, $n \in \mathbb N$ has
an equation of the same kind, with coefficients $(A'(n), B'(n),
C'(n))$ given by the equation
\begin{displaymath}
\left(
\begin{array}{c}
A'(n) \\
B'(n)\\
C'(n)
\end{array}
\right) =\left( \left(
\begin{array}{ccc}
\lambda ^n & n \lambda ^{n-1} & 0 \\
0 & \lambda ^n & 0 \\
0 & 0 & \lambda ^{-2n}
\end{array}
\right) ^{T} \right) ^{-1} \left(
\begin{array}{c}
A \\
B\\
C
\end{array}
\right)
\end{displaymath}
Thus, $(A'(n), B'(n), C'(n))=(\lambda ^{-n} A, \lambda ^{-n}B -n
\lambda^{-n-1}A, \lambda ^{2n} C)$, in other words, $$(A'(n), B'(n),
C'(n))=( A, B -n \lambda^{-1}A, \lambda ^{3n} C),$$ it follows that
$\gamma ^n (l) \underset{n \to \infty}{\longrightarrow} \Flecha{e_1,
e_3}$, which implies that $\Flecha{e_1, e_3} \subset
\Lambda(\Gamma)$.
If $\Flecha{e_1, e_3} \subset \Lambda(\Gamma)$, let us take a line
$l$ contained in $\Lambda(\Gamma)$ not passing through $[0:0:1]$,
then $l$ has an equation of the form
$$Az_1 +Bz_2 +Cz_3 =0, \quad C \ne 0.$$
It is not hard to check that $\gamma ^{-n} (l)$, $n \in \mathbb N$
has an equation of the same kind, whose coefficients $(A'(-n),
B'(-n), C'(-n))$ are given by the equation
\begin{displaymath}
\left(
\begin{array}{c}
A'(-n) \\
B'(-n)\\
C'(-n)
\end{array}
\right) = \left(
\begin{array}{ccc}
\lambda ^n & 0 & 0 \\
n \lambda ^{n-1} & \lambda ^n & 0 \\
0 & 0 & \lambda ^{-2n}
\end{array}
\right) \left(
\begin{array}{c}
A \\
B\\
C
\end{array}
\right) .
\end{displaymath}
Thus $(A'(-n), B'(-n), C'(-n))=(\lambda ^n A ,n \lambda ^{n-1}A +
\lambda ^n B, \lambda ^{-2n}C)$, equivalently, $$(A'(-n), B'(-n),
C'(-n))=(\lambda ^{3n} A ,n \lambda ^{3n-1}A + \lambda ^{3n} B,
C),$$ since $|\lambda|<1$ we have that $\gamma ^{-n}(l) \underset{n
\to \infty}{\longrightarrow} \Flecha{e_1, e_2}$, which implies that
$\Flecha{e_1, e_2} \subset \Lambda(\Gamma)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{main2}]
First, we notice that Theorem \ref{dissubeq} implies that $\mathbb P_{\mathbb C}^2
\setminus C(\Gamma)\subset Eq(\Gamma)$, because $\mathbb P_{\mathbb C}^2 \setminus
C(\Gamma)$ is a $\Gamma$-invariant open set and $C(\Gamma)$ contains
at least three complex lines in general position. It follows from
Theorem \ref{eqeqdis} that $\mathbb P_{\mathbb C}^2 \setminus C(\Gamma)\subset
Eq(\Gamma)= \Omega(\Gamma)$. Therefore $\Gamma$ acts properly and
discontinuously on $\mathbb P_{\mathbb C}^2 \setminus C(\Gamma)$ and $\Lambda(\Gamma)
\subset C(\Gamma)$.
Given that $\Lambda(\Gamma)$ contains at least three complex lines
in general position, the Lemma \ref{CsubLambda} implies that
$C(\Gamma)\subset \Lambda(\Gamma)$.
In order to prove that $C(\Gamma)$ is a union of complex lines, it
suffices to check that for every $p \in \cup _{\gamma \in \Gamma}
\Lambda (\gamma)$, there exists a complex line contained in
$\Lambda(\Gamma)$ passing through $p$. Hence, let us assume that $p
\in \Lambda(\gamma)$ for some $\gamma \in \Gamma$. We need only to
consider the case when $\Lambda(\gamma)$ consists of one complex
line $l$ and one point $q \notin l$. If $p \in l$ then there is
nothing to prove. If $p=q$ then there exists a complex line $l_1
\subset \Lambda(\Gamma)$, $l_1 \ne l$, so when $n \to \infty$,
$\gamma ^n(l_1)$ or $\gamma ^{-n}(l_1)$ goes to a complex line
passing through $p$ and contained $\Lambda(\Gamma)$.
Since $C(\Gamma)$ contains three complex lines in general position
and $\Gamma$ has no fixed point, there exist three complex lines
$l_1, l_2, l_3$ and three elements $\gamma _1, \gamma _2, \gamma _3
\in \Gamma$ such that $l_1 \subset \Lambda(\gamma _1)$, $l_2 \subset
\Lambda(\gamma _2)$, $l_3 \subset \Lambda(\gamma _3)$. Now, let $U$
be a $\Gamma$-invariant open subset of $\mathbb P_{\mathbb C}^2$ on which $\Gamma$ acts
properly and discontinuously.
If $\gamma _1, \gamma _2, \gamma _3$ are pairwise different, then,
by Theorem \ref{polloteo}, we can assume $l_1, l_2, l_3\subset
\mathbb P_{\mathbb C}^2\setminus U$. It follows from Theorems \ref{dissubeq} and
\ref{eqeqdis} that $U \subset Eq(\Gamma)=\Omega(\Gamma)$.
If $\gamma _1, \gamma _2, \gamma _3$ are not pairwise different,
then without loss of generality we can assume that $\gamma _1 \ne
\gamma _2 = \gamma _3$. By Theorem \ref{polloteo} we can assume that
$l_1 \subset P^2 \setminus U$ and at least one of the lines $l_2$ or
$l_3$ is contained in $\mathbb P_{\mathbb C}^2 \setminus U$, therefore $\mathbb P_{\mathbb C}^2 \setminus
U$ contains at least two complex lines in general position. If there
were not more than two complex lines in general position contained
in $\mathbb P_{\mathbb C}^2 \setminus U$, then the intersection point of $l_1$ and
$l_2$ or the intersection point of $l_1$ and $l_3$ would be a
$\Gamma$-fixed point, and it cannot happen. Therefore $\mathbb P_{\mathbb C}^2
\setminus U$ contains at least three lines in general position, and
it follows from Theorems \ref{dissubeq} and \ref{eqeqdis} that $U
\subset Eq(\Gamma)=\Omega(\Gamma)$.
\end{proof}
In what follows, we only consider groups $\Gamma \subset PSL(3, \mathbb{C})$
whose action on $\mathbb P_{\mathbb C}^2 $ have no fixed point nor invariant lines.
\begin{lemma} \label{3lineslema}
If $\Gamma \subset PSL(3, \mathbb{C})$ is a infinite discrete group, having
no fixed points nor invariant lines then $ C(\Gamma)\cap
\Lambda(\Gamma)$ contains at least three complex lines in general
position.
\end{lemma}
\begin{proof}
Given that $\Gamma \subset PSL(3, \mathbb{C})$ is an infinite and discrete
group, by Theorem \ref{polloteo}, there exists an infinite order
element $\gamma_0 \in \Gamma$, and one complex line, $l_0$, such
that $l_0 \subset \Lambda(\gamma _0) \subset C(\Gamma)$, and $l_0
\subset \Lambda(\Gamma)$. Since $l_0$ is not $\Gamma$-invariant,
there exists an element $\gamma _1 \in \Gamma$ such that $l_1
:=\gamma _1 (l_0)\ne l_0$ (we notice that $l_1 \subset
\Lambda(\gamma _1 \gamma _0 \gamma _1 ^{-1}) \subset C(\Gamma)$).
Let $q_0$ be the intersection point of $l_0$ and $l_1$. If $\gamma
(l_0)$ passes through $q_0$ for every $\gamma \in \Gamma$, then
$q_0$ is a $\Gamma$-fixed point, which cannot happen. Thus, there
exists $\gamma _2 \in \Gamma$ such that $l_2:=\gamma_2 (l_0)$, $l_1$
and $l_0$ are in general position, and we see that $l_2 \subset
\Lambda(\gamma _2 \gamma _0 \gamma _2 ^{-1} ) \subset C(\Gamma)$.
\end{proof}
\begin{example} Let $\alpha$ and $\beta$ in $PSL(3, \mathbb{C})$ be the elements
induced, respectively, by the matrices
\begin{displaymath}
A= \left(
\begin{array}{ccc}
\frac{1}{2} & 0 &0 \\
0 & 1 & 0 \\
0 & 0 & 2
\end{array}
\right), \quad \quad B= \left(
\begin{array}{ccc}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}
\right) .
\end{displaymath}
Let $G$ be the group $\langle A , B \rangle \subset GL(3, \mathbb{C})$, and
$\Gamma = \langle \alpha , \beta \rangle \subset PSL(3, \mathbb{C})$, then
$\Gamma$ is a discrete group, it has no fixed points nor invariant
lines and $\Lambda(\Gamma)= C(\Gamma) = \Flecha{e_1, e_2} \cup
\Flecha{e_2, e_3} \cup \Flecha{e_3, e_1}$.
\end{example}
\begin{proof}
a) The group $G$ is a discrete, so $\Gamma$ is a discrete group. In
order to proof this statement, we use the norm in $\{ \mathbf{g} \in
GL(3, \mathbb{C}): \det \mathbf{g} = \pm 1\}$ given by $\|\mathbf{g}\|= \max
{|g_{i,j}|}$. We notice that each element in $G$ is given by one of
the following matrices:
\begin{displaymath}
\left(
\begin{array}{ccc}
2^{m_1} & 0 & 0 \\
0 & 2^{m_2} & 0 \\
0 & 0 & 2^{m_3}
\end{array}
\right), \quad \left(
\begin{array}{ccc}
0 & 0 & 2^{m_1} \\
2^{m_2} & 0 & 0 \\
0 & 2^{m_3} & 0
\end{array}
\right), \quad
\left(
\begin{array}{ccc}
0 & 2^{m_1} & 0 \\
0 & 0 & 2^{m_2} \\
2^{m_3} & 0 & 0
\end{array}
\right),
\end{displaymath}
where $m_1, m_2, m_3 \in \mathbb Z$ and $m_1+m_2+m_3 =0$, because
the set $\mathbf{H}$ of such matrices, is a group, and $A \in
\mathbf{H}$, $B \in \mathbf{H}$. Since $G \subset \mathbf{H}$, it
suffices to prove that $\mathbf{H}$ is a discrete group, and , for
this, we only need to check that, for every $k>0$, the set $\{
\mathbf{h} \in \mathbf{H} \, | \, \| \mathbf{h} \| < k \}$ is
finite. Let us assume that $\| h \| < k$ and, without loss of
generality, $m_1 = \max\{m_1, m_2 , m_3\}>0$, then
$$0 < m_1 < \log _2 k,$$
$$-\log _2 k < m_2 + m_3 < 2 \log _2 k.$$
We have two cases according to whether $m_2$ and $m_3$ are both
negative or $m_1 m_2 <0$. In any case, there are finitely many
values of $m_1, m_2, m_3$ under these conditions.
(b) The group $\Gamma$ has no fixed points nor invariant lines,
because every fixed point of $\alpha$ is not fixed by $\beta$, and
every $\alpha$-invariant complex line is not $\beta$-invariant.
(c) $\Lambda(\Gamma)= C(\Gamma) = \Flecha{e_1, e_2} \cup
\Flecha{e_2, e_3} \cup \Flecha{e_3, e_1}$.
The Lemma \ref{3lineslema} implies that $C(\Gamma) \cap
\Lambda(\Gamma)$ contains three complex lines in general position,
then Theorem \ref{main2} implies that $\Lambda(\Gamma)= C(\Gamma)$.
It is not hard to check that $\Lambda(\gamma)= \Lambda(\gamma ^3)$,
and $\gamma ^3$ is given by a diagonal matrix, it follows that
$\Lambda(\gamma ^3) \subset \Flecha{e_1, e_2} \cup \Flecha{e_2, e_3}
\cup \Flecha{e_3, e_1}$. Therefore $C(\Gamma) \subset \Flecha{e_1,
e_2} \cup \Flecha{e_2, e_3} \cup \Flecha{e_3, e_1}$, and it is not
hard to see that $\Flecha{e_1, e_2} \cup \Flecha{e_2, e_3} \cup
\Flecha{e_3, e_1} \subset C(\Gamma)$.
\end{proof}
Inspired by the last example we state the following Lemma.
\begin{lemma} \label{4lineslema}
If $\Gamma \subset PSL(3, \mathbb{C})$ is an infinite discrete group without
fixed points nor invariant lines, and $C(\Gamma)$ contains more than
three complex lines (not necessarily in general position) then $
C(\Gamma) \cap \Lambda(\Gamma)$ has at least \emph{four} complex
lines in general position.
\end{lemma}
\begin{proof}
The proof of this Lemma is an extension of the proof of
\ref{3lineslema}, so there exist three complex lines $l_0, l_1, l_2$
(in general position), contained in the set $ C(\Gamma) \cap
\Lambda(\Gamma)$. Let $q_0$, $q_1$, $q_2$ be the intersection points
of the pairs of complex lines $l_0$ and $l_1$, $l_1$ and $l_2$,
$l_2$ and $l_0$, respectively. Now we proceed by contradiction, we
assume that there are not more than three lines (in general
position) contained in $C(\Gamma) \cap \Lambda(\Gamma)$, then for
every $\gamma \in \Gamma$, the complex lines $\gamma (l_j)$ for
$j=0,1,2$ necessarily pass through one of the points $q_0, q_1,
q_2$. We can suppose, without loss of generality, that there exists
one such line, $l_3$, distinct from $l_0$ and $l_1$, passing through
$q_0$. If there are no more lines of the form $\gamma (l_j),$
$j=0,1,2$, then $q_0$ is fixed by the whole group $\Gamma$ which
cannot happen. Therefore, there exists another complex line $l_4$
such that $l_4$ is equal to $\gamma (l_j)$ for some $\gamma \in
\Gamma$ and for some $j=0,1,2$,
and $l_4$ passes through a point $q_1$ distinct from $q_0$ (because
otherwise $q_0$ is fixed by $\Gamma$), then $l_0, l_2, l_3, l_4$ are
in general position.
\end{proof}
\begin{lemma} \label{lineatractora} Let $\gamma \in PSL(3, \mathbb{C})$ be any
element. Let us assume that $l_1$ is a complex line contained in
$\Lambda(\gamma)$, then for every complex line $l$ different from
$l_1$ (except, maybe, for a family of complex lines in a pencil of
complex lines), some of the sequences of \textbf{\emph{distinct}}
complex lines, $(\gamma ^n(l)) _{n \in \mathbb N}$ or $(\gamma
^{-n}(l))_{n \in \mathbb N}$ goes to $l_1$, as $n \to \infty$.
\end{lemma}
\begin{proof}
We split the proof in several cases according to the classification
of the elements in $PSL(3, \mathbb{C})$ given in \cite{Na2}.
(a) In the first case, we can assume that $\gamma$ has a lift of the
form
\begin{displaymath}
\left(
\begin{array}{ccc}
1 & 1& 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array}
\right),
\end{displaymath}
then $\Lambda(\gamma)$ is the complex line $\Flecha{e_1, e_3}$,
which is represented, in $(\mathbb P_{\mathbb C}^2 )^*$, by $[0:1:0]$. If $l$ is a
complex line represented in $(\mathbb P_{\mathbb C}^2)^*$ by $[A:B:C]$, then $\gamma ^n
(l), n \in \mathbb N$ is represented in $(\mathbb P_{\mathbb C}^2)^*$ by
\begin{displaymath}
[A(n): B(n): C(n)]:= [(A, B, C)\left(
\begin{array}{ccc}
1 & -n & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array}
\right) ] = [A: -nA+B: C].
\end{displaymath}
We obtain that $[A(n): B(n): C(n)] \underset{n \to
\infty}{\longrightarrow} [0:1:0]$, whenever $A \ne 0$, which implies
that $\gamma ^n(l) \underset{n \to \infty}{\longrightarrow}
\Flecha{e_1, e_3}$, whenever $l$ is a complex line not passing
through the point $[1:0:0]\in \mathbb P_{\mathbb C}^2$.
b) We can assume that $\gamma$ has a lift of the form:
\begin{displaymath}
\left(
\begin{array}{ccc}
\zeta & 1& 0 \\
0 & \zeta & 0 \\
0 & 0 & \zeta ^{-2}
\end{array}
\right), \quad |\zeta| =1,
\end{displaymath}
then $\Lambda(\gamma)$ is the complex line $\Flecha{e_1, e_3}$ which
is represented, in $(\mathbb P_{\mathbb C}^2 )^*$, by $[0:1:0]$. If $l$ is a complex
line represented in $(\mathbb P_{\mathbb C}^2 )^*$ by $[A:B:C]$, then $\gamma ^n (l), n
\in \mathbb N$ is represented in $(\mathbb P_{\mathbb C}^2 )^*$ by
\begin{displaymath}
[A(n): B(n): C(n)]:= [(A, B, C)\left(
\begin{array}{ccc}
\zeta ^{-n} & -n \zeta ^{-(n+1)} & 0 \\
0 & \zeta ^{-n} & 0 \\
0 & 0 & \zeta^{2n}
\end{array}
\right)]
\end{displaymath}
Hence, $$[A(n): B(n): C(n)]= [A \zeta ^{-n}: -n \zeta ^{-(n+1)}A +
\zeta ^{-n}B:\zeta ^{2n} C ].$$
We obtain that $[A(n): B(n): C(n)] \underset{n \to
\infty}{\longrightarrow} [0:1:0]$, whenever $A \ne 0$, which means
that $\gamma ^n(l) \underset{n \to \infty}{\longrightarrow}
\Flecha{e_1, e_3}$, whenever $l$ is a complex line not passing
through the point $[1:0:0]\in \mathbb P_{\mathbb C}^2$.
(c) We can assume that $\gamma$ has a lift of the form:
\begin{displaymath}
\left(
\begin{array}{ccc}
1 & 1& 0 \\
0 & 1 & 1 \\
0 & 0 & 1
\end{array}
\right).
\end{displaymath}
In this case, $\Lambda(\gamma)$ is the complex line $\Flecha{e_1,
e_2}$ which is represented, in $(\mathbb P_{\mathbb C}^2)^*$, by $[0:0:1]$. If $l$ is a
complex line represented in $(\mathbb P_{\mathbb C}^2)^*$ by $[A:B:C]$, then $\gamma ^n
(l), n \in \mathbb N$ is represented in $(\mathbb P_{\mathbb C}^2)^*$ by
\begin{displaymath}
[A(n): B(n): C(n)]:= [(A, B, C)\left(
\begin{array}{ccc}
1 & -n & \frac{n(n+1)}{2} \\
0 & 1 & -n \\
0 & 0 & 1
\end{array}
\right)] .
\end{displaymath}
Hence, $$[A(n): B(n): C(n)]=[A:-nA+B: \frac{n(n+1)}{2}A-nB+C],$$and
it follows that $[A(n): B(n): C(n)] \underset{n \to
\infty}{\longrightarrow} [0:0:1]$, which means that $\gamma ^n(l)
\underset{n \to \infty}{\longrightarrow} \Flecha{e_1, e_2}$ for any
line $l$.
(d) We can assume $\gamma$ has a lift of the form
\begin{displaymath}
\left(
\begin{array}{ccc}
\zeta & 1& 0 \\
0 & \zeta & 0 \\
0 & 0 & \zeta ^{-2}
\end{array}
\right), \quad |\zeta| >1.
\end{displaymath}
In this case $\Lambda(\gamma)$ is equal to the union of the complex
lines $\Flecha{e_1, e_2}$ and $\Flecha{e_1, e_3}$ which are
represented in $(\mathbb P_{\mathbb C}^2)^*$, by $[0:0:1]$ and $[0:1:0]$, respectively.
If $l$ is a complex line represented in $(\mathbb P_{\mathbb C}^2)^*$ by $[A:B:C]$,
then $\gamma ^n (l), n \in \mathbb N$ is represented in $(\mathbb P_{\mathbb C}^2)^*$
by
\begin{displaymath}
[A(n): B(n): C(n)]:= [(A, B, C)\left(
\begin{array}{ccc}
\zeta ^{-n} & -n \zeta ^{-(n+1)} & 0 \\
0 & \zeta ^{-n} & 0 \\
0 & 0 & \zeta^{2n}
\end{array}
\right) ]
\end{displaymath}
Hence,
\begin{eqnarray*}
[A(n): B(n): C(n)] & = & [A \zeta ^{-n}: -n \zeta ^{-(n+1)}A +
\zeta ^{-n}B:\zeta ^{2n} C ] \\
& = & [A \zeta ^{-3n}: -n \zeta ^{-3n-1}A+ \zeta ^{-3n}B : C ].
\end{eqnarray*}
So that $[A(n): B(n): C(n)] \underset{n \to \infty}{\longrightarrow}
[0:0:1]$, whenever $C \ne 0$, which means that $\gamma ^n(l)
\underset{n \to \infty}{\longrightarrow} \Flecha{e_1, e_2}$,
whenever $l$ is a complex line not passing through the point
$[0:0:1]\in \mathbb P_{\mathbb C}^2$.
Now, if $l$ is a complex line represented in $(\mathbb P_{\mathbb C}^2)^*$ by
$[A:B:C]$, then $\gamma ^{-n} (l), n \in \mathbb N$ is represented
in $(\mathbb P_{\mathbb C}^2)^*$ by
\begin{displaymath}
[A(-n): B(-n): C(-n)]:= [(A, B, C)\left(
\begin{array}{ccc}
\zeta ^{n} & n \zeta ^{n-1} & 0 \\
0 & \zeta ^{n} & 0 \\
0 & 0 & \zeta^{-2n}
\end{array}
\right) ]
\end{displaymath}
Hence,
\begin{eqnarray*}
[A(-n): B(-n): C(-n)] & = & [\zeta ^{n}A : n \zeta ^{n-1}A +
\zeta ^{n}B: \zeta ^{-2n}C ] \\
& = & [A : n \zeta ^{-1}A+ B : \zeta ^{-3n} C ] \\
& = & [A /n : A \zeta ^{-1}+ B/n : \zeta ^{-3n} C /n].
\end{eqnarray*}
Thus, $[A(-n): B(-n):C(-n)] \underset{n \to \infty}{\longrightarrow} [0:1:0]$,
whenever $A \ne 0$ or $B \ne 0$, which implies that $\gamma ^{-n}(l)
\underset{n \to \infty}{\longrightarrow} \Flecha{e_1, e_3}$,
whenever $l$ is a complex line different from $\Flecha{e_1, e_2}$.
The remaining cases are proved in a similar way.
\end{proof}
We define the set $\mathcal{E}(\Gamma)$ as the subset of $(\mathbb P_{\mathbb C}^2 )^*$
consisting of all the complex lines $l$ for which there exists an
element $\gamma \in \Gamma$ such that $ l \subset \Lambda(\gamma)$.
Also we define $E(\Gamma)$ as the subset of $\mathbb P_{\mathbb C}^2$ given by
$E(\Gamma) = \overline{\bigcup _{l \in \mathcal{E}(\Gamma)} l}$. It
is not hard to see that $E(\Gamma) \subset C(\Gamma)$ and
$E(\Gamma)=\bigcup _{l \in \overline{\mathcal{E}(\Gamma)}} l$.
\begin{remark}
If $\mathcal{E}(\Gamma)$ contains at least two distinct complex
lines, then
$$C(\Gamma) = E(\Gamma)=\overline{\bigcup _{l \in \mathcal{E}(\Gamma)} l}
=\bigcup _{l \in \overline{\mathcal{E}(\Gamma)}} l .$$
\end{remark}
\begin{proof}
We need only to consider the case when $\Lambda(\gamma)$ has the
form $l \cup\{p\}$, where $l$ is a complex line and $p$ is a point.
By hypothesis, there exists a complex line $l_1 \in
\mathcal{E}(\Gamma)$ such that $l_1 \ne l$. If $l_1$ passes through
$p$ then $p \in E (\Gamma)$, so we can assume that $l_1$ does not
pass through $p$ and it follows that one of the sequences of lines
$\gamma ^n (l_1)$ or $\gamma ^{-n}(l_1)$ goes to a complex line in
$\overline{\mathcal{E}(\Gamma)}$ passing through $p$, as $n \to
\infty$. Therefore $p \in E(\Gamma)$ and $C(\Gamma) \subset
E(\Gamma)$.
\end{proof}
\begin{lemma}\label{Esperfecto}
If $\Gamma \subset PSL(3, \mathbb{C})$ is a discrete group and
$\mathcal{E}(\Gamma)$ contains at least four complex lines in
general position, then $\overline{\mathcal{E}(\Gamma)} \subset
(\mathbb P_{\mathbb C}^2)^*$ is a perfect set. Hence, $C(\Gamma)$ is a non-numerable
union of complex lines.
\end{lemma}
\begin{proof}
It suffices to prove that that each complex line in
$\overline{\mathcal{E}(\Gamma)}$ is an accumulation line of lines
lying in $\mathcal{E}(\Gamma)$. Furthermore, it is sufficient to
prove that each complex line in $\mathcal{E}(\Gamma)$ is an
accumulation line of lines lying in $\mathcal{E}(\Gamma)$. Let $l_1$
be a complex line in $\mathcal{E}(\Gamma)$, then there exists
$\gamma \in \Gamma$ such that $l_1 \subset \Lambda(\gamma)$. By the
Lemma \ref{lineatractora} and given that $\mathcal{E}(\Gamma)$
contains at least four lines in general position, we have that there
exists a line $l$ in $\mathcal{E}(\Gamma)$ such that some of the
sequences of distinct lines in $\mathcal{E}(\Gamma)$, $(\gamma
^{n}(l)) _{n \in \mathbb N }$ or $(\gamma ^{-n}(l)) _{n \in \mathbb
N }$ goes to $l_1$ as $n \to \infty$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{main3}]
a) By Lemma \ref{3lineslema}, the set $\Lambda(\Gamma) \cap
C(\Gamma)$ contains at least three complex lines in general
position, then by Theorem \ref{eqeqdis} we have that
$Eq(\Gamma)=\Omega(\Gamma)$.
If $U \subset \mathbb P_{\mathbb C}^2$ is a $\Gamma$-invariant open set on which
$\Gamma$ acts properly and discontinuously, then, by Theorem
\ref{polloteo}, there exists a complex line $l$ contained in $\mathbb P_{\mathbb C}^2
\setminus U$. Given that $\Gamma$ does not have invariant lines nor
fixed points, then $\mathbb P_{\mathbb C}^2 \setminus U$ contains at least three
complex lines in general
position.
Hence, by Theorem \ref{dissubeq}, $U \subset
Eq(\Gamma)=\Omega(\Gamma)$. Therefore $\Omega(\Gamma)$ is the
maximal open set on which $\Gamma$ acts properly and
discontinuously. The proof that every connected component of
$\Omega(\Gamma)$ is complete Kobayashi hyperbolic is obtained
imitating the proof of Lemma 2.3 in \cite{BN}, and of the main
Theorem in \cite{BN}.
(b) The equality $\Lambda(\Gamma)= C (\Gamma)$ follows from the
Lemma \ref{3lineslema} and the Theorem \ref{main2}. Conexity follows
from the facts that two distinct complex lines always intersect and
a complex line is path-connected.
(c) The Lemma \ref{Esperfecto} implies that
$\overline{\mathcal{E}(\Gamma)}$ is a perfect set. Now, if
$\mathcal{D} \subset (\mathbb P_{\mathbb C}^2)^*$ is a non-empty, closed
$\Gamma$-invariant subset, then there is a complex line $l \in
\mathcal{D}$ and the set $\{ \gamma (l) \, | \, \gamma \in \Gamma
\}$ contains at least three complex lines in general position. It
cannot happen that $\{ \gamma (l) \, | \, \gamma \in \Gamma \}$
contains only three complex lines because in such case
$\Lambda(\Gamma)$ would consist of only three complex lines, a
contradiction. Therefore, there are more than three complex lines in
the set $\{ \gamma (l) \, | \, \gamma \in \Gamma \}$ and imitating
the proof of the Lemma \ref{4lineslema}, we obtain that $\{ \gamma
(l) \, | \, \gamma \in \Gamma \}$ contains four lines in general
position. Therefore, $\mathcal{D}$ contains four lines in general
position. Let $\gamma$ be an element in $\Gamma$, applying the lemma
\ref{lineatractora} we deduce that every complex line contained in
$\Lambda(\gamma)$ is contained in $\mathcal{D}$, then
$\overline{\mathcal{E}(\Gamma)} \subset \mathcal{D}$.
\end{proof}
We would like to thank Professor Jos\'e Seade for enlightening
discussions during the process of this work and the kind hospitality
received by the three authors at IMATE-UNAM -Cuernavaca, and the
Facultad de Matem\'aticas, UADY. Also, during this time, the second
author was in a postdoctoral year at IMPA, and he is grateful to
this institution and its people, for their support and hospitality.
|
\section{Introduction}
In recent years, the study of long range correlations in experimental data has
moved into the focus of interest. Whereas a large amount of empirical results
has been obtained using detrended fluctuation analysis \cite{PengHavlinStanleyEtAl95},
the same information is, in principle, also contained in the power spectrum.
Observations of nontrivial power law decays of the power spectrum have a long
tradition and are usually denoted by $1/f$-noise, or, more precisely,
$1/f^\beta$ with $\beta\approx 1$. Examples include noise in electric
resistors, in semiconductor devices (flicker noise),
and fluctuations in geophysical data
\cite{Weissman88,HoogeKleinpenningVandamme81,BalascoLapennaTelesca02}.
By weak ergodicity breaking one describes the behavior of a system
which is ergodic in the sense that a single trajectory is able to explore
completely some invariant component of the phase-space (it is indecomposable),
but where the invariant measure is not normalizable, so that the time scales
for this exploration might diverge (in other words, there is no finite
microscopic time scale). This concept goes back to Bouchaud \cite{Bouchaud92}.
A consequence of this is that
time averages computed on a single (infinite) trajectory will not converge to
the ensemble mean. Instead, the time average is a random variable itself, as a
function of the initial condition of the trajectory, which is drawn from a
distribution with non-zero variance. It has been shown that
Continuous-Time Random Walks \cite{BelBarkai05}, but also specific deterministic dynamical
systems \cite{BelBarkai06a} exhibit this behavior. A natural consequence of it
is the non-existence of a finite correlation time, reflected by a power law
decay of the auto-correlation function, which translates itself into
a $1/f^\beta$-behavior of the power spectrum. This type of spectrum
has been derived for a two state process of this type by Margolin and Barkai \cite[Eq.~(33)]{MargolinBarkai06}.
As said, the crucial aspect of systems with weak ergodicity breaking is that
a time average in the limit of an infinite trajectory does not
converge to a sharp value but is a random variable with a distribution of
finite width. This should also hold for the power spectrum. Therefore,
it is not evident that an estimation of $\beta$ from a numerically computed
power spectrum of a single trajectory is reliable. We therefore investigate
the estimation of the power spectrum from single realizations of a
specific process with weak ergodicity breaking.
A relevant remark is necessary: If one estimates the power spectrum by
a discrete Fourier transform from a time series with sampling interval
$\Delta t$ (i.e., uses the periodogram), then the variance of
the estimator $f_k$ for the power contained in the $k$th discrete frequency is
$f_k^2$. However, if the correlations in the time domain decay exponentially
fast, then the errors in adjacent frequencies $f_k$ and $f_{k\pm 1}$ are
sufficiently weakly correlated such that the error of a {\sl binned}
spectrum where one takes averages over $M$ adjacent frequencies decays
like $1/\sqrt{M}$ \cite[Sect.13.4]{PressTeukolskyVetterlingEtAl86}.
In the limit of an infinite time series, the
estimation error of the binned power spectrum of an ergodic process therefore
decays to zero, i.e., the spectrum assumes uniquely defined values.
In the following, we will
discuss results for binned power spectra of processes with weak ergodicity
breaking. The binning implies that
in the analytical calculations we are allowed to
ignore the time discretization which is present in every numerical data
analysis.
The next Sections are devoted to an analytical derivation of the
fact that the binned power spectrum of a weakly non-ergodic process
is a random variable whose distribution has a non-zero variance.
In addition, we compute the correlations between different bins.
We will show that because of these correlations, further binning cannot
reduce the uncertainty about the true power, and moreover, that
the whole uncertainty about the power spectrum reduces to
a random normalization factor. This has the important consequence that
the power law exponent $\beta$ can indeed be estimated numerically regardless
of the difficulties to estimate the spectrum itself.
In Section \ref{sectionNumericalSimulations} we illustrate the analytical (asymptotic) results
by numerical simulations, also opposing time averages to ensemble averages.
Note that because of the long range correlations, cutting
a long trajectory into pieces and treat these as an ensemble is
not valid, since this ensemble would not represent an independent sample of
initial conditions.
\section{The model and basic approach}
Rebenshtok and Barkai introduced the following model for a thermodynamic
system showing weak ergodicity breaking \cite{RebenshtokBarkai07,RebenshtokBarkai08}:
In its simplest form it is characterized by two distributions, one waiting time
distribution with density $\phi(t)$ and one distribution of an observable
$x \in \mathbb{R}$ according to the probability density $\kappa(x)$.
In the model let $\chi_0, \chi_1, \dotsc$ be i.i.d. random variables
distributed according to $\kappa(x)$ and let $\tau_0,\tau_1,\dotsc$ be i.i.d. waiting times
distributed according to $\phi(t)$. The process $X(t)$ takes the value
$\chi_0$ in the time $0\leq t < \tau_0$ and $\chi_1$ for the next time interval of length $\tau_1$.
In general
\begin{equation}
X(t)= \chi_i \quad \text{ for } T_{i-1} \leq t < T_i
\end{equation}
with
\begin{equation}
T_i = \sum_{j=0}^{i-1} \tau_j.
\end{equation}
We assume that the first four moments of $\kappa(x)$ are
finite, i.e.,
\begin{equation}
\mu_i = \int \dop x \, x^i \kappa(x) < \infty \quad \text{for } i=1,2,3,4.
\label{definitionMoments}
\end{equation}
Moreover, we assume that the distribution is centered, i.e., $\mu_1 = 0$.
The waiting time distribution $\phi(t)$ should be in the domain of normal attraction
of an one-sided L\'{e}vy stable distribution with exponent $\alpha$ ($0 < \alpha < 1$).
Therefore, the Laplace transform $\hat{\phi}(\lambda)$ of $\phi(t)$,
\begin{equation}
\hat{\phi}(\lambda) = \int \dop t \, e^{-\lambda t} \phi(t),
\end{equation}
can be expanded as
\begin{equation}
\hat{\phi}(\lambda) = 1 - ( \timeconstant \lambda )^\alpha + o(\lambda^\alpha) \quad \text{as } \lambda \to 0+
\label{expansionLaplaceWaitingTime}
\end{equation}
with the scaling parameter $\timeconstant>0$.
The Fourier transform of the time series $X(t)$ up to time $T$ is given by
\begin{equation}
F_T(\omega) = \int_0^T \dop t \, e^{i\omega t} X(t).
\end{equation}
The spectrum may be estimated from
\begin{equation}
S_T(\omega) = \frac{1}{n(T)} F_T(\omega) F_T(-\omega).
\label{spectralEstimateSimple}
\end{equation}
The function $n(T)$ denotes a normalization. Normally, it takes the form
$n(T)=T$, but this model requires to take $n(T)= T^\alpha$ for a non trivial
spectrum. For finite time series one would have to add correction terms for
the fact that the basis for Eq.~\eqref{spectralEstimateSimple} is a biased
estimate of the correlation function (e.g., see \cite[chapter 8.1]{KammeyerKroschel89}).
These terms vanish for the asymptotic case $T \to \infty$ which
we are considering here. If we have an ensemble of realizations, one can look
at the unbinned spectrum
\begin{equation}
S_{\text{ub}}(\omega) = \lim_{T \to \infty} S_T(\omega).
\end{equation}
Denoting the ensemble average by $\langle \cdot \rangle$, we are interested
in the expectation value of the spectrum
\begin{equation}
\langle S_{\text{ub}}(\omega) \rangle = \lim_{T \to \infty} \frac{1}{T^\alpha} \langle F_T(\omega) F_T(-\omega) \rangle
\label{unbinnedExpectation}
\end{equation}
and the covariances
\begin{equation}
\begin{split}
\langle &S_{\text{ub}}(\omega_1) S_{\text{ub}}(\omega_2) \rangle \\
&= \lim_{T \to \infty} \frac{1}{T^{2\alpha}} \langle F_T(\omega_1) F_T(-\omega_1) F_T(\omega_2) F_T(-\omega_2) \rangle.
\end{split}
\label{unbinnedCovariance}
\end{equation}
The limits Eqs.~\eqref{unbinnedExpectation} and \eqref{unbinnedCovariance} can
be rewritten to
\begin{equation}
\begin{split}
\langle &S_{\text{ub}}(\omega) \rangle
= \biggl. \lim_{r \to \infty} \frac{1}{r^\alpha} \langle F_{rT_1}(\omega) F_{rT_2}(-\omega) \rangle \biggr|_{T_1=T_2=1}, \\
\langle &S_{\text{ub}}(\omega_1) S_{\text{ub}}(\omega_2) \rangle \\
&= \biggl. \lim_{r \to \infty} \frac{1}{r^{2\alpha}} \begin{aligned}[t] \langle &F_{rT_1}(\omega_1) F_{rT_2}(-\omega_1)
F_{rT_3}(\omega_2) F_{rT_4}(-\omega_2) \rangle
\biggr|_{\substack{T_1=T_2=1 \\ T_3=T_4=1}}. \end{aligned}
\end{split}
\label{unbinnedLimitMulti}
\end{equation}
In the following, it is helpful to define the two- and four point correlations
\begin{equation}
\begin{split}
&C_2(t_1,t_2) = \langle X(t_1) X(t_2) \rangle, \\
&C_4(t_1,t_2,t_3,t_4) = \langle X(t_1) X(t_2) X(t_3) X(t_4) \rangle
\end{split}
\end{equation}
and their (double resp. quadruple) Laplace transforms
\begin{equation}
\begin{split}
\hat{C}_2(\lambda_1,\lambda_2) &= \Laplacetransform \left[ C_2(t_1,t_2) \right] \\
&= \int \dop t_1 \, \int \dop t_2 \, e^{-\lambda_1 t_1 -\lambda_2 t_2} C_2(t_1,t_2),
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\hat{C}_4(\lambda_1,\lambda_2,\lambda_3,\lambda_4)
&= \Laplacetransform \left[ C_4(t_1,t_2,t_3,t_4) \right] \\
&= \int \dop^4 t \, e^{-\bl \mathbf{t}} C_4(\mathbf{t}).
\end{split}
\end{equation}
The double resp. quadruple Laplace transforms of the expressions in Eq.~\eqref{unbinnedLimitMulti}
are
\begin{equation}
\begin{split}
\Laplacetransform &\bigl[ \langle F_{T_1}(\omega) F_{T_2}(-\omega) \rangle \bigr]
= \frac{1}{\lambda_1 \lambda_2} \hat{C}_2(\lambda_1 - i \omega, \lambda_2 + i\omega), \\
\Laplacetransform &\bigl[ \langle F_{T_1}(\omega_1) F_{T_2}(-\omega_1) F_{T_3}(\omega_2) F_{T_4}(-\omega_2) \rangle \bigr] \\
&= \frac{1}{\lambda_1 \lambda_2 \lambda_3 \lambda_4}
\hat{C}_4(\lambda_1 - i \omega_1, \lambda_2 + i \omega_1, \lambda_3 - i \omega_2, \lambda_4 + i \omega_2).
\end{split}
\end{equation}
The limits in Eq.~\eqref{unbinnedLimitMulti} can also be expressed in
Laplace space using a formulation following the multidimensional Tauberian theorem by Drozhzhinov and Zav'jalov
\cite{DrozhzhinovZavjalov80,Drozhzhinov83} as
\begin{align}
\begin{split}
\Laplacetransform &\Bigl[ \lim_{r \to \infty} \frac{1}{r^\alpha} \langle F_{rT_1}(\omega) F_{rT_2}(-\omega) \rangle \Bigr] \\
&= \lim_{\zeta \to 0+} \frac{\zeta^\alpha}{\lambda_1 \lambda_2}
\hat{C}_2(\zeta \lambda_1 - i \omega, \zeta\lambda_2 + i\omega),
\end{split} \label{limitUnbinnedExpectation} \displaybreak[0] \\
\begin{split}
\Laplacetransform &\Bigl[ \lim_{r \to \infty} \frac{1}{r^{2\alpha}} \langle F_{rT_1}(\omega_1) F_{rT_2}(-\omega_1)
F_{rT_3}(\omega_2) F_{rT_4}(-\omega_2) \rangle \Bigr] \\
&= \lim_{\zeta \to 0+} \frac{\zeta^{2\alpha}}{\lambda_1 \lambda_2 \lambda_3 \lambda_4}
\hat{C}_4(\begin{aligned}[t] \zeta&\lambda_1 - i \omega_1, \zeta\lambda_2 + i \omega_1, \\
&\zeta\lambda_3 - i \omega_2, \zeta\lambda_4 + i \omega_2). \end{aligned}
\end{split} \label{limitUnbinnedCovariance}
\end{align}
In the next section, we will show how one can calculate the Laplace transforms
$\hat{C}_n(\bl)$ of the multi point correlations $C_n(\mathbf{t})$.
However, when only one time series is given,
taking $S_T(\omega)$ from Eq.~\eqref{spectralEstimateSimple} as estimate
of the spectrum will lead to unsatisfactory results as the value of $S_T(\omega)$
will fluctuate except for some pathological cases. One often relies to the concept
of binning. Instead of considering a single frequency $\omega$, one averages all
available frequencies in an given interval
$[\omega - \frac{1}{2} \Delta \omega, \omega + \frac{1}{2} \Delta \omega]$.
For a time series of length $T$ the Fourier transform returns values at
frequencies which are evenly spaced with distance $\frac{2\pi}{T}$. For large $T$ the
sum can be approximated by an integral and
we can consider as observable for the binned spectrum
\begin{equation}
S_T(\omega,\Delta \omega) = \frac{1}{\Delta \omega}
\int_{\omega - \frac{1}{2} \Delta \omega}^{\omega + \frac{1}{2} \Delta \omega}
\dop \omega^\prime \, S_T(\omega^\prime).
\label{spectralEstimateBinning}
\end{equation}
The spectrum can be estimated from
\begin{equation}
S_{\text{bin}}(\omega,\Delta \omega) = \lim_{T\to \infty} S_T(\omega,\Delta \omega).
\end{equation}
Analytically, it is helpful to let the bin size go to zero as a last step
\begin{equation}
S_{\text{bin}}(\omega) = \lim_{\Delta \omega \to 0} S(\omega, \Delta \omega).
\end{equation}
It is important to take the last limit at the very end. In most of the important cases, the
value of $S_{\text{bin}}(\omega)$ will converge almost surely to the
value of the spectrum.
In this paper, we want to consider the situation where the stochastic
process $X(t)$ is the model by Rebenshtok and Barkai. The estimation of
the spectrum is connected with averaging over the time series,
therefore the question arises
how the weak ergodicity breaking affects the observable $S_{\text{bin}}(\omega)$.
Does the value of $S_{\text{bin}}(\omega)$ converge to a single value or does it
converge to a non trivial probability distribution? It will turn out
that the latter is the case. Unlike the observable discussed by
Rebenshtok and Barkai \cite{RebenshtokBarkai07,RebenshtokBarkai08},
it seems not to be possible to determine directly the probability
distribution of $S_{\text{bin}}(\omega)$. But one can obtain important
information from the moments $\langle S_{\text{bin}}(\omega) \rangle$ and
$\langle S_{\text{bin}}(\omega_1) S_{\text{bin}}(\omega_2) \rangle$.
The first step is to look at $\langle S(\omega,\Delta \omega) \rangle$
with $2|\omega| > \Delta \omega$:
\begin{equation}
\langle S_{\text{bin}}(\omega,\Delta \omega) \rangle
= \lim_{T \to \infty}
\frac{T^{-\alpha}}{\Delta \omega} \int_{\omega - \frac{1}{2} \Delta \omega}^{\omega + \frac{1}{2} \Delta \omega}
\dop \omega' \, \langle F_T(\omega') F_T(-\omega') \rangle.
\end{equation}
We proceed similarly to the unbinned case and formulate the limit in
Laplace space (note that the Laplace transform and the $\omega$-integral
commute by Fubini's theorem and the estimate
$|\langle F_{T_1}(\omega) F_{T_2}(-\omega) \rangle | \leq \mu_2 T_1 T_2$
where $\mu_2$ is defined in Eq.~\eqref{definitionMoments})
\begin{equation}
\begin{split}
\Laplacetransform&\left[ \lim_{r \to \infty} r^{-\alpha}
\frac{1}{\Delta \omega} \int_{\omega - \frac{1}{2} \Delta \omega}^{\omega + \frac{1}{2} \Delta \omega} \dop\omega' \,
\langle F_{r T_1}(\omega') F_{rT_2}(-\omega') \rangle \right] \\
&\! = \lim_{\zeta \to 0+} \frac{\zeta^\alpha}{\Delta \omega} \int_{\omega - \frac{1}{2} \Delta \omega}^{\omega + \frac{1}{2} \Delta \omega}
\dop \omega' \,
\frac{1}{\lambda_1 \lambda_2} \hat{C}_2(\zeta \lambda_1 - i \omega', \zeta \lambda_2 + i \omega').
\end{split}
\label{LaplaceTwoPointLimes}
\end{equation}
Similarly, $\langle S(\omega_1,\Delta \omega_1) S(\omega_2,\Delta \omega_2) \rangle$
with $2|\omega_1| > \Delta \omega_1$ and $2|\omega_2| > \Delta \omega_2$ can
be obtained from
\begin{equation}
\begin{split}
\langle &S_{\text{bin}}(\omega_1,\Delta \omega_1) S_{\text{bin}}(\omega_2,\Delta \omega_2) \rangle \\
&= \begin{aligned}[t] \biggl. \lim_{r \to \infty}& r^{-2\alpha}
\frac{1}{\Delta \omega_1 \Delta \omega_2}
\int_{\omega_1 - \frac{1}{2} \Delta \omega_1}^{\omega_1 + \frac{1}{2} \Delta \omega_1} \dop \omega_1' \,
\int_{\omega_2 - \frac{1}{2} \Delta \omega_2}^{\omega_2 + \frac{1}{2} \Delta \omega_2} \dop \omega_2' \\
&\langle F_{r T_1}(\omega_1') F_{rT_2}(-\omega_1') F_{r T_3}(\omega_2') F_{rT_4}(-\omega_2') \rangle
\biggl|_{\substack{T_1=T_2=1 \\ T_3=T_4=1}}. \end{aligned}
\end{split}
\end{equation}
One gets
\begin{equation}
\begin{split}
\Laplacetransform &\biggl[\begin{aligned}[t]
\lim_{r \to \infty} \frac{r^{-2\alpha}}{\Delta \omega_1 \Delta \omega_2}
&\int_{\omega_1 - \frac{1}{2} \Delta \omega_1}^{\omega_1 + \frac{1}{2} \Delta \omega_1} \dop \omega_1' \,
\int_{\omega_2 - \frac{1}{2} \Delta \omega_2}^{\omega_2 + \frac{1}{2} \Delta \omega_2} \dop \omega_2' \\
&\langle F_{r T_1}(\omega_1') F_{rT_2}(-\omega_1') F_{r T_3}(\omega_2') F_{rT_4}(-\omega_2') \rangle
\biggr] \end{aligned} \\
&= \begin{aligned}[t] \lim_{\zeta \to 0+} &\frac{\zeta^{2\alpha}}{\Delta \omega_1 \Delta \omega_2}
\int_{\omega_1 - \frac{1}{2} \Delta \omega_1}^{\omega_1 + \frac{1}{2} \Delta \omega_1} \dop \omega_1' \,
\int_{\omega_2 - \frac{1}{2} \Delta \omega_2}^{\omega_2 + \frac{1}{2} \Delta \omega_2} \dop \omega_2'
\frac{1}{\lambda_1 \lambda_2 \lambda_3 \lambda_4} \\
&\hat{C}_4(\zeta \lambda_1 - i \omega_1',\zeta \lambda_2 + i \omega_1',\zeta \lambda_3 - i \omega_2',\zeta \lambda_4 + i \omega_2').
\end{aligned}
\end{split}
\label{fourPointLimes}
\end{equation}
Therefore the problem can be reduced to the determination of the Laplace transforms $\hat{C}_n(\bl)$
of the $n$ point correlations functions. We describe in the next section how to obtain them.
\section{A diagrammatic approach to the multi point correlation functions}
In \cite{NiemannKantz08} a diagrammatic method was introduced how to determine the joint probability
distributions and multi point correlation functions of a continuous-time random walk.
In this section we are adapting this method to determine $\hat{C}_n(\bl)$.
For any set of natural numbers $q_1,\dotsc,q_n \in \mathbb{N}_0$, one defines
\begin{equation}
C_n[\mathbf{q}](\mathbf{t})
= \left\langle \prod_{k=1}^n \chi_{q_k} 1_{[T_{q_k},T_{q_k+1}[}(t_k) \right\rangle
\end{equation}
with the indicator function ($A$ being any subset $A \subseteq \mathbb{R}$)
\begin{equation}
1_A(t) = \begin{cases} 1 & \text{for } t \in A, \\
0 & \text{for } t \notin A.
\end{cases}
\end{equation}
The $C_n[\mathbf{q}](\mathbf{t})$ can be considered as partial correlations which
only contribute, when the time $t_i$ is in the $(q_i+1)$th waiting time (for $i=1,\dotsc,n$).
Therefore,
\begin{equation}
\begin{split}
C_n(\mathbf{t}) &= \langle X(t_1) \dotsm X(t_n) \rangle \\
&= \sum_{q_1,\dotsc,q_n=0}^\infty C_n[\mathbf{q}](\mathbf{t}).
\end{split}
\end{equation}
Define the contribution of one step as
\begin{equation}
\eta^{(j)}_n[\mathbf{q}](\mathbf{t})
= \prod_{i:q_i=j} \chi_j 1_{[0,\tau_j[}(t_i) \prod_{i:q_i>j} \delta(t_i - \tau_j) \prod_{i:q_i<j} \delta(t_i).
\end{equation}
One obtains by induction
\begin{equation}
C_n[\mathbf{q}](\mathbf{t}) = \left\langle \underset{j}{\bigstar} \eta^{(j)}_n[\mathbf{q}](\mathbf{t}) \right\rangle
\end{equation}
where $\star$ denotes the convolution in the time variable $t_1, \dotsc, t_n$ and $j$ runs
from $0$ to any natural number larger than $\max(q_1,\dotsc,q_n)$.
Transforming into Laplace space gives
\begin{equation}
\begin{split}
\hat{C}_n[\mathbf{q}](\bl) &= \Laplacetransform\left[ C_n[\mathbf{q}](\mathbf{t})\right] \\
&= \prod_j \langle \hat{\eta}^{(j)}_n [\mathbf{q}] (\bl) \rangle.
\end{split}
\end{equation}
Using the following notations from \cite{NiemannKantz08} (suppressing the dependence on $\mathbf{q}$)
\begin{equation}
\singlestepset{V}_j = \{i : q_j = i \}, \quad
\singlestepset{E}_j = \{i:q_i<j\}, \quad
\singlestepset{L}_j = \{i:q_i>j\}.
\end{equation}
The set $\singlestepset{V}_j$ (``vertex'') contains the indices which are
at position $j$, the set $\singlestepset{E}_j$ (``earlier'') contains the indices
which are before that position and the set $\singlestepset{L}_j$ (``later'')
contains the indices which are after that position.
With the notation
\begin{equation}
\Lambda_J = \sum_{j \in J} \lambda_j
\end{equation}
one gets
\begin{equation}
\begin{split}
\langle \hat{\eta}^{(j)}_n [\mathbf{q}] (\bl) \rangle
&= \biggl\langle \int \mathrm{d}^n t \, e^{- \bl \mathbf{t}}
\prod_{i \in \singlestepset{V}_j} \chi_j 1_{[0,\tau_j[}(t_i)
\prod_{i \in \singlestepset{E}_j} \delta(t_i - \tau_j)
\prod_{i \in \singlestepset{L}_j} \delta(t_i)
\biggr\rangle \\
&= \frac{\mu_{|\singlestepset{V}_j|}}{\prod_{v \in \singlestepset{V}_j} \lambda_v}
\sum_{J \in \powerset(\singlestepset{V}_j)} (-1)^{|J|} \hat{\phi}(\Lambda_{J\cup \singlestepset{L}_j}). \\
\end{split}
\end{equation}
Here, the power set of $\singlestepset{V}_j$ is denoted by $\powerset(\singlestepset{V}_j)$ and
$|J|$ is the number of elements of $J$. Therefore, the last sum goes over all
subsets $J$ of indices of $\singlestepset{V}_j$.
\begin{figure}
\includegraphics[width=85mm]{fig1.eps}
\caption{A generic diagram}
\label{genericDiagram}
\end{figure}
As in \cite{NiemannKantz08} we introduce diagrams which denote the
relative ordering of the $q_i$s.
A generic diagram is given in Fig.~\ref{genericDiagram}.
The main idea behind this consists in grouping all steps $i$
with the same $\singlestepset{V}_i$, $\singlestepset{E}_i$ and $\singlestepset{L}_i$.
If $\singlestepset{V}_i=\{\}$ then there can be several successive steps with
the same $\singlestepset{E}_i$ and $\singlestepset{L}_i$. These successive steps
are combined to a horizontal line. If $\singlestepset{V}_i\neq\{\}$ then there
can be only one step with these sets of indices. This is represented by a vertex
with the indices $\singlestepset{V}_i$ leaving. Therefore the diagram Fig.~\ref{genericDiagram}
represents all $\mathbf{q}$ with the following properties: First there is any
number of steps $i$ with $\singlestepset{E}_i = \{\}$, $\singlestepset{V}_i=\{\}$ and
$\singlestepset{L}_i = \diagramset{L}_0$. A single step with $\singlestepset{E}_i=\{\}$,
$\singlestepset{V}_i = \diagramset{V}_1$ and $\singlestepset{L}_i=\diagramset{L}_1$ follows.
At next is any number of steps $i$ with $\singlestepset{E}_i = \diagramset{V}_1$,
$\singlestepset{V}_i=\{\}$ and $\singlestepset{L}_i=\diagramset{L}_1$. The rest in
interpreted analogously. For each $n$ point correlation there is only a finite
number of diagrams and summing over all $\mathbf{q}$ represented by a diagram
turns out to be easy. It turns out that one can associate to each part of
the diagram a factor (for a more comprehensive derivation of these results,
we refer to \cite{NiemannKantz08}).
One obtains the following rules (where $\mu_i$ is defined in Eq.~\eqref{definitionMoments})
\begin{equation}
\begin{split}
&\gamma_{\text{vertex } i}=\frac{\mu_{|\diagramset{V}_i|}}{\prod_{v \in \diagramset{V}_i \lambda_v}}
\sum_{J \in \powerset(\diagramset{V}_i)} (-1)^{|J|} \hat{\phi}(\Lambda_{J\cup \diagramset{L}_i}), \\
&\gamma_{\text{line } i}= \sum_{j=0}^\infty \hat{\phi}^j (\Lambda_{\diagramset{L}_i})
=\frac{1}{1 - \hat{\phi}(\Lambda_{\diagramset{L}_i})}.
\end{split}
\end{equation}
Since we assumed that $\mu_1 = 0$, we do not need to consider diagrams which
contain vertices with only one leaving line. Therefore a single diagram
remains for $\hat{C}_2(\lambda_1,\lambda_2)$ which drawn in Fig.~\ref{diagramsCorrelations}a. This
gives
\begin{equation}
\begin{split}
\hat{C}_2(\lambda_1,\lambda_2) &= \gamma_{\text{line}, \diagramset{L}=\{1,2\}} \gamma_{\text{vertex}, \diagramset{V}=\{1,2\},\diagramset{L}=\{\}} \\
&= \frac{\mu_2}{\lambda_1 \lambda_2} \frac{1 - \hat{\phi}(\lambda_1) - \hat{\phi}(\lambda_2) + \hat{\phi}(\lambda_1+\lambda_2)}{1-\hat{\phi}(\lambda_1+\lambda_2)}.
\end{split}
\label{resultTwoPointCorrelation}
\end{equation}
\begin{figure}
\includegraphics[width=85mm]{fig2.eps}
\caption{The types of diagrams relevant for a) the two point correlation and b) four point correlation}
\label{diagramsCorrelations}
\end{figure}
For the four point correlation two types of diagrams are relevant, which are
drawn in Fig.~\ref{diagramsCorrelations}b. The contributions are
\begin{equation}
\begin{split}
&D_{\text{I}}(\bl) \begin{aligned}[t]
&= \gamma_{\text{line}, \diagramset{L}=\{1,2,3,4\}} \gamma_{\text{vertex}, \diagramset{V}=\{3,4\},\diagramset{L}=\{1,2\}} \\
&\phantom{=} \times \gamma_{\text{line}, \diagramset{L}=\{1,2\}} \gamma_{\text{vertex}, \diagramset{V}=\{1,2\},\diagramset{L}=\{\}}
\end{aligned} \\
&= \frac{\mu_2^2}{\lambda_1 \lambda_2 \lambda_3 \lambda_4}
\frac{\hat{\phi}(\Lambda_{\{1,2\}}) - \hat{\phi}(\Lambda_{\{1,2,3\}}) - \hat{\phi}(\Lambda_{\{1,2,4\}}) + \hat{\phi}(\Lambda_{\{1,2,3,4\}})}
{1 - \hat{\phi}(\Lambda_{\{1,2,3,4\}})} \\
&\phantom{=} \times \frac{1 - \hat{\phi}(\lambda_1) - \hat{\phi}(\lambda_2) + \hat{\phi}(\lambda_1+\lambda_2)}
{1 - \hat{\phi}(\lambda_1+\lambda_2)}
\end{split} \label{resultD1}
\end{equation}
\begin{equation}
\begin{split}
D_{\text{II}}(\bl)
=& \gamma_{\text{line}, \diagramset{L}=\{1,2,3,4\}} \gamma_{\text{vertex}, \diagramset{V}=\{1,2,3,4\},\diagramset{L}=\{\}} \\
=& \frac{\mu_4}{\lambda_1 \lambda_2 \lambda_3 \lambda_4} \frac{\sum_{\diagramset{J} \in \diagramset{P}(\{1,2,3,4\})}
(-1)^{|\diagramset{J}|} \hat{\phi}(\Lambda_\diagramset{J})}
{1 - \hat{\phi}(\Lambda_{\{1,2,3,4\}})}.
\end{split} \label{resultD2}
\end{equation}
The function $\hat{C}_4(\bl)$ is obtained by summing over all possible
combinations of indices at the vertices, i.e.,
\begin{equation}
\begin{split}
\hat{C}_4(\bl)
=& D_{\text{I}}(\lambda_1,\lambda_2,\lambda_3,\lambda_4) + D_{\text{I}}(\lambda_1,\lambda_3,\lambda_2,\lambda_4) \\
&+ D_{\text{I}}(\lambda_1,\lambda_4,\lambda_2,\lambda_3) + D_{\text{I}}(\lambda_2,\lambda_3,\lambda_1,\lambda_4) \\
&+ D_{\text{I}}(\lambda_2,\lambda_4,\lambda_1,\lambda_3) + D_{\text{I}}(\lambda_3,\lambda_4,\lambda_1,\lambda_2) \\
&+D_{\text{II}}(\lambda_1,\lambda_2,\lambda_3,\lambda_4).
\end{split}
\label{resultFourPointCorrelation}
\end{equation}
\section{Evaluating the spectral observables}
\subsection{The unbinned observables}
We can use the results of the last section to determine the spectral
observables. Using the expansion Eq.~\eqref{expansionLaplaceWaitingTime}
and the two point correlations Eq.~\eqref{resultTwoPointCorrelation}
to calculate the limit Eq.~\eqref{limitUnbinnedExpectation}
\begin{equation}
\begin{split}
&\lim_{\zeta \to 0+} \frac{\zeta^\alpha}{\lambda_1 \lambda_2}
\hat{C}_2(\zeta \lambda_1 - i \omega, \zeta\lambda_2 + i\omega) \\
&\quad = \frac{1}{\lambda_1\lambda_2 (\lambda_1 + \lambda_2)^\alpha}
\frac{\mu_2}{\timeconstant^\alpha} \frac{2 - \hat{\phi}(i\omega) - \hat{\phi}(-i\omega)}{\omega^2}
\end{split}
\end{equation}
The Laplace transform can be inverted by using
\begin{equation}
\Laplacetransform \left[ \frac{\min(T_1,T_2)^\alpha}{\Gamma(1+\alpha)} \right] = \frac{1}{\lambda_1 \lambda_2(\lambda_1+\lambda_2)^\alpha}.
\end{equation}
Therefore, one has
\begin{equation}
\langle S_{\text{ub}}(\omega) \rangle
= \frac{\mu_2}{\timeconstant^\alpha \Gamma(1+\alpha)} \frac{2 - \hat{\phi}(i\omega) - \hat{\phi}(-i\omega)}{\omega^2}
\label{expectationUnbinnedSpectrum}
\end{equation}
with the behavior near $\omega \to 0+$
\begin{equation}
\langle S_{\text{ub}}(\omega) \rangle
= \frac{2 \mu_2 \cos\left( \frac{\pi}{2} \alpha \right)}{\timeconstant^\alpha \Gamma(1+\alpha)}
\frac{1}{|\omega|^{2-\alpha}} + o \left( \frac{1}{|\omega|^{2-\alpha}} \right).
\label{expectationSpectrumOrigin}
\end{equation}
This last equation shows that the spectrum has a $1/f^{2-\alpha}$ form near the origin.
Following \cite[Lemma XV.1.3]{Feller71}, there exists an $\Omega \geq 0$
such that
\begin{equation}
\hat{\phi}(i\omega) = 1 \quad \Leftrightarrow \quad \omega \in \Omega \mathbb{Z}.
\label{explanationOmega}
\end{equation}
The case $\Omega \neq 0$ appears if and only if the waiting times $\tau_i$ can take
only the values $0, \frac{2\pi}{\Omega}, \frac{4\pi}{\Omega}, \dotsc$. Therefore,
more generally,
\begin{equation}
\hat{\phi}(z) = \hat{\phi}(z + i\omega) \quad \text{for }
\operatorname{Re}(z) \geq 0 \text{ and } \omega \in \Omega \mathbb{Z}.
\end{equation}
In this paper we concentrate on the spectrum for frequencies $\omega \notin \Omega \mathbb{Z}$.
The frequencies $\omega \in \Omega \mathbb{Z}$ behave differently, e.g., the behavior of
$S_{\mathrm{ub}}(0)$ is completely described by the result of Rebenshtok and Barkai \cite{RebenshtokBarkai07, RebenshtokBarkai08}.
Now, we can proceed to calculate the limit Eq.~\eqref{limitUnbinnedCovariance} by looking at the
terms Eqs.~\eqref{resultD1} and \eqref{resultD2} with the parameters given in Eq.~\eqref{resultFourPointCorrelation}.
We have
\begin{equation}
\begin{split}
&\lim_{\zeta \to 0+} \zeta^{2\alpha} \begin{aligned}[t]
\Bigl( &
D_{\text{I}}(\zeta \lambda_1 - i \omega_1,\zeta \lambda_2 + i \omega_1,\zeta \lambda_3 - i \omega_2,\zeta \lambda_4 + i \omega_2) \\
&+ D_{\text{I}}(\zeta \lambda_3 - i \omega_2,\zeta \lambda_4 + i \omega_2,\zeta \lambda_1 - i \omega_1,\zeta \lambda_2 + i \omega_1)
\Bigr) \end{aligned} \\
&= \Gamma(1+\alpha)^2
\frac{1}{\Lambda_{\{1,2,3,4\}}^\alpha} \left( \frac{1}{\Lambda_{\{1,2\}}^\alpha} + \frac{1}{\Lambda_{\{3,4\}}^\alpha} \right)
\langle S_{\text{ub}}(\omega_1) \rangle \langle S_{\text{ub}}(\omega_2) \rangle.
\end{split}
\end{equation}
Using the fact $\hat{\phi}(i\omega_1) 1_{\Omega \mathbb{Z}}(\omega_1 - \omega_2) = \hat{\phi}(i\omega_2) 1_{\Omega \mathbb{Z}}(\omega_1 - \omega_2)$,
yields similarly
\begin{equation}
\begin{split}
&\lim_{\zeta \to 0+} \zeta^{2\alpha} \begin{aligned}[t]
\Bigl( &
D_{\text{I}}(\zeta \lambda_1 - i \omega_1,\zeta \lambda_4 + i \omega_2,\zeta \lambda_2 + i \omega_1,\zeta \lambda_3 - i \omega_2) \\
&+ D_{\text{I}}(\zeta \lambda_2 + i \omega_1,\zeta \lambda_3 - i \omega_2, \zeta \lambda_1 - i \omega_1,\zeta \lambda_4 + i \omega_2)
\Bigr) \end{aligned} \\
&= \Gamma(1+\alpha)^2
\frac{1}{\Lambda_{\{1,2,3,4\}}^\alpha} \left( \frac{1}{\Lambda_{\{2,3\}}^\alpha} + \frac{1}{\Lambda_{\{1,4\}}^\alpha} \right)
\langle S_{\text{ub}}(\omega_1) \rangle \langle S_{\text{ub}}(\omega_2) \rangle \\
&\phantom{=}\times 1_{\Omega \mathbb{Z}}(\omega_1 - \omega_2)
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
&\lim_{\zeta \to 0+} \zeta^{2\alpha} \begin{aligned}[t]
\Bigl( &
D_{\text{I}}(\zeta \lambda_1 - i \omega_1,\zeta \lambda_3 - i \omega_2,\zeta \lambda_2 + i \omega_1,\zeta \lambda_4 + i \omega_2) \\
&+ D_{\text{I}}(\zeta \lambda_2 + i \omega_1,\zeta \lambda_4 + i \omega_2, \zeta \lambda_1 - i \omega_1,\zeta \lambda_3 - i \omega_2)
\Bigr) \end{aligned} \\
&= \Gamma(1+\alpha)^2
\frac{1}{\Lambda_{\{1,2,3,4\}}^\alpha} \left( \frac{1}{\Lambda_{\{1,3\}}^\alpha} + \frac{1}{\Lambda_{\{2,4\}}^\alpha} \right)
\langle S_{\text{ub}}(\omega_1) \rangle \langle S_{\text{ub}}(\omega_2) \rangle \\
&\phantom{=}\times 1_{\Omega \mathbb{Z}}(\omega_1 + \omega_2).
\end{split}
\end{equation}
Finally
\begin{equation}
\lim_{\zeta \to 0+} \zeta^{2\alpha}
D_{\text{II}}(\zeta \lambda_1 - i \omega_1,\zeta \lambda_4 + i \omega_2,\zeta \lambda_2 + i \omega_1,\zeta \lambda_3 - i \omega_2)
=0.
\end{equation}
In \ref{appendixLaplace}, we derive the following Laplace transform
\begin{equation}
\begin{split}
\Laplacetransform &\begin{aligned}[t] \biggl[ \min(T_1,&T_2,T_3,T_4)^\alpha \min(T_1,T_2)^\alpha \\
&\times F(\alpha,-\alpha; 1+\alpha; \frac{\min(T_1,T_2,T_3,T_4)}{\min(T_1,T_2)} ) \biggr] \end{aligned} \\
&= \frac{\Gamma(1+\alpha)^2}{\lambda_1\lambda_2\lambda_3\lambda_4} \frac{1}{\Lambda_{\{1,2,3,4\}}^\alpha} \frac{1}{\Lambda_{\{1,2\}}^\alpha}
\end{split}
\label{LaplaceTransformFourTimes}
\end{equation}
with the hypergeometric function $F(\alpha,-\alpha;1+\alpha;x)$.
As
\begin{equation}
F(\alpha,-\alpha;1+\alpha;1) = \frac{\Gamma(1+\alpha)^2}{\Gamma(1+2\alpha)}
\end{equation}
we obtain for the limit Eq.~\eqref{unbinnedLimitMulti}
\begin{equation}
\begin{split}
\langle S_{\text{ub}}(\omega_1) S_{\text{ub}}(\omega_2) \rangle
=& 2 \frac{\Gamma(1+\alpha)^2}{\Gamma(1+2\alpha)} \langle S_{\text{ub}}(\omega_1) \rangle \langle S_{\text{ub}}(\omega_2) \rangle \\
&\times \left( 1 + 1_{\Omega \mathbb{Z}}(\omega_1 - \omega_2) + 1_{\Omega \mathbb{Z}}(\omega_1 + \omega_2) \right).
\end{split}
\label{unbinnedCovarianceGeneral}
\end{equation}
Eq.~\eqref{unbinnedCovarianceGeneral} will play an important role in
determining the properties of the binned case.
In particular, the last equation
determines the variance of $S_{\text{ub}}(\omega)$ (for $2 \omega \notin \Omega \mathbb{Z}$)
\begin{equation}
\operatorname{Var}[S_{\text{ub}}(\omega)]
= V_{\text{ub}}(\alpha) \langle S_{\text{ub}}(\omega) \rangle^2
\label{resultUnbinnedVariance}
\end{equation}
with the function
\begin{equation}
V_{\text{ub}}(\alpha) = 4 \frac{\Gamma(1+\alpha)^2}{\Gamma(1+2\alpha)} - 1.
\label{definitionVUnbinned}
\end{equation}
As already mentioned in the introduction, even in the ergodic limit $\alpha \to 1$,
the value $V_{\text{ub}}(1) = 1$ is larger than zero and therefore the
observable $S_{\text{ub}}(\omega)$ does fluctuate.
The correlation coefficient between $S_{\text{ub}}(\omega_1)$ and $S_{\text{ub}}(\omega_2)$
(for $2 \omega_1 \notin \Omega \mathbb{Z}$ and $2 \omega_2 \notin \Omega \mathbb{Z}$; here
$\Omega$ denotes the periodicity of $\hat{\phi}(\lambda)$, see Eq.~\eqref{explanationOmega})
\begin{equation}
\rho[S_{\text{ub}}(\omega_1), S_{\text{ub}}(\omega_2)]
= \begin{cases}
1 & \text{for } \omega_1 - \omega_2 \in \Omega \mathbb{Z} \\
R(\alpha) & \text{for } \omega_1 - \omega_2 \notin \Omega \mathbb{Z}
\end{cases}
\label{casesUnbinnedCorrelation}
\end{equation}
with
\begin{equation}
R(\alpha) = \frac{2\Gamma(1+\alpha)^2 - \Gamma(1+2\alpha)}{4\Gamma(1+\alpha)^2 - \Gamma(1+2\alpha)}.
\label{resultUnbinnedCorrelation}
\end{equation}
In many cases we have $\Omega = 0$, i.e.,
\begin{equation}
\rho[S_{\text{ub}}(\omega_1), S_{\text{ub}}(\omega_2)]
= R(\alpha) \quad \text{for } |\omega_1| \neq |\omega_2|.
\end{equation}
In the ergodic limit, we have $R(1) = 0$. Therefore, the
observations $S_{\text{ub}}(\omega_1)$
and $S_{\text{ub}}(\omega_2)$ of the spectrum at two different
frequencies $\omega_1$ and $\omega_2$ are uncorrelated.
\subsection{The binned observables}
Here, we are going to use the results from the unbinned case to
evaluate the limits Eqs.~\eqref{LaplaceTwoPointLimes} and
\eqref{fourPointLimes}. The limits on the right hand sides of these
equations can be commuted with the frequency integrals describing
the binning. As this statement is not obvious, we are giving an argument.
First, we are considering $\langle S_{\text{bin}}(\omega, \Delta \omega) \rangle$
with $2 |\omega| > \Delta \omega$. Then we have the simple estimate
\begin{equation}
|\hat{C}_2(\zeta \lambda_1 - i \omega, \zeta \lambda_2 + i \omega)|
\leq \frac{1}{\omega^2} \frac{4}{1- \hat{\phi}(\zeta(\lambda_1 + \lambda_2))}.
\end{equation}
The commutativity follows by dominated convergence. Therefore, we get
\begin{equation}
\langle S_{\text{bin}}(\omega, \Delta \omega) \rangle
= \frac{1}{\Delta \omega}
\int_{\omega - \frac{1}{2} \Delta \omega}^{\omega + \frac{1}{2} \Delta \omega}
\dop \omega^\prime \, \langle S_{\text{ub}}(\omega') \rangle
\end{equation}
and accordingly
\begin{equation}
\langle S_{\text{bin}}(\omega) \rangle
= \langle S_{\text{ub}}(\omega) \rangle.
\end{equation}
It is not surprising that the binned estimate yields also the
correct expectation value for an ensemble of realizations.
The argument for $\langle S_{\text{bin}}(\omega_1,\Delta \omega_1)
S_{\text{bin}}(\omega_2,\Delta \omega_2)\rangle$ (with
$2 |\omega_1| > \Delta \omega_1$ and $2 |\omega_2| > \Delta \omega_2$)
needs more work. First notice that we have the following inequality
\begin{equation}
|1 - \hat{\phi}(\lambda + i \omega)| \geq 1 - |\hat{\phi}(\lambda + i\omega)|
\geq 1 - \hat{\phi}(\lambda).
\end{equation}
This gives the following estimates (replace the tilded variables by
the corresponding permutations of the $\lambda_i$s and $\omega_i$s)
\begin{align}
\begin{split}
&\bigl| D_{\text{I}}(\zeta \tilde{\lambda}_1 + i \tilde{\omega}_1, \zeta \tilde{\lambda}_2 + i \tilde{\omega}_2,
\zeta \tilde{\lambda}_3 + i \tilde{\omega}_3, \zeta \tilde{\lambda}_4 + i \tilde{\omega}_4) \bigr| \\
&\quad \leq \frac{1}{| \tilde{\omega}_1 \tilde{\omega}_2 \tilde{\omega}_3 \tilde{\omega}_4 |}
\frac{4}{1 - \hat{\phi}(\zeta \tilde{\Lambda}_{\{1,2,3,4\}})}
\frac{4}{1 - \hat{\phi}(\zeta \tilde{\Lambda}_{\{1,2\}})}
\end{split} \displaybreak[0] \\
\intertext{and}
\begin{split}
&\bigl| D_{\text{II}}(\zeta \lambda_1 - i \omega_1, \zeta \lambda_2 + i \omega_1,
\zeta \lambda_3 - i \omega_2, \zeta \lambda_4 + i \omega_2) \bigr| \\
&\quad \leq \frac{1}{\omega_1^2 \omega_2^2}
\frac{16}{1 - \hat{\phi}(\zeta \Lambda_{\{1,2,3,4\}})}.
\end{split}
\end{align}
By dominated convergence and Eq.~\eqref{unbinnedCovarianceGeneral}, this leads to
\begin{equation}
\begin{split}
&\langle S_{\text{bin}}(\omega_1, \Delta \omega_1) S_{\text{bin}}(\omega_2, \Delta \omega_2) \rangle \\
&= \frac{1}{\Delta \omega_1 \Delta \omega_2} \int_{\omega_1 - \frac{1}{2} \Delta \omega_1}^{\omega_1 + \frac{1}{2} \Delta \omega_1}
\dop \omega_1^\prime \,
\int_{\omega_2 - \frac{1}{2} \Delta \omega_2}^{\omega_2 + \frac{1}{2} \Delta \omega_2}
\dop \omega_2^\prime \, \langle S_{\text{ub}}(\omega_1') S_{\text{ub}}(\omega_2') \rangle \\
&= 2 \frac{\Gamma(1+\alpha)^2}{\Gamma(1+2\alpha)} \frac{1}{\Delta \omega_1 \Delta \omega_2}
\int
\dop \omega_1^\prime \,
\int
\dop \omega_2^\prime \, \langle S_{\text{bin}}(\omega_1') \rangle \langle S_{\text{bin}}(\omega_2') \rangle
\end{split}
\end{equation}
or shorter
\begin{equation}
\langle S_{\text{bin}}(\omega_1) S_{\text{bin}}(\omega_2) \rangle
= 2 \frac{\Gamma(1+\alpha)^2}{\Gamma(1+2\alpha)}
\langle S_{\text{bin}}(\omega_1) \rangle \langle S_{\text{bin}}(\omega_2) \rangle
\label{binnedCovarianceGeneral}
\end{equation}
for all $\omega_1, \omega_2 \neq 0$. The additional factor present
in Eq.~\eqref{unbinnedCovarianceGeneral} for the values
$\omega_1 - \omega_2 \in \Omega \mathbb{Z}$ and $\omega_1 + \omega_2 \in \Omega \mathbb{Z}$
has vanished due to the binning.
The variance of $S_{\text{bin}}(\omega)$ turns out to be
\begin{equation}
\operatorname{Var}\left[ S_{\text{bin}}(\omega) \right]
= V_{\text{bin}}(\alpha) \langle S_{\text{bin}}(\omega) \rangle^2
\label{varianceBinned}
\end{equation}
with the function
\begin{equation}
V_{\text{bin}}(\alpha) = 2 \frac{\Gamma(1+\alpha)^2}{\Gamma(1+2\alpha)} - 1.
\label{definitionV}
\end{equation}
For $0<\alpha<1$, the value of $V_{\text{bin}}(\alpha)$ is positive,
i.e., the variance of $S_{\text{bin}}(\omega)$ does not vanish and
the observable $S_{\text{bin}}(\omega)$ is a proper probability distribution.
$V_{\text{bin}}(\alpha)$ describes the part of the fluctuation of the spectrum
which is connected with the weak ergodicity breaking, i.e., the
part that remains after binning. Correspondingly, for the ergodic
limit $\alpha \to 1$, the function $V_{\text{bin}}(\alpha)$ vanishes
($V(1) = 0$) and the observable $S_{\text{bin}}(\omega)$ does
not fluctuate.
In general, the whole spectrum is estimated. The question arises, how
the weak ergodicity breaking affect different frequencies. Is it possible
that one would estimate a wrong exponent $\beta$ of the $1/f^\beta$ noise?
This can be answered by using Eq.~\eqref{binnedCovarianceGeneral} to
calculate the correlation coefficient for all $0 < \alpha < 1$ and $\omega_1,\omega_2 \neq 0$:
\begin{equation}
\rho\left[ S_{\text{bin}}(\omega_1), S_{\text{bin}}(\omega_2) \right] = 1.
\label{binnedCorrelation}
\end{equation}
This implies that $S_{\text{bin}}(\omega_1)$ and $S_{\text{bin}}(\omega_2)$ are linearly coupled
\begin{equation}
S_{\text{bin}}(\omega_2) = \frac{\langle S_{\text{bin}}(\omega_2) \rangle}{\langle S_{\text{bin}}(\omega_1) \rangle}
S_{\text{bin}}(\omega_1).
\end{equation}
Therefore, as soon as one value of the spectrum is evaluated, the other values
follows from this one. In other words, the weak ergodicity breaking does only
affect a random prefactor which is the same for all frequencies.
\section{Numerical Verification}
\label{sectionNumericalSimulations}
In order to verify the presented analytical results we have
performed numerical realizations of the process described above.
To this purpose we have generated time series according to the
following procedure:
\begin{itemize}
\item Pick random numbers, $\tau_i$, from a totally asymmetric L{\'{e}}vy
$\alpha$-stable distribution with a given $\alpha$. These numbers
represent the waiting times. The random numbers were generated by
means of the procedure introduced in \cite{Weron01} with the scaling
parameter set to one.
\item Pick additional random numbers, $\chi_i$, from a normal
distribution. These numbers are the values of the process during
the respective time intervals.
\item The values $\chi_i$ are written into the entries of an
array, $A_j$, of total length $L$, for $j\in [\Delta t
\sum_{k=1}^{i-1}\tau_k, \Delta t \sum_{k=1}^{i}\tau_k[$. In all
simulations presented here we chose $\Delta t=0.25$, defining the
minimal resolution of the process.
\end{itemize}
The unbinned spectra, $S_\mathrm{ub}(\omega)$, can than be
determined straight forwardly by means of the fast Fourier
transform of the array $\mathbf{A}$. Of course, the spectra
obtained for these processes are discrete. In order to obtain the
binned spectra, $S_{\mathrm{bin}} (\omega,\Delta \omega)$, we
average $S_\mathrm{ub}$ over various consecutive frequencies and
assign the obtained value to the average of these frequencies.
\begin{figure}
\centerline{\includegraphics[width=75mm]{fig3.eps}}
\caption{Power spectra corresponding to a single realization of
the waiting time process are plotted for $\alpha=0.5$ for the
unbinned case as well as for two different choices of the bin
size. In the inset we show spectra corresponding to different
realizations of the same process for a fixed choice of the binning
for $\alpha=0.5$. Note that the curves exhibit a shift with respect to
one another.\label{fig3}}
\end{figure}
In Fig.~\ref{fig3} we plot the unbinned spectrum of a typical
realization of a process of length $L=2^{21}$ for $\alpha=0.5$
together with the binned spectra for two choices of the binning,
$\Delta \omega=2^{-16} \pi$, and $\Delta \omega= 2^{-10}\pi$,
respectively. It should be observed how the spectra get smoother
when increasing the bin size, while the average slope of the
spectra is not changed by the binning. Deviations from the
power law behavior at small and large frequencies stem from the
finite size of array $\mathbf{A}$ and the minimal resolution of
the process, respectively.
The inset of Fig.~\ref{fig3} shows the binned spectrum with
$\Delta \omega= 2^{-10}\pi$ for various realizations of the
processes with length $L=2^{21}$ for $\alpha=0.5$. Note that,
while the exponent characterizing the decay of the spectrum does
not depend on the realization, the normalization factor of the
spectrum does. We would like to emphasize that the normalization
factor will neither converge by increasing the length of the
process nor by choosing larger bins. This probabilistic property
of the normalization factor is a consequence of the weak
ergodicity breaking.
\begin{figure}
\centerline{\includegraphics[width=75mm]{fig4.eps}} \caption{The
averaged unbinned spectra corresponding to various choices of
$\alpha$ are plotted. Note that the spectra show an algebraic
decay over various decades of $\omega$ with an exponent $\beta$
depending on $\alpha$. The curves have been shifted for better
visibility. In the inset we compare the theoretical prediction for
the exponent $\beta$ (solid line) to the numerical findings
(dots).\label{fig4}}
\end{figure}
To verify our predictions with respect to the dependence of the
spectrum on $\alpha$, we have considered the unbinned spectrum
averaged over 10000 realizations of the process, $\langle
S_\mathrm{ub}(\omega)\rangle$. In Fig.~\ref{fig4} we plot $\langle
S_\mathrm{ub}(\omega)\rangle$ for processes of length $L=2^{21}$
and $\alpha=0.1,0.4,0.7,0.9$ (bottom to top), respectively. The
curves have been shifted with respect to one another for better
visibility. Note that the spectra show a $\sim 1/\omega^\beta$
behavior over various decades with an an exponent $\beta$ that
decreases with increasing $\alpha$. In the inset of
Fig.~\ref{fig4} we compare our prediction for the value of the
exponent $\beta(\alpha)=2-\alpha$ (solid line), see
Eq.~\eqref{expectationSpectrumOrigin}, to the values obtained
numerically by means of measuring the slope of the spectra in
log-log scale at the intermediate region. A fairly good agreement
is observed.
\begin{figure}
\centerline{\includegraphics[width=75mm]{fig5.eps}} \caption{The
theoretical predictions for $V(\alpha)$ are plotted for the
unbinned (solid line) and binned (dashed line) case. Numerical
results for the unbinned and the binned case are represented by
the square and circular dots, respectively. In the binned case,
various choices of process length have been considered,
$L=2^{13},2^{15},2^{17},2^{23}$ (top to bottom). The
inset shows the proportionality factor between the spectrum and the
square root of the variance for the unbinned case,
$\upsilon_{\mathrm{ub}}(\omega)=\mathrm{Var}[S_{\mathrm{ub}}(\omega)]/\langle
S_{\mathrm{ub}}(\omega)\rangle^2$, for $\alpha=0.5$. Note that it
is constant over various decades.\label{fig5}}
\end{figure}
Let us now check the predictions with respect to the variance of
the spectra. To calculate the variance of the unbinned spectra we
have considered processes of length $L=2^{21}$ and averaging was
carried out over 10000 realizations of the process. In accordance
with Eq.~\eqref{resultUnbinnedVariance} it is found that the
variance is proportional to the squared expectation value of the
spectrum over various decades with deviations at very high and
very small frequencies only (see inset of Fig.~\ref{fig5}). To
determine the proportionality factor we have averaged this factor
over the intermediate frequency region. In Fig.~\ref{fig5} the
theoretical prediction Eq.~\eqref{definitionVUnbinned} for the
proportionality factor $V_\mathrm{ub}(\alpha)$ (solid line) is
compared to the numerical results (square dots), for various
choices of $\alpha$. We find a reasonable correspondence with
deviations at very small $\alpha$. However, these deviations become
smaller when considering longer process lengths.
To verify the predictions for the asymptotic behavior of binned
spectra we have considered processes of various lengths,
$L=2^{13},2^{15},2^{17},2^{23}$, while the size of
the bins in frequency space was kept fixed at $\Delta \omega=
2^{-8}\pi$. Note that, by augmenting the process length while
keeping the bin size fixed, the number of frequencies of the discrete
spectrum falling into a specific bin increases. Again, the
averaging was carried out over 10000 realizations of each process.
In Fig.~\ref{fig5} the theoretical prediction for $V_{\text{bin}}(\alpha)$ for
the binned case (dashed line), see Eq.~\eqref{definitionV}, is
compared to the numerical data obtained for the binned processes
(circular dots). It should be observed that, as the length of the
process is increased (top to bottom), the values of $V_{\text{bin}}$ depart
from the predictions of the unbinned case (solid line) and
approach the predictions for the binned case (dashed line).
\begin{figure}
\centerline{\includegraphics[width=75mm]{fig6.eps}}
\caption{$\rho[S_\mathrm{ub}(\omega_1),S_\mathrm{ub}(\omega_2)]$
is plotted for $\omega_1=2\pi$ and various choices of $\alpha$.
The curves shift upwards as $\alpha$ is decreased. The peak
appears at $\omega_1=\omega_2$. For very small $\alpha$ the peak
extends over the whole frequency domain.\label{fig6}}
\end{figure}
Next, we have verified our predictions with respect to the
correlation coefficient of the spectra. For the unbinned case we
have again considered processes of length $L=2^{21}$ and averaged
over 10000 realizations. In Fig.~\ref{fig6}, the correlation
coefficient of the unbinned spectrum,
$\rho[S_\mathrm{ub}(\omega_1),S_\mathrm{ub}(\omega_2)]$, is
plotted for the fixed frequency $\omega_1=2\pi$ and various
choices of $\alpha$. Note that the correlation coefficient for not
too small $\alpha$ is indeed flat (apart from the trivial peak at
$\omega_1=\omega_2$) and that its value increases for decreasing
$\alpha$. For very small values of $\alpha$ the peak around
$\omega_1=\omega_2$ gets extended over the whole frequency domain
but this effect can be remedied by considering longer process
lengths.
\begin{figure}
\centerline{\includegraphics[width=75mm]{fig7.eps}} \caption{The
theoretical predictions for $\rho_{|\omega_1|\neq
|\omega_2|}(\alpha)$ are plotted for the unbinned (solid line) and
binned (dashed line) case. Numerical results for the unbinned and
the binned case are represented by the square and circular dots,
respectively. In the binned case, various choices of process
length have been considered,
$L=2^{13},2^{15},2^{17},2^{19},2^{23}$ (bottom to top).
\label{fig7}}
\end{figure}
In Fig.~\ref{fig7} we compare the predicted value for the
correlation coefficient in the unbinned case (solid line),
$\rho_{|\omega_1|\neq |\omega_2|}(\alpha)=R(\alpha)$, see
Eq.~\eqref{resultUnbinnedCorrelation}, with the numerically
obtained data (squared dots). For the numerical data the
correlation coefficient was averaged over $2^{16}$ choices of the
pairs $\omega_1$ and $\omega_2$. Reasonable agreement between
theory and numerical data is found with deviations at small
$\alpha$ values stemming from the broadening of the peak around
$|\omega_1|=|\omega_2|$.
Finally, we repeated the analysis of the correlation coefficient
for the binned spectra. Again, we considered processes of various
lengths, $L=2^{13},2^{15},2^{17},2^{19},2^{23}$, while keeping
the bin size fixed in frequency space with $\Delta \omega=
2^{-8}\pi$. In Fig.~\ref{fig7} the theoretical prediction for the
correlation coefficient, $\rho_{|\omega_1|\neq
|\omega_2|}(\alpha)=1$, see Eq.~\eqref{binnedCorrelation}, is
plotted as dotted line while the numerical data is represented by
the circular dots. As the length of the process is increased, the
numerical values for $\rho_{|\omega_1|\neq |\omega_2|}$ depart
from the theoretical prediction for the unbinned case (solid line)
and approach the predicted line for the binned case (dashed line).
\section{Summary}
In this paper, we have determined the spectral properties of
a model by Rebenshtok and Barkai which shows weak ergodicity
breaking. The analytical results were verified by numerical
simulations. Near the origin, the spectrum shows a typical
$1/f^{2-\alpha}$ behavior with $\alpha$ characterizes the
tail behavior of the waiting time. Using a single time series will result
in fluctuations of the unbinned spectrum which are also seen in the
ergodic case. Therefore, one commonly uses binning to determine
reliable values. While the fluctuations of the binned spectral
observable vanish in the ergodic case, this is not the case
for weak ergodicity breaking. However, the fluctuation does
only affect a common prefactor for the whole spectrum.
Consequently, the measurement of the exponent $\beta$ of the $1/f^\beta$ behavior
is not hindered by the weak ergodicity breaking.
Recently, the emergence of universal fluctuation for processes
with weak ergodicity breaking has been discussed by He et al.~\cite{HeBurovMetzlerEtAl08},
Sokolov et al.~\cite{SokolovHeinsaluHanggiEtAl09} and Esposito et
al.~\cite{EspositoLindenbergSokolov10}. It has been shown that for
several time averaged observables as the average mean square displacement
the fluctuations are described by a Mittag--Leffler distribution.
As can be infered from Eqs.~\eqref{varianceBinned} and \eqref{definitionV},
the first two moments of the binned spectrum are also in agreement with
the assumption of a Mittag--Leffler distribution. However, it remains to
be seen in a future work if the binned spectrum does also belong
to this universality class.
The analytical results for the binned spectrum are valid in the limit
of infinite time series length. In a numerical realization of
this, binning means averaging only over a finite number of discrete frequencies
due to the finite process time.
It is therefore a nontrivial observation that these binned spectra
converge for numerical manageable process lengths
towards the theoretical prediction.
\section*{Acknowledgments}
We want to thank Eli Barkai und the referees for many helpful comments,
especially for pointing out the
connection to the universal fluctuations.
|
\section{introduction}
Since the relativistic and higher-order $\alpha_s$ corrections are
less important for bottomonia than for any other $q\bar q$ systems,
study on bottomonia may offer more direct information about the
hadron configuration and perturbative QCD. The spin-triplet state of
bottomonia $\Upsilon(nS)$ and the P-states $\chi_b(nP)$ were
discovered decades ago, however its partner the singlet state
$\eta_b$ evaded detection for long time, even though much efforts
were made. Many phenomenological researches on $\eta_b$ have been
done by some
groups \cite{Hao:2007rb,Ebert:2002pp,Motyka:1997di,Liao,Recksiegel:2003fm,Gray:2005ur,Eichten:1994gt,Ke:2007ih}.
Different approaches result in different level splitting $\Delta
M=\Upsilon(1S)-\eta_b(1S)$. In Ref. \cite{Recksiegel:2003fm} the
authors used an improved perturbative QCD approach to get $\Delta
M=44$ MeV; using the potential model suggested
in \cite{Buchmuller:1980su} Eichten and Quigg estimated $\Delta M=87$
MeV \cite{Eichten:1994gt}; in Ref. \cite{Motyka:1997di} the authors
selected a non-relativistic Hamiltonian with spin dependent
corrections to study the spectra of heavy quarkonia and got $\Delta
M$=57 MeV; the lattice prediction is $\Delta M$=51 MeV \cite{Liao},
whereas the lattice result calculated in Ref. \cite{Gray:2005ur} was
$\Delta M=64\pm14$MeV. Ebert $et\, al.$ \cite{Ebert:2002pp} directly
studied spectra of heavy quarkonia in the relativistic quark model
and gave $m_{\eta_b}=9.400$ GeV.
The Babar Collaboration \cite{:2008vj} first measured
$\mathcal{B}(\Upsilon(3S)\rightarrow\gamma\eta_b)=(4.8\pm0.5\pm0.6)\times10^{-4}$,
$M(\eta_b)=9388.9^{+3.1}_{-2.3}\pm2.7$ MeV and $\Delta M=
71.4^{+3.1}_{-2.3}\pm2.7$ MeV in 2008 and new data were released in
2009\cite{:2009pz}\footnote{We thank Dr. Michael Roney for telling
us the discovery history of $\eta_b$ as the Babar collaboration
published its 10-sigma discovery of the $\eta_b$ in 2008 whereas in
2009 the CLEO collaboration released a re-analysis of their data and
reported a 4-sigma confirmation of the Babar discovery.}. More
recently the CLEO Collaboration \cite{Bonvicini:2009hs} confirmed
the observation of $\eta_b$ using the database of 6 million
$\Upsilon(3S)$ decays and assuming $\Gamma(\eta_b)\approx$10 MeV,
they obtained
$\mathcal{B}(\Upsilon(3S)\rightarrow\gamma\eta_b)=(7.1\pm1.8\pm1.1)\times10^{-4}$,
$M_{\eta_b}=9391.8\pm6.6\pm2.0$ MeV and the hyperfine splitting
$\Delta M= 68.5\pm6.6\pm2.0$ MeV, whereas using the database with 9
million $\Upsilon(2S)$ decays they obtained
$\mathcal{B}(\Upsilon(2S)\rightarrow\gamma\eta_b)<8.4\times10^{-4}$
at 90\% confidential level It is noted that the data of the two
Collaborations are in accordance on $M_{\eta_b}$, but the central
values of $\mathcal{B}(\Upsilon(3S)\rightarrow\gamma\eta_b)$ are
different. However, if the experimental errors are taken into
account, the difference is still within one standard deviation.
Some theoretical
work \cite{Radford:2009qi,Colangelo:2009pb,Seth:2009ba} is devoted
to account the experimental results.
In Ref. \cite{Ebert:2002pp} the authors studied these radiative
decays and estimated
$\mathcal{B}(\Upsilon(3S)\rightarrow\eta_b+\gamma)=4\times10^{-4}$,
$\mathcal{B}(\Upsilon(2S)\rightarrow\eta_b+\gamma)=1.5\times10^{-4}$
and
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)=1.1\times10^{-4}$
with the mass $m_{\eta_b}$ = $9.400$ GeV. Their results about
$m_{\eta_b}$ and $\mathcal{B}(\Upsilon(3S)\rightarrow\eta_b+\gamma)$
are close to the data. The authors of Ref. \cite{Choi:2007se}
systematically investigated the magnetic dipole transition
$V\rightarrow P\gamma$ in the light-front quark
model (LFQM) \cite{Jaus:1999zv,Cheng:2003sm,Hwang:2006cua,Wei:2009nc}. In
the QCD-motivated effective Hamiltonian there are several free
parameters, i.e., the quark mass and $\beta$ in the wavefunction (the
notation of $\beta$ was given in the aforementioned literatures)
which are fixed by the variational principle, then
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ was calculated
and the central value is $8.4\,({\rm or}\,7.7)\times
10^{-4}$. \footnote{The different values correspond to the deferent
potentials adopted in calculations.}
\begin{center}
\begin{figure}[htb]
\begin{tabular}{c}
\scalebox{0.9}{\includegraphics{DeltaM.eps}}
\end{tabular}
\caption{$\Delta M$ coming from different experimental measurement and
theoretical work.}\label{DM}
\end{figure}
\end{center}
It is also noted that the mass of $m_{\eta_b}=9.657\,({\rm or}
\,9.295)$ GeV presented in Ref.\cite{Choi:2007se} deviates from the
data ($m_{\eta_b}=9391.8\pm6.6\pm2.0$ MeV) and the fitted $\beta$
values are different for singlet and triplet \cite{Cheng:2003sm}.
Here we are going to take an alternative way to fix the values of
$\beta$.
Since experimentally, $m_{\eta_b}$ is determined by
$\mathcal{B}(\Upsilon(nS)\rightarrow\eta_b+\gamma)$ and a study on
the radiative decays can offer us much information about the
characteristics of $\eta_b$, one should carefully investigate the
transition within a more reliable theoretical framework. That is
the aim of the present work, namely we will use the LFQM to evaluate the hadronic matrix element which
is governed by the non-perturbative QCD. The method is proven to
be successful for calculating the transition rates of the
processes where light hadrons exist in the final states.
In this work, we first fix $\beta_n$'s for $\Upsilon(nS)$ in terms
of their decay constants. Then, using the data of
$\mathcal{B}(\Upsilon(3S)\rightarrow\eta_b+\gamma)$ and $m_{\eta_b}$,
we determine $\beta_{\eta_b}$. With the parameters being fixed, we
are able to estimate the rates
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$,
$\mathcal{B}(\Upsilon(2S)\rightarrow\eta_b+\gamma)$,
$\mathcal{B}(\Upsilon(4S)\rightarrow\eta_b+\gamma)$ and
$\mathcal{B}(\Upsilon(5S)\rightarrow\eta_b+\gamma)$. Since
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ is sensitive to
$\Delta M$, the measurement of
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ would be helpful
for accurately determining the mass of $\eta_b$.
Recently the transition rates of
$\Upsilon(5S)\rightarrow\Upsilon(1S,2S)+\pi\pi$ were measured by the
BELLE Collaboration \cite{Belle-1} and it was found that the widths
exceed by more than two orders of magnitude the previously measured
partial widths between lower $\Upsilon$ resonances. The authors
\cite{Chao-1} suggested that the re-scattering processes of
$\Upsilon(5S)\to B^{(*)}\bar B^{(*)}\to
\Upsilon(mS)+\sigma/f_0(980)\to \Upsilon(mS)+\pi\pi$ make
substantial contributions to the observable rate of the dipion
transition of $\Upsilon(5S)$ because its mass exceeds the production
threshold of $B^{(*)}\bar B^{(*)}$, so that the intermediate bosons
$B^{(*)}$ and $\bar B^{(*)}$ are on their mass-shell. They apply the
same mechanism to study the transition of $\Upsilon(5S)\to
\Upsilon(1S)+\eta$ \cite{Chao-2} and find that the re-scattering
processes would enhance its width by almost two orders of magnitude.
There are different interpretations for the anomalous enhancement of
the $\Upsilon(5S)$ decays that the measured resonance
$\Upsilon(10870)$ is a mixture of $b\bar b$ bound state in the $5S$
state with a hybrid $b\bar bg$ or a tetraquark $b\bar bq\bar
q$ \cite{Hou}.
Thus, we would like to further test the mechanism in the radiative
decays of $\Upsilon(5S)$. In the radiative decay of
$\Upsilon(5S)\to\eta_b+\gamma$, the re-scattering processes
$\Upsilon(5S)\to B^{(*)}\bar B^{(*)}\to \Upsilon(mS)+\gamma$ also
exist and one only needs to replace the effective vertex of
$B^{(*)}\bar B^{(*)}\Upsilon(1S)$ and $B^{(*)}\bar B^{(*)}\eta$ by
the electromagnetic vertex $B^{(*)}\bar B^{(*)}\gamma$ and
$B^{(*)}\bar B^{(*)}\eta_b$ respectively in the diagrams given in
Ref. \cite{Chao-1}. Thus one can expect that the corresponding
mechanism should enhance the ratio of $\Upsilon(5S)\to
\eta_b+\gamma$. Our calculations show that the theoretical
estimation on the enhancement factor strong depends on the parameter
$g_{B^{(*)}B^{(*)}\eta_b}$ (see below for more details). The future
measurements on the radiative decays of $\Upsilon(10870)$ can help
to determine if it is the $\Upsilon(5S)$ state as long as the rate
of $\Upsilon(5S)\to \eta_b+\gamma$ is obviously larger than that of
lower resonances of the family, otherwise other mechanisms may be
more favored. Anyhow, the radiative decays would provide a decisive
probe for the re-scattering mechanism.
This paper is organized as follows: after the introduction, in
section II we present our calculations of the form factors for
$V\rightarrow P\gamma$ in the LFQM
and the corresponding numerical results. In the section III we
study the possible re-scattering effects on $\Upsilon(5S)\to \eta_b+
\gamma$. The section IV is devoted to our conclusion and
discussion.
\section{$\Upsilon(nS)\to \eta_b +\gamma$ in the LFQM}
\subsection{Description of $\Upsilon(nS)\to \eta_b +\gamma$ in the LFQM}
The Feynman diagrams describing $\Upsilon(nS)\to \eta_b+ \gamma$
are presented in Fig. \ref{fig:LFQM}. In this work, we calculate
the transition rate of the radiative decays $\Upsilon(nS)\to
\eta_b +\gamma$ in the LFQM.
\begin{center}
\begin{figure}[htb]
\begin{tabular}{c}
\scalebox{0.9}{\includegraphics{t1.eps}}
\end{tabular}
\caption{Feynman diagrams depicting the radiative decay $\Upsilon(nS)\to \eta_b+\gamma$.\label{fig:LFQM}}
\end{figure}
\end{center}
The transition amplitude of $\Upsilon(nS)\to \eta_b+ \gamma$ can be
expressed in terms of the form factor $\mathcal{
F}_{\Upsilon(nS)\to\eta_b}(q^2)$ which is defined as
\cite{Choi:2007se,Hwang:2006cua}
\begin{eqnarray}\label{2S1}
&&\langle
\eta_b(\mathcal{P}')|J_{em}^\mu|\Upsilon(\mathcal{P},h)\rangle\nonumber\\&&=ie\,\varepsilon^{\mu\nu\rho\sigma}\epsilon_\nu(\mathcal{P},h)q_\rho
\mathcal{P}_\sigma\mathcal{ F}_{\Upsilon(nS) \to\eta_b}(q^2),
\end{eqnarray}
where $\mathcal{P}$ and $\mathcal{P}'$ are the four-momenta of
$\Upsilon(nS)$ and $\eta_b$. $q=\mathcal{P}-\mathcal{P}'$ is the
four-momentum of the emitted photon and
$\epsilon_\nu(\mathcal{P},h)$ denotes the polarization vector of
$\Upsilon(nS)$ with helicity $h$. For applying the LFQM, we first
let the photon be virtual, i.e. leave its mass-shell $q^2=0$ into
the un-physical region of $q^2<0$. Then $\mathcal{ F}_{\Upsilon(nS)\to
\eta_b}(q^2)$ can be obtained in the $q^+=0$ frame with $q^2=q^+q^-
- {\bf q}^2_\perp=-{\bf q}^2_\perp<0$. Then we just analytically
extrapolate $\mathcal{ F}_{\Upsilon(nS) \to\eta_b}({\bf q}^2_\perp)$
from the space-like region to the time-like region ($q^2\geq 0$). By
taking the limit $q^2\rightarrow 0$, one obtains $\mathcal{
F}_{\Upsilon(nS)\to \eta_b}( q^2=0)$.
By means of the light front quark model, one can obtain the
expression of form factor $\mathcal{ F}_{\Upsilon(nS)\to \eta_b}(q^2)$
\cite{Choi:2007se}:
\begin{eqnarray}\label{21}
\mathcal{ F}_{\Upsilon(nS)\to\eta_b}(q^2)= e_bI(m_1,m_2,q^2) + e_{b}
I(m_2,m_1,q^2),
\end{eqnarray}
where $e_{b}$ is the electrical charge for bottom quark ,
$m_1=m_2=m_b$ and
\begin{eqnarray}\label{22}
I(m_1,m_2,q^2) &=&\int^1_0 \frac{dx}{8\pi^3}\int d^2{\bf k}_\perp
\frac{\phi(x, {\bf k'}_\perp)\phi(x,{\bf k}_\perp)}
{x_1\tilde{M_0}\tilde{M'_0}}
\nonumber\\
&&\times \biggl\{{\cal A} + \frac{2} {{\cal M}_0} [{\bf k}^2_\perp -
\frac{({\bf k}_\perp\cdot{\bf q}_\perp)^2}{{\bf q}^2_\perp}]
\biggr\}.
\nonumber\\
\end{eqnarray}
${\cal A}=x_2m_1+x_1m_2$, $x=x_1$ and the other variables in Eq.
(\ref{22}) are defined in Appendix. In the covariant light-front
quark model the authors of Ref. \cite{Hwang:2006cua} obtained the
same form factor $\mathcal{ F}_{\Upsilon(nS)\to
\eta_b}(\mathbf{q}^2)$. The decay width for
$\Upsilon(nS)\rightarrow \eta_b+\gamma$ is easily achieved
\begin{eqnarray}\label{23}
\Gamma(\Upsilon(nS)\rightarrow
\eta_b+\gamma)=\frac{\alpha}{3}\bigg[\frac{m_{\Upsilon(nS)}^2-m_{\eta_b}^2}{2m_{\Upsilon(nS)}}\bigg]^3
\mathcal{ F}^2_{\Upsilon(nS) \to\eta_b}(0),
\end{eqnarray}
where $\alpha$ is the fine-structure constant and
$m_{\Upsilon(nS)},\; m_{\eta_b}$ are the masses of $\Upsilon(nS)$
and $\eta_b$ respectively.
\subsection{Numerical results}
In Ref.\cite{Choi:2007se} the authors
fixed the parameter $\beta_{b\bar b}=1.145$ or 1.803 \footnote{The
authors of Ref. \cite{Choi:2007se} used the harmonic oscillator and
linear potential forms for the confinement term in their
computations, thus they obtained two different values for the
$\beta$ parameter. Since later in our calculations we do not
evaluate the spectra of the concerned hadrons, the concrete value of
$\beta$ does not influence our numerical results.} in the Gaussian
wavefunction using the variational method with $m_b=5.2$ GeV.
However their predictions on the mass $m_{\eta_b}=9.657\,({\rm or}
\,9.295)$ GeV and $m_{\Upsilon(1S)}=9.691\,({\rm or} \,9.558$) GeV (the values in the brackets are obtained with a different potential
form), which obviously deviate from data ($m_{\eta_b}=9300\pm 20\pm
20$ GeV and $m_{\Upsilon(1s)}=9460.30\pm 0.26$ GeV \cite{PDG08}), thus
we are going to fix $\beta$ in an alternative way.
First we set the b quark mass as $m_b=4.64$ GeV which was used in
Ref. \cite{Cheng:2003sm}. Then we extract the decay constant
$f_{\Upsilon(nS)}$ of $\Upsilon(nS)$ from the data
$\Gamma(\Upsilon(nS)\rightarrow e^+e^- )$ \cite{PDG08} with
\begin{eqnarray}\label{31}
\Gamma(V\rightarrow
e^+e^-)=\frac{4\pi}{27}\frac{\alpha^2}{M_V}f^2_{V}.
\end{eqnarray}
In parallel, we can calculate the constant in the LFQM using the
formula \cite{Jaus:1999zv,Cheng:2003sm}
\begin{eqnarray}\label{31}
f_V&=&\frac{\sqrt{N_c}}{4\pi^3M}\int dx\int
d^2k_\perp\frac{\varphi}{\sqrt{2x(1-x)}\tilde M_0}
\biggl[xM_0^2\nonumber\\&&-m_1(m_1-m_2)-k^2_\perp+\frac{m_1+m_2}{M_0+m_1+m_2}k^2_\perp\biggl],\nonumber\\
\end{eqnarray}
equating the two results. $\beta_{\Upsilon(nS)}$ is determined (see
Table \ref{tab:etab1}). Then, we use the formula presented in
Section II to calculate
$\mathcal{B}(\Upsilon(3S)\rightarrow\eta_b+\gamma)$ and compare it
with the central value of the experimental data
$\mathcal{B}(\Upsilon(3S)\rightarrow\eta_b+\gamma)=(7.1\pm1.8\pm1.1)\times10^{-4}$ \cite{Bonvicini:2009hs}
or
$\mathcal{B}(\Upsilon(3S)\rightarrow\eta_b+\gamma)=(4.8\pm0.5\pm0.6)\times10^{-4}$ \cite{:2009pz}
to fit $\beta_{\eta_b}$ and the corresponding values are presented
in Table \ref{tab:etab2}. At last with all the parameters we
estimate $\mathcal{B}(\Upsilon(1S)\rightarrow\gamma\eta_b)$,
$\mathcal{B}(\Upsilon(2S)\rightarrow\eta_b+\gamma)$,
$\mathcal{B}(\Upsilon(4S)\rightarrow\eta_b+\gamma)$ and
$\mathcal{B}(\Upsilon(5S)\rightarrow\eta_b+\gamma)$ which are shown
in Table \ref{tab:etab2}. It is noted that at this step, we only
consider the direct decay modes, but for
$\Upsilon(5S)\rightarrow\eta_b+\gamma$, the re-scattering effect may
play a dominant role as mentioned in the introduction and we will
discuss the details in next section.
In order to illustrate the dependence of our results on $m_b$, we
re-set $m_b=5.2$ GeV, which was adopted by the authors of
Ref. \cite{Choi:2007se} and fitted $\beta_{\Upsilon(nS)}$ and
$\beta_{\eta_b}$ again. Using the new parameter the
$\mathcal{B}(\Upsilon(nS)\rightarrow\eta_b+\gamma)$ is computed and
the result is also listed in Table \ref{tab:etab2}. From the
Table \ref{tab:etab2}, we find
1. The predicted branching ratios
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ and
$\mathcal{B}(\Upsilon(4S)\rightarrow\eta_b+\gamma)$ are not
sensitive to $m_b$ and $\beta$, but
$\mathcal{B}(\Upsilon(2S)\rightarrow\eta_b+\gamma)$ and
$\mathcal{B}(\Upsilon(5S)\rightarrow\eta_b+\gamma)$ slightly change
as $m_b$ and $\beta$ vary within certain ranges;
2. Our results about
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ and
$\mathcal{B}(\Upsilon(2S)\rightarrow\eta_b+\gamma)$ are somehow
larger than that given in Ref. \cite{Ebert:2002pp};
3. If the final state interaction is not taken into account, the
branching ratio of $\Upsilon(5S)\rightarrow\eta_b+\gamma$ is not
anomalous compared to $\Upsilon(mS)\rightarrow\eta_b+\gamma\,
(m=1,2,3,4)$. Because the BELLE Collaboration \cite{Abe:2007tk} found
the rate of
$\Upsilon(5S)\rightarrow\Upsilon(1S,2S)+\pi\pi$ is
anomalously large compared to the similar dipion transitions between
lower $\Upsilon$ resonances, one has reason to doubt if such anomaly
would appear in $\Upsilon(5S)\rightarrow\eta_b+\gamma$.
It is noted that $\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$
is sensitive to $m_{\eta_b}$ (or\, $\Delta M$) since the decay width is
proportional to $(\Delta M)^3$, thus as $\Delta M$ is small, i.e.
the masses of initial and daughter mesons are close to each other,
any small changes of $m_{\eta_b}$ which has not been accurately
measured yet \cite{PDG08}, can lead to a remarkable difference. In
Fig. \ref{fig:LFQM2} we display the dependence of
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ on $\Delta M$.
Thus the accurate measurement on
$\mathcal{B}(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ will be a great
help to determine the mass of $m_{\eta_b}$.
\begin{widetext}\begin{center}
\begin{table}
\caption{ The $b$ quark mass and the parameters $\beta$ in the wavefunction in the unit of GeV.} \label{tab:etab1}
\begin{tabular}{c|ccccc}\toprule[1pt]
$m_b$ & $\beta_{\Upsilon(1S)}$ & $\beta_{\Upsilon(2S)}$&
$\beta_{\Upsilon(3S)}$&$\beta_{\Upsilon(4S)}$&$\beta_{\Upsilon(5S)}$
\\\midrule[1pt]
4.64&1.301&0.924&0.795&0.657&0.673
\\ 5.20&1.257&0.894&0.769&0.635&0.652\\\bottomrule[1pt]
\end{tabular}
\end{table}
\begin{table}
\caption{ The branching ratio of
$\Upsilon(nS)\rightarrow\gamma\eta_b$.}
\label{tab:etab2}
\begin{tabular}{c|c|cccc}\toprule[1pt]
$m_b$ & $\beta_{\eta_b}$ &
$\mathcal{B}{(\Upsilon(1S)\rightarrow\eta_b+\gamma)}$&$\mathcal{B}{(\Upsilon(2S)\rightarrow\eta_b+
\gamma)}$&$\mathcal{B}{(\Upsilon(4S)\rightarrow\eta_b+\gamma)}$
&$\mathcal{B}{(\Upsilon(5S)\rightarrow\eta_b+\gamma)}$\\\midrule[1pt]
4.64&0.915\footnote{fitted from CLEO
data}&2.25$\times10^{-4}$&4.91$\times 10^{-4}$&4.66$\times
10^{-6}$&9.27$\times10^{-8}$\\
& 0.901\footnote{fitted from BaBar
data}&2.52$\times10^{-4}$&9.21$\times 10^{-4}$&3.86$\times
10^{-6}$&6.97$\times10^{-8}$
\\\hline
5.20&0.885$^a$&1.83$\times10^{-4}$&2.27$\times10^{-4}$&4.31$\times10^{-6}$&8.57$\times10^{-8}$
\\ & 0.874$^b$& 2.05$\times 10^{-4}$&5.31$\times10^{-4}$&3.53$\times 10^{-6}$&6.32$\times10^{-8}$\\
\bottomrule[1pt]
\end{tabular}
\end{table}\end{center}
\end{widetext}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=9cm]{1s.eps}
\end{tabular}
\end{center}
\caption{The dependance of
$\mathcal{B}(\Upsilon(1S)\rightarrow\gamma\eta_b)$ on $\Delta M$.}
\label{fig:LFQM2}
\end{figure}
\section{Possible re-scattering effects inducing a large $\mathcal{B}(\Upsilon(5S)\to \eta_b +\gamma)$ }
\begin{center}
\begin{figure}[htb]
\begin{tabular}{c}
\scalebox{0.45}{\includegraphics{fsi.eps}}
\end{tabular}
\caption{The diagrams for $\Upsilon(nS)\rightarrow
B^{(*)+}B^{(*)-}\to \eta_b+\gamma$. Other diagrams can be obtained by
a charge conjugation transformation $B^{(*)}\rightleftarrows \bar{B}^{ (*)}$. \label{fig:fsi}}
\end{figure}
\end{center}
As aforementioned, the re-scattering of hadrons may remarkably
enhance the rate of $\Upsilon(nS)\,(n>4)\rightarrow \eta_b+ \gamma$.
Similar to the re-scattering effects on in the branching ratios of
$\Upsilon(5S)\rightarrow \Upsilon(1S)+\pi\pi $ \cite{Chao-1} and
$\Upsilon(nS)(n\geq4)\rightarrow \Upsilon(1S)\eta $ \cite{Chao-2},
the transitions $\Upsilon(nS)\rightarrow\eta_b + \gamma$ can occur
via re-scattering sub-processes with the intermediate states being
$B^{(*)}\bar{B}^{(*)}$ where another $B^{(*)}$ meson is exchanged at
the t-channel. The corresponding diagrams are depicted in
Fig. \ref{fig:fsi}. Since the other diagrams can be obtained by a
charge conjugation transformation of $B^{(*)}\rightleftarrows \bar{B}^{ (*)}$ the contribution of each diagram in
Figs. \ref{fig:fsi} (a-f) should be multiplied by a factor 2.
Now let us calculate the amplitudes of the re-scattering processes
which occur at the hadron level. Following Refs.
\cite{Chao-1,Chao-2} we will not account for the contribution from
the dispersive parts of the diagrams where $B^{(*)}$ from
$\Upsilon(ns)$ are off-shell, but only concern the absorptive parts
where $B^{(*)}$ are real particles on their mass-shells. The
off-shell effect of the meson exchanged at t-channel is compensated
by a monepole form factor which is also reflects the inner
structures of the mesons at the effective vertex
\begin{eqnarray}\label{eq:monopole}
\mathcal{F}(m_i,q^2)=\frac{(\Lambda+m_i)^2-m_i^2}{(\Lambda+m_i)^2-q^2},
\end{eqnarray}
where $q$ and $m_i$ are the momentum and mass of the exchanged
meson respectively. And the cutoff is set as $\Lambda=600$ GeV
\cite{Chao-1,Chao-2}.
The absorptive part of the amplitude is read as
\begin{eqnarray}\label{eq:amplitude}
Abs_i=&&\frac{|\mathbf{p}_1|}{32\pi^2m_{\Upsilon(nS)}}\int d \Omega
\mathcal{A}_i[\Upsilon(nS)\rightarrow B^{(*)}\bar{B}^{(*)}]
\nonumber\\&&\times \mathcal{C}_i[B^{(*)}\bar{B}^{(*)}\rightarrow
\eta_b+\gamma]\times\mathcal{F}(m_i,q^2),
\end{eqnarray}
with $i=a,b,c,d,e,f$. Here, $d\Omega$ and $\mathbf{p}_1$ are the
solid angle and linear momentum of the on-shell $B^{(*)}$ in the
rest frame of $\Upsilon(nS)$, respectively.
The effective couplings for $\Upsilon BB$, $\Upsilon\! B^*\!B$ and
$\Upsilon B^*B^*$ we adopt in this work are directly borrowed from
Refs. \cite{Chao-1,Chao-2}, and some discussions will be made in the
last section.
\begin{subequations} \label{effective-Lagrangians1}
\begin{eqnarray}
\mathcal{L}_{\Upsilon BB}&=& g_{\Upsilon
BB}\Upsilon_\mu(\partial^\mu
B{B}^{\dagger}-B\partial^\mu {B}^{\dagger}),\label{L-YBB}\\
\mathcal{L}_{\Upsilon B^*B}&=& \!\frac{g_{\Upsilon\!
B^*\!B}}{m_{\Upsilon}}\varepsilon^{\mu\nu\alpha\beta}\partial_\mu
\!\Upsilon_\nu
\!\nonumber\\
&& \times(B^*_\alpha\overleftrightarrow{\partial}_\beta
{B}^{\dagger}\!\! - \!\!
B\overleftrightarrow{\partial}_\beta{B}^{*\dagger}_\alpha\!),\label{L-YB*B}\\
\mathcal{L}_{\Upsilon B^*B^*}&=& g_{\Upsilon B^* B^*} (
-\Upsilon^\mu
B^{*\nu}\overleftrightarrow{\partial}_\mu {B}_\nu^{*\dagger} \nonumber\\
&&+ \Upsilon^\mu B^{*\nu}\partial_\nu{B}^{*\dagger}_{\mu} -
\Upsilon_\mu\partial_\nu B^{*\mu}
{B}^{*\nu\dagger}),\label{L-YB*B*}
\end{eqnarray}
\end{subequations}
where
$\overleftrightarrow{\partial}=\overrightarrow{\partial}-\overleftarrow{\partial}$
and the coupling constants were fixed as \cite{Chao-1,Chao-2}
\begin{eqnarray}
g_{\Upsilon BB}&&=2.5\nonumber\\g_{\Upsilon B^*B}&&=1.4\pm
0.3\nonumber\\g_{\Upsilon B^*B^*}&&=2.5\pm0.4.\,\label{HQS:g-YBB}
\end{eqnarray}
Following the strategy of Ref.\cite{Chao-2}, we obtain
\begin{subequations} \label{effective-Lagrangians2}
\begin{eqnarray}
\mathcal{L}_{\gamma BB}&=& g_{\gamma BB}A_\mu(\partial^\mu
B{B}^{\dagger}-B\partial^\mu {B}^{\dagger}),\label{L-YBB}\\
\mathcal{L}_{\gamma B^*B}&=& \!\frac{g_{\gamma\!
B^*\!B}}{m_{B^*}}\varepsilon^{\mu\nu\alpha\beta}\partial_\mu
\!A_\nu
\!\nonumber\\
&& \times(B^*_\alpha\overleftrightarrow{\partial}_\beta
{B}^{\dagger}\!\! - \!\!
B\overleftrightarrow{\partial}_\beta{B}^{*\dagger}_\alpha\!),\label{L-YB*B}\\
\mathcal{L}_{\gamma B^*B^*}&=& g_{\gamma B^* B^*} ( -A^\mu
B^{*\nu}\overleftrightarrow{\partial}_\mu {B}_\nu^{*\dagger} \nonumber\\
&&+ A^\mu B^{*\nu}\partial_\nu{B}^{*\dagger}_{\mu} -
A_\mu\partial_\nu B^{*\mu} {B}^{*\nu\dagger}),\label{L-YB*B*}\\
\mathcal{L}_{B^*B\eta_b}&=& ig_{B^*B\eta_b}B^*_{\mu}\partial^\mu\eta_b{B}^{\dagger},\label{L-B*Beta}\\
\mathcal{L}_{B^*B^*\eta_b}&=&
i\frac{g_{B^*B^*\eta_b}}{m_{B^*}}\varepsilon^{\mu\nu\alpha\beta}\partial_{\mu}B^*_{\nu}{B^*}^{\dagger}_{\alpha}
\partial_\beta\eta_b.\label{L-B*B*eta}
\end{eqnarray}
\end{subequations}
With the heavy quark spin symmetry\cite{Isgur:1989vq}, we have
\begin{eqnarray}
g_{\gamma BB}&=&g_{\gamma B^*B}=g_{\gamma B^*B^*},\nonumber\\
g_{\eta_b B^*B}&=&g_{\eta_b B^*B^*}=g_{\Upsilon(1S)
BB}.\,\label{HQS:g-etaBB}
\end{eqnarray}
In terms of the theoretically evaluated value of
$\Gamma(B^{*+}\rightarrow B^+ \gamma)=0.40\pm0.03$
keV \cite{Choi:2007se} one can fix $g_{\gamma B^*B}\approx 3.5$. The
coupling $g_{\eta_b B^*B}$ should be at the order of $O(1)$, but may
vary within a reasonable range. If we choose $g_{\eta_b
B^*B}=g_{\eta_b B^*B^*}\sim 1$ the contributions of the diagrams
(a), (c), (d), (f) in Fig. \ref{fig:fsi} to
$\mathcal{B}(\Upsilon(5S)\rightarrow\gamma \eta_b)$ are
approximately a few of $10^{-7}$ GeV and that of the diagrams (b),
(e) are slightly smaller than the direct transition. If the
interference between the contributions of the direct decay and that
through the re-scattering is constructive, the total width would be
about a few times larger, but if it is destructive, the width would
be very suppressed. However, the coupling $g_{\Upsilon(1S)BB}$ may
be as large as 15 as was estimated in Refs. \cite{Chao-1,Chao-2}. If
we use this value for the coupling constants $g_{\eta_b B^*B}$,
$g_{\eta_b B^*B^*}$ the contributions of the diagrams in
Fig. \ref{fig:fsi} would enhance the total width by more than two
orders. In principle, the diagram (a) in Fig. \ref{fig:fsi} can
contribute to $\Upsilon(4S)\rightarrow \eta_b+\gamma$, but the mass
of $\Upsilon(4S)$ is just above the threshold of $B^+B^-$, so that
the phase space would greatly suppress the contribution of the
diagram.
\section{Conclusion}
By studying the radiative decay of $\Upsilon(nS)\to
\eta_b+\gamma$, we can learn much about the hadronic structure of
$\eta_b$. Efforts have been made to explore the spin singlet
$\eta_b$, in the Data-book of 2008, $\eta_b$ was still omitted
from the summary table \cite{PDG08}. In fact, determination of
the mass of $\eta_b$ is made via the radiative decays of
$\Upsilon(nS)\to \eta_b+\gamma$ \cite{:2008vj}, and the recent
data show that $m_{\eta_b}=9388.9^{+3.1}_{-2.3}(stat)\pm
2.7(syst)$ by the $\Upsilon(3S)$ data and
$m_{\eta_b}=9394.2^{+4.8}_{-4.9}(stat)\pm 2.0(syst)$ by the
$\Upsilon(2S)$ data \cite{:2009pz}. Recently, Penin
\cite{Penin:2009wf} reviewed the progress for determining the mass
of $\eta_b$ and indicated that the accurate theoretical prediction
of $m_{\eta_b}$ would be a great challenge. Indeed, determining
the wavefunction of $\eta_b$ would be even more challenging. In
this work, we are not going to obtain the wavefunction or even the
mass of $\eta_b$ based on the fundamental theories, such the
non-perturbative QCD, but using the radiative transition to
testify the phenomenologically determined wavefunction as long as
the mass is well measured. We carefully study the transition rates
of the radiative decays which would help experimentalists to
extract information about $m_{\eta_b}$. The transition rate of
$\Upsilon(1S)\to\eta_b+\gamma$ is very sensitive to the mass
splitting $\Delta M=m_{\Upsilon(1S)}-m_{\eta_b}$ (see the text
above), thus an accurate measurement of the radiative decay may be
more useful to learn the spin dependence of the bottominia.
Recently our experimental colleagues have made some progress. CLEO
and BaBar collaborator measure the
$\Upsilon(3S)\rightarrow\eta_b+\gamma$ and the mass $m_{\eta_b}$
which offer an opportunity for us to more deeply study $\eta_b$.
Following Ref. \cite{Choi:2007se} we systematically study
$\Upsilon(nS)\rightarrow\eta_b+\gamma$ in the LFQM. We take an
alternative way which is different from that adopted in
Ref. \cite{Choi:2007se} to fix the parameters $i.e.$, namely we use the
data to fit $\beta_{\Upsilon(nS)}$ and $\beta_{\eta_b}$ in the
wavefunctions. Then using these parameters we estimate
$\Upsilon(1S)\rightarrow\eta_b+\gamma$,
$\Upsilon(2S)\rightarrow\eta_b+\gamma$,
$\Upsilon(4S)\rightarrow\eta_b+\gamma$ and
$\Upsilon(5S)\rightarrow\eta_b+\gamma$. Our result indicates
$\Upsilon(2S)\rightarrow\eta_b+\gamma$ and
$\Upsilon(5S)\rightarrow\eta_b+\gamma$ are more sensitive to the
parameters than $\Upsilon(1S)\rightarrow\eta_b+\gamma$ and
$\Upsilon(4S)\rightarrow\eta_b+\gamma$.
Since the value of $\Upsilon(1S)\rightarrow\eta_b+\gamma$ is
sensitive to $\Delta M$, we show the dependance of
$B(\Upsilon(1S)\rightarrow\eta_b+\gamma)$ on $\Delta M$. We hope that
our experimental colleagues will conduct accurate measurements in
the near future to determine the precise value of $m_{\eta_b}$.
The anomalous largeness of the branching ratio of
$\Upsilon(5S)\rightarrow\Upsilon(1S,2S)+\pi\pi$ motivates
a hot surf of theoretical studies \cite{Simonov}. It was suggested
that the re-scattering effects may explain the unusual large
branching ratio. This mechanism should be tested somewhere else. In
Ref. \cite{Chao-2} the authors evaluated the effect induced by the
mechanism for
$\Upsilon(5S)\rightarrow\Upsilon(1S,2S)+\eta$ and found
that the corresponding branching ratio is also greatly enhanced
compared to the transition among lower resonances. We suggest to
further test the mechanism at the radiative decays where the
effective electromagnetic vertex is relatively simple. Our result
which is obtained in terms of the LFQM, indicates the branching
ratio of $\Upsilon(5S)\rightarrow\eta_b+\gamma$ is not anomalous
compared to $\Upsilon(mS)\rightarrow\eta_b+\gamma\, (m=1,2,3,4)$ as
long as the re-scattering is not taken into account. However, there
could be a two-order enhancement in magnitude for
$\Upsilon(5S)\rightarrow\eta_b+\gamma$ which is induced by the
re-scattering effects. Thus measurement of
$\Upsilon(5S)\rightarrow\eta_b+\gamma$ would be an ideal probe for
the re-scattering mechanism which successfully explains the data of
$\Upsilon(5S)\rightarrow\Upsilon(1S,2S)+\pi\pi$. This is
one of the tasks of the LHCb which will be operating very soon.
\section*{Acknowledgments}
This project is supported by the National Natural Science Foundation of
China (NSFC) under Contracts Nos. 10705001 and 10775073; the Foundation for
the Author of National Excellent Doctoral Dissertation of P.R. China
(FANEDD) under Contracts No. 200924; the Doctoral Program Foundation of Institutions of
Higher Education of P.R. China under Grant No. 20090211120029; the
Special Grant for the Ph.D. program of Ministry of Eduction of P.R.
China; the Program for New Century Excellent Talents in University (NCET) by Ministry of Education of P.R. China; the Special Grant for New Faculty from
Tianjin University.
\section*{Appendix}
The incoming (outgoing) meson in Fig. \ref{fig:LFQM} has the
momentum ${P}^({'}^)={p_1}^({'}^)+p_2$ where ${p_1}^({'}^)$ and
$p_2$ are the momenta of the off-shell quark and antiquark and
\begin{eqnarray}\label{20}
p^+_1&=&x_1 P^+,\qquad p^+_2 = x_2 P^+,
\nonumber\\
{ p}_{1\perp}&=& x_1{ P}_\perp + { k}_\perp,\qquad { p}_{2\perp}= x_2{
P}_\perp - { k}_\perp,
\nonumber\\
p'^+_1&=&x_1 P^+,\qquad p'^+_2 = x_2 P^+,
\nonumber\\
{ p'}_{1\perp}&=& x_1{ P'}_\perp + { k'}_\perp, \qquad{ p'}_{2\perp}=
x_2{ P'}_\perp - { k'}_\perp\nonumber
\end{eqnarray}
with $x_1+x_2=1$, where $x_i$ and $k_\perp(k'_\perp)$ are internal
variables. $M_0$ and $\tilde {M_0}$ are defined
\begin{eqnarray}\label{app2}
&&M_0^2=\frac{k^2_\perp+m^2_1}{x_1}+\frac{k^2_\perp+m^2_2}{x_2},\nonumber\\&&
\tilde {M_0}=\sqrt{M_0^2-(m_1-m_2)^2}.\nonumber
\end{eqnarray}
The radial wavefunctions $\phi$ related to $\Upsilon(nS)$ are defined
\begin{eqnarray*}\label{app4}
\phi(1S)&=&4\Big(\frac{\pi}{\beta^2}\Big)^{3/4}\sqrt{\frac{\partial
k_z}{\partial x}}{\exp}\Big(-\frac{k^2_z+k^2_\perp}{2\beta^2}\Big),\nonumber\\
\phi(2S)&=&4\Big(\frac{\pi}{\beta^2}\Big)^{3/4}\sqrt{\frac{\partial
k_z}{\partial x}}{\exp}\Big(-\frac{k^2_z+k^2_\perp}{2\beta^2}\Big)\nonumber\\&&\times\frac{1}{\sqrt{6}}
\Big(-3+2\frac{k^2_z+k^2_\perp}{\beta^2}\Big),
\nonumber\\
\phi(3S)&=&4\Big(\frac{\pi}{\beta^2}\Big)^{3/4}\sqrt{\frac{\partial
k_z}{\partial x}}{\exp}\Big(-\frac{k^2_z+k^2_\perp}{2\beta^2}\Big)\nonumber\\&&\times\frac{1}{{2\sqrt{30}}}
\Big(-15-20\frac{k^2_z+k^2_\perp}{\beta^2}+4\frac{(k^2_z+k^2_\perp)^2}{\beta^4}\Big),
\nonumber\\
\phi(4S)&=&4\Big(\frac{\pi}{\beta^2}\Big)^{3/4}\sqrt{\frac{\partial
k_z}{\partial x}}{\rm
exp}\Big(-\frac{k^2_z+k^2_\perp}{2\beta^2}\Big)\frac{1}{{12\sqrt{35}}}
\nonumber\\&&\times
\Big(-105+210\frac{k^2_z+k^2_\perp}{\beta^2}-84\frac{(k^2_z+k^2_\perp)^2}{\beta^4}\nonumber\\&&+8\frac{(k^2_z+k^2_\perp)^3}{\beta^6}\Big),
\nonumber\\
\phi(5S)&=&4\Big(\frac{\pi}{\beta^2}\Big)^{3/4}\sqrt{\frac{\partial
k_z}{\partial x}}{\rm
exp}\Big(-\frac{k^2_z+k^2_\perp}{2\beta^2}\Big)\frac{1}{{72\sqrt{70}}}
\nonumber\\&&\times
\Big(945-2520\frac{k^2_z+k^2_\perp}{\beta^2}+1512\frac{(k^2_z+k^2_\perp)^2}{\beta^4}
\nonumber\\&&-288\frac{(k^2_z+k^2_\perp)^3}{\beta^6}+
16\frac{(k^2_z+k^2_\perp)^4}{\beta^8}\Big).
\end{eqnarray*}
with
\begin{eqnarray}
k_z&=&\frac{x_2M_0}{2}-\frac{m_2^2+k^2_\perp}{2x_2M_0},\\
\frac{\partial k_z}{\partial
x}&=&\frac{M_0}{4x_1x_2}\Big[1-\Big(\frac{m_1^2-m_2^2}{M_0^2}\Big)^2\Big].
\end{eqnarray}
More information can be found in Ref. \cite{Cheng:2003sm}.
\vfil
|
\section*{Introduction}
Since the original paper by Ising\cite{Ising1925}, in which it was proved that the Ising model on an infinite linear chain would not show a phase transition, a huge amount of research has been conducted on the subject. The first phase of this fruitful line investigated regular systems, and after the articles by Peierls\cite{Peierls1936} and Onsager\cite{Onsager1944} it became clear that regular lattices in $d$ dimensions would magnetize when $d\ge2$.\par
In a second phase, a large number of fractals was investigated\cite{YevalGefen1980,MonHsi2003,Achiam1985,CarMarRui1998,BaFaAl2008}, mainly via the so-called renormalization group techniques, to discover that, although no rigorous theorem has been proven,
those (and only those) fractals which have an infinite minimum order of ramification display spontaneous magnetization.\par
In the same years, fundamental analytical results were obtained for disordered structures embedded in Euclidean lattices, applying percolation theory concepts\cite{ChaFro1985}.\par
More general graphs\cite{Jullien1979,ReggeZecchina1996,Ceresole1998,Herrero2002,DogGolMen2002} have become increasingly popular in the last twenty years:
the main difference with the previous cases is that the metric structure
of the embedding space ceases to play an essential role, as in general a graph is a topological structure which is not necessarily embeddable in a finite dimensional Euclidean space.
The absence of translational invariance and scale invariance makes general graphs very difficult to study, as \emph{ad hoc} techniques must be employed, that usually admit no straightforward generalization.\par
An important result would be the identification of a simple parameter, capable of determining whether the Ising model on a given graph exhibits a phase transition: we present here a theorem stating a sufficient condition for a graph to exhibit spontaneous magnetization, which is a generalization of the classic Peierls-Griffiths theorem\cite{Peierls1936,griffithspeierls} for the square lattice.
While interesting works, employing the same basic techniques as the Griffiths theorem, have been proposed for higher-dimensional lattices, typically stemming from the paper of Dobrushin\cite{Dobrushin1964rus,Dobrushin1964}, such as the profound contribution by Isakov\cite{Isakov1984} and the extensions to non-symmetric situations treated in Pirogov-Sinai theory\cite{Sinai1982}, or directly from the paper by Griffiths, as in Lebowitz and Mazel\cite{Lebowitz1998}, nothing applying to inhomogeneous networks and arbitrary graphs has yet emerged, and our contribute aims essentially at filling this gap.\par
The reason why the modulus of magnetization $\langle \left\vert M\right\vert\rangle$ is considered is that it is indissolubly tied to the long range order of the graph: it is easy to prove that, when the external field is zero, stating $\langle \left\vert M\right\vert\rangle \;\ge \epsilon >0$ is equivalent to the existence of a non-zero measure subset of all the correlation functions such that all of its members are greater than a small constant $\epsilon'>0$.\par
In the following, we first present the concepts of open and closed borders in a graph for later use; we then define the ferromagnetic Ising model on a general graph and derive the equivalence of the sum over configurations and the sum over different borders. Next we prove a theorem stating a sufficient condition for a graph to exhibit spontaneous magnetization. Because of the technical nature of the theorem, we thoroughly examine its more and less immediate consequences for a wide range of different graphs. Lastly, we discuss our results and the current comprehension of the mechanism of spontaneous magnetization on graphs for the Ising model.
\section*{Open and Closed Borders in a Graph}
A {\bf graph} $\cal G$ is a pair $\left({\cal P},{\cal L}\right)$, where $\cal P$ is a countable collection of vertices and ${\cal L}\subset {\cal P}\times{\cal P}$ is a set of unoriented bonds between points.
Any pair ${\cal G}'=({\cal P}',{\cal L}')$, such that ${\cal P}'\subset {\cal P}$, ${\cal L}'$ contains only links between elements of ${\cal P}'$ and ${\cal L}'\subset {\cal L}$, is called a {\bf subgraph}, and it's denoted ${\cal G}'\subset {\cal G}$.
We will restrict our attention to those graphs whose coordination number $z_i$, representing the number of bonds in ${\cal L}$ having one extremum in $i$, is uniformly limited: an integer $z_{Max}>0$ exists such that $z_i\le z_{Max}$ for all $i\in{\cal P}$.\par
We now define a {\bf path} $\gamma$, between two points $i$ and $j$, as a collection of consecutive bonds of $\cal L$,
where consecutive means that each pair shares a vertex with the next one:
\[ \gamma = \left\{ (il_1), (l_1 l_2), \dots , (l_{D-1},j) \right\}. \]
Directly associated to the concept of path, the \textbf{chemical distance} between two points $i$ and $j$ is defined as the length of the shortest path connecting them. The \emph{chemical distance} straightforwardly induces the so-called intrinsic metric of the graph.
The \textbf{intrinsic fractal dimension} $d_{frac}$ of a graph, defined as the minimum $d$ such that $N_r$, the maximum number of vertices included in a Van Hove sphere\cite{BurioniRWAG} of radius $r$ (i.e. the set of points within a chemical distance from a given point of no more than $r$ bonds), satisfies $N_r\le r^{d}$ as $r\rightarrow\infty$. It differs from the usual \emph{fractal dimension} in that it refers to the topological nature of the graph (i.e. on its natural - chemical - distance), and not on the metric structure of the space into which the graph is embedded.
To proceed we declare what will be considered a border from now on.
\begin{definition}
Given a connected graph $\ensuremath{ {\cal G} } = \left( {\cal P}, {\cal L} \right) $,
we can define a {\bf border} $\cal B$ as a set of bonds that separates exactly two connected subgraphs. It means that two sets ${\cal P}_1,{\cal P}_2\subset{\cal P}$ exist such that
\begin{itemize}
\item ${\cal P}_1 \cap {\cal P}_2 =\emptyset$ and ${\cal P}_1 \cup {\cal P}_2 = {\cal P}$,
\item any path on \ensuremath{ {\cal G} } from a point of ${\cal P}_1$ to a point of ${\cal P}_2$ must contain at least one bond of $\cal B$,
\item a path exists between any two points in ${\cal P}_i$ $(i = 1,2)$ that doesn't contain any bond of $\cal B$.
\end{itemize}
\end{definition}
It is noteworthy that the union of two disjoint borders is not a border itself under this definition, as it divides the graph into three subgraphs. This is a feature we'll later need to avoid overcounting different configurations.\par
The intuitive idea of open and closed border
is actually an artifact created by our visualizing regular lattices as immersed in a finite dimensional real space: the seeming adjacency of the vertices creates a contour of the graph, which we use to define closed and open borders. The fact is that this contour is heavily dependent on what particular immersion we employ, and ceases to exist when we consider the graph for itself.
The border in itself has no geometry whatsoever, since it is just a collection of links, and even the notion of "continuous" border, without further specifications, makes no sense from a graph-theoretic point of view: in a general graph, a border is just a collection of links that splits it into two parts.
We now define open and closed borders with respect to an external set of points, as it will be useful later.
\begin{definition}
Given a border $\cal B$ and a set of points ${\cal E}\subset{\cal P}$, we say that $\cal B$ is {\bf closed} with respect to the \emph{external points set} $\cal E$ if either ${\cal P}_1\cap{\cal E}=\emptyset$ or ${\cal P}_2\cap{\cal E}=\emptyset$, otherwise $\cal B$ is {\bf open}.
\end{definition}
For any finite subgraph ${\cal G}_N$ of a given graph $\cal G$, we choose the natural set of external points $\cal E$:
\begin{eqnarray*}
{\cal E} \equiv \left\{ i\in\Gn{} : (i,j)\in{\cal L}\;\; for\;\, some\,\; j\in\ensuremath{ {\cal G} } \setminus\Gn{} \right\}.
\end{eqnarray*}
Now, given a border ${\cal B}_i$ that divides \Gn{} into two subgraphs $A_i$ and $C_i$, we define $A_i$ as
\begin{itemize}
\item \emph{internal} if
\begin{itemize}
\item ${\cal B}_i$ is closed and $A_i\cap{\cal E}=\emptyset$, or
\item ${\cal B}_i$ is open and $A_i$ contains fewer elements than $C_i$, or
\item ${\cal B}_i$ is open, $A_i$ has the same size as $C_i$ and the points in $A_i$ linked to ${\cal B}_i$ have negative spin;
\end{itemize}
\item \emph{external} if
\begin{itemize}
\item ${\cal B}_i$ is closed and $A_i\cap{\cal E}\ne\emptyset$, or
\item ${\cal B}_i$ is open and $A_i$ has more elements than $C_i$, or
\item ${\cal B}_i$ is open, $A_i$ has the same size as $C_i$ and the points in $A_i$ linked to ${\cal B}_i$ have positive spin.
\end{itemize}
\end{itemize}
The reason why we had to select a finite subgraph ${\cal G}_N$ is that we need to be able to count the number of spins in the graph for the previous definitions to make sense.
\section*{The Ferromagnetic Ising Model on a Graph}
Let now $\sigma_i=\pm 1$ be a spin variable for each vertex $i \in {\cal P}$.
We define the {\bf Ising Hamiltonian} on a graph as
\begin{equation}
{\cal H} = -\sum_{(i,j)\in {\cal P}\times{\cal P}} J_{ij} \sigma_i \sigma_j - \sum_{i\in {\cal P}} \sigma_i h_i,
\end{equation}
where the couplings $J_{ij}=J_{ji}$ must satisfy $0\le J_{ij} < J_{Max}<\infty$ for some $J_{Max}$, and $J_{ij}>0$ if and only if $(i,j)\in{\cal L}$.
In the following we will set the external field to zero everywhere ($h_i\equiv 0$).\par
Now that we have presented the terminology we'll be using, we are going to study the equilibrium statistical mechanics of the Ising model at inverse temperature $\beta$, and in particular the modulus of the magnetization
\begin{equation}
\langle |M|\rangle = {\cal Z}^{-1} \sum_{\left\{ \sigma_i \right\} } { \left\vert \sum_{j\in{\cal P}} \sigma_j \right\vert \over \left\vert{\cal P}\right\vert } e^{-\beta {\cal H}(\left\{ \sigma_i \right\})},
\end{equation}
where ${\cal Z} = \sum_{\left\{ \sigma_i \right\} } e^{-\beta {\cal H}(\left\{ \sigma_i \right\})}$ is the partition function, and $\left\vert {\cal P} \right\vert$ is the cardinality of $\cal P$.\par
Since $\Delta |M| = \langle |M|^2\rangle - \langle |M|\rangle\langle |M|\rangle\;$ is a variance,
\[\langle M^2\rangle \;\ge (\langle |M|\rangle)^2,\]
so stating $\langle |M|\rangle\;=\epsilon>0$ implies $\langle M^2\rangle\;\ge \epsilon^2>0$. On the other hand, since $M^2\le |M|$, the converse is true, so $\langle M^2\rangle\;>0 $ and $\langle |M|\rangle\;>0$ are equivalent.
As we have now defined the main quantities we'll be studying, our next step is to prove that we can substitute the sum over configurations of the graph with a sum over possible \emph{border classes}, that we now define.
\section*{Equivalence between sets of borders and spin configurations}
\begin{definition}
A {\bf border class} is a class $C=\{C^i\}$ of border sets $C^i = \left\{ B^i_1,B^i_2,\dots,B^i_{N_i} \right\}$, where $i=1,\dots,N_C$, such that
\begin{itemize}
\item $B^i_u\cap B^i_v = \emptyset$ for all $i=1,\dots,N_C$ and $u,v=1,\dots,N_i$,
\item $\cup^{l=1,\dots,N_i} B^i_l = \cup^{m=1,\dots,N_j} B^j_m$ for all $i,j=1,\dots,N_C$.
\end{itemize}
\end{definition}
\begin{theorem}
To any given border class corresponds one and only one configuration of spins on $\cal P$, once we set the value of a single spin.
\end{theorem}
\begin{proof}
To prove that, for any border class and a given spin $p\in{\cal P}$, we can construct a single spin configuration, we first choose an arbitrary representative $C^i = \left\{ B^i_1,\dots,B^i_N \right\}$ of $C$ and set all the spins to the value of $p$, then for each $B^i_k \in C^i$ we flip all the spins of the subgraph which doesn't contain $p$. The result is independent of the order in which we choose the $B^i_k$, since each spins changes sign once for every border that separates it from the fixed spin $p$, and is independent of the specific $i$.\\
To prove that for any given spin configuration we can create a single border class, we proceed as follows: let $R^{\pm}$ be the sets of all plus (minus) spins,
\begin{eqnarray*}
R^{\pm} \equiv \left\{ i\in\ensuremath{ {\cal G} } : \sigma_i=\pm 1 \right\};
\end{eqnarray*}
We now choose the subsets $R^\pm_i$ of $R^{\pm}$, so that each $R^\pm_i$ is connected, while for all $i\ne j$ $R^\pm_i$ and $R^\pm_j$ are disconnected; moreover we require that
\begin{eqnarray*}
&&R^\pm \equiv R^\pm_1 \cup R^\pm_2 \cup \dots \cup R^\pm_S,\\
&&R^\pm_i\cap R^\pm_j = \emptyset \;\;\;\forall_{i\ne j}.
\end{eqnarray*}
We are selecting individual clusters of homogeneous spins, so satisfying the above requisites is always possible.
Setting now
\begin{eqnarray*}
\partial R^\pm_i \equiv \left\{ (a,b)\in{\cal L} : a\in R^\pm_i, b\notin R^\pm_i \right\},
\end{eqnarray*}
the sets $B^\pm \equiv \partial R^\pm_1 \cup \dots \cup \partial R^\pm_S$ are a collection of links each defined unambiguously, and furthermore $B^+\equiv B^-$.
It may happen that for some $i$ the subgraph $\ensuremath{ {\cal G} } \setminus R^\pm_i$ is made of two disconnected subgraphs (e.g. when a ring of plus spins is surrounded by minus spins); as a consequence $\partial R^\pm_i$ is not a border according to our definition. In that case it is possible to split $\partial R^\pm_i$ into subsets, so that each of them divides \ensuremath{ {\cal G} } into two connected subgraphs. After dealing in this way whith all the $R^\pm_i$, we are left with a collection of well-defined borders $\partial T^+_j$, with $j=1,\dots,U_+$, and $\partial T^-_k$, with $k=1,\dots,U_-$ where $U_+,U_->S$.
It is still possible that some of the borders $T^+_i$, while defining exactly the same zones, have no correspective in $T^-_i$ but, since they nevertheless verify $B^+\equiv B^-$, they belong to the same border class, completing the proof.
\end{proof}
The main consequence of this result is that we can substitute a sum over border classes for a sum over configurations whenever needed, and we can infer from the structure of the borders some limiting properties for the spins distributions, as we'll see soon.
It is worthwhile to explicitly notice that, when we pass from a sum over configurations to one over borders, and not border classes, we overcount some borders, as there are more than one representative of each border class: this is not going to be a problem in the use we'll make of this result.
\section*{Generalized Peierls-Griffiths' Theorem}
We can divide the set of all configurations on \Gn{} into two classes:
\begin{itemize}
\item all the negative spins are internal to some border (class $\cal N$),
\item at least a negative spin exists that is external to all borders (class $\cal P$).
\end{itemize}
The second case implies that every positive spin lies inside some border, since it must lie on the opposite side of the negative spin which is always external.\par
We now restrict our attention to the configurations belonging to the first class, denoting by a subscript $\cal N$ the quantities that pertain to it; we can obtain a good estimate of the number of negative spins, $\langle N_-\rangle_{\cal N}$, as follows: the sign of a spin $p$ is negative if it is contained inside an odd number of borders, positive otherwise; we obtain a very naive, yet effective, approximation if we consider any spin contained inside at least one border as negative: letting $I^p$ be $1$ if $p$ is inside at least a border, $0$ otherwise, we can write
\begin{eqnarray*}
\langle N_-\rangle_{\cal N}\;\; \le \sum_{p\in{\cal G}} \;\langle I^p\rangle.
\end{eqnarray*}
We are now to give a reasonable estimate of $\langle I^p\rangle$: take all the configurations $\cal C$ with at least one border containing $p$, call $b_{min}$ the length of the shortest border in $\cal C$ containing $p$ and let $k$ be the number of borders containing $p$, so as to write
\begin{eqnarray*}
\langle I^p\rangle\;\; = \ensuremath{{\cal Z}}^{-1}\sum_{ {\cal C}|_{p\;inside} } \ensuremath{e^{-\beta{\cal H}}} = \ensuremath{{\cal Z}}^{-1}\sum_{b_{min} \ge 1} \sum_{k \ge 1} \sum_{{\cal C}|^{b_{min}}_{k\;borders}} \ensuremath{e^{-\beta{\cal H}}} .
\end{eqnarray*}
Now fix $b_{min}$ and consider the configurations containing $k$ borders: if we remove the shortest border from such a configuration $\cal C$, we obtain a new configuration ${\cal C}'$ with $(k-1)$ borders containing $p$, each of them at least $b_{min}$ long. ${\cal C}'$ will be present in the partition function $\cal Z$, but different configurations with $k$ borders $\cal C$ may give the same ${\cal C}'$;
defining now $\mu^p(b)$ as the number of possible borders of $b$ links containing $p$, the degeneration induced by removing the shortest border is not greater than $\mu^p(b_{min})$.
The energy of a configuration $\cal C$ and the corresponding Boltzmann factor obey
\begin{eqnarray*}
&&E_{\cal C} \ge E_{{\cal C}'} + 2\beta J_{min} b_{min},\\
&&e^{-\beta{\cal H}_{{\cal C}}} \le e^{-{\cal H}_{{\cal C}'}}e^{-2\beta J_{min} b_{min}};
\end{eqnarray*}
if we now limit the sum in the partition functions to those configurations obtained by removing a border from the numerator, we can write
\begin{eqnarray*}
{\cal Z}^{-1} \sum_{k \ge 1} \sum_{{\cal C}|_{k\; borders}} \ensuremath{e^{-\beta{\cal H}}} \le \mu^p(b_{min})\; e^{-2\beta J_{min} b_{min}}.
\end{eqnarray*}
In this way the average number of minus spins is bounded by a function depending only on the number of borders encircling a given spin:
\begin{eqnarray*}
\langle N_-\rangle_{\cal N}\;\; \le \sum_{p\in{\cal G}} \sum_{b_{min}} \mu^p(b_{min})\;e^{-2 \beta J_{min} b_{min}};
\end{eqnarray*}
this result states that, no matter what the maximum number of spins you can isolate inside a border is, as long as $\mu^p(b)$ grows at most exponentially the value of $\langle N_-\rangle_{\cal N}$ can be limited at low enough temperatures.
An analogous result holds for configurations of class $\cal P$ when exchanging the roles of positive and negative spins:
\[ \langle N_+\rangle_{\cal P}\;\; \le \sum_{p\in{\cal G}} \sum_{b_{min}} \mu^p(b_{min})\;e^{-2 \beta J_{min} b_{min}}. \]
Let now $\mu(b)) = \sup_p\mu^p(b)$. We can now prove the following theorem:
\begin{theorem}
If on an infinite graph $\mu(b)\le A^{b}$, definitely for $b\ge\bar{b}$ and for some for some $A>0$, then the graph exhibits spontaneous magnetization at large enough $\beta$ (low enough temperatures).
\end{theorem}
\begin{proof}
The average modulus of magnetization is $|M| = N^{-1}\left( N_+-N_- \right)$; writing the Boltzmann factor for a configuration $\cal C$ as $P_{\cal C} = {\cal Z}^{-1} e^{-\beta {\cal H}({ \cal C })}$, we can write the following:
\begin{multline*}
\langle |M|\rangle\;\;=\;\sum_{\cal C \in P \cup N}\;|M_{\cal C}|\;P_{\cal C} \;= \sum_{\cal C \in (P \cup N)^+} M_{\cal C}\;P_{\cal C} - \sum_{\cal C \in (P \cup N)^-} M_{\cal C}\;P_{\cal C}\\
= \left( \sum_{\cal C \in N^+}\;M_{\cal C} P_{\cal C} - \sum_{\cal C \in N^-}\;M_{\cal C} P_{\cal C} \right) + \left( \sum_{\cal C \in P^+}\;M_{\cal C} P_{\cal C} - \sum_{\cal C \in P^-}\;M_{\cal C} P_{\cal C} \right) \\
= \left( \sum_{\cal C \in N}\;M_{\cal C} P_{\cal C} - \sum_{\cal C \in P}\;M_{\cal C} P_{\cal C} \right) + 2 \left( \sum_{\cal C \in P^+}\;M_{\cal C} P_{\cal C} - \sum_{\cal C \in N^-}\;M_{\cal C} P_{\cal C} \right)\\
\ge \left( \sum_{\cal C \in N}\;M_{\cal C} P_{\cal C} - \sum_{\cal C \in P}\;M_{\cal C} P_{\cal C} \right) = 1 - \frac{2}{N} \left( \sum_{\cal C \in N}\;(N_-)_{\cal C} P_{\cal C} + \sum_{\cal C \in P}\;(N_+)_{\cal C} P_{\cal C} \right)\\
\ge 1-\frac{4}{N} \sum_{p\in V} \sum_{b_{min}} \mu^p(b_{min})\;e^{-2 \beta J_{min} b_{min}}\\
\ge 1- 4 e^{-2\beta J_{min}} \sum_{b\ge 1} \mu(b_{min})\,e^{-2\beta (b-1) J_{min}},
\end{multline*}
When the sum on the last line converges, the equation tells us that, for large enough $\beta$ (low temperatures), $\langle |M|\rangle$ is greater than a positive constant, so that spontaneous magnetization on an infinite graph is achieved, while in general $\langle |M| \rangle$ is finite for every $N$, but can tend to zero as $N\rightarrow\infty$.
\end{proof}
The main problem in employing the previous theorem is determining bounds on $\mu(b)$.
The simplest case in which the hypothesis does not hold is a situation in which for some finite $b$ the number of borders surrounding a given point is infinite. As a sound check of the validity of the theorem, all the \emph{weakly separable}\cite{camparicassi2009} graphs, which do not magnetize, belong to this category.\par
To further our understanding of the result, we need to present a new parameter.
Given a subgraph $A\subset {\cal G}$, we define its external boundary $\partial A$ as the set of points in ${\cal G}\setminus A$ that have a bond to a point in $A$; denoting the number of vertices in $A$ as $\left\vert A \right\vert$ we now present the {\bf isoperimetric dimension} $d_{iso}$ as the minimum $d$ such that $\partial A \ge C\cdotp \left\vert A \right\vert^{d -1\over d}$. The largest set of points encompassable with $b$ links is thus smaller than $b^{d_{iso}\over d_{iso}-1}$; since these points are connected, $b^{d_{iso}\over d_{iso}-1}$ is also the maximum radius of a set including $i$ with a border $b$, so the set of reachable points, $V(b)\subset {\cal P}$, has a cardinality $\left\vert V(b)\right\vert \le b^{d_{iso} d_{frac}\over d_{iso}-1}$.\par
Given a point $p$ and for each border $\cal B$, consider $\bar{\cal B}$, the collection of vertices contributing to $\cal B$ which are on the inside of $\cal B$ with respect to $p$. As the two are in biunivocal relation once $p$ is chosen, counting the borders is the same as counting the vertex borders.
As a consequence of the previous paragraphs, the following holds:
\begin{proposition}
In a graph with isoperimetric dimension $d_{iso}>1$, the number of possible borders surrounding $p$ is bounded by
\[ \mu^p(b) \le \sum_q^{\left\vert V(b)\right\vert} N_q(b) \le b^{d_{iso} d_{frac}\over d_{iso}-1} \cdotp N_{Sup}(b), \]
where the sum is over the points $q$ which can be enclosed in a border of size $b$, $N_q(b)$ is the maximum number of vertex borders of length $b$, containing $p$, which can be created starting from the point $q$, and $N_{Sup}(b)=\sup_{q\in V(b)} N^p(b)$.
\end{proposition}
A border is \emph{connected} if the corresponding vertex border is a connected set.
We will need the following proposition regarding $N^p(b)$ to obtain a general result.
\begin{proposition}
The number of connected vertex borders starting from a given point $p$ grows at most exponentially with $b$: $N^p(b)\le const\cdotp C^b$.
\end{proposition}
\begin{proof}
A {\bf tree} is a graph that has no loops.
From each connected subgraph $A$ we can draw a number of different spanning trees, i.e. trees having the same set of points $\cal P$ as the original $A$. For any spanning tree we can construct a path visiting all its vertices in no less than $b-1$ steps, and in no more than $2 z_{Max} b$ steps.
While the former statement is obvious, we now prove the latter using the following algorithm:
starting from $i$, choose link and cross it; at each vertex on the path, choose a link not yet crossed; if there is no free link, step back through the link from which the path first arrived at the vertex. With this algorithm, each link is crossed no more than twice, so the path is of no more than $2 z_{Max} b$; furthermore, all the links are crossed, so, since the boundary is connected, all the points are visited.
As a consequence, all the spanning trees of $b$ points starting from a vertex $i$ can be constructed as paths of $b-1,b,\dots,2 z_{Max} b$ steps. Since each step can be chosen among at most $z_{Max}$ links, the total number of possible spanning trees, starting from $i$ and made of $b$ vertices, is less than
\[ z_{Max}^{b-1} + z_{Max}^{b} + \dots + z_{Max}^{2 z_{Max} b} = { z_{Max}^{2 z_{Max} b + 1 } - z_{Max}^{b-1} \over z_{Max}-1} , \]
and the thesis follows:
\[ N^p(b) \le const \cdotp C^b . \]
\end{proof}
When a vertex border is made of more disconnected parts, $\mu(b)$ grows exponentially if, for all borders, it's possible to connect all the parts using no more than $l\cdotp b$ vertices, where $l$ is a constant of the graph: in fact in this case to each border of length $b$ corresponds one connected vertex border of length between $b$ and $b\cdotp l$, so that
\[ N(b) \le C^{b}+C^{b+1}+\dots+C^{l\cdotp b} = {C^{ l\cdotp b + 1 } - C^b\over C-1}.\]
\begin{figure}
\includegraphics{Sierpinski_carpet.eps}%
\caption{\label{sierpinskicarpet} The Sierpinski carpet allows for disconnected borders, but they can be connected with no more with $l\cdotp b$ links, so the number of possible borders grows no faster than an exponential with $b$, and the graph magnetizes.}
\end{figure}
Noting that $d_{iso}>1$ implies that there is no border length $b$ for which $\mu^p(b)$ is infinite, the previous results can be combined to form the following theorem.
\begin{theorem}
For all graphs with isoperimetric dimension $d_{iso} > 1$ and vertex borders which are connectable with no more than $l\cdotp b$ vertices, a finite critical $\beta_c<\infty$ exists such that for all $\beta>\beta_c$ spontaneous magnetization is achieved.
\end{theorem}
To the latter category belong the regular lattices in $d\ge 2$ dimensions and crystals with any kind of elementary cells; we explicitly note that for an Euclidean lattice in $d=2$ dimensions we recover the result by Griffiths\cite{griffithspeierls}.
In addition, each vertex border can be connected with no more than $l\cdotp b$ vertices in the Sierpinski carpet too, which therefore magnetizes, in accord with the existing literature\cite{Vezzani2003} (see Fig.\ref{sierpinskicarpet}).
\par
Consider now the ladders of infinitely growing height (see Fig.\ref{growingladder}): they are structures described, at any offset $n$ on the semi-infinite base line, by a non-decreasing integer function $h(n)$; as long as the isoperimetric dimension of the ladder is strictly greater than one ($h(n)\ge A_0 n^\alpha + B_0$, for $\alpha>0$, $A_0>0$ and $B_0>0$) the previous arguments apply, so the ladder magnetizes; on the other hand, when $d_{iso}=1$ a little more work is required: the total number $V(b)$ of vertices which can be included in at least one border of length $b$ can not grow faster than the number of points on the left of the rightmost border of length $b$; the latter is at at offset $n(b)=\max\{ n' \vert b=h^{-1}(n') \}$, so that $V(b)$ satisfies $V(b) \sim \sum_{i=1}^{n(b)} h(i)$ for large $b$;
when
\[ B_0 + A_0 \log i \le h(i) \le B_1 + A_1 i^{\alpha} \]
for some $A_0,A_1 >0$, $B_0,B_1\ge 0$ and $\alpha\ge1$, the volume satisfies
\[ V(b) \le \int_1^{e^{b/A_0}} di\;\left( B_1 + A_1 i^\alpha \right) \sim {b\over A_0} {A_1\over 2 \left( \alpha+1 \right)} \; e^{2\left( \alpha+1 \right)b/A_0},\]
and since the borders are all connected $\mu(b)$ is exponential and the graph magnetizes.
When instead $\lim_{n\rightarrow\infty} {h(n)\over \log n} = 0$, for all $\epsilon>0$ and $n$ large enough $h(n)\le\epsilon \log n$ holds; as a consequence,
\[ V(b) \ge \int_0^{b\over\epsilon} di\;h(i) \;> \int_0^{b\over\epsilon}di = e^{b\over\epsilon} \]
holds for all $\epsilon>0$; in this case the sum $\sum_b \mu(b) e^{-2\beta \left( b-1 \right) J_{min}}$ diverges for all temperatures, so the hypotheses of our theorem are not fulfilled.
These results are in agreement with a result by Chayes and Chayes\cite{Chayes1986} about more general structures called $d$-wedges, where it is proved that $h(n)\ge \log n$ is both a sufficient and a necessary condition for spontaneous magnetization.
\begin{figure}
\includegraphics{growingladders.eps}%
\caption{\label{growingladder} Two examples of growing ladder graphs: in (a) all the borders are connected (remember that, following the definition we use here, a border divides a graph into exactly two subgraphs, each \emph{connected}), and as a consequence the graph will magnetize. In (b) too the borders are connected, and even if the growth in width of the ladder is very slow (logarithmic) the same result holds.}
\end{figure}
\par
\section*{Discussion}
To give a more intuitive interpretation of the theorem we proved, we can proceed as follows: if the number of borders grows less than exponentially, we can argue that all of these borders will contain a number of spins increasing slowly with the length of the border; as a consequence, the formation of large
clusters of spins in a magnetized graph will be energetically unfavoured, so that the latter will result a stable state.
On the other hand, if $\mu^i(b)$ grows very fast with $b$, we expect that some of the borders will be far from the vertex $i$, so that more and more vertices will be enclosed in short (low $b$) borders; this in turn means that large clusters of spins can be flipped spending a small amount of energy, so that a magnetized graph may be unstable with regard to thermal fluctuations.\par
The condition of our theorem is a strong one, in that it investigates a global property of the graph. For this reason it can not be a necessary condition for achieving spontaneous magnetization: if a graph has a part, which has zero measure in the thermodynamic limit, for which the number of boundaries $\mu^p(b)$ is greater than any exponential (e.g. a semi-infinite line connected to a point on a plane), the hypothesis of the theorem is false but the graph as a whole can still magnetize.\par
An important, yet straightforward, observation is that whenever a subgraph of non zero measure exists that is magnetizable, all the graph is magnetizable: in fact all the correlation functions, as computed on the subgraph, are smaller than or equal to the corresponding ones in the complete graph; when, on the other hand, the graph is formed by a collection of zero measure, weakly connected, magnetizable subgraphs (e.g. an infinite collection of parallel planes, each connected via a single link to the next one), there is no guarantee that $\langle \vert M \vert \rangle>0$.\par
Our result about the Ising model on graphs is a further step towards a full comprehension of the mechanism of phase transitions on general networks: together with a sufficient condition for the lack of spontaneous magnetization\cite{camparicassi2009}, it allows to ascertain the magnetizability of a large number of structures with a minimal amount of computation.\par
Further steps extending this work should aim at closing the gap between magnetizable and non-magnetizable graphs under the $\langle \vert M \vert\rangle$ definition, in order to identify a condition both necessary and sufficient for spontaneous magnetization; another direction of development could be to treat non symmetric situations, as in Pirogov-Sinai theory.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
In a previous work \cite{cct} it was shown that the Wigner's Quantization \cite{wig}, unlike the ordinary quantization based on creation and annihilation operators acting on a Fock vacuum, is compatible with a Hopf algebra structure of its Universal Enveloping (graded)-Lie algebra; it can therefore be regarded as the natural framework to investigate Hopf-algebra preserving, twist-deformations of quantum mechanical systems\footnote{We recall that the Wigner's Quantization
is based on super-Lie algebra valued operators acting on a vacuum state which corresponds to a lowest-weight representation; the ordinary quantization is recovered for a specific value of the lowest weight, which is nothing else than the Wigner's vacuum energy, see \cite{cct} and references therein for details.}. Due to the fact that the ordinary quantization is recovered for a special choice of the Wigner's vacuum energy it is quite important to understand whether
and under which prescription a Hopf algebra structure can be implemented for the ordinary quantization (creation and annihilation operators) as well. This is the viewpoint we are adopting in this paper. Essentially, the first of our results here can be stated as follows: composite operators entering the Hamiltonian and made with Heisenberg algebra operators (coordinates, momenta, the constant $\hbar$) have to be treated as primitive elements (generators) of a dynamical Lie algebra. Their ``composite" nature has to be disregarded and only their commutation relations with respect to the other primitive elements (generators) of the dynamical Lie algebra have to be retained. Within this framework the Universal Enveloping Algebra of the dynamical Lie algebra is endowed with a Hopf algebra structure.
\par
The next topic consists in applying an abelian Drinfeld twist which deforms the Universal Enveloping Algebra while preserving its Hopf algebra structure. Since the twist is defined in terms of the three momenta $p_i$, for consistency these generators have to be counted among the primitive elements of the dynamical Lie algebra. A deformed Universal Enveloping Algebra expressed in terms of the twist-deformed primitive elements and their twist-deformed brackets follows from this construction. The next point consists in investigating the behavior of $3D$ non-relativistic quantum mechanical systems which are originally (i.e., in the undeformed case) rotationally invariant. One is guaranteed that the $so(3)$ algebra is preserved by the twisted angular momenta under the twisted brackets. On the other hand those operators which, at the undeformed level, are rotationally invariant (since they commute with the ordinary angular momentum generators under the ordinary brackets) can become anomalous. This means that their twisted commutators with respect to the twisted angular momenta
can be non-vanishing. The non-zero result, called the deformation ``anomaly", vanishes when the deformation parameter goes to zero. The anomalous operators are expanded in Taylor-series of the deformation parameter ${\vec{\rho}}$. A specific choice of the higher-order contributions can be made in order to minimize the overall anomaly (a similar feature is also encountered for standard quantum anomalies). These considerations apply for both the deformation of (undeformed) rotationally invariant primitive elements, as well as (undeformed) rotationally invariant composite operators (the operator ${\vec L}^2$, which is a Casimir of the $so(3)$ subalgebra, but not a Casimir of the whole Euclidean algebra $e(3)$, is perhaps the most obvious example).
\par In an Appendix we provide some motivations for the special role played by both the twist-deformed generators and the twist-deformed brackets. On the other hand the connection between the abelian twist deformation and the non-commutative quantum mechanics results from the fact that the ordinary commutator between twist-deformed coordinates gives a constant matrix $\theta_{ij}$. This is a constant element of the Universal Enveloping Algebra and depends on the deformation
parameter ${\vec \rho}$.
Our work is naturally motivated by the recent upsurge of interest in Noncommutative (NC) theories, both from the condensed matter physics and quantum
gravity point of view. In the former case it has been known for a very long time that the guiding center coordinates of an
electron moving in a plane, but subjected to a constant (i.e. uniform and static) magnetic field, give rise to noncommutativity \cite{dun}.
This can have important consequences for example in QHE \cite{myle}. Besides, it can also arise due to Berry curvature effects
appearing from the breaking of time-reversal symmetry in ferromagnetic systems or from the breaking of spatial-inversion symmetry
in materials like GaAs crystals, as it has been been shown by Xiao et. al. \cite{xiao}.
In both these cases, the noncommutativity is of Moyal type, with time being the ordinary c-numbered variable.
On the other hand, it has been argued by Doplicher et. al. \cite{dfr}, by bringing in considerations of both
general relativity and quantum physics, that the nature of space-time is expected
to be fuzzy at the Planck-length scale. Similar conclusions were also drawn by \cite{sw} from low energy considerations of string theory.
Moyal type of Noncommutativity is one of the simplest types where these features can be realized.
\par
On the other hand, we have still to face the perennial problem of rotational/Lorentz symmetry in NC
theories defined in more than 2D. As already recalled in this paper we shall be basically considering the $3D$
problem, where the basic NC-ty among the spatial coordinates only is given by
\bea
[x_i,x_j]&=&i\theta_{ij}.
\eea
Clearly, the vector dual to $\theta_{ij}$, i.e.
\bea
\theta_i&=&{1\over 2}\epsilon_{ijk}\theta_{jk},
\eea
is pointing towards a particular direction, thereby violating the $SO(3)$ symmetry (note that the $2D$ case is safe from this problem).
Nevertheless,
it has been shown in the literature \cite{wes, abj} that this symmetry can be restored in a Hopf algebraic
setting by using a Drinfeld's twist, such that $\theta_{ij}$ remains invariant under the twisted action
of the rotation. This is in conformity with the usual philosophy of the twisted approach, where the matrix $\Theta= \{\theta_{ij}\}$
is regarded as a (matrix-valued) new constant of Nature like $\hbar, G, c $, etc. This is in contrast to other approaches followed in
literature (see for example \cite{aar}).
In the relativistic field theory this implies, in a similar manner, that the Poincar\'e
symmetry itself is restored, so that the usual Wigner classification of particles remains unchanged. Since these results
there has been a flurry of activities in this direction. Despite that,
it was still however not clear how one could investigate the simple QM in this framework\footnote{It should be recalled,
in this context, that the $x_i$'s are operators in QM, while they are mere c-numbered labels for the
continuous degrees of freedom in QFT and are not counted as members of the configuration space
of variables. Consequently, we have to impose the Heisenberg algebra between the coordinates
and the conjugate momenta in QM, i.e. $[x_i,p_j]=i\hbar \delta_{ij}$, but not in the case of
QFT. This distinction carries over to NC-QFT where, although the coordinates are promoted
to the level of operators, they are certainly not valued in the same space as the
field or other composite operators. Moreover, there is no conjugate momentum $p_i$ to the coordinate $x_i$.}.
For instance it was not clear how to define a rotationally invariant potential
even in the above mentioned framework of the twisted Hopf algebra. This is an important question,
considering the fact that the exact solution of the energy-spectrum of a particle, confined in
a noncommutative disc, has already been worked out using the method of piece-wise constant potential \cite{scgv},
which was subsequently used
to study the thermodynamics of a system of particles confined in such a disc \cite{sg}. This analysis had the additional virtue of being
carried on in a purely operatorial level, avoiding the pitfalls associated with the inequivalences between Moyal or Voros star product, which are currently debated in the
literature \cite{glv}.
This gives the main motivation for the present work, whose main results have been sketched before.\par
The scheme of the paper is as follows. In Section {\bf 2} we link (undeformed) Hopf algebras and Second Quantization, pointing out why some
operators should be regarded as ``primitive elements", while other operators should keep their ``composite" property. In Section {\bf 3} we review the needed facts and formulas concerning the abelian Drinfeld twist. In Section {\bf 4} we discuss the twisted rotations presenting general formulas for
the twisted brackets of the twisted angular momentum. In Section {\bf 5} we present the anomalous twist-deformed commutators for (twisted) primitive elements
such as the quadratic (harmonic), quartic and Coulomb potentials and for the deformation of the ${\vec L}^2$ composite operator. In the Appendix we give heuristic considerations motivating the use of both twisted generators and twisted brackets. Finally, in the Conclusions we make some extra comments on the results here found.
\section{Undeformed Hopf Algebras and Second Quantization}
Before addressing the problem of twisting Hopf algebras in association to NC Quantum Mechanics, we need to learn how to apply undeformed Hopf algebras
to Second Quantization. We work in the framework of the Hopf algebra structure of the Universal Enveloping Algebra of a Lie algebra (the Lie algebra itself is regarded as a dynamical symmetry of a quantum mechanical system). Our discussion has a general validity. For simplicity it will be illustrated with the basic examples of the Euclidean Lie algebras $e(2)$ and $e(3)$.\par
Additive operators (whose eigenvalues in a multi-particle state are the sum of the single-particle eigenvalues) have to be
assumed as ``primitive elements" of the dynamical symmetry algebra (i.e., as generators of the Lie algebra). This is because the additivity of the eigenvalues is encoded
in the undeformed coproduct. Indeed,
\bea
\Delta(\Omega)&=&\Omega\otimes{\bf 1}+{\bf 1}\otimes \Omega
\eea
encodes the additivity of the eigenvalues ($\omega_{1+2}=\omega_1+\omega_2$) for the operator $\Omega$.\\
This remark is still valid if the additive operator under consideration is a Casimir operator.
The Hamiltonian $H$ of a free-particle system is an additive operator. Therefore, $H=\frac{(\vec{p})^2}{2m}$ has to be regarded as a primitive element of the dynamical symmetry algebra despite its ``composite" nature. It is also a Casimir operator of the Euclidean algebra.
From physical considerations we are forced to reject the Hopf algebra equivalence $H={\vec p}^2$ (for simplicity we set $m=\frac{1}{2}$)
which would amount to consider ${\vec p}$ as an element of the Lie algebra, with $H$ beloging to the Enveloping algebra. This Hopf algebra equality
would imply the unphysical coproduct rule for $H$
\bea\label{wrongcop}
\Delta ( H)&=&\Delta({\vec p}^2)= \Delta({\vec p})\cdot\Delta({\vec p}) = {\vec p}^2\otimes {\bf 1} + {\bf 1}\otimes {\vec p}^2 + 2{\vec p}\otimes {\vec p}\neq H\otimes{\bf 1}+{\bf 1}\otimes H.\nonumber\\
&&
\eea
Note that a relation like (\ref{wrongcop}) makes perfect sense (in physical, as well as in mathematical considerations)
by replacing both ${\vec p}$ with the $3D$ angular momenta ${\vec L}$ and the free Hamiltonian $H$ with the $so(3)$
Casimir operator ${\vec L}^2$. Assuming (as it has to be done) that the components of ${\vec L}$
are Lie-algebra primitive elements, the coproduct
\bea
\Delta ({\vec L}^2)&=& \Delta({\vec L})\cdot\Delta({\vec L}) = {\vec L}^2\otimes {\bf 1} + {\bf 1}\otimes {\vec L}^2 + 2{\vec L}\otimes {\vec L}
\eea
reflects the fact that ${\vec L}^2$ is not an additive operator since, for a composite system, we have that
$({\vec L}_{1+2})^2=({\vec L}_1+{\vec L}_2)^2$.\par
Unlike $H$, which has to be assumed as a primitive element, ${\vec L}^2$ is a genuine composite operator.
As this example shows, the distinction between a ``primitive operator" versus a ``composite operator" cannot be done in purely
mathematical terms. Rather, the mathematical setting has to be accommodated to grasp the physical properties of the system under
investigation.\par
Additive operators have a direct interpretation in terms of their primitive coproducts. Composite operators, such as ${\vec L}^2$,
have no such direct interpretation. In this particular example, the eigenvalues of the composite system are obtained by decomposing
into direct sums the tensor products of the subsystems with the help of the Clebsch-Gordan coefficients.\par
When dealing with the Second Quantization we have to specify at first the single-particle states. This can be done by giving
a complete set
of mutually commuting observables. One should note that these observables can be either ``primitive", as well as ``composite" operators in the sense specifed above.
The discussion can be done in general. It is however useful to work out some specific examples that will be used in the following.
Let us consider the Euclidean Lie algebras $e(2)$ and $e(3)$, respectively.\par
$e(2)$ admits the three generators $p_1,p_2, L$, satisfying the commutation relations
\bea
\relax [p_1,L]&=& -ip_2,\nonumber\\
\relax [p_2,L]&=&ip_1,\nonumber\\
\relax [p_1,p_2]&=&0.
\eea
$e(2)$ admits only one Casimir operator, ${\cal C}\equiv{\vec p}^2={p_1}^2+{p_2}^2$. Indeed $\relax[{\vec p}^2,L]=0$. \par
The Casimir corresponds to the energy $E$ of a non-relativistic, free, two-dimensional particle (whose mass has been normalized, as before, to $m=\frac{1}{2}$).
Since the free energy is an additive operator, the Casimir operator ${\cal C}$ has to be added to the dynamical symmetry Lie algebra. For that we need to enlarge $e(2)$
by defining
\bea
{\overline{e(2)}}&=& e(2)\oplus u(1),
\eea
whose primitive generators are $\{p_1,p_2,L,{\cal C}\}$. ${\cal C}$ can be consistently identified with ${\vec p}^2$ as far as Lie-algebra and single-particle eigenvalues are concerned. We force this identification by imposing that the set of mutually commuting operators $p_1, p_2, {\cal C}$ admits the compatible set of respective eigenvalues $(\sqrt{E}\cos\alpha, \sqrt{E}\sin\alpha, E)$. This is not enough to completely specify the states in the Hilbert space because we still need to take into account the information carried on by the angular momentum $L$ (whose eigenvalues are the integers $m$). This can be done as follows. At first, without loss of generality, we fix the ``reference frame" specified by the eigenvalues $p_1=0$, $p_2=\sqrt{E}$
(recovered by setting $\alpha=\frac{\pi}{2}$). Next, we consider the little (Lie) group of transformations respecting the reference frame and their associated Lie-algebra, Hermitian, operators. We can now find a complete set of observable operators which are mutually ``weakly commuting" when the
reference frame constraint is taken into account. In the example above, mutually ``weakly commuting" observables are given by $p_2, L$, since
$\relax [p_2,L]=ip_1\approx 0$, when $p_1\equiv 0$ is taken into account.
\par
We can extend these considerations to the less trivial case of the three-dimensional Euclidean algebra $e(3)$, whose generators ($p_1,p_2,p_3,L_1,L_2,L_3$) satisfy the commutation relations
\bea\label{eucl3}
\relax [p_i,p_j]&=&0,\nonumber\\
\relax [p_i,L_j]&=& i\epsilon_{ijk}p_k,\nonumber\\
\relax [L_i,L_j]&=&i\epsilon_{ijk}L_k
\eea
(the $L_i$'s are the generators of the $so(3)$ subalgebra). \par
$e(3)$ admits two Casimir operators, ${\cal C}_1$, ${\cal C}_2$, given respectively by
\bea\label{twocasimir}
{\cal C}_1&=& {\vec p}^2,\nonumber\\
{\cal C}_2 &=& {\vec L}{\vec p}.
\eea
One should note that ${\vec L}^2$ is a Casimir operator of the $so(3)$ subalgebra; on the other hand it is not a Casimir operator for $e(3)$.\par
We can repeat the same construction as in the $e(2)$ case, enlarging the algebra to ${\overline{ e(3)}}$, by the addition of ${\cal C}_1$, ${\cal C}_2$ as primitive elements,
\bea\label{exte3}
{\overline {e(3)}}&=&e(3)\oplus u(1)\oplus u(1).
\eea
The identification (\ref{twocasimir}) is assumed to hold in the Lie algebra sense, but not in the Hopf algebra sense.\par
By setting $E$ (the energy) to be the eigenvalue of the ${\cal C}_1$ Casimir operator, without loss of generality we can work within the
$p_1=0, p_2=0, p_3=\sqrt{E}$ reference frame. In this reference frame ${\vec L}{\vec p}\equiv L_3p_3$, such that its eigenvalues are expressed by
$m\sqrt{E}$. A set of mutually ``weakly commuting" observables, respecting the given reference frame, is given by
$p_3, L_3, {\vec L}^2$, with eigenvalues $\sqrt{E}$,$ m$, $l(l+1)$, respectively. Indeed
\bea
\relax [p_3, L_3]&=&0,\nonumber\\
\relax[ L_3, {\vec L}^2]&=& 0,\nonumber\\
\relax [p_3, {\vec L}^2]&\approx&0.
\eea
A state of the system is uniquely specified in terms of its free energy $E$, the orbital angular momentum $l$ and its component along the third axis $m$.\par
In the set of three, mutually weakly-commuting operators which specify the state of the (single-particle) system, two of them ($p_3, L_3$) are primitive operators, while the remaining one (${\vec L}^2$) is a composite operator.
\section{The abelian Drinfeld twist}
In this Section we will recall the basic formulas concerning the abelian Drinfeld twist deformation of
the Universal Enveloping Algebra ${\cal U}({\bf g})$ of a given Lie algebra ${\bf g}$. For our purposes the twist is expressed by
${\cal F}\in {\cal U}({\bf g})\otimes {\cal U}({\bf g})$, such as
\bea\label{abeliantwist}
{\cal F}&=& \exp (i\rho_{ij} p_i\otimes p_j),\nonumber\\
\rho_{ij}&=& \epsilon_{ijk}\rho_k,
\eea
where ${\vec \rho}$ is a dimensional $c$-number and $p_i$ ($i,j=1,2,3$) are the three-dimensional momenta.
It is obviously required that $p_i\in {\bf g}$.\par
The twist induces a deformation in the Hopf algebra ${\cal U}({\bf g})\rightarrow {\cal U}^{\cal F}({\bf g})$ (see \cite{asc}). Particularly, the co-structures are deformed. The deformed co-structures (coproduct, counit and antipode), applied to an element $g\in {\bf g}$, are respectively given by
\bea
\Delta^{\cal F} (g)&=& {\cal F} \Delta (g) {\cal F}^{-1},\nonumber\\
\varepsilon^{\cal F} (g) &=& \varepsilon (g),\nonumber \\
S^{\cal F} (g) &=& \chi S(g) \chi^{-1},
\eea
where
\bea
\chi &=& f^\alpha S(f_\alpha)\in {\cal U}({\bf g})
\eea
(we are denoting, as usual, ${\cal F}=f^\alpha\otimes f_\alpha$, ${\cal F}^{-1} = {\overline f}^\alpha\otimes {\overline f}_\alpha$).\par
The generators of ${\cal U}^{\cal F}({\bf g})$ are expressed as
\bea\label{defgen}
g^{\cal F}&=& {\overline f}^\alpha (g){\overline f}_\alpha.
\eea
The ${\cal F}$-deformed brackets in ${\cal U}^{\cal F}({\bf g})$ are defined through
\bea\label{Fbrackets}
\relax [g^{\cal F}, h^{\cal F}]_{\cal F} &=& {g^{\cal F}}_1 h^{\cal F}S(g^{\cal F})_2,
\eea
where the Sweedler's notation
\bea\label{sweed}
\Delta^{\cal F}(g^{\cal F}) &=& (g^{\cal F})_1\otimes ({g}^{\cal F})_2
\eea
has been used.\par
The ${\cal F}$-deformed brackets satisfy the Jacobi identity.\par
A more complete list of the properties of the twist-deformed Hopf algebra ${\cal U}^{\cal F}({\bf g})$ is encountered in \cite{asc}.
\par
The Universal Enveloping Algebra of the following Lie algebras can be deformed in terms of the (\ref{abeliantwist}) abelian twist. We have\par
{\em i}) the Heisenberg algebra $h_B(3)$, whose generators are $\hbar, x_i, p_i$ (for $i=1,2,3$). $\hbar$
is a central element and the only non-vanishing commutation relations are given by
\bea\label{heiscomm}
\relax [x_i,p_j]&=& i\delta_{ij}\hbar;
\eea
{\em ii}) the Euclidean algebra $e(3)$ (considered in the previous Section), whose generators are $p_i, L_i$.
Its non-vanishing commutators are given by
\bea
\relax [p_i, L_j]&=& i\epsilon_{ijk} p_k,\nonumber\\
\relax [L_i,L_j]&=& i\epsilon_{ijk}L_k.
\eea
This algebra can be induced by the $h_B(3)$ Heisenberg algebra after setting
\bea\label{angmom}
L_i &=& \frac{1}{\hbar}\epsilon_{ijk}x_jp_k
\eea
and interpreting the $L_i$'s as primitive elements\footnote{As recalled in the previous Section, the notion of ``primitive elements'' is used to underline the fact that the generators of the Lie algebra should not be regarded as composite operators of the Hopf algebra structure.}. Similarly, the extended algebra ${\overline {e(3)}}$ introduced in (\ref{exte3}) can be twist-deformed under (\ref{abeliantwist});
\par
{\em iii}) the algebra $g$, whose primitive elements are the Heisenberg algebra generators ${\hbar, x_i, p_i}$ and the angular momentum generators $L_i$ whose commutation relations, as before, can be induced by the
(\ref{angmom}) position;
\par
{\em iv}) the ``oscillator'' algebra ${osc}$, given by the set of primitive elements \\
$\hbar, x_i, p_i, L_i,
H,K,D$. The commutators involving the generators $H,K,D$ can be read from the positions
\bea
H&=& \frac{1}{\hbar} {\vec p}^2,\nonumber\\
D&=& \frac{1}{2\hbar}({\vec x}{\vec p}+{\vec p}{\vec x}),\nonumber\\
K&=& \frac{1}{\hbar}{\vec x}^2
\eea
(the $\hbar$ at the denominator in the r.h.s. expressions is required in order to compensate the corresponding term coming from the (\ref{heiscomm}) commutators).
$H,K,D$ defines the $sl(2)$ subalgebra. The complete set of non-vanishing commutators among primitive elements of $osc$ is given by
\bea\label{ocomm}
\relax [x_i,p_j]&=&i\delta_{ij}\hbar,\nonumber\\
\relax [x_i, L_j] &=&i\epsilon_{ijk}x_k,\nonumber\\
\relax [p_i, L_j]&=& i\epsilon_{ijk}p_k,\nonumber\\
\relax [x_i, H] &=& 2ip_i,\nonumber\\
\relax [x_i,D]&=& ix_i,\nonumber\\
\relax [p_i, D] &=& -i p_i,\nonumber\\
\relax [p_i,K]&=& - 2 i x_i,\nonumber\\
\relax [H, D]&=& -2i H,\nonumber\\
\relax [H,K]&=& -4 i D,\nonumber\\
\relax [D,K]&=& - 2i K.
\eea
The Hamiltonian of the harmonic oscillator is given by a linear combination of $H$ and $K$;
{\em v}) a finite Lie algebra $g_b$ of hermitian operators which can all be regarded as primitive elements and recovered from at most bilinear combinations in ${\vec x}$ and ${\vec p}$, is given by
the set of generators $\hbar, x_i, p_i, P_{ij}, X_{ij}, {M^+}_{ij}, {M^-}_{ij}$. The commutation relations
involving $P_{ij}, X_{ij}, {M^+}_{ij}, {M^-}_{ij}$ can be read by assuming
\bea
P_{ij} &=& \frac{1}{\hbar}p_ip_j,\nonumber\\
X_{ij} &=& \frac{1}{\hbar}x_ix_j,\nonumber\\
{M^+}_{ij} &=&\frac{1}{\hbar}( x_ip_j+p_jx_i),\nonumber\\
{M^-}_{ij}&=& \frac{i}{\hbar}(x_ip_j-p_jx_i);
\eea
{\em vi}) the above construction can be further generalized. Any Lie algebra containing $h_B(3)$ as a subalgebra and at least one primitive element which is expressed as a trilinear (or $k$-linear, for $k\geq 3$) combination in ${\vec x}$ and ${\vec p}$ is necessarily infinite-dimensional. Indeed, the closure of the commutation relations of this generator with the previous ones requires that new higher-order multilinear terms have to be included as primitive elements. This procedure never stops, leading to an infinite-dimensional Lie algebra. This algebra can be regarded as the unfolded algebra
of primitive elements (the multilinear combinations in terms of ${\vec x}$ and ${\vec p}$ is its folded version, in analogy of what happens, in a different context, with finite $W$-algebras \cite{dBT} or the unfolded version of higher-spin algebras,
see \cite{vas}). One should note that a primitive element which is $k$-linear in ${\vec x}, {\vec p}$, requires the $\frac{1}{\hbar^{k-1}}$ factor (for instance, a primitive element can be associated to $\frac{1}{\hbar^3}({\vec x^2})^2$).
\section{Twisted rotations}
The abelian Drinfeld twist (\ref{abeliantwist}) induces, through eq. (\ref{defgen}), the following deformation of the space coordinates
\bea\label{xf}
{x_i}^{\cal F} &=& x_i-\epsilon_{ijk}\rho_k \hbar p_j.
\eea
This deformation corresponds to the Bopp shift and one should note that the second term in the r.h.s. is quadratic in the Heisenberg algebra generators.
This result was also obtained in \cite{cct}.
The shift maps $x_i\in h_B(3)$ into ${x_i}^{\cal F}\in {\cal U} (h_B(3))$.
Concerning the $p_i$ momenta, they undergo no deformation: $p_i^{ \cal F}=p_i$.
\par
The non-commutative quantum mechanics (for a constant operator $\theta_{ij}$) is recovered from the abelian twist. Indeed
\bea\label{ncemer}
[{x_i}^{\cal F},{x_j}^{\cal F}] &=& i\theta_{ij},
\eea
where
\bea\label{nctwist}
\theta_{ij}&=& 2\hbar^2\epsilon_{ijk}\rho_k,
\label{theta-rho}
\eea
with $\theta_{ij}$ an operator belonging to ${\cal U}(h_B(3))$.
Similarly, but in the ``opposite'' direction, the ${\cal F}$-commutator of the ordinary coordinates produces
\bea
[x_i,x_j]_{\cal F} &=& -\frac{1}{2}i\theta_{ij}.
\eea
The ${\cal F}$-commutator among twisted space coordinates is vanishing
\bea
\relax [{x_i}^{\cal F},{x_j}^{\cal F}]_{\cal F} &=& 0.
\eea
The twisted coproduct of the space coordinates and of the twisted space coordinates is respectively given by
\bea
\Delta^{\cal F} (x_i) &=& x_i\otimes {\bf 1}+{\bf 1}\otimes x_i +\epsilon_{ijk}\rho_k(\hbar\otimes p_j-p_j\otimes \hbar), \nonumber\\
\Delta^{\cal F} ({x_i}^{\cal F}) &=& {x_i}^{\cal F}\otimes {\bf 1}+{\bf 1}\otimes {x_i}^{\cal F} -2\epsilon_{ijk}\rho_k p_j\otimes \hbar.
\eea
If the algebra admits as primitive elements, besides the $p_i$'s, the angular momentum operators $L_i$, their deformation ${L_i}^{\cal F}$, induced by the (\ref{abeliantwist}) twist is given by
\bea\label{defL}
{L_i}^{\cal F} &=& L_i+K_i,\nonumber\\
K_i&=& \rho_kp_ip_k-\rho_ip_kp_k.
\eea
The extra-term $K_i$ can also be written as
\bea
K_i&=& -\rho_j{\vec p}^2\Pi_{ij},
\eea
in terms of the $\Pi_{ij}$ projector
\bea
\Pi_{ij}&=& (\delta_{ij}-\frac{p_ip_j}{{\vec p}^2}).
\eea
The twisted coproduct of the (twisted) angular momentum reads as
\bea
\Delta^{\cal F} (L_i) &=& L_i\otimes {\bf 1}+{\bf 1}\otimes L_i +\rho_{k}(p_i\otimes p_k-p_k\otimes p_i),
\nonumber\\
\Delta^{\cal F} ({L_i}^{\cal F}) &=& {L_i}^{\cal F}\otimes {\bf 1}+{\bf 1}\otimes {L_i}^{\cal F} +2\rho_{k}p_i\otimes p_k-2\rho_i p_k\otimes p_k.
\eea
The original $su(2)$ rotational algebra is recovered in terms of the ${\cal F}$-commutator of the twisted angular momentum. We have indeed
\bea
[{L_i}^{\cal F},{L_j}^{\cal F}]_{\cal F} &=& i\epsilon_{ijk}{L_k}^{\cal F}.
\eea
As a consequence we get the first result, namely that the rotational symmetry is preserved by the (\ref{abeliantwist}) twist-deformation.\par
One can also check that
\bea
\relax [x_i^{\cal F},L_j^{\cal F}]_{\cal F}&=&i\epsilon_{ijk}x_k^{\cal F}, \nonumber\\
\relax [p_i^{\cal F},L_j^{\cal F}]_{\cal F}&=&[p_i,L_j^{\cal F}]_{\cal F}=
i\epsilon_{ijk}p_k^{\cal F},
\eea
showing that both $x_i^{\cal F}$ and $p_i$ have vectorial transformation properties under the deformed brackets. The situation is thus different
from the case of the undeformed brackets where $x_i^{\cal F}$, unlike $p_i$, fails to transform as a vector \cite{cct}.\par
The next important point is to check whether the operators which are rotationally invariant in the undeformed case, keep the rotational invariant property even in the deformed case or otherwise acquire an anomalous term which disappears in the limit ${\vec \rho}\rightarrow 0$.
We investigate, specifically, the commutation relations
\bea\label{defcomm}
&\relax [ {L_i}^{\cal F}, B^\sharp]_{\cal F}&
\eea
for an operator $B^\sharp$ belonging to the Universal Enveloping Algebra of a Lie algebra containing the Euclidean algebra $e(3)$ as a subalgebra and such that $B^\sharp$ is expanded in ${\vec \rho}$ Taylor series:
\bea\label{taylor}\label{taylorexp}
B^\sharp&=& B_0+B_1+B_2+\ldots,
\eea
with $B_k$ $k$-linear in ${\vec \rho}$. Here $B_0\equiv B$ denotes the undeformed limit for ${\vec \rho}\rightarrow 0$ of $B^\sharp$
(we can therefore say that the operator $B^\sharp$ is the deformation of $B$).\par
The rotational invariance in the undeformed limit requires that the following relation involving ordinary commutators and angular momentum operators has to be satisfied
\bea
\relax [L_i, B_0] &=& 0.
\eea
With a little algebra one can easily prove that the deformed commutator (\ref{defcomm}) can be expressed in terms of
ordinary commutators:
\bea\label{defcommexp}
\relax [ {L_i}^{\cal F}, B^\sharp]_{\cal F}&=& [L_i-K_i,B^\sharp] + M_{ik}[p_k,B^\sharp],
\eea
where $K_i$ enters (\ref{defL}) and $M_{ik}$ is given by
\bea
M_{ik} &=& 2\rho_kp_i-2\rho_ip_k.
\eea
The r.h.s. in (\ref{defcommexp}) is a consequence of the equality
\bea
2K_i-M_{ik}p_k&=&0.
\eea
The meaning of the ${\cal F}$-deformed brackets for non-commutative theories is discussed in the Appendix.\par
It is worth pointing out that the Taylor-expanded series (\ref{taylor}) starting with $B_0$ does not necessarily coincide with the ${\cal F}$-deformed operator ${B_0}^{\cal F}$ (which can also be understood as
Taylor-expanded). In the next Section we will discuss this point in more detail.\par
For completeness we write here the ${\cal F}$-deformed operators $H^{\cal F}, D^{\cal F}, K^{\cal F}$ obtained by applying the
(\ref{abeliantwist}) twist to the $H,D,K$ primitive elements of the oscillator algebra $osc$ given in (\ref{ocomm}).
We have
\bea
H^{\cal F} &=& H,\nonumber\\
D^{\cal F} &=& D, \nonumber\\
\relax K^{\cal F} &=& K-2\epsilon_{ijk}\rho_kx_ip_j+\hbar [{\vec \rho}^2{\vec p}^2 -({\vec \rho}{\vec p})^2].
\eea
\section{Anomalous operators}
For our purposes it is useful to set
\bea
\relax L_i (B_n) &=& [L_i, B_n],\nonumber\\
\relax T_i(B_n) &=& -[K_i,B_n]+M_{ik}[p_k,B_n].
\eea
An undeformed rotationally invariant operator $B$ such that
\bea
\relax [{L_i} , B] &=& 0
\eea
can develop, under deformation, an anomaly $A_i$ which is expressed through
\bea\label{anomaly}
\relax [{L_i}^{\cal F} , B^\sharp]_{\cal F} &=& A_i
\eea
(as discussed in the previous Section, $B^\sharp$ is the deformation of $B$).\par
The anomaly $A_i$ can be expanded in powers of the deformation parameter ${\vec \rho}$. We have
\bea\label{iteranom}
\relax (A_i)_0&=& L_i(B_0)=0,\nonumber\\
\relax (A_i)_n &=& L_i(B_n) + T_i(B_{n-1}),
\eea
with $(A_i)_n$ the $n$-th order contribution in ${\vec \rho}$.\par
Let us consider now the deformation of the rotationally invariant primitive elements $H,D,K$ of
the ${{osc}}$ oscillator algebra (\ref{ocomm}). We get that
\bea
&H^\sharp=H^{\cal F}=H&
\eea
is rotationally invariant under twist-deformed rotations since
\bea
\relax [{L_i}^{\cal F} , H^{\cal F}]_{\cal F} &=& 0.
\eea
Similarly,
\bea
&D^\sharp=D^{\cal F}=D&
\eea
is rotationally invariant under twist-deformed rotations since
\bea
\relax [{L_i}^{\cal F} , D^{\cal F}]_{\cal F} &=& 0.
\eea
On the other hand we get that $K$ gets anomalous since
\bea
K^\sharp&=&K-2\epsilon_{ijk}\rho_kx_ip_j-\hbar ({\vec \rho}{\vec p})^2+\gamma\hbar
{\vec \rho}^2{\vec p}^2,
\eea
which coincides with $K^{\cal F}$ for the special value $\gamma=1$, namely
\bea
{K^\sharp}|_{\gamma=1}&=&K^{\cal F},
\eea
is such that
\bea\label{constanom}
\relax [{L_i}^{\cal F} , K^{\sharp}]_{\cal F} &=& 4\hbar \rho_i.
\eea
The r.h.s. term $4\hbar\rho_i$, which is independent of $\gamma$, is the (constant) anomalous operator.\footnote{In this context
it should be recalled that $\rho_i$,
despite its appearance,
transforms as a scalar under rotations. In the same spirit, despite its appearance, $({\vec \rho} {\vec p})$
is not a scalar under
rotations.}
\par
The analysis of the anomaly can be performed also for composite operators. Let us consider the
Euclidean algebra $e(3)$ defined in (\ref{eucl3}). We investigate the ${({\vec L}^2)}^\sharp$ deformation of the composite operator
${\vec L}^2$, which is the Casimir of the $so(3)$ subalgebra. In accordance with the (\ref{taylorexp}) expansion and the (\ref{defcommexp})
equation, the ${({\vec L}^2)}^\sharp$ Taylor-expansion in ${\vec\rho}$ stops at the second order for a minimal anomaly. We indeed get
\bea
({\vec L}^2)^\sharp &=& ({\vec L}^2)_0 +({\vec L}^2)_1 +({\vec L}^2)_2,
\eea
with
\bea
({\vec L}^2)_0&=& {\vec L}^2,\nonumber\\
({\vec L}^2)_1&=& \alpha_1 ({\vec \rho}{\vec p})({\vec p}{\vec L}) +\alpha_2 {\vec p}^2({\vec \rho}{\vec L}),\nonumber\\
({\vec L}^2)_2&=& \beta_1{\vec \rho}^2({{\vec p}^2})^2+\beta_2 ({\vec \rho}{\vec p})^2{\vec p}^2.
\eea
The anomalous deformed commutator, given by
\bea\label{anomcas}
\relax [{L_i}^{\cal F} , ({\vec L}^2)^\sharp]_{\cal F} &=& 4 \rho_i{\vec p}^2 +2i\epsilon_{kjl}\rho_kp_lL_j-2i({\vec \rho}{\vec p}) \epsilon_{ijl}p_lL_j +\nonumber\\&&
\alpha_1 i\epsilon_{ijk}\rho_jp_k({\vec p}{\vec L})+\alpha_2 i\epsilon_{ijk}({\vec p}^2)\rho_jL_k+
(2\beta_2-\alpha_2)i\epsilon_{ijk}\rho_jp_k({\vec \rho}{\vec p}){\vec p}^2,\nonumber\\
&&
\eea
does not depend on the $\beta_1$ coefficient. The minimal anomaly is recovered by setting
\bea
&\alpha_1=\alpha_2=\beta_2=0.&
\eea
The minimal anomaly is therefore given by the first line in the r.h.s. of (\ref{anomcas}).\par
The deformed rotational anomaly can be discussed for more general potential terms. Let us consider the addition of an
an anharmonic quartic term, given by
\bea\label{anham}
B= B_0 &=& \lambda \frac{({\vec x}^2)^2}{\hbar^3},
\eea
to the harmonic oscillator potential. In the above formula $\lambda$ is a positive coupling constant which, for simplicity, will be set equal to $1$. The $\frac{1}{\hbar^3}$ factor is introduced, as recalled at the end of Section {\bf 3}, in order to make (\ref{anham}) a primitive element of an infinite-dimensional Lie algebra $g$ containing the oscillator algebra $osc$ (\ref{ocomm}) as a finite subalgebra (since $\hbar$ is a central element of the Lie algebra $g$ one can evaluate it by setting, as usual, $\hbar=1$).\par
The $B^\sharp$ expansion starting from $B_0$ given in (\ref{anham}) produces a minimal anomaly for $B_k=0$ for $k\geq 5$.
By using the iterative procedure (\ref{iteranom}), we obtain that
\bea
B_1\equiv\left(\frac{({\vec x}^2)^2}{\hbar^3}\right)_1&=&\frac{ 4i}{\hbar^2}(p_i({\vec \rho}{\vec x})-({\vec \rho}{\vec x})x_i){\vec x}^2
\eea
gives the minimal anomaly of the first order, expressed by
\bea
(A_i)_1 &=& \frac{8}{\hbar}(2\rho_i{\vec x}^2-x_i({\vec \rho}{\vec x})).
\eea
At the next order we have that
\bea
B_2\equiv\left(\frac{({\vec x}^2)^2}{\hbar^3}\right)_2&=&\frac{ 2}{\hbar}\left(
2(\epsilon_{ijk}\rho_ip_jx_k)(\epsilon_{lmn}\rho_lp_mx_n)-({\vec \rho}{\vec p})^2{\vec x}^2\right)+
i\alpha ({\vec \rho}{\vec p})({\vec \rho}{\vec x})
\eea
produces an anomalous term $(A_i)_2$, given by
\bea\label{ambanom}
\left(A_i\right)_2 &=& \frac{16 i}{\hbar} \rho_i(\epsilon_{ljk}\rho_l p_jx_k) ({\vec p}{\vec x})-
16 \rho_i (\epsilon_{ljk}\rho_l p_j x_k)
+\nonumber\\
&&
-(4+\alpha )({\vec \rho}{\vec p})(\epsilon_{ijk} \rho_j x_k)
-(12+\alpha )(\epsilon_{ijk} \rho_j p_k)({\vec \rho}{\vec x})
.
\eea
The choice $\alpha=-4$ (respectively, $\alpha=-12$) makes disappear the third (fourth) term in the right hand side.\par
In the final example we consider the deformation of the Coulomb potential
\bea
B=B_0&=& \frac{1}{r},
\eea
with $r=\sqrt{{\vec x}^2}$.\footnote{It is worth to mention that it is consistent to produce the infinite-dimensional Lie algebra $g$ of primitive elements obtained by repeatedly applying the commutation relations to the generating elements $\hbar, x_i, p_i, \frac{1}{r}$.} \par
Unlike the expansion for the (\ref{anham}) anharmonic oscillator potential, the iterative procedure in this case never stops ($B_k\neq0$ at all orders).\par
Due to
\bea\label{tibo}
T_i(B_0)&=& i\hbar\rho_k(p_ix_k-p_kx_i)\frac{1}{r^3}+\hbar^2(\frac{\rho_i}{r^3}-3x_i\frac{{\vec\rho}{\vec x}}{r^5}),
\eea
we can set
\bea
B_1=\left(\frac{1}{r}\right)_1 &=& \alpha \hbar \epsilon_{ljk}\rho_lp_jx_k\frac{1}{r^3},
\eea
so that
\bea
L_i(B_1) &=& i \alpha\hbar (-{\vec \rho}{\vec p} x_i+p_i{\vec \rho}{\vec x})\frac{1}{r^3}.
\eea
By choosing
\bea
\alpha&=& -1,
\eea
the term proportional to $\hbar$ in (\ref{tibo}) can be reabsorbed. On the other hand, one can easily see that the anomalous term
$\hbar^2(\frac{\rho_i}{r^3}-3x_i\frac{{\vec\rho}{\vec x}}{r^5})$ cannot be reabsorbed by the contributions
coming from the higher order terms $\left(\frac{1}{r}\right)_k$ for $k>1$.\par
Summarizing, we get
\bea
\left(\frac{1}{r}\right)^\sharp &=& \frac{1}{r}-\hbar\epsilon_{ljk}\rho_lp_jx_k\frac{1}{r^3} + O(\hbar^2),
\eea
satisfying the anomalous twist-deformed commutator (with minimal anomaly)
\bea
\relax \left[{L_i}^{\cal F} , \left(\frac{1}{r}\right)^\sharp\right]_{\cal F} &=& \hbar^2(\frac{\rho_i}{r^3}-3x_i\frac{{\vec\rho}{\vec x}}{r^5})+O(\hbar^3).
\eea
\section{Conclusions}
In this work we investigated the Non-commutative Quantum Mechanics as a result of an abelian Drinfeld twist provided by formula (\ref{abeliantwist}). The twist deforms the Hopf algebra defined on the Universal Enveloping Algebra of a suitable Lie algebra. The Lie algebra under consideration, named {\it dynamical Lie algebra}, contains composite operators of the Heisenberg algebra operators $x_i, p_i, \hbar$. Nevertheless, these composite operators should be treated
as primitive elements, i.e. as generators, of the dynamical Lie algebra. Besides the Hamiltonian, the dynamical Lie algebra also includes, among its generators, the momenta $p_i$
in order to have a well-defined action of the (\ref{abeliantwist}) twist on its Enveloping Algebra endowed with the Hopf algebra structure. The dynamical Lie algebra
is either finite if its primitive elements result from composite operators at most quadratic in $x_i,p_i$; it is an infinite Lie algebra otherwise. \par
We gave motivations for the use of ${\cal F}$-deformed generators and ${\cal F}$-deformed commutators in dealing with the twist-deformed Enveloping Algebra and pointed out the connection between twist and non-commutativity, which is provided by the equations (\ref{ncemer}) and (\ref{nctwist}), with $\theta_{ij}$ a constant operator belonging
to ${\cal U}(h_B(3))$. \par
The ${\cal F}$-deformed angular momenta ${L_i}^{\cal F}$ close the $so(3)$ algebra under ${\cal F}$-deformed brackets. On the the other hand, several operators which
at the undeformed level are rotationally invariant are anomalous in the deformed case, with the anomaly expressed by equation (\ref{anomaly}). We discussed various examples of anomalous operators. For the (deformed) $3D$ harmonic oscillator potential the anomaly is a linear constant operator, see (\ref{constanom}). In more general cases the anomaly is an operator which belongs to the Enveloping Algebra and is not necessarily constant. The concept of ``minimal anomaly" can be introduced. It corresponds to the specific choice, made so that to minimize the r.h.s. of equation (\ref{anomaly}), of the higher-order terms in the ${\vec \rho}$ Taylor-expansion of the deformed operator $B^\sharp$. In some cases the notion of minimal anomaly becomes ambiguous. This is the case for instance of formula (\ref{ambanom}). A (different) term contributing to the anomaly is eliminated by choosing the arbitrary parameter $\alpha$ to be either $\alpha=-4$ or $\alpha=-12$.\par
We also discussed the anomalous property of the twist-deformed Coulomb potential and of the twist-deformed $so(3)$ composite Casimir operator ${\vec L}^2$,
regraded as belonging to ${\cal U}(e(3))$, the Enveloping Algebra of the three-dimensional Euclidean algebra.\par
In a different (not involving the Drinfeld twist) context from ours, non-commutative Quantum Mechanics has been studied in several works, see e.g. \cite{cst},
where the non-commutative hydrogen atom was discussed, and \cite{glr}. In \cite{bg} (see also {\cite{der} and \cite{dh}) the classical counterpart of the non-commutative quantum mechanics is shown to be a constrained system. In \cite{gk} investigations of a dynamical (i.e. non-constant) non-commutative matrix
$\theta_{ij}$ were made.
\\ ~
\\ ~
\par {\large{\bf Acknowledgments}}{} ~\\{}~\par
We have profited of clarifying discussions with P. Aschieri on the properties of the ${\cal F}$-deformed commutators.
We acknowledge several useful discussions with P. G. Castro.
Z. K. and F. T. are grateful to the S. N. Bose National Center for Basic Sciences of Kolkata for hospitality.
B. C. acknowledges a TWAS-UNESCO associateship appointment at CBPF
and CNPq for financial support.
The work was supported by Edital Universal CNPq, Proc. 472903/2008-0 (Z.K., F.T).
~\\{}~\par
{{\Large \bf Appendix}}
~\par
~\par
The use of the ${\cal F}$-deformed brackets (\ref{Fbrackets}) in association with Non-commutative theories can be argued to appear naturally according to the following heuristic considerations. Let us consider a general Lie algebra ${\bf g}$ corresponding to a group
$G$ and its Universal Enveloping Lie algebra ${\cal U}({\bf g})$. Let the generators satisfy the commutation relations
\bea
[T_a,T_b]&=&if_{ab}^cT_c.
\eea
The adjoint representation of the group $G$ is obtained by the adjoint action of the group element,
according to
\bea\label{adjact}
T_a \rightarrow gT_ag^{-1}&=& D_{ab}T_b.
\eea
The matrices $D$ provide the adjoint representation for $g\in G$ and satisfy $D(g_1)D(g_2)=D(g_1g_2)$. In the infinitesimal version, near the identity, we can express $g$ as
\bea
g&\approx& {\bf 1}+i\omega_aT_a,
\eea
so that the above transformation (\ref{adjact}) reads as
\bea
T_a \rightarrow T'_a&=&T_a+\delta T_a,
\eea
with $\delta T_a=i\omega_b[T_b,T_a]$.
Before we generalize this construction to the NC case, we need to recast the commutative case itself in a
bit different setting. For that we can work with the group $G$, endowed with a Hopf group-algebra structure \cite{asc}. In the undeformed case, the undeformed
coproduct of $g$ is $\Delta_0(g)=g\otimes g$ and the antipode is $S(g)=g^{-1}$. Near the identity, it is equivalent to the following assignments of antipodes to the Lie algebra generators
$T_a$ and the identity ${\bf 1}$:
\bea
&S(T_a)=-T_a,\quad\quad S({\bf 1})={\bf 1}.
\eea
Now one can easily see that the finite form of the adjoint action on the group element is
\bea\label{adjactant}
T_a\rightarrow gT_ag^{-1}&=&gT_aS(g).
\eea
It is therefore associated with the coproduct $\Delta_0(g)=g\otimes g$.\par
When considering again $g$ to be close to the identity, the corresponding infinitesimal version of the coproduct is given by
\bea
\Delta_0(g)&=&{\bf 1}\otimes{\bf 1}+i\omega_b \Delta_0 (T_b),
\eea
where
\bea
\Delta_0 (T_b)&=& T_b\otimes {\bf 1}+{\bf 1}\otimes T_b
\eea
is the undeformed coproduct of the Lie-algebra generator $T_b$.
Just like ({\ref{adjactant}) the infinitesimal transformation of $T_a$ can be written with the help of the antipode
\bea
T_a \rightarrow T'_{a}&=&{\bf 1} T_a S({\bf 1})+i\omega_b(T_b T_a S({\bf 1})+{\bf 1} T_a S(T_b))= T_a+i\omega_b [T_b,T_a].
\eea
From this perspective the commutator brackets are associated with the undeformed coproduct.
Once this standpoint is adopted, it is natural to expect that the deformed coproduct arising in NC theories should be associated with the deformed
brackets.
To that end, let us consider the deformed coproduct obtained by the (\ref{abeliantwist}) Drinfeld abelian twist expressing the usual Moyal type of noncommutativity. We have
\bea
\Delta_0 (g)\rightarrow \Delta_{\cal F}(g) &=&{\cal F}\Delta_0(g) {\cal F}^{-1}
\eea
with ${\cal F}$ given by (\ref{abeliantwist}). Considering again a group element close to the identity, we
can write
\bea
\Delta_{\cal F}(g)&=& {\bf 1}\otimes{\bf 1}+ i\omega_b \Delta_{\cal F}(T_b).
\eea
Let the corresponding deformed coproduct for the Lie algebra generator $T_b$ be denoted in the Sweedler's notation (\ref{sweed}),
as
\bea
\Delta_{\cal F}(T_b)&=&\xi_1 \otimes \xi_2.
\eea
Repeating the same steps as before it can be easily seen that, in its infinitesimal form, the transformation rule corresponding to the deformed coproduct of a generic element $A\in {\cal U({\bf G})}$ is
\bea
A\rightarrow A'&=&A+ \delta A,
\eea
with
\bea
\delta A&=&i\omega_b [T_b,A]_{\cal F}.
\eea
The deformed bracket reads as
\bea
[T_b, A]_{\cal F}&=& \xi_1 A S(\xi_2).
\eea
However, it can be easily seen that the algebra will not close under the deformed brackets, unless the original generators $T_a$'s are deformed further as \cite{asc} (see also formula (\ref{defgen}))
\bea
T_a^{\cal F}&=& {\overline f}^\alpha (T_a){\overline f}_\alpha.
\eea
These deformed generators span a linear subspace of the deformed Hopf algebra ${\cal U}^{\cal F}({\bf g})$ and their deformed brackets induce on them a Lie-algebraic structure. One has to note at least two important differences of these deformed generators and brackets in contrast to the undeformed ones.
Firstly, the exponentiation of the deformed generators does not yield elements of the Lie group, so that only the infinitesimal
version of the symmetry transformations are considered. Secondly, the deformed brackets are not manifestly antisymmetric, as it can be easily verified.
|
\section{Introduction}
Many writers have proposed that some puzzles of quantum mechanics (QM) might find elegant resolution, if we allow that there is retrocausality in the quantum world. Perhaps most importantly, retrocausality has been claimed to offer an attractive path to a Lorentz-invariant explanation of Bell correlations, and other manifestations of quantum nonlocality, without action-at-a-distance (see, e.g., [\ref{ref:costa}]--[\ref{ref:cramer86}], [\ref{ref:hokk}], [\ref{ref:miller96}]--[\ref{ref:price96}], [\ref{ref:suth83}]--[\ref{ref:wharton09}]). In general, it is a central assumption of Bell's Theorem~[\ref{ref:bell}] and other No Hidden Variable theorems in QM that there is no retrocausality: that hidden variables (HVs) are independent of future measurement settings. Prima facie, then, admitting retrocausality might be expected to make a big difference to the prospects for HV approaches to QM -- i.e., for the view that QM provides only an \emph{incomplete} description of the quantum world.
Some advocates of retrocausality in QM (see, e.g., [\ref{ref:costa76}], [\ref{ref:costa77}], [\ref{ref:miller96}], [\ref{ref:price97}]) have suggested that considerations of time-symmetry count in its favour: that if QM were made ``more time-symmetric,'' retrocausality would be a natural consequence. Against this, critics object that there is complete time-symmetry in classical physics, and yet no apparent retrocausality. Why should QM be any different?\footnote{I don't know an example of this objection in print, but I have often heard it in person.}
The main purpose of this note is to call attention to a respect in which QM is demonstrably different, under some assumptions about quantum ontology. Under these assumptions, a simple (and, so far as I know, previously unnoticed) argument shows that a time-symmetric ontology necessarily puts the properties of a system
\emph{prior to a measurement} under the control of an experimenter who chooses a setting for the measurement in question. Moreover, the argument turns on a feature of QM always thought to constitute one of the fundamental differences between classical and quantum physics, viz., the role in the latter of the new kind of discreteness introduced by quantisation.
However, the assumptions required by the argument are far from uncontroversial, and are easily seen to fail in a range of familiar interpretations of QM. So the argument does not support the sweeping claim that time-symmetry in QM requires retrocausality. But it does suggest that retrocausality is ``close to the surface'' in QM, in the sense it is a direct consequence of time-symmetry, under assumptions which many writers (past and present) would regard as well-motivated, even if not indisputable.
Roughly speaking, the interpretations in which the assumptions fail are of two kinds: (i) instrumentalist views of QM, which reject the kind of ontology ``between measurements'' which is needed to provide a domain of retro-effects; and (ii) all no-collapse versions of ontic realism about the wave function, which remove the particular kind of discreteness on which the argument depends.
Leaving instrumentalism to one side, the crucial issue is whether the required discreteness of measurement outcomes is regarded as ``all there is,'' or whether continuity is provided elsewhere in the ontology of one's model, e.g., in the form of an ontic wave function (without collapse). In the former case -- i.e., roughly, if we are any sort of realist other than an Everettian or a Bohmian -- the argument shows that we do need to make a choice between time-asymmetric ontology and retrocausality in QM, in a manner not true of classical physics.\footnote{The situation for Everettian and Bohmian views is less clear. There may be other reasons to think that they cannot entirely combine time-symmetry and one-way causality. (I discuss the case of the Bohm view in \S7 below.) But they escape a particularly sharp argument for the incompatibility of these options which applies to other realist views.}
In this case, the usual assumption that HVs are independent of future measurement settings is incompatible with time-symmetry.
\begin{center\vspace{12pt}
\begin{center\vspace{6pt}
\section{Polarization: classical and quantum}
\subsection{The classical case}
We first consider the standard description in classical electromagnetism (CEM) of the apparatus shown in Figure 1. A beam of light, linearly polarized in direction $\tau_R$, is directed towards an ideal polarizing cube set at angle $\sigma_R$. CEM predicts that the beam will split into two output beams. The \emph{transmission} beam, here labelled $\textup{\textbf{R}}=1$, will have an intensity $\cos^{2}(\tau_R - \sigma_R)$ times that of the input beam. The \emph{reflection} beam, here labelled $\textup{\textbf{R}}=0$, will have an intensity {$\sin^{2}(\tau_R - \sigma_R)$ times that of the input beam. Thus in the case in which $\tau_R = \sigma_R$, all the energy goes on the transmission beam; in the case in which $\tau_R = \sigma_R + \pi/2$, all the energy goes on the reflection beam; and other cases are distributed between these extremes, in a continuous fashion. The transmission ($\textup{\textbf{R}}=1$) and reflection ($\textup{\textbf{R}}=0$) beams have linear polarization in directions $\sigma_R$ and $\sigma_R + \pi/2$, respectively.
Figure 1 may be taken to represent a class of possible experiments, as $\tau_R$ and $ \sigma_R$ are allowed to vary. It follows from the time-symmetry of CEM that for any instance of this class there is a corresponding ``reverse'' experiment, in the intuitive sense of ``reverse'' in which moving from A to B is the reverse of moving from B to A. In an experiment of this reversed form, two beams of linearly polarized light, with appropriate polarization angles, intensities and phases, are combined into a \emph{single} beam, on passage through a polarizing cube.
Figure 2 depicts a typical case of this reversed kind. The diagram has been mirror-reversed compared to Figure 1 (thus preserving the convention that inputs come from the left and outputs go to the right), in order to make it easy (below) to depict the case in which the \emph{output} from an experiment of this reversed kind is used as \emph{input} for a case of the first kind. The mathematical constraints are exactly as one would expect from time-symmetry: if the input on the transmission channel, $\textup{\textbf{L}}=1$, is polarized in direction $\sigma_L$ with intensity $\cos^{2}(\tau_L - \sigma_L)$, the input on the reflection channel, $\textup{\textbf{L}}=0$, is polarized in direction $\sigma_L + \pi/2$ with intensity $\sin^{2}(\tau_L - \sigma_L)$, and the phases of the two beams are related exactly as in the original case, then there is an output beam with intensity $1$ and polarization $\tau_L$, in the direction of the input beam in the corresponding experiment of the original form.\footnote{If this seems surprising, keep the symmetry of the two classes of experiments in mind. Each instance of the new kind is simply the time-reverse of an instance of the old, in the intuitive sense mentioned above. If you were shown a video of such an experiment, with any dissipative effects suppressed, you would be unable to tell whether the video was being played forwards or backwards.}
\subsection{The quantum case}
Figures 3 and 4 show the quantum versions of these kinds of experiments, in the single photon case. For the moment, to highlight the feature which will be of interest, we assume that the outputs at $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$ and the inputs at $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$ are now ``discrete,'' in the sense that the photon leaves or enters the apparatus on one channel or other. (More later on the case in which the input and/or output may be a superposition.)
The factors $\cos^{2}(\tau_R - \sigma_R)$ and $\cos^{2}(\tau_L - \sigma_L)$ now represent probabilities, rather than intensities.
In Figure 3, a photon with polarization $\tau_R$ has a probability $\cos^{2}(\tau_R - \sigma_R)$ of being detected on the $\textup{\textbf{R}}=1$ channel, and a probability $\sin^{2}(\tau_R - \sigma_R)$ of being detected on the $\textup{\textbf{R}}=0$ channel. (The intensity interpretation is recovered in the limit, as the numbers of photons goes to infinity.)
There are some subtleties about what these probabilities amount to, on the input side. Intuitively, we want to say that in Figure 4, a photon emerging with polarization $\tau_L$ has a probability $\cos^{2}(\tau_L - \sigma_L)$ of \emph{having entered the apparatus} on the $\textup{\textbf{L}}=1$ channel, and a probability $\sin^{2}(\tau_L - \sigma_L)$ of having entered on the $\textup{\textbf{L}}=0$ channel. But such retrodictive probabilities are unreliable in many circumstances. If we know that the only photon source lies on the
$\textup{\textbf{L}}=1$ channel, for example, this information would incline us to say that the probability was close to one that an emerging photon with polarization $\tau_L$ had come from that channel, independently of $\cos^{2}(\tau_L - \sigma_L)$ (at least unless $\cos^{2}(\tau_L - \sigma_L)\approx 0$)
We can deal with this issue by stipulating that we lack such information: e.g., by requiring that the choice of input channel is chosen at random, with equal probability. In these circumstances, the factors $\cos^{2}(\tau_L - \sigma_L)$ and $\sin^{2}(\tau_L - \sigma_L)$ do supply reliable retrodictive probabilities -- specifically, the conditional probabilities of input on the $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$ channels, respectively, given an output of a photon with polarization $\tau_L$, in the direction shown. (Again, the intensity interpretation is then recovered in the limit, as the numbers of photons goes to infinity.)\vspace{12pt}
\begin{center\vspace{12pt}
\begin{center
\begin{center
\section{Experimental control}
We now consider the question of what can be controlled, in the CEM and quantum cases, by experimenters who control \emph{only the polarizer settings,} $\sigma_L$ and $\sigma_R$. The reason for this restriction on what the experimenters control will become clear as we proceed.
\subsection{The classical case}
Figure 5 shows the CEM version of experiment we obtain by combining a pair of experiments of the two previous classes. What interests us is the question of what control, if any, the lefthand experimenter, Lena, has over the intermediate polarization, \emph{if she has no control over the inputs at} $\textup{\textbf{L}}=1$ \emph{and} $\textup{\textbf{L}}=0$.
To put this question in stark form, imagine that the inputs $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$ are controlled by a Demon, who knows what setting $\sigma_L$ Lena has chosen for her polarizer. It is not difficult to see that under these conditions the Demon can produce any intermediate polarization $\tau$ he wishes, by an appropriate selection of inputs. This follows directly from the time-symmetry CEM: the inputs the Demon needs on the left are exactly the outputs Nature produces on the right, with intermediate polarization $\tau$ and right setting $\sigma_R=\sigma_L$.
Lena's lack of control of $\tau$ on the left, in the case in which the Demon controls the inputs, is exactly mirrored for her sister experimenter, Rena, on the right. The intuitive reason why Rena cannot control $\tau$ by varying the right polarizer setting $\sigma_R$ is that (as we just noted) Nature is able to make up for any difference in $\sigma_R$ by a difference in intensities of the output beams $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$. In this case, $\tau$ doesn't shift, no matter how much Rena wiggles the setting $\sigma_R$. This is why there need be no retrocausality in this case, of course.\footnote{For future reference, note that it does not follow that there \emph{could not be} retrocausality. To block retrocausality altogether, Nature must behave as the mirror image of the perfect obstructive Demon. Imperfect Demons, who block some but not all forward influence on the left, correspond to ways in which Nature \emph{might be,} to allow some retrocausality on the right.
To characterise this situation, I shall say that Lena has no \emph{input-independent} control of $\tau$, and that Rena has no \emph{output-independent} control of $\tau$. The importance of the notion of input-independence is that it mimics on the front end of the experiment precisely the question we need to consider on the back end, to think about the possibility of retrocausality. For on the back end, on the right, Rena has control of the \emph{setting,} $\sigma_R$, but not the \emph{outcome.} As we have just seen, the issue of retrocausality turns on the question as to what else Rena can influence, if her control is limited in this way. Input-independence enables us to think about the analogous issue at the front end of the apparatus, and time-symmetry enables us to move from one case to the other.
The next step is to consider these issues in the quantum case.\vspace{12pt}
\begin{center
\subsection{The quantum case}
Figure 6 shows the analogous single photon case. (It will become clear in a moment why we need to distinguish between $\tau_L$ and $\tau_R$, in this case.
For the moment, we are still assuming that the photon enters the apparatus on the left on one channel or other. (In other words, that the Demon cannot supply photons in a superposition of $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$.) It is a striking fact that in this case, Lena has more control over the left-polarization $\tau_L$ than in the classical case, even if the Demon retains control of the inputs at $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$. Of course, the Demon can refuse to provide any photons at all, in which case nothing emerges from the Lena's apparatus. But if we assume that the Demon shares with Lena the goal of emitting photons to the right, Lena now has complete control of $\tau_L$, $\textup{mod\ }\pi/2$. For if Lena sets her polarizer to angle $\sigma_L=\tau_L$, the Demon's only choices are to send in a photon via $\textup{\textbf{L}}=1$, in which case it emerges with polarization $\tau_L$; or to send in a photon via $\textup{\textbf{L}}=0$, in which case it emerges with polarization $\tau_L + \pi/2$.\footnote{The physical direction of the $\textup{\textbf{L}}=0$ beam depends on $\sigma_L$, but this creates no additional difficulty for the Demon, under our assumption that he knows what setting Lena has chosen.}
Thus in the notation introduced above, Lena now has \emph{input-independent} control of $\tau_L$, up to an additive factor of $\pi/2$, by means of her control of the left setting, $\sigma_L$. Note that this additional degree of control on Lena's part results directly from restriction on the Demon's options imposed by the assumption the input photon be on one channel or other. In effect, Lena is attempting to control a continuously-valued output variable ($\tau_L$) by manipulating a continuously-valued input variable ($\sigma_L$). The Demon can counteract any such influence, if he has an appropriate continuously-valued input variable of his own to play with, as in the classical case above. But if the Demon is restricted to a finite, \emph{discrete} range of inputs -- $n,$ say, in the general case -- then the best he can do is to group the possible outputs into $n$-tuples, in such a way that Lena cannot control which member of a particular $n$-tuple the output turns out to be; but can control from which $n$-tuple the choice is to be made.
This restriction on the Demon's abilities, and a corresponding restriction on Nature at the other end of the experiment, will play a crucial role in what follows; I shall call it \textbf{Discreteness}. I want to emphasise two points about Discreteness, at this stage. The first is that although it is an idea which -- at least naively -- seems at home in the quantum context, it doesn't depend on the quantum framework: we could impose Discreteness in the classical case, too, by stipulating that the Demon can only use one input channel at a time. The second is that I am not claiming that Discreteness is a realistic restriction, even in the quantum case: on the contrary, as we shall see, we can restore continuity in the quantum case, by allowing the Demon to input superpositions of $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$.
The real interest of Discreteness in the quantum case stems from the fact that it does turn out to be a realistic restriction, at least under some views of quantum ontology, at the other end of the experiment.\footnote{To avoid a possible confusion, let me also stress that my use of the terms \emph{discrete} and \emph{Discreteness} is not derived from the familiar distinction between discrete and continuous spectra, in relation to quantum observables. The observables we shall be discussing are certainly discrete (in fact, $2$-valued) in the latter sense. As we shall see, however, this is compatible, in some views of quantum ontology, with the outputs not being discrete in the sense relevant to my argument.}
We have seen that Discreteness gives Lena some input-independent control of $\tau_L$, on the left side of our experiment.
What does it entail on the other side of the experiment? In the classical case, we noted that Lena's lack of input-independent control of $\tau$ is mirrored for Rena, on the right. Rena has no output-independent control of $\tau$, so long as Nature absorbs any change she (Rena) makes to the setting $\sigma_R$ in changes in the output intensities (thereby requiring no change in $\tau$ itself). As we noted, this is why there need be no retrocausality in the classical case (despite complete time-symmetry).
To ask the corresponding question in the quantum case, it will be helpful to make explicit another assumption, which I shall call \textbf{Realism}. It is that $\tau_L$ is a \emph{beable,} in John Bell's sense. As Bell (\ref{ref:bell2}, p.~174) puts it:
``The beables of the theory are those elements which might correspond to elements of reality, to things which exist.''
Thus Realism is the assumption that $\tau_L$ labels something \emph{real} -- some aspect of the \emph{ontology} of the quantum world, as philosophers would say. Some writers reject this assumption, in the quantum domain, at least in the sense that they think that the quantum state should not be understood as playing this role -- more on this option in \S5.1 below -- but for now, I assume it.
With Realism in place, we can now formulate a further assumption, viz., \textbf{Time-symmetry}. It stipulates that the ontology of the system depicted in Figure 6 be time-symmetric, in the intuitive sense introduced above: if a process is permitted by the theory, then so, equally, is the process we describe by replacing $t$ by $-t$ in a description of the original process -- by ``playing the video in reverse,'' in the familar analogy.
In the quantum version of our experiment, Time-symmetry thus requires that there be an element of reality, $\tau_R$, that stands to $\sigma_R$, $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$ in precisely the way that $\tau_L$ stands to $\sigma_L$, $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$. Why? Because if there were not, the absence of such a $\tau_R$ in a complete description of the experiment would easily allow us to tell whether a video of a particular run of the experiment was being shown to us ``forwards'' or ``backwards.'' When it is shown forwards, we see $\tau_L$ correlated with $\sigma_L$, $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$ on the left; when it is shown backwards, we appear to see $\tau_R$ correlated with $\sigma_R$, $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$ on the right. If Time-symmetry does not hold -- and there is no such $\tau_R$, in fact -- then the latter case cannot be veridical, but must result from reversal of the movie.\footnote{It is not a relevant objection that no video could actually be made of such processes, due to the obvious limits on observability. We simply need to imagine a computer-generated movie, depicting the evolution of the assumed ontology.}
Note that the textbook model of QM, which combines evolution of the state function in accordance with Schr\"odinger's Equation with collapse in accordance with the Projection Postulate, is manifestly \emph{not} time-symmetric, in this sense. As is well-known, it is the Projection Postulate that introduces a time-asymmetry. (It requires that the state be an eigenvector of the operator corresponding to an observable \emph{after} but not \emph{before} the measurement in question.) I shall say a little more about the connection between this feature of the textbook picture and our present concerns in \S5.3 below, but for the moment, it means that the standard picture is simply off the map, for present purposes -- of no direct relevance to the issue as to whether if we \emph{do} assume Time-symmetry in QM, we are committed to retrocausality.
The assumptions Realism and Time-symmetry thus together make sense of the presence of the term $\tau_R$ in Figure 6, and entitle us to regard it as a beable, denoting an element of reality in just the way that (Realism entails that) $\tau_L$ does. However, we saw that if we assume Discreteness on the left, Lena has some input-independent control over $\tau_L$, by manipulating $\sigma_L$. If we now assume Discreteness on the right -- i.e., if we assume that a single photon emerges either at $\textup{\textbf{R}}=1$ or at $\textup{\textbf{R}}=0$ -- then it follows by symmetry that Rena has precisely the same output-independent control over the value of $\tau_R$ as her sister has over the value of $\tau_L$. She, too, can determine its value up to an additive factor of $\pi/2$, no matter what Nature does with the outputs on $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$.
In particular, this means that if Rena changes the setting $\sigma_R$ by any amount $\rho\neq\pi/2$, this will result in a \emph{different} value of $\tau_R$ for any subsequent photons. In counterfactual terms, it seems reasonable to say of any particular case that if she \emph{had} chosen $\sigma_R+\rho$ rather than $\sigma_R$, the value of $\tau_R$ would have been different. Intuitively, Rena thus has retrocausal control over $\tau_R$, $\textup{mod\ }\pi/2$.
As we wanted to show, then, time-symmetry does require retrocausality in this case, for reasons not present in CEM.\footnote{As before, we could duplicate the result in CEM by imposing the restriction that the output energy be entirely on one channel or other. However, without the flexibility provided by indeterminism in the QM case, a time-symmetric CEM model of this kind would lead to inconsistencies, in the sense that Lena and Rena could choose incompatible values of $\tau$. The distinction between $\tau_L$ and $\tau_R$ is needed to avoid this possibility in the QM case.} Once again, the heavy lifting is done by Discreteness, for it is this that ensures that Nature lacks the range of variability required to absorb the difference of a change in
$\sigma_R$ entirely in the future. Given the time-symmetric ontology required by the combination of Realism and Time-symmetry, Discreteness makes retrocausality a simple consequence of the dynamical laws.
Let us label this piece of reasoning the \emph{Time-Symmetry Argument for Retrocausality} (TSAR), and summarise its conclusions in the following form:
\begin{equation*}
{\textup{{Realism}}}+{\textup{Time-symmetry}}+{\textup{{Discreteness}}}\Rightarrow{\textup{{Retrocausality.}}
\end{equation*}
\section{Discussion}
What is TSAR's significance? At a minimum, I think, the argument provides an interesting toy model, built from very familiar materials, that casts some useful light on the complex relationship between time-symmetry and retrocausality. The model begins with a case in which time-symmetry clearly does not require retrocausality; and then shows how a simple restriction on the space of possibilities available to Nature (in Her response to an agent's choices) can have the effect of imposing retrocausality, if time-symmetry is to be preserved. The model seems likely to throw some light, too, on the consistency constraints of a retrocausal framework -- as suggested above, the indeterminism in QM seems to play a crucial role at this point.
But does TSAR have any direct application to QM? In particular, can it be used to support the claim that time-symmetry in QM requires retrocausality? The toy model shows how time-symmetry \emph{can} require retrocausality, but is the world described by QM the kind of world in which this is true; or more like the world of CEM, in which it is not true? I want to conclude with some discussion of these questions.
To focus the discussion, let me put on the table the following claim:
\begin{description}
\item[Symmetry-Requires-Retrocausality Thesis (SRRT):] TSAR establishes\\ that a time-symmetric ontology for photon polarization is necessarily retrocausal. Hence, if the quantum world is time-symmetric, it must be retrocausal.
\end{description}
I stress that I am not endorsing this thesis, but merely putting it on the table for discussion. I am interested in what counter-arguments are available to its opponents.
The most powerful objections to SRRT are those that question the status of the Discreteness assumption on which TSAR relies, arguing that it is unrealistic, at least in some interpretations of QM. I turn to such arguments in \S6. Before that, however, I want briefly to mention a couple of responses which take a different tack, questioning the ontological assumptions of TSAR.
\section{Ontological options}
\subsection{Avoid the ontology altogether}
The symmetry argument does not go through if we deny that the usual quantum polarization $\tau_L$ is a beable, or element of reality. Hence SRRT can certainly be evaded by a sufficiently thoroughgoing instrumentalism about the quantum world. This loophole is not available to proponents of the Hidden Variable program, of course.\footnote{It is not even available to \emph{opponents} of HVs, in so far as they are interested in the project of exhibiting supposedly undesirable consequences of the HV approach -- in this context, they cannot begin by denying their opponents' basic premise.}
The costs of this loophole should not be underestimated, however. One way to see what some of these costs are is to note that the apparatus in Figure 6 can easily be used to signal, by an experimenter who controls both the input channel and the left polarizer setting $\sigma_L$. Under these circumstances, $\tau_L$ would normally be thought to be what ``carries'' the signal, or causal influence, from one side of the apparatus to the other. Denying that there is any beable to play this role seems to amount to accepting a kind of action-at-a-distance (though timelike, rather than spacelike).\footnote{In [\ref{ref:evans}], Evans, Wharton and I argue that this option should be regarded as analogous to the standard view of spacelike nonlocality.}
\subsection{Avoid the specific ontology of the example}
A more subtle objection would be that the grounds for treating the usual quantum polarization $\tau_L$ as a beable rest on the assumption that there is no retrocausality. If the photon ``already knows'' $\sigma_R$, then it doesn't seem to need the full information carried by $\tau_L$, in order to explain the correlations we find in the full experiment, for variable $\sigma_L$ and $\sigma_R$ -- perhaps there is no such beable as $\tau_L$.
One way to see this is to exploit the symmetry of the experiment. This symmetry entails that if the dependence of the correlation between $\textup{\textbf{L}}$ and $\textup{\textbf{R}}$ on $\sigma_L$ and $\sigma_R$ can be explained (as in the orthodox picture) by postulating a beable $\tau_L$ to ``carry'' information from left to right, it could equally be explained by postulating a beable $\tau_R$ carrying information from right to left.\footnote{A reason for preferring one explanation to the other would have to come ``from outside'', for the symmetry implies that it cannot rest on anything internal to the experiment itself.} Since in the former case the explanation of the correlation does not require $\tau_R$, the explanation of the same correlation by the latter means would not require $\tau_L$.
The latter explanation, by means of $\tau_R$, provides an extreme illustration of the general point: by making information about the ``future'' end of the experiment available at the ``past'' end, we thereby reduce the need for a beable to carry information from past to future.
This is an important point, in my view. Among other things, it raises the question as to whether a time-symmetric model should involve the two beables $\tau_L$ and $\tau_R$, or rather some single beable, dependent both on $\sigma_L$ and $\sigma_R$ (which would combine the role of both $\tau_L$ and $\tau_R$, so to speak). On the face of it, models with two beables seem to fail to exploit the ontological efficiency possible in a retrocausal theory -- a fact which might be held to be a disadvantage of retrocausal models (e.g., those of [\ref{ref:aharonov}, \ref{ref:sutherland}, \ref{ref:wharton07}]) which proceed by adding an additional wave function (``coming from the future'') to standard QM, and interpreting both wave functions in an ontic manner.\footnote{Cramer's Transactional Interpretation [\ref{ref:cramer86}] seems to evade this objection, because its ``offer'' and ``confirmation'' waves do not have the same ontic status as the two wave functions in the models just cited.}
As an illustration of the ontological economy possible in a retrocausal theory, consider the following toy model. Replace the continuously-valued beables $\tau_L$ and $\tau_R$ with a single two-bit binary variable -- i.e., one of the four possibilities $\textbf{00}$, $\textbf{01}$, $\textbf{10}$ and $\textbf{11}$, where the first and second bits
represent the left (past) and right (future) input/output channels,
respectively. Let the probability distribution over the four possible values of this two-bit variable depend on $\sigma_L$ and $\sigma_R$, in accordance with Malus' Law:
\begin{equation*}
\textup{Pr}(\textbf{00}\lor\textbf{11})=\cos^{2}(\sigma_L - \sigma_R)\textup{;\ }
\textup{Pr}(\textbf{01}\lor\textbf{10})=\sin^{2}(\sigma_L - \sigma_R).
\end{equation*}
This model reproduces the predictions of QM (assuming Discreteness), and hence demonstrates that if retrocausality is allowed,
we need far less ontology for the intuitive causal
work of conveying information from one end of the experiment to the other. In
particular, we no longer need a beable to carry the continuous information
corresponding to the setting $\sigma_L$ -- in so far as it needs to be (given retrocausality), that variability can be entirely absorbed into the
statistics governing a discrete variable.\footnote{Indeed, we could economise even further, to a \emph{one-bit} variable representing the ``parity'' of the input--output channel comparison. If we map $\textbf{00}$ and $\textbf{11}$ in the previous model to $\textbf{1}$, and $\textbf{01}$ and $\textbf{10}$ to $\textbf{0}$, and now stipulate that $\textup{Pr}(\textbf{0})=\cos^{2}(\sigma_L - \sigma_R)$, we again get the correct predictions. The difference is that in this simpler model the hidden variable ``tells'' the photon in the future whether it should take the same or a different output channel, compared to its input channel in the past, without telling it \emph{which} input channel it took in the past. The resulting correlations might then seem to involve ``spooky action at a time-like distance,'' in comparison to those of the richer model.}
I think it is an interesting question whether the information structure and hence the ontological economy of this toy model could be replicated in a realistic physical model of the photon polarization case, or indeed of more complex cases. For the moment, however, I simply want to emphasise the relevance of this possibility to the present argument. On the one hand, it does reveal a significant lacuna in TSAR, as I presented it above, at least in the sense that it makes Realism seem ill-formulated: someone attracted to the approach described here will claim to be a realist by disposition, in the sense that they take seriously the question as to what exists ``between measurements,'' even though they do {not} -- as our version of Realism requires -- assume that $\tau_L$ labels an element of reality. On the other hand, however, this point carries no weight at all as an objection to SRRT. An opponent of SRRT who sought to rely on it would simply shoot herself in the foot, invoking retrocausality in order to block an argument in favour of retrocausality.
\subsection{Make the ontology time-asymmetric}
Another way to escape the consequences of SRRT -- i.e., to avoid a commitment to retrocausality -- is to make one's ontology time-asymmetric. It is important to be clear that this option is not an \emph{objection} to SRRT. SRRT merely claims that TSAR establishes that \emph{if} the ontology of QM is time-symmetric, then it is retrocausal. In other words, SRRT does not deny that we can avoid retrocausality by abandoning time-symmetry. It simply maintains that we need to make a choice between the two. This option concedes that we need to make a choice, but chooses time-asymmetry in preference to retrocausality
As I remarked earlier (\S3.2), the textbook version of QM, with the Projection Postulate, is an example of a model that exploits this option. I want to make two notes in passing at this point. First, it is important to be clear that avoiding an argument requiring retrocausality is not the same thing as demonstrating that there is no retrocausality. I shall explore this distinction below with reference to the de Broglie-Bohm model, but -- as I shall note there (\S7.1) -- the same point applies to the textbook picture: its case for thinking that the quantum world is not retrocausal seems remarkably weak.
Secondly, our discussion above hints at an additional cost of the textbook model, on top of the cost (if cost it be) of time-asymmetry itself. In \S5.2 we exploited the fact a model with a time-asymmetric boundary-independent ontology (i.e., in that case, with a beable $\tau_L$ to carry information from past to future) could easily be transformed into a different model, the time-reverse of the first, which would do an equally good job of explaining the correlations observed on the experiment in question. Hence the threat: unless there is some independent reason to prefer one model to the other -- which would be, \emph{ipso facto,} a reason to expect time-asymmetry in the domain in question -- then the time-asymmetry will have introduced an undetectable fact of the matter, about which of two kinds of worlds we actually inhabit. (In the terminology of \S5.2, is it a $\tau_L$ world or a $\tau_R$ world?) We break the symmetry at the cost of making it undetectable which way it is broken. A time-symmetric model avoids this consequence.
\subsection{Conclusion}
The prospects for rejecting SRRT on ontological grounds are slim. TSAR can be avoided by a thoroughgoing instrumentalism, prepared to deny the existence of a physical reality\footnote{Or, perhaps more subtly, to deny sense to the \emph{question} of the existence of such a reality.} between quantum measurements. Otherwise, however, TSAR's ontological prerequisites do not seem especially demanding. Quantum realists will need to look elsewhere for an objection to SRRT.
\section{Avoiding Discreteness?}
So the best prospect for rejecting SRRT seems to lie with role of Discreteness in TSAR. And on the face of it, a basis for challenge on these grounds is not hard to find.
As noted above, input-independent control disappears from the QM case, if we allow the Demon the option of introducing a photon in a superposition of the $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$ cases. In this case, by an appropriate choice of input amplitudes on each channel, the Demon can produce any $\tau_L$ he wishes, whatever Lena's choice of $\sigma_L$. (As in the classical case, this follows immediately from a consideration of the time-reversed case, given the time-symmetry of the relevant quantum dynamics.)
To avoid output-independence, and hence retrocausality, this option needs to be available to Nature in Rena's case. However, it may seem that Rena can frustrate Nature, simply by placing photon detectors on the output channels. Since such a measurement is guaranteed to yield either $\textup{\textbf{R}}=1$ or $\textup{\textbf{R}}=0$, rather than a superposition of both, it seems that Rena can easily deprive Nature of the flexibility needed to evade the argument for retrocausality.
But this argument is too quick. In effect, it depends on the assumption that the measurement outcomes constitute the entirety of relevant output ontology -- in other words, the sole available reservoir, from Nature's point of view, for non-retrocausal variability, in the light of changes Rena may make to $\sigma_R$. Some views of QM will reject this assumption, however. Everettians will certainly do so, for example. In the Everett picture, there is in reality no single definite outcome, whatever the appearances in any single branch. Rather, the photon emerges from Rena's polarizer in a superposition of $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$, with the mod-squared amplitudes for the two components provided by the factors $\cos^{2}(\tau_R - \sigma_R)$ and $\sin^{2}(\tau_R - \sigma_R)$, respectively. Once again, these factors restore the continuous variability of the classical case.
More on this kind of challenge to SRRT in a moment. First, it might be objected that even without the Everett view (or similar), the option of producing a superposition of $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$ is still available to Nature; a definite outcome only being needed when Rena makes a measurement.
This is true, but I think of limited use at this point, given the constraints of the problem. The task is to explain the correlations observed between inputs and outputs, in a quantum device of the kind depicted in Figure 6; and the issue is whether an explanation can be given using a time-symmetric intermediate dynamics and ontology, without admitting retrocausality. We have noted that in the analogous classical problem in Figure 5, the solution depends on the fact that the outputs may be continuously distributed between $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$. In the quantum case, it is no help that superposition may provide continuity ``inside the black box,'' so long as the experimental outputs themselves remain discrete. The Everett view gets off the hook by denying that the experimental outputs really \emph{are} discrete, at the global level. Where there is discreteness in the final conditions of the experiment, however, it doesn't seem to make any difference whether it appears at the time the photon encounters the polarizer, or at some later time.
For the Everett view, the wave function provides Nature with all the flexibility She needs to absorb the consequences of changes in $\sigma_R$, within a time-symmetric ontology, without requiring that they show up at earlier times. And the trick is available to any view that takes an ontological view of the wave function -- so long as it does not throw the advantage away, by allowing the wave function to collapse! In particular, therefore, it is available to the de Broglie-Bohm (dBB) view,\footnote{Or at least to most versions of dBB: in some versions, the wave function may not be ontic, in the required sense.} despite the fact that this theory also provides discrete outputs and inputs. Like Everettians, Bohmians can consistently combine time-symmetric dynamics and ontology with one-way causality, at least so far as TSAR is concerned.\footnote{I am setting aside the question as to whether the dBB approach can be applied to photons. One justification for this concession is that the argument of \S3.2 could be recast in terms of spin.}
However, to say that this combination is consistent, from a Bohmian point of view, is not yet to recommend it -- even by Bohmian lights. Do Bohmians actually have good reason to think that causality in these experiments is entirely one-way? And if so, echoing a concern we raised in \S5.3, are there good reasons to break the symmetry in one direction rather than the other? I shall return to this issue in \S7 below.
\subsection{Making the best case for TSAR?}
We have noted that SRRT can be rejected, without resorting to instrumentalism, so long as quantum ontology provides the continuity required to ensure that all changes Rena makes to $\sigma_R$ can be taken up in the future (thereby avoiding the need for retrocausal changes). The wave function in Everettian and (most) dBB models provides this continuity in the ontology; and hence TSAR has no direct bite, in these cases.
Addressing the issue from the other end, what is the best way to make a case for TSAR? In my view, it is to begin with a naive operational viewpoint, treating measurement outcomes as real and definite, and as the primary data to be explained; and being at best sceptical concerning claims that in some far-reaching sense, the ontology of the macroscopic world is anything other than it seems to be. When this starting point is combined with realism -- that is, for these purposes, with the project of trying to understand what \emph{microscopic} reality could account for the observed correlations at the level of measurements -- then TSAR demonstrates that retrocausality emerges in a natural way from time-symmetry and intuitive causality. The beable structure required for a local causal explanation of correlations between sequential measurements inevitably introduces retrocausality, when made time-symmetric.
Once again, this case for retrocausality can be avoided by adding an element to the ontology which is not there in the naive picture, and which departs a considerable distance from the naive picture -- a real wave function, surviving measurement. But since it has not previously been appreciated just how directly retrocausality arises from any time-symmetric naive model, it seems premature, to say the least, to take the fact that TSAR \emph{can} be evaded by these means as a reason for thinking that it \emph{should} be ignored -- and all the more so, because most of the standard reasons for thinking that a naive realist picture won't work in QM (i.e., the No Hidden Variable theorems) have themselves \emph{assumed} that there is no retrocausality.
This case for not rushing hastily from ``can'' to ``should'' (with respect to the possibility of avoiding the implications of TSAR) might be strengthened if we could challenge the presumption that an Everettian or dBB view does avoid retrocausality. If it does not avoid it, after all, then this way of escaping the implications of TSAR would be a case of out of the frying pan, into the fire, from the point of view of an opponent of retrocausality.
Indeed, there is a prospect of an interesting dilemma for opponents of retrocausality, if the extra ``non-naive'' ontology needed to block TSAR automatically makes it impossible to exclude retrocausality on observational grounds (by putting it ``out of sight,'' as it were). With this prospect in mind, I now turn to an issue deferred above. What grounds, if any, can an advocate of the dBB view offer that such a theory is \emph{not} retrocausal, in its usual form?\footnote{I set aside for the present the corresponding issue for the Everettian case.}
\section{Retrocausality in de Broglie-Bohm?}
TSAR shows that ``intuitive'' one-way causality must fail, given time-symmetry, in some versions of a QM account of photon polarization. While the dBB theory is not one of those versions, escaping an argument for the conclusion that the intuitive view is mistaken is not the same as offering an argument that it is not mistaken. It is unclear what form the latter argument might take, for a proponent of the dBB view. To address these issues, we need to step back a bit, and think about what is at stake. One of the attractions of examples like that of Figure 6 is that they make it easy to pose the relevant questions, without surreptitiously introducing a bias in one direction or other. The symmetries of the model make any bias easily visible.
Intuitive causality clearly involves such a bias: we find it natural to say that if Rena had chosen a different setting, that difference would have made no difference to the photon arriving from the left, but might have affected the result to the right; but very unnatural to say that if Lena had chosen a different setting, that difference would have made no difference to the photon departing to the right, but might have affected the input from the left. We have seen that if the ontology is time-symmetric, and the outputs at $\textup{\textbf{R}}=1$ and $\textup{\textbf{R}}=0$ are discrete in the relevant sense, this asymmetry is unsustainable -- we are forced to revise our view about Rena's case. But where does the intuitive bias come from in the first place? And what reason, if any, do we have in continuing to take it for granted, in a case such as the dBB theory (where abandoning it is not forced on us)?
Our simple model helps with the first of these questions. In the classical case, the obvious difference between Lena and Rena is that Lena controls both the $\sigma_L$ \emph{and} the inputs $\textup{\textbf{L}}=1$ and $\textup{\textbf{L}}=0$; whereas Rena controls only the setting $\sigma_R$. We introduced the Demon in order to restore symmetry, but what breaks symmetry, normally, is the fact that there is no Demon: on the contrary, experimenters often control their inputs, too.\footnote{A fact which relies on the availability of low entropy sources, presumably.}
But if this is the right story about the source of the bias in the classical case, it seems to cast doubt on our entitlement to retain the bias, in the single-photon case. Intuitively, Lena does have enough control over the classical input beams to ensure that they \emph{would have been the same} (from the same directions, with the same intensities and the same polarizations), if she had chosen a different measurement setting. She does not have enough control over the behaviour of a single-photon source to ensure that \emph{the same photon} would have entered the apparatus, with \emph{the same properties,} if she had chosen a different measurement setting. Intuitive causality tells us that a different setting wouldn't have made any difference; but in this case, unlike in the classical case, the intuition isn't backed up by anything that Lena can actually \emph{do.}
It might be objected that Lena can do something almost as useful, even in the QM case. She can ensure that no photons enter the apparatus through (say) the $\textup{\textbf{L}}=0$ channel, by simply blocking that channel. Or, if it is objected that this does not exclude thermal photons emitted from the screen itself, she can provide a photon source on the $\textup{\textbf{L}}=1$ of sufficient intensity to guarantee that \emph{most} photons arrive via that channel.
But if this explains the intuitive causal asymmetry, then the asymmetry should disappear if we deprive Lena of this ability -- e.g., if we assume that the inputs to Lena's apparatus are provided by thermal radiation, with no built-in preference for one channel or the other. Yet intuition still seems to tell us that Lena's choices still make a difference to the photon between the polarizers, whereas Rena's do not. Is this a mistake? And if not, where does this persistent asymmetry come from?
Furthermore, note that it would not be an option at this point to try to restore symmetry by denying that Lena's choices make a difference to the photon between the polarizers. In that case, the only place to absorb the differences associated with Lena's choices would be in the inputs, and so we would be granting Lena retrocausal influence \emph{over} those inputs. The only way to avoid such retrocausal influence on Lena's part is to insist that her choices make a difference exclusively in the future. But then symmetry would imply that Rena's choices would make a difference exclusively in the past -- so we still face the problem of grounding an \emph{asymmetry,} if we are to avoid retrocausality, in one place or other.
If we wish to insist on this intuitive causal asymmetry, even in the case in which we explicitly exclude the kind of bias introduced by allowing Lena access to low-entropy sources, then there seem to be three possible explanations of its origins:
\begin{enumerate}
\item Some asymmetry in \emph{initial conditions,} other than the usual thermodynamic asymmetry in the sources Lena would normally be taken to control.
\item Some \emph{dynamical} asymmetry in the theory itself.
\item Some \emph{conventional} asymmetry, governing the causal or counterfactual reasoning we use in considering what depends on what.
\end{enumerate}
Let us explore these options in reverse order.
The third option has some plausibility, in my view, as an account of the time-asymmetry of much of the kind of causal and counterfactual reasoning we employ in ordinary life (see, e.g., [\ref{ref:price96}], [\ref{ref:price07}], [\ref{ref:pricewes}]). But it is entirely irrelevant here, where the question is whether there are any good grounds for excluding HV models for QM involving the kind of lawlike correlations between HVs and later measurement settings that have been called ``retrocausal.'' Option (3) might reveal that this is a poor name for the view, but it provides no reason whatsoever to deny the existence of such correlations themselves. In particular, it provides no reason for a proponent of the standard dBB view to the deny the existence of such correlations, \emph{within the standard dBB framework.} (In other words, a proponent of dBB who is convinced that her view excludes such correlations must have something other than Option (3) in mind.)
But Option (2) is off the table, in the case of the standard dBB view, because the dynamics is not time-asymmetric. So the only remaining option is (1) -- i.e.,
to appeal to some further asymmetry of initial conditions, of a kind not excluded by the elimination of special initial conditions related to the thermodynamic asymmetry. But how are we to justify such conditions? And to show that they are at all in tension with violations of Independence? (It won't do to appeal to the familiar causal asymmetry, if the familiar causal asymmetry relies on something else entirely, namely, the low entropy initial conditions available in typical experiments.)
So far as I can see, there is nowhere else for the dBB view to turn at this point, in defence of the claim that the theory is not retrocausal, in its standard form. Once the role of special initial conditions related to the thermodynamic asymmetry is factored out, any remaining prohibition on retrocausality is merely put in by hand, with no substantial justification whatsoever. Given the distinctive ontology of the dBB view, the required assumption needs to be put in at two points, apparently: it needs to be assumed that \emph{neither} the wave function \emph{nor} the initial positions of the particles are affected by later measurement choices.
In the absence of any apparent justification for these assumptions, I conclude that while dBB blocks TSAR, it doesn't thereby exclude retrocausality, in the sense of interest in QM. It blocks an argument \emph{from} time-symmetry \emph{to} retrocausality, but this is not the same excluding retrocausality itself.\footnote{For present purposes, I am setting aside the suggestion that retrocausality can be excluded on the grounds that it conflicts with free will. See [\ref{ref:price96}], [\ref{ref:evans}], for some critical discussion of this point. In the present context, I don't know any proponent of the dBB view who maintains that it is our knowledge of our own free will that establishes that the dBB theory is entitled to the additional assumptions required to ensure that it is not retrocausal.}
Intriguingly, there are hints of an argument for the opposite conclusion -- for the view that the dBB theory should \emph{reject} intuitive causality at the fundamental level, in favour of a symmetric picture. Goldstein and Tumulka [\ref{ref:gold03}] offer a dBB ``toy model'' in which microcausal retrocausality provides a Lorentz-invariant explanation of the Bell correlations. This model differs in some ways from other retrocausal proposals (e.g., [\ref{ref:costa}]--[\ref{ref:cramer86}], [\ref{ref:hokk}], [\ref{ref:miller96}]--[\ref{ref:price96}], [\ref{ref:suth83}]--[\ref{ref:wharton09}]) for reconciling these correlations with special relativity, but the underlying strategy is exactly the same: zig-zag causality, retrocausal on one arm, provides a decomposition of Bell's spacelike correlations into a product of timelike correlations.\footnote{The main difference is that in the proposal in [\ref{ref:gold03}], the zig-zag goes initially via the future, rather than initially via the past (see [\ref{ref:gold03}], Figure 4, p.~563).}
Suppose it were to turn out that abandoning intuitive causality at the fundamental level, in favour of a causally symmetric picture, provided a successful route to a Lorentz-invariant formulation of the dBB theory; and that we were convinced that this was the only way to make the theory Lorentz-invariant. We would then have a basis for an analogue of the argument of \S3.2, within the dBB framework, in which Lorentz-invariance played the role of Discreteness: if Nature is constrained to be Lorentz-invariant, She simply does not have the option of absorbing all counterfactual changes of $\sigma_R$ in the future, but must allow Rena a degree of output-independent control of the past.\footnote{For a friend of retrocausality, it would be an enormously satisfying result if two of the great lessons of Einstein's \emph{annus mirabilis,} the quantisation of light and Lorentz-invariance, were to converge in this way, in support of a realist view of the quantum world. }
\subsection{The textbook model, again}
Finally, I would like to note that most of these considerations apply to the textbook model of QM, with the Projection Postulate, as well as to dBB. We observed in \S5.3 that because the textbook model is explicitly time-asymmetric, it evades TSAR. However, as we have just emphasised with respect to dBB, avoiding an argument that entails retrocausality is not the same as establishing that there is no retrocausality -- it is necessary but not sufficient, as it were.
The discussion in this section seems to apply with almost equal force to the textbook model as it does to dBB. The one difference, I think, is that because the dynamics of the textbook model are not time-symmetric, we cannot immediately exclude option 2 above, as an explanation of the intuitive causal asymmetry. But a little reflection suggests that the dynamical time-asymmetry of the textbook model -- i.e., the Projection Postulate itself -- is of entirely the wrong character for the purpose. To rule out retrocausality, we need to show that changing a measurement setting -- e.g., in our experiment, the value of $\sigma_R$ -- does not change the earlier state of the system measured. The Projection Postulate is simply silent on this matter: it specifies the state \emph{after} a measurement, but says nothing about the state \emph{before} a measurement. As in the case of dBB, then, the textbook model's assumption that the state function is independent of future measurement settings seems simply to be put in by hand, in practice.\footnote{See [\ref{ref:schulman97}] for a retrocausal approach that exploits this fact, by allowing that the initial wave function may indeed depend on what happens to a system in the future.}
\section{Summary}
We began by noting the claim that time-symmetry favours retrocausality in QM; and the counter-argument that classical physics seems to permit time-symmetry without retrocausality and that it isn't clear why QM should be different. We have now offered an argument that shows that QM \emph{is} different, under some assumptions about quantum ontology. However, we have also noted that these assumptions are far from compulsory: they are rejected, in one way or another, by several interpretations of QM.
To put this conclusion in perspective, it is worth stressing that the relative merits of the views of QM that reject these assumptions -- e.g., instrumentalist views, on one hand, or the Everett and dBB views, on the other -- depend on the demerits of alternative approaches. In particular, they depend on the issue of the viability of the HV program, in the ``just the particles'' sense -- views which combine the HV program with an epistemic view of the quantum state.
The present result contributes to clarifying this issue, in the following sense. As we have noted, Bell's Theorem and other No Hidden Variable results rely on the assumption that HVs are independent of future measurement settings -- an assumption with considerable intuitive appeal, against the background of the one-way causality familiar in ordinary life.
However, the present argument shows that within the framework to which such an HV program is committed -- a framework in which discreteness is \emph{not} offset by non-epistemic continuity at the level of the wave function -- a blanket prohibition on retrocausality is incompatible with time-symmetry, in a manner specific to the QM case. This provides a new reason for re-examining the assumption in question, and for exploring HV approaches that relax it. The photon polarization experiment discussed in \S3.2 may offer a useful model for investigating these questions, under the assumptions of time-symmetry and discrete outputs.\footnote{See [\ref{ref:evans}] for further relevant discussion of this case, developed by means of an analogy with a photon polarization version of the EPR-Bohm experiment; and [\ref{ref:price08}] for a more abstract approach to the general project of retrocausal modelling.}
\section*{Acknowledgements}
The ideas in this paper owe a great deal to Ken Wharton, having emerged in discussion about related issues for our joint paper with Peter Evans [\ref{ref:evans}]. I am also very grateful for comments from Richard Healey, Ruth Kastner, Owen Maroney, David Miller and Rob Spekkens, from an audience in Cambridge in May 2010, and from an anonymous referee. My research is supported by the Australian Research Council and the University of Sydney.
\section*{References}
\renewcommand{\theenumi}{\arabic{enumi}}
\renewcommand{\labelenumi}{[\theenumi]}
\begin{enumerate}
\item Aharonov, Y., Bergmann, P.G.~\& Lebowitz, J.L., 1964. ``Time symmetry in the quantum process of measurement'', \textit{Physical Review}, 134, B1410--B1416. [\href{http://prola.aps.org/abstract/PR/v134/i6B/pB1410_1}{DOI: 10.1103/PhysRev.134.B1410}] \label{ref:aharonov}
\item Bell, J.S., 1964. ``On the Einstein Podolsky Rosen Paradox'', \emph{Physics} 1, 195--200. \label{ref:bell}
\item Bell, J.S., 1984. ``Beables for quantum field theory'', CERN-TH 4035/84, Aug.~2, 1984. Reprinted in \emph{Speakable and Unspeakable in Quantum Mechanics,} 2nd edn., Cambridge: Cambridge University Press, pp.~173--180.\label{ref:bell2}
\item Costa de Beauregard, O., 1953. ``M\'echanique quantique'', \emph{Comptes Rendus de l'Acad\'emie des Sciences}, T236, 1632--1634. \label{ref:costa}
\item Costa de Beauregard, O.,~1976. ``Time Symmetry and Interpretation of Quantum Mechanics'', \textit{Foundations of Physics}, 6, 539--559. \\\hspace{0pt}
[\href{http://www.springerlink.com/content/k8l386556135g250/}{DOI: 10.1007/BF00715107}]\label{ref:costa76}
\item Costa de Beauregard, O., 1977. ``Time symmetry and the Einstein paradox'', \textit{Il Nuovo Cimento}, 42B, 41--63. [\href{http://www.springerlink.com/content/x755748406278g65/}{DOI: 10.1007/BF02906749}]\label{ref:costa77}
\item Cramer, J.G., 1980. ``Generalized absorber theory and the Einstein-Podolsky-Rosen paradox'', \textit{Physical Review D}, 22, 362--376. \\\hspace{0pt}
[\href{http://prd.aps.org/abstract/PRD/v22/i2/p362_1}{DOI: 10.1103/ PhysRevD.22.362}]
\item Cramer, J.G., 1986: ``The transactional interpretation of quantum mechanics'', \textit{Reviews of Modern Physics}, 58, 647--687. \\\ [\href{http://rmp.aps.org/abstract/RMP/v58/i3/p647_1}{DOI: 10.1103/RevModPhys.58.647}]
\label{ref:cramer86}
\item Evans, P., Price, H. and Wharton, K.B., 2010. ``New slant on the EPR-Bell experiment'', \emph{British Journal for the Philosophy of Science,} forthcoming. [\href{http://arxiv.org/abs/1001.5057}{arXiv:1001.5057v3 [quant-ph]}] \label{ref:evans}
\item Hokkyo, N., 1988. ``Variational formulation of transactional and related interpretations of quantum mechanics'', \emph{Foundations of Physics Letters,} 1, 293--299. [\href{http://www.springerlink.com/content/p81r939075727216/}{DOI: 10.1007/BF00690070}] \label{ref:hokk}
\item Goldstein, S.~and Tumulka, R., 2003. ``Opposite arrows of time can reconcile relativity and nonlocality'', \emph{Classical and Quantum Gravity,} 20, 557--564. [\href{http://arxiv.org/abs/quant-ph/0105040v5}{arXiv:quant-ph/0105040v5}] [\href{http://www.iop.org/EJ/abstract/0264-9381/20/3/311/}{DOI: 10.1088/0264-9381/20/3/311}]\label{ref:gold03}
\item Miller, D.J., 1996. ``Realism and time symmetry in quantum mechanics'', \emph{Physics Letters,} A222, 31--36. [\href{http://linkinghub.elsevier.com/retrieve/pii/0375960196006202}{DOI: 10.1016/0375-9601(96)00620-2}] \label{ref:miller96}
\item Miller, D.J., 1997. ``Conditional probabilities in quantum mechanics from a time-symmetric formulation'', \emph{Il Nuovo Cimento,} 112B, 1577--1592.
\item Price, H., 1984. ``The philosophy and physics of affecting the past'', \textit{Synthese}, 16, 299--324. [\href{http://www.springerlink.com/content/p94624888311jv41/}{DOI: 10.1007/BF00485056}]
\item Price, H., 1994. ``A neglected route to realism about quantum mechanics'', \textit{Mind}, 103, 303--336. [\href{http://arxiv.org/abs/gr-qc/9406028}{arXiv:gr-qc/9406028v1}]
\item Price, H., 1996. \textit{Time's Arrow and Archimedes' Point,} New York:
Oxford University Press.
\label{ref:price96}
\item Price, H., 1997. ``Time symmetry in microphysics'', \emph{Philosophy of Science,} 64, S235--244. [\href{http://arxiv.org/abs/quant-ph/9610036}{arXiv:quant-ph/9610036v1}] \label{ref:price97}
\item Price, H., 2007. ``Causal perspectivalism'', in H.~Price and R.~Corry, eds., \emph{Causation, Physics and the Constitution of Reality: Russell's Republic Revisited,} Oxford: Oxford University Press, pp.~250--292.\label{ref:price07}
\item Price, H., 2008. ``Toy models for retrocausality'', \emph{Studies in Hist.~and Phil.~of Mod.~Phys.,} 39, 752--776. [\href{http://arxiv.org/abs/0802.3230}{arXiv:0802.3230v1 [quant-ph]}]\label{ref:price08}
\item Price, H. \& Weslake, B., 2010: ``The time-asymmetry of causation'', in H.~Beebee, C.~Hitchcock and P.~Menzies, eds., \emph{The Oxford Handbook of Causation,} Oxford: Oxford University Press, pp.~414--443.\label{ref:pricewes}
\item Schulman, L.S., 1997: \emph{Time's Arrows and Quantum Measurement,} Cambridge: Cambridge University Press.\label{ref:schulman97}
\item Sutherland, R.I., 1983. ``Bell's theorem and backwards-in-time causality'', \textit{Int.~J.~Th.~Phys.}, 22, 377--384. [\href{http://www.springerlink.com/content/k8g22512n5w38xu1/}{DOI: 10.1007/BF02082904}]\label{ref:suth83}
\item Sutherland, R.I., 1998. ``Density formalism for quantum theory'', \emph{Foundations of Physics.} 28, 1157--1190. [\href{http://www.springerlink.com/content/u551710rp1x27723/}{DOI: 0.1023/A:1018850120826}]
\item Sutherland, R.I., 2008. ``Causally symmetric Bohm model'', \emph{Studies in Hist.~and Phil.~of Mod.~Phys.,} 39, 782--805. [\href{http://arxiv.org/abs/quant-ph/0601095}{arXiv:quant-ph/0601095v2}] \label{ref:sutherland}
\item Wharton, K.B., 2007. ``Time-symmetric quantum mechanics'', {\em Foundations of Physics,} 37, 159--168. [\href{http://www.springerlink.com/content/v765h37416011n63/?p=877db523fa624fd582e512cfb9a5bebb&pi=1}{DOI: 10.1007/s10701-006-9089-1}] \label{ref:wharton07}
\item Wharton, K.B., 2009. ``A novel interpretation of the Klein-Gordon equation'', \emph{Foundations of Physics,} 40, 313--332. \\\ [\href{http://www.springerlink.com/content/l14jr4x24pk82521/}{DOI: 10.1007/s10701-009-9398-2}] [\href{http://arxiv.org/abs/0706.4075}{arXiv:0706.4075v3 [quant-ph]}].
\label{ref:wharton09}
\end{enumerate}
\end{document}
|
\section{Introduction}\label{sec:intro}
The past decade has seen an exponential increase in the number of
parallax measurements to radio pulsars. Whereas only 14 such
measurements were known by the end of the year 2000, currently
52 radio pulsars have their distance determined either through VLBI or
pulsar timing measurements (see Tables \ref{tab:curpxs} and
\ref{tab:oldpxs}).
The importance of radio pulsar parallaxes arises from the wide variety
of highly accurate investigations to which pulsars lend
themselves. For example, the combination of an accurate parallax with
the pulsar dispersion measure (DM: the integrated electron density
between the pulsar and Earth) provides an average electron density
measurement along the line of sight, which can be used to construct
Galactic electron density models \citep[see,
e.g.,][]{cl02}. Also, the highly polarised nature of pulsars allows
measurement of the Faraday rotation which - in combination with a
distance - can be used to map out the component of the Galactic
magnetic field parallel to the line of sight \citep{hml+06}. Finally,
in pulsar timing, distances are essential in correcting spindown and
orbital period derivatives for the Shklovskii effect caused by proper
motion \citep{shk70}. Consequentially, certain pulsar timing tests of
general relativity are dependent on accurate distance measurements, as
described by \citet{dt91} and used in \citet{nss+05} and
\citet{dvtb08}, amongst others.
The large increase in the number of radio pulsar parallaxes and their
unique applications warrant an investigation into potential biases. As
pointed out by \citet{lk73}, in an homogeneous field of stars the
number of stars per unit of distance increases as $D^2$, where $D$ is
distance. This makes it statistically more likely to find an object at
larger distances where more volume is sampled and, hence, more sources
lie. \citet{lk73} showed analytically that this statistical
underestimate of the stellar distance depends on the precision of the
parallax measurement. Specifically, they derived the following
proportionality:
\begin{equation}
p(\varpi | \varpi_{\rm 0}) \propto \left( \frac{\varpi_{\rm
0}}{\varpi}\right)^4 \exp \left( -
\frac{\left(\varpi-\varpi_{\rm 0}\right)^2}{2 \sigma^2}\right),
\end{equation}
where $\varpi_{\rm 0}$ is the measured parallax, $\sigma$ is the
standard deviation of the measurement and $p(\varpi | \varpi_{\rm 0})$
is the probability distribution of the actual parallax, $\varpi$,
given the measurement. \citet{bm98} expanded this analysis by using
the intrinsic luminosity function for the star's spectral class as a
further source of prior information to use in the estimate of a more
accurate parallax value.
Two points of confusion have pervaded the literature on the topic of
Lutz-Kelker bias. First, the effect we describe as Lutz-Kelker bias is
commonly referred to as Malmquist bias in extragalactic astronomy. As
\citet{gf97} point out, the Malmquist bias originally referred to a
positive bias in luminosity of magnitude-limited samples and has only
recently acquired the meaning of the geometric bias described by
\citet{lk73}. \citet{smi03b} stresses a second point of confusion and
difference between the original Malmquist bias and Lutz-Kelker bias,
namely the fact that the bias described by \citet{lk73} and discussed
in the present paper, refer to individual parallax measurements rather
than to the overall average of a sample of objects.
In this paper we revisit the analysis by \citet{bm98} with a
particular focus on radio pulsars. To account for the analytically
complex but realistic pulsar luminosity function and Galactic pulsar
distribution, we describe a Monte-Carlo approach to correct previously
published parallax measurements for the Lutz-Kelker bias. Our basic
analytic derivation and a description of the simulations are given in
\S\ref{sec:theory}. The resulting corrections to published
measurements are discussed in \S\ref{sec:revise}. Our conclusions are
summarised in \S\ref{sec:conc}.
\section{Theory and Simulations}\label{sec:theory}
As an illustration of the bias caused by volumetric and intrinsic
luminosity arguments, it is helpful to consider an analytic derivation
along the lines of those presented by \citet{lk73} and
\citet{bm98}. Such an analysis is presented in \S\ref{ssec:Basis}
below. For any practical purposes, however, a more realistic approach
that incorporates a Galactic distribution model for neutron stars as
well as a more realistic intrinsic luminosity function for radio
pulsars, is required. Because such realistic models are analytically
complex, we will implement them only as part of a Monte-Carlo
simulation, which is described in \S\ref{ssec:sims}.
\subsection{Basic Theory}\label{ssec:Basis}
For a pulsar at distance $D = \varpi^{-1}$, Bayes' theorem gives the
probability that the true parallax is $\varpi$, given the measurement
$\varpi_{\rm 0}$, as
\begin{equation}\label{eq:Bayes}
p(\varpi | \varpi_{\rm 0}) = \frac{p(\varpi_{\rm 0} | \varpi)
p(\varpi)} {p(\varpi_{\rm 0})},
\end{equation}
where $p(\varpi_{\rm 0})$ is the prior probability of the measurement
which is taken to be constant without loss of generality. Assuming the
measurement is dominated by Gaussian noise with standard deviation
$\sigma$, we know furthermore that
\begin{equation}\label{eq:Gauss}
p(\varpi_{\rm 0}|\varpi) = \frac{1}{\sqrt{2
\pi}\sigma}\exp\left[-\frac{1}{2}\left(\frac{\varpi-\varpi_{\rm
0}}{\sigma}\right)^2\right].
\end{equation}
The final term in Eq. \ref{eq:Bayes} is the prior probability
distribution of the parallax, $p(\varpi)$, which is composed of two
terms: one defined by the sampled volume and another defined by the
intrinsic pulsar luminosity function. Specifically, assuming a
homogeneous distribution of pulsars in space, we have
\[
p(D) \propto \rho D^2,
\]
where $\rho$ is the spatial density of pulsars. Translating this
probability to a probability in $\varpi$, we get
\begin{equation}\label{eq:volume}
p_{\rm D}(\varpi) = \left|\frac{\partial D}{\partial \varpi}\right| p(D)
\propto \varpi^{-4}.
\end{equation}
Note this is effectively the prior \citet{lk73} determined. We will
henceforth refer to this prior as the ``volumetric'' prior
distribution.
The second contribution to the parallax prior is determined by the
intrinsic radio pulsar luminosity function. Assuming, for simplicity,
the power-law luminosity function
\[
p(L) \propto L^{\beta},
\]
where $L$ is the intrinsic pulsar luminosity and $\beta$ is the slope
of the luminosity function, we obtain
\begin{equation}\label{eq:lumin}
p_{\rm L}(\varpi) = \left|\frac{\partial L}{\partial \varpi}\right|
p(L) \propto \varpi^{-2\beta-3},
\end{equation}
which we will refer to as the ``luminosity'' prior distribution.
Combination of Eqs. \ref{eq:volume} and \ref{eq:lumin} results in the
prior probability distribution of the pulsar parallax
\[
p(\varpi) = p_{\rm D}(\varpi) p_{\rm L}(\varpi) \propto \varpi^{-2\beta-7}.
\]
Inserting this along with Eq. \ref{eq:Gauss} into Eq. \ref{eq:Bayes},
we obtain the resulting probability distribution of the true parallax:
\begin{equation}\label{eq:basres}
p(\varpi | \varpi_{\rm 0}) \propto \varpi^{-2\beta-7} {\rm
e}^{-\frac{1}{2}\left(\frac{\varpi_{\rm
0}-\varpi}{\sigma}\right)^2}.
\end{equation}
The (local) maximum of this probability distribution can easily be
determined by requiring
\[
\frac{\partial p(\varpi |\varpi_{\rm 0})}{\partial \varpi} = 0,
\]
which gives:
\begin{multline*}
(-2\beta-7)\varpi^{-2\beta-8}{\rm e}^{-\frac{1}{2}\left(
\frac{\varpi_{\rm 0}-\varpi}{\sigma}\right)^2} \\
+ \varpi^{-2\beta-7}\frac{\varpi_{\rm
0}-\varpi}{\sigma^2}{\rm e}^{-\frac{1}{2}\left(
\frac{\varpi_{\rm 0}-\varpi}{\sigma}\right)^2} = 0.
\end{multline*}
Rearranging this expression results in the analytic statement of the
bias:
\begin{equation}\label{eq:bassol}
\frac{\varpi}{\sigma} = \frac{1}{2}\left[\frac{\varpi_{\rm
0}}{\sigma} \pm \sqrt{\frac{\varpi_{\rm
0}^2}{\sigma^2}-4(2\beta+7)}\right].
\end{equation}
This result clearly demonstrates that the bias is a function of the
significance of the parallax measurement and not of the parallax
itself.
Equation \ref{eq:bassol} also shows the effect of different intrinsic
luminosity functions: for steep power laws ($\beta < -3.5$) the
luminosity function dominates the prior and the measured parallax
value is likely to be underestimated. For shallow luminosity functions
($\beta > -3.5$) the reverse is true and the parallax is expected to
be overestimated. Note that in this case the prior in Equation
\ref{eq:basres} goes to infinity as $\varpi$ goes to $0$. A luminosity
function with spectral index $\beta = -3.5$ results in a luminosity
prior that exactly cancels out the volumetric prior, so that no bias
is observed (as can easily be seen in Equation \ref{eq:basres}). For
the observed pulsar population, $\beta \approx -1.7$ \citep{lfl+06},
which would imply measured parallaxes to be larger than their actual
values.
\subsection{Realistic Monte-Carlo Simulation}\label{ssec:sims}
There are two clear improvements to be made to the analysis presented
in the previous section. First, \citet{fk06} have shown that the
intrinsic luminosity function of radio pulsars is not a simple
power-law but more likely log-normal in form. Second, pulsars are not
distributed homogeneously in space, but are mostly confined to the
Galactic disk. The analytic complications implied by these
improvements require that we proceed to evaluate the bias through
Monte-Carlo simulations. Practically, we simulate the Gaussian
measurement uncertainty and the two prior distributions independently,
creating three normalised probability distributions\footnote{Since
both the scale height and intrinsic luminosity may be argued to
depend on pulsar age, a more advanced, joint, simulation might be
called for. Given current uncertainties in pulsar luminosities and
the luminosity-age relation, however, we exclude this extension from
the present analysis.}. Multiplication of these distributions
provides the final, bias-corrected, distribution,
$p(\varpi|\varpi_{\rm 0})$ (see for example Figure
\ref{fig:sample}). We will report the peak of this combined
probability distribution as the bias-corrected parallax value,
$\varpi_{\rm Corr}$.
In simulating the volumetric prior, we model a Galaxy of pulsars with
a radius of $15\,$kpc and with the radial density profile derived by
\citet{lfl+06} (their Equation 10, Model C fit):
\[
\rho(R) = A\left(\frac{R}{\rm R_{\odot}}\right)^B
\exp\left[-C\frac{R-{\rm R_{\odot}}}{\rm R_{\odot}}\right],
\]
with constants $A = 41\,$kpc$^{-2}$, $B = 1.9$, $C = 5$ and
R$_{\odot}=8.5\,$kpc (i.e. the distance between the Sun and the
Galactic centre). The density distribution above the Galactic plane is
modelled by an exponential distribution $p(z) \propto \exp(-|z|/E)$ where
$E$ is the scale height, taken to be 330\,pc for common pulsars
\citep[in accordance with the findings of][]{lfl+06} and 500\,pc for
millisecond pulsars \citep{lor95,cc97}. To optimise the volumetric
prior, we use the simulated Galaxy of pulsars to determine a
volumetric prior probability density function for sky sectors of
10$^{\circ}$ in Galactic latitude and 15$^{\circ}$ in Galactic
longitude and for the polar regions, defined as having a Galactic
latitude in excess of 85$^{\circ}$.
The prior information of the intrinsic radio pulsar luminosity
function is simulated through random realisations of pulsar
luminosities from the log-normal distribution derived by \citet{fk06},
with mean intrinsic luminosity 0.07\,mJy\,kpc$^2$ (i.e. $\langle \log
L\rangle = -1.1$) and standard deviation $\sigma_{\log L} = 0.9$ in
base-10 logarithms. This luminosity distribution is subsequently
converted into a distribution of parallaxes using the known pulsar
flux, $S$, via: $S \varpi^{-2} = L$. The flux values we used are given
in column 3 of Table \ref{tab:curpxs}. Notice that, unlike the
power-law luminosity function described in Section \ref{ssec:Basis},
the log-normal distribution does have a maximum. This implies that,
depending on the pulsar flux, parallaxes can be biased towards larger
as well as smaller values.
Our software for estimation of the bias is available
on-line\footnote{http://psrpop.phys.wvu.edu/LKbias}, both through a
web interface and for download and off-line execution.
\begin{figure}
\psfig{angle=0.0,width=8.5cm,figure=NewFig1.eps}
\caption{Example output from our Monte-Carlo simulations, applied to
the PSR J1713+0747 parallax measurement of \citet{cfw94}
($\varpi_{\rm 0} = 0.9 \pm 0.3\,$mas). The top figure shows the
volumetric and luminosity prior distributions (dashed and
dot-dashed lines respectively) as well as their normalised
product, the total prior distribution (solid line). The bottom
figure shows the original measurement (dotted line), the prior
distribution (dashed line) and their normalised product: the
bias-corrected probability distribution (solid line). Vertical
lines show the peak and 1$\,\sigma$ error bars of the measured
(dotted) and bias-corrected (solid line) probability
distributions.}
\label{fig:sample}
\end{figure}
\section{Parallax Measurement Revisions}\label{sec:revise}
We have determined the bias-corrected confidence intervals for all
known VLBI and timing parallax measurements found in the ATNF pulsar
catalogue \citep{mhth05}, on Shami Chatterjee's dedicated
webpage\footnote{Shami Chatterjee's pulsar parallax webpage can be
found at: http://www.astro.cornell.edu/research/parallax.} and
through NASA's Astrophysics Data System Bibliographic Services. These
results can be found in Tables \ref{tab:curpxs} and
\ref{tab:oldpxs}. Note that in these tables and throughout the paper,
we quote 1\,$\sigma$ error bars.
In order to evaluate the potential of our bias-correction method, we
compare historic measurements and their corrected values to more
precise, recent measurements in Section \ref{ssec:history}. Next, we
discuss the most significant revisions presented, in Section
\ref{ssec:Significant}. Possible effects on models of the interstellar
medium density and distances derived from these models are discussed
in Section \ref{ssec:ISM}. In Section \ref{ssec:specific} the
potential impact of Lutz-Kelker bias on systems of particular interest
for fundamental tests of gravitational theories is described.
Finally, bias corrections for some non-radio pulsars are considered in
Section \ref{ssec:nonRadio}.
\input{TableA1}
\input{TableA2}
\subsection{Historic Parallax Measurements}\label{ssec:history}
As discussed in \S\ref{ssec:Basis} and shown by Equation
\ref{eq:bassol}, the size of the Lutz-Kelker bias is strongly
dependent on the significance of the input parallax measurement. This
implies that, as increasingly precise parallax measurements are
published, those parallax values will converge towards decreasing
bias. Comparing the historic parallax measurements collated in Table
\ref{tab:oldpxs} to their most precise updated values, we can now
assess if these corrections have been towards the peak of the prior
distribution, or away from it, respectively indicated by ``T'' and
``A'' in column 5 of Table \ref{tab:oldpxs}. (The peak of the prior
probability density function is determined from the pulsar radio flux
and the volumetric probability density function towards the pulsar and
is provided as $\varpi_{\rm Prior}$ in column 5 of Table
\ref{tab:curpxs}.)
This inspection shows that 31 out of 43 revisions were in the
direction predicted by the prior distributions. The chance of this
happening at random is less than $0.5\%$ so this confirms the
described bias effects are present in the data. It does not, however,
imply bias-correction can resolve all these variations. Specifically,
measurements with underestimated uncertainties will be
under-corrected, as is supposedly the case for the PSR J0953+0755
measurement by \citet{gtwr86} \citep[see \S4 in][for a discussion of
this measurement]{bbb+00}. We note that nearly half of all revisions
listed in Table \ref{tab:oldpxs} are significant at the 1\,$\sigma$
level, irrespective of bias-correction. This strongly suggests more
wide-spread underestimation of parallax uncertainties. Also, in cases
where stronger systematic biases are present, application of the
Lutz-Kelker correction will result in equally unreliable results. For
PSR J0437$-$4715 for example, \citet{vbv+08} have demonstrated that
the parallax value published by \citet{vbb+01} differed from later
measurements because of inaccuracies in the Solar System Ephemeris
models used.
\subsection{Significant Parallax Revisions}\label{ssec:Significant}
As follows from Table \ref{tab:curpxs}, out of the 56 pulsars that
currently have parallax measurements from radio or optical data, only
9 have a bias larger than one standard deviation. In the case of PSRs
J0633+1746 and J0720$-$3125 this is mostly due to the absence of
intrinsic luminosity information, as described in Section
\ref{ssec:nonRadio}, while for PSR J0751+1807 the parallax bias is
mostly caused by the low significance of the measurement. For the
remaining 6 pulsars the significance of the bias is caused by the fact
that the luminosity and volumetric prior peak on the same side of the
measured value. This causes the two components to the prior to
reinforce each other, resulting in more significant bias corrections.
The bias correction of PSR J0108$-$1431 is of particular interest,
since this pulsar was once assumed to be the closest known neutron
star \citep{tnj+94}, based on its DM distance of $D_{\rm 1994} =
130\,$pc as derived from the Galactic electron density model of
\citet{tc93}. The updated model of \citet{cl02} increased the distance
to $D_{\rm 2002} = 200\,$pc and a recent VLBI parallax \citep{dtbr09}
further increased this value to $D_{\rm 2009} =
240^{+124}_{-61}\,$pc. Our bias-corrected parallax $\varpi_{\rm Corr}
= 1.6^{+1.6}_{-0.6}$ places the pulsar at a distance of $D_{\rm
J0108-1431} = 625_{-313}^{+375}\,$pc, strongly suggesting that the
3\,$\sigma$ parallax measurement of \citet{dtbr09} is still
underestimating the distance.
The bias-corrected parallax of PSR B1541+09 is 1.5\,$\sigma$ larger
than the VLBI value. This pulsar has a Galactic latitude of
46$^{\circ}$, which in combination with the parallax measurement of
$\varpi_{\rm 0} = 0.13 \pm 0.02$ \citep{cbv+09} would place it at a
height of 5.5\,kpc above the Galactic disk. This is a highly unlikely
height given the scale height of 330\,pc for normal pulsars,
suggesting that either the parallax measurement is strongly
underestimated because of bias effects or systematic errors, or the
scale height in our Galactic model is underestimated near the Galactic bulge.
\subsection{ISM Density Models}\label{ssec:ISM}
As mentioned in Section \ref{sec:intro}, models of the Galactic
electron distribution are based on pulsar parallax measurements
\citep{cl02}. Since the estimated biases for most measurements are
small (as follows from Table \ref{tab:curpxs}) and the uncertainties
in the electron density models are relatively large, bias correction
is not expected to significantly affect such models. In future
modelling efforts, however, bias-corrected parallax values should
result in more reliable models.
Conversely, Galactic electron density models are used to estimate
pulsar distances based on the DM. While these distance estimates are
not trigonometric, the same reasoning as presented in Section
\ref{ssec:Basis} holds and DM-derived distances can be expected to be
affected by Lutz-Kelker bias. In order to investigate the presence of
Lutz-Kelker bias in DM distances, we consider all pulsars with
trigonometric parallaxes that were not contained in the most recent
Galactic electron density model and that have measurement
uncertainties lower than 20\% \citep[their Figure 12]{cl03}. The DM
distances and parallax values for these 21 pulsars are compiled in
Table \ref{tab:DM}, assuming a 20\% uncertainty on the DM
distance. Comparison of the DM distances with the more precise
parallax measurements from pulsar timing and VLBI compiled in Table
\ref{tab:curpxs} shows that in 15 of the 21 cases the more precise
measurement lies closer to the peak of the prior distribution
$\varpi_{\rm Prior}$, indicating the presence of Lutz-Kelker
bias. Bias-corrected DM-derived parallaxes are also provided, though
interpretation of these values is complicated by the fact that the
uncertainties of DM distances are rough estimates. This analysis
demonstrates that DM distances are likely to be underestimated, a fact
that will need to be taken into account in research that makes use of
such distances, for example in the determination of pulsar velocities
as in \citet{hllk05}. \input{Table3}
\subsection{Binary Pulsars}\label{ssec:specific}
One particular application where bias correction may prove crucial in
future research, is the determination of gravitational wave emission
from binary pulsar systems. This emission is measured through orbital
period decay, predicted by general relativity to be \citep{tw82}:
\[
\dot{P}_{\rm b}^{\rm GR} =
- \frac{192\pi}{5}
\left(
\frac{2 \pi T_{\odot}}
{P_{\rm b}}
\right)^{5/3}
\frac{M_{\rm c} M_{\rm p}}{M_{\rm tot}^{1/3}}
\frac{1+\frac{73}{24}e^2+\frac{37}{96}e^4}{\left(1-e^2\right)^{7/2}},
\]
with $\dot{P}_{\rm b}$ the first derivative of the orbital period,
$P_{\rm b}$, $T_{\odot} = 4.925490947 \mu$s, $M_{\rm p}$ the pulsar
mass, $M_{\rm c}$ the companion mass, $M_{\rm tot} = M_{\rm p}+M_{\rm
c}$ the total system mass and $e$ the orbital eccentricity. Given an
independent measurement of $M_{\rm c}$ and $M_{\rm p}$, a measurement
of $\dot{P}_{\rm b}$ can be used to test gravitational wave emission
theories, as first done by \citet{tw82}. Alternatively, it can be used
to determine the pulsar and companion masses, as in
\citet{nss+05}. Depending on the pulsar's proper motion and position
in the Galaxy, however, certain kinematic terms may contaminate the
$\dot{P}_{\rm b}$ measurement, as first described in detail by
\citet{dt91}. They showed that three extra terms need to be considered
and corrected for: the Shklovskii term based on the pulsar proper
motion and distance \citep{shk70}; the apparent acceleration caused by
the gravitational potential of the Galaxy, dependent on the pulsar's
Galactic latitude and height above the Galactic plane; and the
apparent acceleration caused by differential Galactic rotation, which
is also dependent on the Galactic latitude and therefore the distance
of the pulsar. Since all three non-relativistic contributions to the
measured orbital period derivative $\dot{P}_{\rm b}$ are dependent on
the pulsar distance, accurate estimation of these contaminating
factors and subsequent determination of gravitational wave emission or
pulsar mass estimates, is dependent on an accurate parallax
determination.
Presently, the orbital decay $\dot{P}_{\rm b}$ has only been
determined for 16 binary pulsar systems, of which only four have a
parallax measurement. In the cases of PSRs J0437$-$4715 and
J1909$-$3744 the parallax measurement is precise enough to remain
unaffected by bias effects and in the cases of PSRs J0737$-$3039A/B
and J0751+1807 the uncertainty in the measured $\dot{P}_{\rm b}$ value
is substantially larger than the expected contributions from kinematic
and Galactic effect \citep{ksm+06,nss+05}.
\subsection{Optical Pulsar Parallax Measurements}\label{ssec:nonRadio}
A total of four neutron stars have had their parallaxes measured
though optical observations with the Hubble Space Telescope (PSRs
J0633+1746, J0720$-$3125, J0835$-$4510 and J1856$-$3754). Of these
four, only the Vela pulsar (PSR J0835$-$4510) has been detected at
radio wavelengths. Since an optical intrinsic luminosity function of
neutron stars has thus far not been determined, we disregard the
luminosity prior for the remaining three optical neutron
stars. Because surveys for radio pulsars are generally biased towards
bright pulsars, most known radio pulsars lie at the bright end of the
intrinsic luminosity function. This implies that the luminosity prior
tends to counterbalance the volumetric prior, which is biased towards
small parallax values (as seen in Figure \ref{fig:sample}). The
exclusion of the luminosity prior for non-radio neutron stars implies
that the effect of the volumetric prior becomes much stronger, which
means higher measurement precision is required before anything
definite can be said about the bias in the neutron star distance. A
clear example of this is PSR J1856$-$3754 for which the 3.5\,$\sigma$
measurement of \citet{kva02} is insufficient to overcome the steep
volumetric prior. For pulsars with luminosity information, a
3.5\,$\sigma$ measurement generally \emph{does} suffice to restrict
the bias within the 68\% confidence interval, as illustrated by the
PSR J0613$-$0200 measurement of \citet{hbo06}. We note that the most
precise parallaxes for PSRs J0633+1746 and J0720$-$3125 are only
3.6\,$\sigma$ and 3.1\,$\sigma$ measurements respectively, which is
insufficient to claim confidence in the inferred distance because of
the poor estimate of the bias.
X-ray luminosities of neutron stars are commonly used to derive
neutron star radii which in turn are used to place constraints on
equations of state for dense nuclear matter. One of the limiting
factors in such analyses are the high distance uncertainties since
distances are mainly derived from mostly inprecise methods such as
interstellar dispersion, supernova remnant associations and
interstellar absorption lines, while only few parallax measurements
are available \citet{plps04}. This lack of precise distances
causes even 2 and 3\,$\sigma$ parallax measurements to
carry significant weight in these analyses. As our work shows, without
prior information on the intrinsic luminosity distribution of these
sources, it cannot be confidently claimed that parallax measurements
with such low significance are accurate enough to be used because they
are likely to underestimate the distance to the neutron star. Based on
equations 56 and 58 as well as Figure 2A of \citet{lp07}, these
underestimated distances can be expected to result in underestimations
of neutron star masses and radii, slightly biasing the analysis to
prefer softer equations of state.
\section{Conclusions}\label{sec:conc}
We have reanalysed bias effects first discussed by \citet{lk73} and
described Monte-Carlo simulations that aim to correct for these
biases. Comparison of historic parallax values to the most recent
measurements confirms that the bias effects are present in
observations, though the correction is complicated by systematic
measurement errors and underestimation of parallax
uncertainties. Correction for the described biases may improve some
pulsar-timing tests of gravitational wave emission and may slightly
influence future density models of the interstellar medium. Finally,
we conclude that optical parallax measurements should be used with
caution since we cannot acurately quantify the possible bias in the
inferred distance estimates.
{\bf \noindent Acknowledgments.}
JPWV, DRL and MAM acknowledge support from a WVEPSCoR research
challenge grant held by the WVU Center for Astrophysics. MAM is an
Alfred P. Sloan Fellow. This work has made extensive use of the ATNF
pulsar catalogue \citep{mhth05} and Shami Chatterjee's webpage of
pulsar parallax measurements:
http://www.astro.cornell.edu/research/parallax. The authors thank the
referee Shami Chatterjee for insightful and constructive
comments. JPWV thanks Adam Deller and George Hobbs for useful
discussions on the subject.
\bibliographystyle{mn2e}
|
\section{Introduction}
Tilting theory has been a central topic in the representation theory of algebras.
This theory was originally introduced in the study of module categories over finite-dimensional algebras, and is now an important notion in a lot of areas of mathematics, including finite and algebraic group theory, commutative and noncommutative algebraic geometry, and algebraic topology; see \cite{hand} for the details.
Cohen-Macaulay approximation theory due to Auslander and Buchweitz \cite{ABu} is an aspect of tilting theory for commutative algebra.
Dualizing modules, which play a critical role in Cohen-Macaulay approximation theory, are special types of cotilting module.
The notion of a contravariantly finite subcategory (of the category of finitely generated modules), which is also called a (pre)covering class, was first introduced over artin algebras by Auslander and Smal{\o} \cite{AS} in connection with studying the problem of which subcategories admit almost split sequences.
The notion of a resolving subcategory was introduced by Auslander and Bridger \cite{ABr} in the study of totally reflexive modules, which are also called modules of Gorenstein dimension zero or finitely generated Gorenstein projective modules.
There is an application of contravariantly finite resolving subcategories to the study of the finitistic dimension conjecture \cite{AR}.
For other details of contravariantly finite subcategories and resolving subcategories, see \cite{aus0,AS2,AR1,AR2,ASo,Marz,KS,MR,Belig,Reiten} in addition to the articles introduced above.
On the other hand, there are a lot of articles which study contravariantly finite subcategories of the category of all modules; see \cite{Enochs,EJ,ATT,BEE} for example.
It was found out by Auslander and Reiten \cite{AR} that the notion of a contravariantly finite resolving subcategory is closely related to tilting theory; they classified contravariantly finite resolving subcategories over an artin algebra of finite global dimension in terms of cotilting modules.
\begin{thm}[Auslander-Reiten]\label{ausrei}
Let $R$ be an artin algebra of finite global dimension.
Then the assignment $T\mapsto{}^\perp T$ makes a bijection from the set of isomorphism classes of basic cotilting $R$-modules to the set of contravariantly finite resolving subcategories of $\operatorname{mod} R$.
\end{thm}
\noindent
Here, $\operatorname{mod} R$ denotes the category of finitely generated $R$-modules, and ${}^\perp T$ the full subcategory of $\operatorname{mod} R$ consisting of all finitely generated modules $M$ with $\operatorname{Ext}_R^i(M,T)=0$ for $i>0$.
This paper deals with contravariantly finite resolving subcategories over commutative rings.
The main result of this paper is the following classification theorem.
\begin{thm}\label{goren}
Let $R$ be a commutative henselian (e.g. complete) Gorenstein local ring.
Then all the contravariantly finite resolving subcategories of $\operatorname{mod} R$ are:
\begin{itemize}
\item
the full subcategory of free modules,
\item
the full subcategory of maximal Cohen-Macaulay modules, and
\item
$\operatorname{mod} R$ itself.
\end{itemize}
\end{thm}
The assumption of finiteness of global dimension in Theorem \ref{ausrei} is essential to obtain the bijection in the theorem.
Without this assumption, in general there exist so many contravariantly finite resolving subcategories, and it is extremely difficult to classify them; see \cite{MR}.
Theorem \ref{goren} should be remarkable since in it the ring $R$ is not assumed to have finite global dimension.
By the way, over an arbitrary commutative noetherian local ring, all the three subcategories in Theorem \ref{goren} are resolving, and the first and third subcategories are contravariantly finite.
The second subcategory is contravariantly finite if the base ring is a Cohen-Macaulay local ring admitting a dualizing module.
This fact is known as the Cohen-Macaulay approximation theorem due to Auslander and Buchweitz \cite{ABu}.
From now on, we explain our main results more minutely.
Throughout the rest of this section, let $R$ be a commutative noetherian henselian local ring.
We will give a sufficient condition for a given resolving subcategory to be a test for injective dimension:
\begin{thm}\label{mainlem}
Let ${\mathcal X}$ be a resolving subcategory of $\operatorname{mod} R$ such that the residue field of $R$ has a right ${\mathcal X}$-approximation.
Assume that there exists an $R$-module $G\in{\mathcal X}$ of infinite projective dimension with $\operatorname{Ext}_R^i(G,R)=0$ for $i\gg 0$.
Let $M$ be an $R$-module such that each $X\in{\mathcal X}$ satisfies $\operatorname{Ext}_R^i(X,M)=0$ for $i\gg 0$.
Then $M$ has finite injective dimension.
\end{thm}
Using this result, we will prove the following theorem.
Theorem \ref{goren} will be obtained from this.
\begin{thm}\label{cohen}
Let ${\mathcal X}\ne\operatorname{mod} R$ be a contravariantly finite resolving subcategory of $\operatorname{mod} R$.
Suppose that there is an $R$-module $G\in{\mathcal X}$ of infinite projective dimension such that $\operatorname{Ext}_R^i(G,R)=0$ for $i\gg 0$.
Then $R$ is Cohen-Macaulay and ${\mathcal X}$ consists of all maximal Cohen-Macaulay $R$-modules.
\end{thm}
On the other hand, Theorem \ref{mainlem} also yields the following corollary as an immediate consequence.
The assertion of this corollary is a main result of \cite{CPST}.
(Our method to get the corollary is quite different from the proof given in \cite{CPST}.)
\begin{cor}[Christensen-Piepmeyer-Striuli-Takahashi]\label{ccpst}
Suppose that there is a nonfree totally reflexive $R$-module.
If the full subcategory of $\operatorname{mod} R$ consisting of all totally reflexive $R$-modules is contravariantly finite, then $R$ is Gorenstein.
\end{cor}
A totally reflexive module was defined by Auslander \cite{aus1} as a common generalization of a free module and a maximal Cohen-Macaulay module over a Gorenstein local ring.
Auslander and Bridger \cite{ABr} proved that the full subcategory of totally reflexive modules over a left and right noetherian ring is resolving.
Many other properties of totally reflexive modules are stated in \cite{ABr} and \cite{chri}.
If $R$ is Gorenstein, then the totally reflexive $R$-modules are precisely the maximal Cohen-Macaulay $R$-modules, and so the full subcategory of totally reflexive $R$-modules is contravariantly finite by virtue of the Cohen-Macaulay approximation theorem.
Thus, Corollary \ref{ccpst} can be viewed as the converse of this fact.
Corollary \ref{ccpst} implies a geometric result: let $R$ be a homomorphic image of a regular local ring.
Suppose that there is a nonfree totally reflexive $R$-module and are only finitely many nonisomorphic indecomposable totally reflexive $R$-modules.
Then $R$ is an isolated simple hypersurface singularity.
For the details, see \cite{CPST}.
\section*{Conventions}
In the rest of this paper, we assume that all rings are commutative and noetherian, and that all modules are finitely generated.
Unless otherwise specified, let $R$ be a henselian local ring.
The unique maximal ideal of $R$ and the residue field of $R$ are denoted by $\mathfrak m$ and $k$, respectively.
We denote by $\operatorname{mod} R$ the category of finitely generated $R$-modules.
By a {\em subcategory} of $\operatorname{mod} R$, we always mean a full subcategory of $\operatorname{mod} R$ which is closed under isomorphism.
Namely, in this paper, a subcategory ${\mathcal X}$ of $\operatorname{mod} R$ means a full subcategory such that every $R$-module which is isomorphic to some $R$-module in ${\mathcal X}$ is also in ${\mathcal X}$.
\section{Contravariant finiteness of totally reflexive modules}
In this section, we will state background materials which motivate the main results of this paper.
We start by recalling the definition of a totally reflexive module.
\begin{defn}
We denote by $(-)^\ast$ the $R$-dual functor $\operatorname{Hom}_R(-,R)$.
An $R$-module $M$ is called {\em totally reflexive} (or {\em of Gorenstein dimension zero}) if
\begin{enumerate}[\rm (1)]
\item
the natural homomorphism $M\to M^{\ast\ast}$ is an isomorphism, and
\item
$\operatorname{Ext}_R^i(M,R)=\operatorname{Ext}_R^i(M^\ast,R)=0$ for any $i>0$.
\end{enumerate}
\end{defn}
We introduce three subcategories of $\operatorname{mod} R$ which will often appear throughout this paper.
\begin{note}
We denote by ${\mathcal F}(R)$ the subcategory of $\operatorname{mod} R$ consisting of all free $R$-modules, by ${\mathcal G}(R)$ the subcategory of $\operatorname{mod} R$ consisting of all totally reflexive $R$-modules, and by ${\mathcal C}(R)$ the subcategory of $\operatorname{mod} R$ consisting of all maximal Cohen-Macaulay $R$-modules.
\end{note}
\begin{rem}\label{basic}
The following are basic properties of the subcategories ${\mathcal F}(R)$, ${\mathcal G}(R)$ and ${\mathcal C}(R)$.
\begin{enumerate}[\rm (1)]
\item
Every free $R$-module is totally reflexive, namely, ${\mathcal F}(R)$ is contained in ${\mathcal G}(R)$.
\item
If $R$ is Cohen-Macaulay, then all totally reflexive $R$-modules are maximal Cohen-Macaulay, namely, ${\mathcal G}(R)$ is contained in ${\mathcal C}(R)$.
\item
If $R$ is Gorenstein, then the totally reflexive $R$-modules are precisely the maximal Cohen-Macaulay $R$-modules, namely, ${\mathcal G}(R)$ coincides with ${\mathcal C}(R)$.
\end{enumerate}
The first and third assertions are immediate from the definition of a totally reflexive module.
The second assertion follows from \cite[Theorem (1.4.8)]{chri}.
\end{rem}
Next, we recall the notion of a right approximation over a subcategory of $\operatorname{mod} R$.
\begin{defn}
Let ${\mathcal X}$ be a subcategory of $\operatorname{mod} R$.
\begin{enumerate}[\rm (1)]
\item
Let $\phi:X\to M$ be a homomorphism of $R$-modules with $X\in{\mathcal X}$.
We say that $\phi$ is a {\em right ${\mathcal X}$-approximation} (of $M$) if the induced homomorphism $\operatorname{Hom}_R(X',\phi):\operatorname{Hom}_R(X',X)\to\operatorname{Hom}_R(X',M)$ is surjective for any $X'\in{\mathcal X}$.
\item
We say that ${\mathcal X}$ is {\em contravariantly finite} (in $\operatorname{mod} R$) if every $R$-module has a right ${\mathcal X}$-approximation.
\end{enumerate}
\end{defn}
By definition, contravariant finiteness can be regarded as a weaker version of the property that the inclusion functor has a right adjoint.
\begin{ex}\label{freecov}
Let $R$ be an arbitrary ring and $n$ a nonnegative integer.
The following subcategories of $\operatorname{mod} R$ are contravariantly finite, assuming in addition that $R$ is artinian in (7) and (8).
\begin{enumerate}[\rm (1)]
\item
$\operatorname{mod} R$.
\item
${\mathcal F}(R)$.
\item
${\mathcal C}(R)$, provided that $R$ is a Cohen-Macaulay local ring admitting a dualizing module.
\item
A subcategory of $\operatorname{mod} R$ closed under finite direct sums and direct summands in which there are only finitely many isomorphism classes of indecomposable $R$-modules.
\item
The subcategory of $\operatorname{mod} R$ consisting of all finitely generated $S$-modules, where $S$ is a fixed module-finite $R$-algebra.
\item
The subcategory of $\operatorname{mod} R$ consisting of all $M$ satisfying $\operatorname{Ext}_R^1(M,R)=0$.
\item
The subcategory of $\operatorname{mod} R$ consisting of all $M$ such that there exists an exact sequence $0\to M\to P^0\to\cdots\to P^{n-1}$ with each $P^i$ projective.
\item
The subcategory of $\operatorname{mod} R$ consisting of all $M$ such that there exists an exact sequence $I_{n-1}\to\cdots\to I_0\to M\to 0$ with each $I_i$ injective.
\end{enumerate}
The subcategory (1) is contravariantly finite since the identity map of every $R$-module is a right $\operatorname{mod} R$-approximation.
As to the fact that (2) is contravariantly finite, for a given $R$-module $M$, a surjective homomorphism from a module in ${\mathcal F}(R)$ to $M$ (i.e. a free cover of $M$) is exactly a right ${\mathcal F}(R)$-approximation of $M$.
As for the facts that the subcategories (3)-(8) are contravariantly finite, see \cite[Theorem 1.1]{ABu}, \cite[Proposition 4.2]{AS} (or \cite[Proof of Theorem 1.4]{dep0}), \cite[Theorem 1.3.1(1)]{Iyama}, \cite[Proposition (2.21)]{ABr} (or \cite[Theorem 7.1]{AS2} or \cite[Corollary 5.8]{tor}), \cite[Theorem 1.2]{AR3} and \cite[Theorem 1.8]{AR2} , respectively.
See also \cite[Theorems 4.3 and 4.4]{HI}.
\end{ex}
The following result is immediately obtained from Example \ref{freecov}(3) and Remark \ref{basic}(3).
\begin{thm}
If $R$ is Gorenstein, then ${\mathcal G}(R)$ is contravariantly finite.
\end{thm}
Recently, it has been proved that the (essential) converse of this theorem holds:
\begin{thm}\cite{CPST}\label{cpst}
Suppose that there is a nonfree totally reflexive $R$-module.
If ${\mathcal G}(R)$ is contravariantly finite, then $R$ is Gorenstein.
\end{thm}
Yoshino \cite{yos1,yos2} and Takahashi \cite{dep0,dep1,dep2,catgp} proved the statement of Theorem \ref{cpst} in several special cases.
Theorem \ref{cpst} is shown in \cite{CPST} by choosing a good maximal $R$-regular sequence $\text{\boldmath $x$}=x_1,\dots,x_t$ and considering the smallest extension closed additive subcategory of $\operatorname{mod}(R/\text{\boldmath $x$} R)$ containing all $R/\text{\boldmath $x$} R$-modules $X/\text{\boldmath $x$} X$ with $X\in{\mathcal G}(R)$.
Theorem \ref{cpst} will be recovered by Theorem \ref{main} stated in the next section, and the proof of Theorem \ref{main} is quite different from that of Theorem \ref{cpst} given in \cite{CPST}.
By the way, the assumption in Theorem \ref{cpst} of existence of a nonfree totally reflexive module is necessary.
Indeed, let $R$ be a non-Gorenstein Golod local ring (e.g. a non-Gorenstein Cohen-Macaulay local ring with minimal multiplicity).
Then ${\mathcal G}(R)$ coincides with ${\mathcal F}(R)$ by \cite[Examples 3.5(2)]{AM}.
In particular, ${\mathcal G}(R)$ is contravariantly finite by Example \ref{freecov}(2).
Theorem \ref{cpst} and Example \ref{freecov}(4) yield the following corollary, which improves \cite[Theorem 1.3]{igp}.
\begin{cor}
Let $R$ be a non-Gorenstein local ring.
If there is a nonfree totally reflexive $R$-module, then there are infinitely many nonisomorphic indecomposable totally reflexive $R$-modules.
\end{cor}
Combining this with \cite[Theorems (8.15) and (8.10)]{yos0} (cf. \cite[Satz 1.2]{herzog} and \cite[Theorem B]{BGS}), one obtains the following result.
\begin{cor}\label{cpst2}
Let $R$ be a homomorphic image of a regular local ring.
Suppose that there is a nonfree totally reflexive $R$-module but there are only finitely many nonisomorphic indecomposable totally reflexive $R$-modules.
Then $R$ is a simple hypersurface singularity.
\end{cor}
\section{Contravariantly finite resolving subcategories}
In this section, we will give the main theorems of this paper.
The most general result is Theorem \ref{main}, which implies all of the other results obtained in this paper.
First of all, we recall the definition of the syzygies of a given module.
Let $M$ be an $R$-module and $n$ a positive integer.
Let
$$
F_\bullet =(\cdots \overset{d_{n+1}}{\to} F_n \overset{d_n}{\to} F_{n-1} \overset{d_{n-1}}{\to} \cdots \overset{d_2}{\to} F_1 \overset{d_1}{\to} F_0 \to 0)
$$
be a minimal free resolution of $M$.
We define the {\em $n$th syzygy} $\Omega^nM$ of $M$ as the image of the homomorphism $d_n$.
We set $\Omega^0M=M$.
Note that the $n$th syzygy $\Omega^nM$ of $M$ is uniquely determined up to isomorphism, since so is a minimal free resolution of $M$.
Now we recall the definition of a resolving subcategory.
\begin{defn}\label{resolv}
A subcategory ${\mathcal X}$ of $\operatorname{mod} R$ is called {\em resolving} if it satisfies the following four conditions.
\begin{enumerate}[\rm (1)]
\item
${\mathcal X}$ contains $R$.
\item
${\mathcal X}$ is closed under direct summands: if $M$ is an $R$-module in ${\mathcal X}$ and $N$ is a direct summand of $M$, then $N$ is also in ${\mathcal X}$.
\item
${\mathcal X}$ is closed under extensions: for an exact sequence $0 \to L \to M \to N \to 0$ of $R$-modules, if $L$ and $N$ are in ${\mathcal X}$, then $M$ is also in ${\mathcal X}$.
\item
${\mathcal X}$ is closed under kernels of epimorphisms: for an exact sequence $0 \to L \to M \to N \to 0$ of $R$-modules, if $M$ and $N$ are in ${\mathcal X}$, then $L$ is also in ${\mathcal X}$.
\end{enumerate}
\end{defn}
A resolving subcategory is a subcategory such that any two ``minimal'' resolutions of a module by modules in it have the same length; see \cite[(3.12)]{ABr}.
\begin{rem}\label{resbas}
Let ${\mathcal X}$ be a resolving subcategory of $\operatorname{mod} R$.
Then
\begin{enumerate}[\rm (1)]
\item
${\mathcal X}$ is closed under finite direct sums: if $M_1,\dots,M_n$ are a finite number of $R$-modules in ${\mathcal X}$, then so is the direct sum $M_1\oplus\cdots\oplus M_n$.
\item
${\mathcal X}$ is closed under syzygies: if $M$ is an $R$-module in ${\mathcal X}$, then so is $\Omega^nM$ for every $n\ge 0$.
\end{enumerate}
Indeed, as to (1), it is enough to check that if $M$ and $N$ are $R$-modules in ${\mathcal X}$, then so is $M\oplus N$.
This follows by applying Definition \ref{resolv}(3) to the natural split exact sequence $0 \to M \to M\oplus N \to N \to 0$.
As to (2), it suffices to show that if $M$ is an $R$-module in ${\mathcal X}$, then so is the first syzygy $\Omega M$.
There is an exact sequence $0 \to \Omega M \to F \to M \to 0$ with $F$ free, and $F$ is in ${\mathcal X}$ by Definition \ref{resolv}(1) and the first assertion of this remark.
Hence $\Omega M$ is in ${\mathcal X}$ by Definition \ref{resolv}(4).
\end{rem}
\begin{ex}\label{rslvx}
Let $I$ be an ideal of $R$, $K$ an $R$-module, and $n$ a nonnegative integer.
The following subcategories of $\operatorname{mod} R$ are resolving.
\begin{enumerate}[\rm (1)]
\item
$\operatorname{mod} R$.
\item
${\mathcal F}(R)$.
\item
${\mathcal G}(R)$.
\item
$\{\,M\mid\operatorname{grade}(I,M)\ge\operatorname{grade}(I,R)\,\}$.
\item
${\mathcal C}(R)$, provided that $R$ is Cohen-Macaulay.
\item
$\{\,M\mid M\text{ has bounded Betti numbers}\,\}$.
\item
$\{\,M\mid M\text{ has finite complexity}\,\}$.
\item
$\{\,M\mid M\text{ has lower complete intersection dimension zero}\,\}$.
\item
$\{\,M\mid\operatorname{Tor}_i^R(M,K)=0\text{ for }i>n\,\}$.
\item
$\{\,M\mid\operatorname{Ext}_R^i(M,K)=0\text{ for }i>n\,\}$.
\item
$\{\,M\mid\operatorname{Tor}_i^R(M,K)=0\text{ for }i\gg 0\,\}$.
\item
$\{\,M\mid\operatorname{Ext}_R^i(M,K)=0\text{ for }i\gg 0\,\}$.
\end{enumerate}
The fact that the subcategory (3) is resolving follows from \cite[(3.11)]{ABr} or \cite[Lemma 2.3]{AM}.
The resolving property of (5) follows from that of (4).
The fact that (7) and (8) are resolving follows from \cite[Proposition 4.2.4]{Avramov98} and \cite[Lemma 6.3.1]{Avramov02}, respectively.
The resolving properties of the other subcategories can be checked directly.
Note that the subcategories (11) and (12) are thick.
(A subcategory ${\mathcal X}$ of $\operatorname{mod} R$ is called {\em thick} provided that for any exact sequence $0 \to L \to M \to N \to 0$ in $\operatorname{mod} R$, if two of $L,M,N$ are in ${\mathcal X}$, then so is the third.)
\end{ex}
Now we state the most general result in this paper.
\begin{thm}\label{main}
Let ${\mathcal X}$ be a resolving subcategory of $\operatorname{mod} R$ such that the residue field $k$ has a right ${\mathcal X}$-approximation.
Assume that there exists an $R$-module $G\in{\mathcal X}$ of infinite projective dimension such that $\operatorname{Ext}_R^i(G,R)=0$ for $i\gg 0$.
Let $M$ be an $R$-module such that each $X\in{\mathcal X}$ satisfies $\operatorname{Ext}_R^i(X,M)=0$ for $i\gg 0$.
Then $M$ has finite injective dimension.
\end{thm}
Let ${\mathcal X}$ be as in Theorem \ref{main}.
Let $M$ be an $R$-module.
Theorem \ref{main} says that if $\operatorname{Ext}_R^{\gg 0}(X,M)=0$ for each $X\in{\mathcal X}$, then $\operatorname{Ext}_R^{\gg 0}(X,M)=0$ for each $X\in\operatorname{mod} R$.
In this sense, such a subcategory ${\mathcal X}$ of $\operatorname{mod} R$ is ``large''.
We shall prove Theorem \ref{main} in the next section.
In the rest of this section, we will state and prove several results by using Theorem \ref{main}.
We begin with two corollaries which are immediately obtained.
\begin{cor}\label{bestcor}
Let ${\mathcal X}$ be a resolving subcategory of $\operatorname{mod} R$ which is contained in the subcategory $\{\,M\mid\operatorname{Ext}_R^i(M,R)=0\text{ for }i\gg 0\,\}$ of $\operatorname{mod} R$.
Suppose that in ${\mathcal X}$ there is an $R$-module of infinite projective dimension.
If $k$ has a right ${\mathcal X}$-approximation, then $R$ is Gorenstein.
\end{cor}
\begin{proof}
Each module $X$ in ${\mathcal X}$ satisfies $\operatorname{Ext}_R^i(X,R)=0$ for $i\gg 0$.
Hence Theorem \ref{main} implies that $R$ has finite injective dimension as an $R$-module.
\end{proof}
\begin{cor}\label{maincor}
Let ${\mathcal X}$ be one of the following.
\begin{enumerate}[\rm (1)]
\item
${\mathcal G}(R)$.
\item
The subcategory $\{\,M\mid\operatorname{Ext}_R^i(M,R)=0\text{ for }i>n\,\}$ of $\operatorname{mod} R$, where $n$ is a nonnegative integer.
\item
The subcategory $\{\,M\mid\operatorname{Ext}_R^i(M,R)=0\text{ for }i\gg 0\,\}$ of $\operatorname{mod} R$.
\end{enumerate}
Suppose that in ${\mathcal X}$ there is an $R$-module of infinite projective dimension.
If $k$ has a right ${\mathcal X}$-approximation, then $R$ is Gorenstein.
\end{cor}
\begin{proof}
As we observed in Example \ref{rslvx}, ${\mathcal X}$ is a resolving subcategory of $\operatorname{mod} R$.
Since ${\mathcal X}$ is contained in the subcategory $\{\,M\mid\operatorname{Ext}_R^i(M,R)=0\text{ for }i\gg 0\,\}$, the assertion follows from Corollary \ref{bestcor}.
\end{proof}
\begin{rem}
Corollary \ref{maincor} implies Theorem \ref{cpst}.
Indeed, any nonfree totally reflexive module has infinite projective dimension by \cite[(1.2.10)]{chri}.
\end{rem}
For a subcategory ${\mathcal X}$ of $\operatorname{mod} R$, let ${\mathcal X}^\perp$ (respectively, ${}^\perp{\mathcal X}$) denote the subcategory of $\operatorname{mod} R$ consisting of all $R$-modules $M$ such that $\operatorname{Ext}_R^i(X,M)=0$ (respectively, $\operatorname{Ext}_R^i(M,X)=0$) for all $X\in{\mathcal X}$ and $i>0$.
Applying Wakamatsu's lemma to a resolving subcategory, we obtain the following lemma.
This will often be used in the rest of this paper.
\begin{lem}\label{wak}
Let ${\mathcal X}$ be a resolving subcategory of $\operatorname{mod} R$.
If an $R$-module $M$ has a right ${\mathcal X}$-approximation, then there is an exact sequence $0 \to Y \to X \to M \to 0$ of $R$-modules with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$.
\end{lem}
\begin{proof}
Since $R$ is henselian, $M$ has a {\em minimal} right ${\mathcal X}$-approximation: there exists a right ${\mathcal X}$-approximation $\phi:X\to M$ such that every endomorphism $f:X\to X$ with $\phi=\phi f$ is an automorphism; see \cite[Corollary 2.5]{dep0}.
Take an element $m\in M$, and define a homomorphism $\rho:R\to M$ by $\rho(1)=m$.
As ${\mathcal X}$ is a resolving subcategory, $R$ is in ${\mathcal X}$.
The homomorphism $\rho$ factors through $\phi$, and we see that $m=\phi(x)$ for some $x\in X$.
Thus $\phi$ is surjective.
Put $Y=\operatorname{Ker}\phi$.
Wakamatsu's lemma (see \cite{wk} or \cite[Lemma 2.1.1]{Xu}) shows that $\operatorname{Ext}_R^1(X',Y)=0$ for any $X'\in{\mathcal X}$.
We have $\operatorname{Ext}_R^i(X',Y)\cong\operatorname{Ext}_R^1(\Omega^{i-1}X',Y)=0$ for any $X'\in{\mathcal X}$ and $i>0$ by Remark \ref{resbas}(2).
Therefore $Y$ is in ${\mathcal X}^\perp$.
\end{proof}
By using Lemma \ref{wak} and the theorem which was formerly called ``Bass' conjecture'', we obtain another corollary of Theorem \ref{main}.
\begin{cor}\label{sbmncr2}
Let ${\mathcal X}$ be a resolving subcategory of $\operatorname{mod} R$ such that $k$ has a right ${\mathcal X}$-approximation and that $k$ is not in ${\mathcal X}$.
Assume that there is an $R$-module $G\in{\mathcal X}$ with $\operatorname{pd}_RG=\infty$ and $\operatorname{Ext}_R^i(G,R)=0$ for $i\gg 0$.
Then $R$ is Cohen-Macaulay and $\dim R>0$.
\end{cor}
\begin{proof}
According to Lemma \ref{wak}, there is an exact sequence $0 \to Y \to X \to k \to 0$ with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$.
It is seen from Theorem \ref{main} that the $R$-module $Y$ has finite injective dimension.
Since $k$ is not in ${\mathcal X}$, we have $Y\ne 0$.
By virtue of \cite[Corollary 9.6.2 and Remark 9.6.4(a)]{BH}, the ring $R$ is Cohen-Macaulay.
Assume that $\dim R=0$.
Then $Y$ is injective (cf. \cite[Theorem 3.1.17]{BH}), and so the exact sequence splits.
Hence $k$ is isomorphic to a direct summand of $X$, and $k$ is in ${\mathcal X}$ as ${\mathcal X}$ is closed under direct summands.
This contradiction shows that $R$ has positive Krull dimension.
\end{proof}
To give the next corollary of Theorem \ref{main}, we need the following three lemmas.
\begin{lem}\label{kx}
Let ${\mathcal X}$ be a contravariantly finite resolving subcategory of $\operatorname{mod} R$.
Then, $k\in{\mathcal X}$ if and only if ${\mathcal X}=\operatorname{mod} R$.
\end{lem}
\begin{proof}
The ``if'' part is clear.
Suppose that $k$ is in ${\mathcal X}$ and let us show that ${\mathcal X}$ coincides with $\operatorname{mod} R$.
Fix an $R$-module $M$.
Then Lemma \ref{wak} yields an exact sequence $0 \to Y \to X \to M \to 0$ of $R$-modules with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$.
We have $\operatorname{Ext}_R^i(k,Y)=0$ for all $i>0$, as $k\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$.
By \cite[Proposition 3.1.14]{BH}, $Y$ is an injective $R$-module, and so the exact sequence splits.
Hence $M$ is isomorphic to a direct summand of $X$.
Since ${\mathcal X}$ is closed under direct summands, $M$ is in ${\mathcal X}$.
This completes the proof of the lemma.
\end{proof}
\begin{lem}\label{prpprp}
Let ${\mathcal X}$ be a resolving subcategory of $\operatorname{mod} R$.
Suppose that every $R$-module in ${}^\perp({\mathcal X}^\perp)$ admits a right ${\mathcal X}$-approximation.
Then ${\mathcal X}={}^\perp({\mathcal X}^\perp)$.
\end{lem}
\begin{proof}
It is easy to see that ${\mathcal X}$ is contained in ${}^\perp({\mathcal X}^\perp)$.
Let $M$ be an $R$-module in ${}^\perp({\mathcal X}^\perp)$.
Then there is an exact sequence $0 \to Y \to X \to M \to 0$ of $R$-modules with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$ by Lemma \ref{wak}.
This exact sequence can be regarded as an element of $\operatorname{Ext}_R^1(M,Y)$, and this Ext module vanishes since $M\in {}^\perp({\mathcal X}^\perp)$ and $Y\in{\mathcal X}^\perp$.
Hence the exact sequence splits, and $M$ is isomorphic to a direct summand of $X$.
Since ${\mathcal X}$ is closed under direct summands, $M$ is in ${\mathcal X}$.
Thus ${\mathcal X}={}^\perp({\mathcal X}^\perp)$.
\end{proof}
\begin{lem}\label{pdidsup}
Let $M$ and $N$ be nonzero $R$-modules.
Assume either that $M$ has finite projective dimension or that $N$ has finite injective dimension.
Then one has an equality
$$
\sup\{\,i\mid\operatorname{Ext}_R^i(M,N)\ne 0\,\}=\operatorname{depth} R-\operatorname{depth}_RM.
$$
\end{lem}
\begin{proof}
Assume that $\operatorname{pd}_RM=s<\infty$.
Then it is easy to see that $\operatorname{Ext}_R^i(M,N)=0$ for $i>s$, so that $\sup\{\,i\mid\operatorname{Ext}_R^i(M,N)\ne 0\,\}\le s$.
There is an exact sequence
$$
0 \to R^{\oplus n_s} \overset{A_s}{\to} R^{\oplus n_{s-1}} \overset{A_{s-1}}{\to} \cdots \overset{A_1}{\to} R^{\oplus n_0} \to M \to 0
$$
such that each $A_i$ is a matrix over $R$ all of whose entries are elements in the maximal ideal $\mathfrak m$.
Dualizing this by $N$, we obtain an exact sequence $N^{\oplus n_{s-1}} \overset{{}^t(A_s)}{\to} N^{\oplus n_s} \to \operatorname{Ext}_R^s(M,N) \to 0$.
It is seen from Nakayama's lemma that $\operatorname{Ext}_R^s(M,N)\ne 0$.
Thus $\sup\{\,i\mid\operatorname{Ext}_R^i(M,N)\ne 0\,\}=s=\operatorname{depth} R-\operatorname{depth}_RM$, where the last equality follows from the Auslander-Buchsbaum formula.
As to the case where $N$ has finite injective dimension, see \cite[Satz 2.6]{ische} or \cite[Exercise 3.1.24]{BH}.
\end{proof}
The above lemma can also be shown by using the formulas \cite[Theorem 4.1]{Foxby} and \cite[Corollary (2.14)]{Christensen2}.
Now we can show the following corollary.
There are only two contravariantly finite resolving subcategories possessing such $G$ as in the corollary.
\begin{cor}\label{maincr2}
Let ${\mathcal X}$ be a contravariantly finite resolving subcategory of $\operatorname{mod} R$.
Assume that there is an $R$-module $G\in{\mathcal X}$ with $\operatorname{pd}_RG=\infty$ and $\operatorname{Ext}_R^i(G,R)=0$ for $i\gg 0$.
Then either of the following holds.
\begin{enumerate}[\rm (1)]
\item
${\mathcal X}=\operatorname{mod} R$,
\item
$R$ is Cohen-Macaulay and ${\mathcal X}={\mathcal C}(R)$.
\end{enumerate}
\end{cor}
\begin{proof}
Suppose that ${\mathcal X}\ne\operatorname{mod} R$.
Then Lemma \ref{kx} says that $k$ is not in ${\mathcal X}$.
By Corollary \ref{sbmncr2}, $R$ is Cohen-Macaulay.
First, we show that ${\mathcal C}(R)$ is contained in ${\mathcal X}$.
For this, let $M$ be a maximal Cohen-Macaulay $R$-module.
We have only to prove that $M$ is in ${}^\perp({\mathcal X}^\perp)$ by Lemma \ref{prpprp}.
Let $N$ be a nonzero $R$-module in ${\mathcal X}^\perp$.
Theorem \ref{main} implies that $N$ is of finite injective dimension.
Since $M$ is maximal Cohen-Macaulay, we have $\sup\{\,i\mid\operatorname{Ext}_R^i(M,N)\ne 0\,\}=0$ by Lemma \ref{pdidsup}.
Therefore $\operatorname{Ext}_R^i(M,N)=0$ for all $N\in{\mathcal X}^\perp$ and $i>0$.
It follows that $M$ is in ${}^\perp({\mathcal X}^\perp)$, as desired.
Next, we show that ${\mathcal X}$ is contained in ${\mathcal C}(R)$.
We have an exact sequence $0 \to Y \to X \to k \to 0$ with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$ by Lemma \ref{wak}.
Since $k$ is not in ${\mathcal X}$, the module $Y$ is nonzero.
By Theorem \ref{main}, $Y$ has finite injective dimension.
According to Lemma \ref{pdidsup}, for a nonzero $R$-module $X'$ in ${\mathcal X}$, we have equalities $0\ge\sup\{\,i\mid\operatorname{Ext}_R^i(X',Y)\ne 0\,\}=\operatorname{depth} R-\operatorname{depth}_RX'=\dim R-\operatorname{depth}_RX'$.
Therefore $X'$ is a maximal Cohen-Macaulay $R$-module, as desired.
\end{proof}
Next, we study contravariantly finite resolving subcategories all of whose objects $X$ satisfy $\operatorname{Ext}_R^{\gg 0}(X,R)=0$.
We start by considering special ones among such subcategories.
\begin{prop}\label{pd}
Let ${\mathcal X}$ be a contravariantly finite resolving subcategory of $\operatorname{mod} R$.
Suppose that every $R$-module in ${\mathcal X}$ has finite projective dimension.
Then either of the following holds.
\begin{enumerate}[\rm (1)]
\item
${\mathcal X}={\mathcal F}(R)$,
\item
$R$ is regular and ${\mathcal X}=\operatorname{mod} R$.
\end{enumerate}
\end{prop}
\begin{proof}
If ${\mathcal X}=\operatorname{mod} R$, then our assumption says that all $R$-modules have finite projective dimension.
Hence $R$ is regular.
Assume that ${\mathcal X}\ne\operatorname{mod} R$.
Then there is an $R$-module $M$ which is not in ${\mathcal X}$.
There is an exact sequence $0 \to Y \to X \to M \to 0$ with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$ by Lemma \ref{wak}.
Note that $Y\ne 0$ as $M\notin{\mathcal X}$.
Fix a nonzero $R$-module $X'\in{\mathcal X}$.
We have $\operatorname{Ext}_R^i(X',Y)=0$ for all $i>0$, and hence $\operatorname{pd}_RX'=\sup\{\,i\mid\operatorname{Ext}_R^i(X',Y)\ne 0\,\}=0$ by Lemma \ref{pdidsup} and the Auslander-Buchsbaum formula.
Hence $X'$ is free.
This means that ${\mathcal X}$ is contained in ${\mathcal F}(R)$.
On the other hand, ${\mathcal X}$ contains ${\mathcal F}(R)$ since ${\mathcal X}$ is resolving.
Therefore ${\mathcal X}={\mathcal F}(R)$.
\end{proof}
Combining Proposition \ref{pd} with Corollary \ref{maincr2}, we can get the following.
\begin{cor}\label{216}
Let ${\mathcal X}$ be a contravariantly finite resolving subcategory of $\operatorname{mod} R$.
Suppose that every module $X\in{\mathcal X}$ is such that $\operatorname{Ext}_R^i(X,R)=0$ for $i\gg 0$.
Then one of the following holds.
\begin{enumerate}[\rm (1)]
\item
${\mathcal X}={\mathcal F}(R)$,
\item
$R$ is Gorenstein and ${\mathcal X}={\mathcal C}(R)$,
\item
$R$ is Gorenstein and ${\mathcal X}=\operatorname{mod} R$.
\end{enumerate}
\end{cor}
\begin{proof}
The corollary follows from Proposition \ref{pd} in the case where all $R$-modules in ${\mathcal X}$ are of finite projective dimension.
So suppose that in ${\mathcal X}$ there exists an $R$-module of infinite projective dimension.
Then Corollary \ref{maincr2} shows that either of the following holds.
\begin{enumerate}[\rm (i)]
\item
${\mathcal X}=\operatorname{mod} R$,
\item
$R$ is Cohen-Macaulay and ${\mathcal X}={\mathcal C}(R)$.
\end{enumerate}
By the assumption that every $X\in{\mathcal X}$ satisfies $\operatorname{Ext}_R^i(X,R)=0$ for $i\gg 0$, we have $\operatorname{Ext}_R^i(k,R)=0$ for $i\gg 0$ in the case (i).
In the case (ii), since $\Omega ^dk$ is in ${\mathcal X}$ where $d=\dim R$, we have $\operatorname{Ext}_R^{i+d}(k,R)\cong\operatorname{Ext}_R^i(\Omega^dk,R)=0$ for $i\gg 0$.
Thus, in both cases, the ring $R$ is Gorenstein.
\end{proof}
Finally, we obtain the following result from Corollary \ref{216}, Example \ref{freecov}(2)(3) and Example \ref{rslvx}(2)(5).
It says that the category of finitely generated modules over a Gorenstein local ring possesses only three contravariantly finite resolving subcategories.
\begin{cor}
Let $R$ be a Gorenstein local ring.
Then all the contravariantly finite resolving subcategories of $\operatorname{mod} R$ are ${\mathcal F}(R)$, ${\mathcal C}(R)$ and $\operatorname{mod} R$.
\end{cor}
\section{Proof of Theorem \ref{main}}
This section is devoted to proving Theorem \ref{main}.
First of all, let us recall the definition and some fundamental properties of the transpose of a given module.
Let $M$ be an $R$-module.
Take a minimal free resolution $F_\bullet =(\cdots \overset{d_2}{\to} F_1 \overset{d_1}{\to} F_0 \to 0)$ of $M$.
We define the {\em transpose} $\operatorname{Tr} M$ of $M$ as the cokernel of the $R$-dual homomorphism $d_1^\ast:F_0^\ast \to F_1^\ast$ of $d_1$.
\begin{rem}\label{trem}
Let $M$ be an $R$-module.
\begin{enumerate}[(1)]
\item
The transpose $\operatorname{Tr} M$ of $M$ are uniquely determined up to isomorphism, since so is a minimal free resolution of $M$.
\item
One has $\operatorname{Tr}(\operatorname{Tr} M)\cong M$ and $M^\ast\cong\Omega^2\operatorname{Tr} M$ up to free summand.
\end{enumerate}
\end{rem}
For an $R$-module $M$, we denote by $\nu_R(M)$ the minimal number of generators of $M$, i.e. $\nu_R(M)=\dim_k(M\otimes_Rk)$.
\begin{lem}\label{tr}
Let $M$ be an $R$-module.
The transpose $\operatorname{Tr} M$ has no nonzero free summand.
\end{lem}
\begin{proof}
Let $F_\bullet = (\cdots \overset{d_2}{\to} F_1 \overset{d_1}{\to} F_0 \to 0)$ be a minimal free resolution of $M$.
Assume that $\operatorname{Tr} M$ has a nonzero free summand.
Then there exists a surjective homomorphism $\pi:\operatorname{Tr} M\to R$.
We obtain a commutative diagram
$$
\begin{CD}
@. 0 @. 0 \\
@. @VVV @VVV \\
F_0^\ast @>{\delta}>> F @>>> N @>>> 0 \\
@| @V{\theta}VV @VVV \\
F_0^\ast @>{d_1^\ast}>> F_1^\ast @>{\varepsilon}>> \operatorname{Tr} M @>>> 0 \\
@. @V{\pi\varepsilon}VV @V{\pi}VV \\
@. R @= R \\
@. @VVV @VVV \\
@. 0 @. 0
\end{CD}
$$
with exact rows and columns, where $F$ is a free $R$-module.
This diagram yields isomorphisms $M=\operatorname{Coker} d_1\cong\operatorname{Coker}(\delta^\ast\theta^\ast)\cong\operatorname{Coker}(\delta^\ast)$, as the homomorphism $\theta^\ast$ is surjective.
Hence there exists a surjective homomorphism $F^\ast\to\Omega M$, and we have $\operatorname{rank} F=\operatorname{rank} F^\ast\ge\nu_R(\Omega M)=\operatorname{rank} F_1=\operatorname{rank} F_1^\ast=\operatorname{rank} F+1$, which is a contradiction.
\end{proof}
For an $R$-module $M$, let $M^\ast M$ be the ideal of $R$ generated by the subset
$$
\{\,f(x)\mid f\in M^\ast,x\in M\,\}
$$
of $R$.
We verify that this ideal has the following property.
\begin{lem}\label{indec}
An $R$-module $M$ has a nonzero free summand if and only if $M^\ast M=R$.
\end{lem}
\begin{proof}
If $M$ has a nonzero free summand, then there exists a surjective homomorphism $\pi:M\to R$.
Hence there is an element $x\in M$ such that $\pi(x)=1$.
We have $1=\pi(x)\in M^\ast M$, so $M^\ast M=R$.
Conversely, suppose that the equality $M^\ast M=R$ holds.
Then there are elements $f_1,\dots,f_n\in M^\ast$ and $x_1,\dots,x_n\in M$ such that $1=f_1(x_1)+\cdots+f_n(x_n)$.
Define homomorphisms $f:M^{\oplus n}\to R$ and $g:R\to M^{\oplus n}$ by $f(y_1,\dots,y_n)=f_1(y_1)+\cdots+f_n(y_n)$ for $y_1,\dots,y_n\in M$ and $g(a)=(ax_1,\dots,ax_n)$ for $a\in R$.
Then the composite map $fg$ is the identity map of $R$, so $g$ is a split monomorphism.
Hence $R$ is isomorphic to a direct summand of $M^{\oplus n}$.
Since $R$ is henselian, $R$ is isomorphic to a direct summand of $M$ by virtue of the Krull-Schmidt theorem, that is, $M$ has a nonzero free summand.
\end{proof}
Let $M$ and $N$ be $R$-modules.
We denote by $\mathfrak P_R(M,N)$ the $R$-submodule of $\operatorname{Hom}_R(M,N)$ consisting of all homomorphisms from $M$ to $N$ which factors through some free $R$-module.
We denote by $\operatorname{\underline{Hom}}_R(M,N)$ the residue $R$-module $\operatorname{Hom}_R(M,N)/\mathfrak P_R(M,N)$.
For $R$-modules $M$ and $N$, there is a natural homomorphism
$$
\lambda_{MN}:M\otimes_RN\to\operatorname{Hom}_R(M^\ast,N)
$$
which is given by $\lambda_{MN}(x\otimes y)(f)=f(x)y$ for $x\in M$, $y\in N$ and $f\in M^\ast$.
The following proposition will be used to prove the essential part of the proof of Theorem \ref{main}.
\begin{prop}\label{key}
Let ${\mathcal X}$ be a subcategory of $\operatorname{mod} R$ and $0 \to Y \overset{f}{\to} X \to M \to 0$ an exact sequence of $R$-modules with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$.
Let $G\in{\mathcal X}$, set $H=\operatorname{Tr}\Omega G$, and suppose that $(H^\ast H)M=0$.
Let $0 \to K \overset{g}{\to} F \overset{h}{\to} H \to 0$ be an exact sequence of $R$-modules with $F$ free.
Then the induced sequence
$$
\begin{CD}
0 @>>> K\otimes_RY @>{g\otimes_RY}>> F\otimes_RY @>{h\otimes_RY}>> H\otimes_RY @>>> 0
\end{CD}
$$
is exact, and the map $h\otimes_RY$ factors through the map $F\otimes_Rf:F\otimes_RY\to F\otimes_RX$.
\end{prop}
\begin{proof}
According to \cite[Proposition (2.6)]{ABr}, there is a commutative diagram
$$
\begin{CD}
@. 0 \\
@. @VVV \\
@. \operatorname{Ext}_R^1(\operatorname{Tr} H,Y) \\
@. @VVV \\
@. H\otimes_RY @>{\delta}>> H\otimes_RX @>{\varepsilon}>> H\otimes_RM @>>> 0 \\
@. @V{\alpha}VV @V{\beta}VV @V{\gamma}VV \\
0 @>>> \operatorname{Hom}_R(H^\ast,Y) @>{\zeta}>> \operatorname{Hom}_R(H^\ast,X) @>{\eta}>> \operatorname{Hom}_R(H^\ast,M) @>>> \operatorname{Ext}_R^1(H^\ast,Y) \\
@. @VVV \\
@. \operatorname{Ext}_R^2(\operatorname{Tr} H,Y) \\
@. @VVV \\
@. 0
\end{CD}
$$
with exact rows and columns, where $\alpha=\lambda_{HY}$, $\beta=\lambda_{HX}$ and $\gamma=\lambda_{HM}$.
Note from Remark \ref{trem}(2) that $\operatorname{Tr} H\cong \Omega G$ and $H^\ast\cong\Omega ^2\operatorname{Tr} H\cong\Omega ^3G$ up to free summand.
Since $G\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$, we have $\operatorname{Ext}_R^1(\operatorname{Tr} H,Y)\cong\operatorname{Ext}_R^2(G,Y)=0$, $\operatorname{Ext}_R^2(\operatorname{Tr} H,Y)\cong\operatorname{Ext}_R^3(G,Y)=0$ and $\operatorname{Ext}_R^1(H^\ast,Y)\cong\operatorname{Ext}_R^4(G,Y)=0$.
Hence $\alpha$ is an isomorphism and $\eta$ is a surjective homomorphism.
Diagram chasing shows that $\delta$ is an injective homomorphism.
From the assumption that $(H^\ast H)M=0$, we see that $\gamma=\lambda_{HM}$ is the zero map.
We thus obtain the following commutative diagram with exact rows.
$$
\begin{CD}
0 @>>> H\otimes_RY @>{\delta}>> H\otimes_RX @>{\varepsilon}>> H\otimes_RM @>>> 0 \\
@. @V{\alpha}V{\cong}V @V{\beta}VV @V{\gamma}V{0}V \\
0 @>>> \operatorname{Hom}_R(H^\ast,Y) @>{\zeta}>> \operatorname{Hom}_R(H^\ast,X) @>{\eta}>> \operatorname{Hom}_R(H^\ast,M) @>>> 0
\end{CD}
$$
As $\eta\beta=\gamma\varepsilon=0$, there exists a homomorphism
$$
\sigma: H\otimes_RX \to \operatorname{Hom}_R(H^\ast,Y)
$$
with $\zeta\sigma=\beta$.
The commutativity $\zeta\alpha=\beta\delta$ and the injectivity of $\zeta$ imply that $\sigma\delta=\alpha$.
Put
$$
\tau=\alpha^{-1}\sigma:H\otimes_RX \to H\otimes_RY.
$$
The map $\delta$ is a split monomorphism with a splitting map $\tau$.
There is a commutative diagram
$$
\begin{CD}
@. 0 \\
@. @VVV \\
@. \operatorname{Tor}_1^R(H,Y) \\
@. @VVV \\
@. K\otimes_RY @>>> K\otimes_RX @>>> K\otimes_RM @>>> 0 \\
@. @V{g\otimes_RY}VV @VVV @VVV \\
0 @>>> F\otimes_RY @>{F\otimes_Rf}>> F\otimes_RX @>>> F\otimes_RM @>>> 0 \\
@. @V{h\otimes_RY}VV @V{\xi}VV @VVV \\
0 @>>> H\otimes_RY @>{\delta}>> H\otimes_RX @>{\varepsilon}>> H\otimes_RM @>>> 0 \\
@. @VVV @VVV @VVV \\
@. 0 @. 0 @. 0
\end{CD}
$$
with exact rows and columns, and we have $h\otimes_RY=(\tau\delta)(h\otimes_RY)=\tau\xi(F\otimes_Rf)$.
Thus, the homomorphism $h\otimes_RY$ factors through the homomorphism $F\otimes_Rf$.
We have an isomorphism $\operatorname{Tor}_1^R(H,Y)=\operatorname{Tor}_1^R(\operatorname{Tr}\Omega G,Y)\cong\operatorname{\underline{Hom}}_R(\Omega G,Y)$ by \cite[Lemma (3.9)]{yos0}.
It remains to prove that $\operatorname{\underline{Hom}}_R(\Omega G,Y)=0$.
There is an exact sequence $0 \to \Omega G \overset{\theta}{\to} P \to G \to 0$ of $R$-modules such that $P$ is a free $R$-module.
Let $\rho\in\operatorname{Hom}_R(\Omega G,Y)$.
We want to show that this map factors through some free $R$-module.
There is a pushout diagram:
$$
\begin{CD}
0 @>>> \Omega G @>{\theta}>> P @>>> G @>>> 0 \\
@. @V{\rho}VV @V{\phi}VV @| \\
0 @>>> Y @>{\psi}>> Q @>>> G @>>> 0
\end{CD}
$$
The second row is a split exact sequence since it can be regarded as an element of $\operatorname{Ext}_R^1(G,Y)$, which vanishes as $G\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$.
Hence there is a homomorphism $\pi:Q\to Y$ such that $\pi\psi=1$, and we have $\rho=(\pi\psi)\rho=\pi\phi\theta$.
Therefore $\rho$ factors through the free $R$-module $P$, as desired.
\end{proof}
The proof of our next result will require the following elementary lemma.
\begin{lem}\label{elmntry}
Let $M$ be an $R$-module, and let
$$
M=M_0 \overset{f_0}{\to} M_1 \overset{f_1}{\to} M_2 \overset{f_2}{\to} \cdots
$$
be a sequence of surjective homomorphisms.
Then $f_i$ is an isomorphism for $i\gg 0$.
\end{lem}
\begin{proof}
For a positive integer $i$, let $K_i$ be the kernel of the composite map $g_i:=f_{i-1}f_{i-2}\cdots f_1f_0$.
Then there is an ascending chain $K_1\subseteq K_2\subseteq K_3\subseteq\cdots$ of submodules of $M$.
Since $M$ is a noetherian $R$-module, there exists an integer $n$ such that $K_i=K_{i+1}$ for any $i\ge n$.
For each positive integer $i$, we have a commutative diagram
$$
\begin{CD}
M/K_i @>{\cong}>> M_i \\
@V{\pi_i}VV @V{f_i}VV \\
M/K_{i+1} @>{\cong}>> M_{i+1}
\end{CD}
$$
where the horizontal isomorphisms are induced by $g_i$ and $g_{i+1}$, and $\pi$ is a natural surjective homomorphism.
Since $\pi_i$ is an isomorphism for $i\ge n$, the map $f_i$ is an isomorphism for $i\ge n$.
\end{proof}
Let ${\mathcal A}$ and ${\mathcal B}$ be abelian categories.
A {\em contravariant cohomological $\delta$-functor} $D$ from ${\mathcal A}$ to ${\mathcal B}$ is defined as a collection of additive contravariant functors $D^n:{\mathcal A}\to{\mathcal B}$ for each integer $n\ge 0$, together with morphisms $\delta^n(s):D^n(X)\to D^{n+1}(Z)$ for each short exact sequence $s:0 \to X \overset{f}{\to} Y \overset{g}{\to} Z \to 0$ in ${\mathcal A}$.
We make the convention that $D^n=0$ for $n<0$.
The following two conditions are imposed:
\begin{enumerate}[\rm (1)]
\item
For each short exact sequence as above, there is a long exact sequence
$$
\cdots \to D^{n-1}(X) \overset{\delta^{n-1}(s)}{\to} D^n(Z) \overset{D^n(g)}{\to} D^n(Y) \overset{D^n(f)}{\to} D^n(X) \overset{\delta^n(s)}{\to} D^{n+1}(Z) \to \cdots.
$$
\item
For each commutative diagram
$$
\begin{CD}
s\phantom{'}:\quad 0 @>>> X @>{f}>> Y @>{g}>> Z @>>> 0 \\
@. @V{\alpha}VV @V{\beta}VV @V{\gamma}VV \\
s':\quad 0 @>>> X' @>{f'}>> Y' @>{g'}>> Z' @>>> 0
\end{CD}
$$
with exact rows, the diagram
$$
\begin{CD}
D^n(X') @>{\delta^n(s')}>> D^{n+1}(Z') \\
@V{D^n(\alpha)}VV @V{D^{n+1}(\gamma)}VV \\
D^n(X) @>{\delta^n(s)}>> D^{n+1}(Z)
\end{CD}
$$
is commutative.
\end{enumerate}
Now we can prove the following, which will play a key role in the proof of Theorem \ref{main}.
\begin{prop}\label{keyprop}
Let ${\mathcal X}$ be a subcategory of $\operatorname{mod} R$ which is closed under syzygies.
Let $0 \to Y \to X \to M \to 0$ be an exact sequence of $R$-modules with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$.
Suppose that there is an $R$-module $G\in{\mathcal X}$ with $\operatorname{pd}_RG=\infty$ and $\operatorname{Ext}_R^i(G,R)=0$ for $i\gg 0$.
Put $H_i=\operatorname{Tr}\Omega(\Omega^iG)$ and assume that $((H_i)^\ast H_i)M=0$ for $i\gg 0$.
Let $D=(D^j)_{j\ge 0}:\operatorname{mod} R\to\operatorname{mod} R$ be a contravariant cohomological $\delta$-functor.
If $D^j(X)=0$ for $j\gg 0$, then $D^j(Y)=D^j(M)=0$ for $j\gg 0$.
\end{prop}
\begin{proof}
Since ${\mathcal X}$ is closed under syzygies, the module $\Omega^iG$ is in ${\mathcal X}$ for $i\ge 0$.
Replacing $G$ with $\Omega^iG$ for $i\gg 0$, we may assume that $\operatorname{Ext}_R^i(G,R)=0$ for all $i>0$ and that $((H_i)^\ast H_i)M=0$ for all $i\ge 0$.
Let
$$
F_\bullet=(\cdots \overset{d_{i+1}}{\to} F_i \overset{d_i}{\to} F_{i-1} \overset{d_{i-1}}{\to} \cdots \overset{d_2}{\to} F_1 \overset{d_1}{\to} F_0 \to 0)
$$
be a minimal free resolution of $G$.
Dualizing this by $R$ yields an exact sequence
$$
0 \to G^\ast \to (F_0)^\ast \overset{(d_1)^\ast}{\to} (F_1)^\ast \overset{(d_2)^\ast}{\to} \cdots \overset{(d_i)^\ast}{\to} (F_i)^\ast \overset{(d_{i+1})^\ast}{\to} (F_{i+1})^\ast \overset{(d_{i+2})^\ast}{\to} (F_{i+2})^\ast \overset{(d_{i+3})^\ast}{\to} \cdots,
$$
and it is easily seen that $H_i\cong(\Omega^{i+3}G)^\ast$ and $\Omega H_i\cong(\Omega^{i+2}G)^\ast$ for $i\ge 0$.
By Proposition \ref{key}, for each integer $i\ge 0$ we have an exact sequence
$$
0 \to (\Omega^{i+2}G)^\ast\otimes_RY \to (F_{i+2})^\ast\otimes_RY \overset{f_i}{\to} (\Omega^{i+3}G)^\ast\otimes_RY \to 0
$$
such that $f_i$ factors through $(F_{i+2})^\ast\otimes_RX$.
There is an integer $a\ge 0$ such that $D^j(X)=0$ for all $j\ge a$.
From the above short exact sequence, we get a long exact sequence
$$
\cdots \overset{D^j(f_i)}{\to} D^j((F_{i+2})^\ast\otimes_RY) \to D^j((\Omega^{i+2}G)^\ast\otimes_RY) \to D^{j+1}((\Omega^{i+3}G)^\ast\otimes_RY) \overset{D^{j+1}(f_i)}{\to} \cdots.
$$
The homomorphism $D^j(f_i)$ factors through $D^j((F_{i+2})^\ast\otimes_RX)$, which vanishes for $j\ge a$ as $(F_{i+2})^\ast\otimes_RX$ is isomorphic to a direct sum of copies of $X$.
Hence $D^j(f_i)=0$ for $j\ge a$, and we obtain an exact sequence
$$
0 \to D^j((F_{i+2})^\ast\otimes_RY) \to D^j((\Omega^{i+2}G)^\ast\otimes_RY) \overset{\varepsilon_{i,j}}{\to} D^{j+1}((\Omega^{i+3}G)^\ast\otimes_RY) \to 0
$$
for $i\ge 0$ and $j\ge a$.
Thus, there is a sequence
$$
D^a((\Omega^2G)^\ast\otimes_RY) \overset{\varepsilon_{0,a}}{\to} D^{a+1}((\Omega^3G)^\ast\otimes_RY) \overset{\varepsilon_{1,a+1}}{\to} D^{a+2}((\Omega^4G)^\ast\otimes_RY) \overset{\varepsilon_{2,a+2}}{\to} \cdots
$$
of surjective homomorphisms of $R$-modules.
Lemma \ref{elmntry} says that there exists an integer $b\ge 0$ such that
$$
\varepsilon_{l,a+l}:D^{a+l}((\Omega^{l+2}G)^\ast\otimes_RY) \to D^{a+l+1}((\Omega^{l+3}G)^\ast\otimes_RY)
$$
is an isomorphism for every $l\ge b$.
It follows from the above short exact sequence that $D^{a+l}((F_{l+2})^\ast\otimes_RY)=0$ for every $l\ge b$.
Since $G$ has infinite projective dimension, each $F_i$ is a nonzero free $R$-module, and the module $(F_{l+2})^\ast\otimes_RY$ is isomorphic to a nonzero direct sum of copies of $Y$.
Thus we have $D^j(Y)=0$ for $j\ge a+b$.
From the exact sequence $0 \to Y \to X \to M \to 0$ we see that $D^j(M)=0$ for $j\ge a+b+1$.
\end{proof}
Now we are in the position to achieve the purpose of this section.
\begin{tpf}
Since $k$ admits a right ${\mathcal X}$-approximation, there exists an exact sequence $0 \to Y \to X \to k \to 0$ of $R$-modules with $X\in{\mathcal X}$ and $Y\in{\mathcal X}^\perp$ by Lemma \ref{wak}.
For an integer $i\ge 0$, put $H_i=\operatorname{Tr}\Omega(\Omega^iG)$.
Lemma \ref{tr} says that $H_i$ has no nonzero free summand.
We have $(H_i)^\ast H_i\ne R$ by Lemma \ref{indec}.
Hence $((H_i)^\ast H_i)k=0$ for $i\ge 0$.
Applying Proposition \ref{keyprop} to the contravariant cohomological $\delta$-functor $D=(\operatorname{Ext}_R^j(\quad,M))_{j\ge 0}$, we obtain $D^j(k)=0$ for $j\gg 0$.
Namely, we have $\operatorname{Ext}_R^j(k,M)=0$ for $j\gg 0$, which implies that $M$ has finite injective dimension.
\qed
\end{tpf}
\section*{Acknowledgments}
The author would like to give his deep gratitude to Shiro Goto and Osamu Iyama for a lot of valuable discussions and helpful suggestions.
The author also thanks Mitsuyasu Hashimoto, Yuji Yoshino and an anonymous referee for important and useful comments.
|
\section{Introduction}
Polynomial identity testing (PIT) ranks as one of the most important open problems
in the intersection of algebra and computer science.
We are provided an arithmetic circuit that computes a polynomial $p(x_1,x_2,\cdots,x_n)$
over a field $\mathbb{F}$, and we wish to test if $p$ is identically zero
(in other words, if $p$ is the zero polynomial).
In the black-box setting, the circuit
is provided as a black-box and we are only allowed to evaluate the polynomial
$p$ at various domain points.
The main goal is to devise a deterministic polynomial time
algorithm for PIT.
Kabanets \& Impagliazzo~\cite{KI04} and Agrawal~\cite{A05, A06}
have shown connections between deterministic algorithms for identity testing and circuit lower bounds,
emphasizing the importance of this problem. To know more about the current state of the general identity
testing problem see the surveys \cite{S09, AS09}.
The first randomized polynomial time PIT algorithm,
which was a black-box algorithm,
was given (independently) by Schwartz~\cite{Sch80} and Zippel~\cite{Z79}.
Randomized algorithms that use less
randomness were given by Chen \& Kao~\cite{CK00}, Lewin \& Vadhan~\cite{LV98},
and Agrawal \& Biswas~\cite{AB03}.
Klivans \& Spielman~\cite{ks01} observed that
even for depth-$3$ circuits for bounded top fanin,
deterministic identity testing was open. Progress
towards this was first made by Dvir \& Shpilka~\cite{DS06},
who gave a quasi-polynomial time algorithm, although with
a doubly-exponential dependence on the top fanin.
The problem was resolved by a polynomial time algorithm given by
Kayal and Saxena~\cite{KS07}, with a running time exponential
in the top fanin. As expected, the current understanding of depth-$4$ circuits is even
more sparse. Identity tests are known only for rather special depth-$4$ circuits
\cite{AM07, S08, SV09, KMSV09}.
Why is progress restricted to such small depth circuits?
Agrawal and Vinay~\cite{AV08} showed that
an efficient black-box identity test for depth-$4$
circuits will actually give a quasi-polynomial black-box test, and subexponential lower bounds,
for circuits of \emph{all depths} (that compute {\em low degree} polynomials).
Thus, understanding depth-$3$ identities seems to be a natural first step towards
the goal of proving more general lower bounds.
For deterministic black-box testing, the first results
were given by Karnin \& Shpilka \cite{KSh08}.
Based on results in~\cite{DS06},
they gave an algorithm for bounded top fanin depth-$3$ circuits having a quasi-polynomial
running time (with a doubly-exponential dependence
on the top fanin). The dependence on the top fanin was later improved (to singly-exponential)
by the rank bound results of Saxena \& Seshadhri \cite{SS09} (for {\em any} $\mathbb{F}$). But the
time complexity also had a quasi-polynomial dependence on the degree of the circuit. This
dependence is inevitable in rank-based methods over finite fields (as shown by \cite{KS07}).
However, over the field of rationals, Kayal \& Saraf \cite{KSar09} showed how to remove this
quasi-polynomial dependence on the degree at the cost of doubly-exponential dependence on
the top fanin, thus giving a polynomial time complexity for bounded top fanin.
In this work we achieve the best of the two works \cite{SS09} and \cite{KSar09}, i.e. we
prove (for rationals) a time complexity that depends only {\em polynomially} on the degree and ``only''
{\em singly}-exponentially on the fanin.
In a quite striking result, Kayal \& Saraf~\cite{KSar09} proved how Sylvester-Gallai theorems
can get better rank bounds over the reals.
We introduce the concept of \emph{Sylvester-Gallai rank bounds} that deals
with the rank of vectors (over some given field) that have some special
incidence properties.
This is a very convenient way to express known Sylvester-Gallai results.
These are inspired by the famous
Sylvester-Gallai theorem about point-line incidences.
We show how this very interesting quantity is tightly connected
to depth-$3$ identities. Sylvester-Gallai rank bounds
over high dimensions were known over the reals, and
are used to prove depth-$3$ rank bounds over reals.
We prove the first ever theorem for high
dimensional Sylvester-Gallai configurations
over \emph{any field}.
\subsection{Definitions and Previous Work}
This work focuses on depth-$3$ circuits. A structural study of
depth-$3$ identities was initiated in \cite{DS06} by defining a notion of {\em rank}
of {\em simple} and {\em
minimal} identities. A depth-$3$ circuit $C$ over a field $\mathbb{F}$ is:
$$C(x_1,\ldots,x_n) = \sum_{i=1}^k T_i$$
where, $T_i$ ({\em a multiplication term}) is a product of $d_i$ linear polynomials
$\ell_{i,j}$ over $\mathbb{F}$. Note that for the purposes of studying identities we can
assume wlog (by {\em homogenization}) that $\ell_{i,j}$'s are linear {\em forms} (i.e.
linear polynomials
with a zero constant coefficient) and that $d_1=\cdots=d_k=:d$.
Such a circuit is referred to as a $\Sigma\Pi\Sigma(k,d)$ circuit (or $\Sigma\Pi\Sigma(k,d,n)$ depending on the context),
where $k$ is the \emph{top fanin} of $C$ and $d$ is the {\em degree} of $C$.
We give a few definitions from~\cite{DS06}.
\begin{definition} {\bf[Simple Circuit]}
$C$ is a {\em simple} circuit if there is no nonzero linear form dividing all
the $T_i$'s.
{\bf[Minimal Circuit]} $C$ is a {\em minimal} circuit if for every proper subset
$S\subset[k]$, $\sum_{i\in S}T_i$ is nonzero.
{\bf[Rank of a circuit]} Every $\ell_{i,j}$ can be seen as an $n$-dimensional vector over $\mathbb{F}$.
The {\em rank} of the circuit, $\text{rk}(C)$, is defined as
the rank of the set of all linear forms $\ell_{i,j}$'s viewed as $n$-dimensional vectors.
\end{definition}
Can all the forms $\ell_{i,j}$ be independent, or must there
be relations between them?
The rank can be interpreted as the minimum number of variables
that are required to express $C$. There exists a linear
transformation converting the $n$ variables of the circuit
into $rank(C)$ independent variables. A trivial upper bound on the rank (for any $\Sigma\Pi\Sigma$-circuit)
is $kd$, since that is the total number of
linear forms involved in $C$.
The rank is a fundamental property of a $\Sigma\Pi\Sigma(k,d)$ circuit
and it is crucial to understand how large this can be
for identities.
A substantially smaller rank bound than $kd$ shows that
identities do not have as many ``degrees of freedom" as general
circuits, and leads to deterministic identity tests.
Furthermore, the techniques used to prove rank bounds
show us structural properties of identities that may
suggest directions to resolve PIT for $\Sigma\Pi\Sigma(k,d)$ circuits.
The rank bounds, in addition to being a natural property of identities, have found
applications in black-box identity testing \cite{KSh08} and learning $\Sigma\Pi\Sigma$ circuits
\cite{Sh09, KSh09}. The result of \cite{KSh08} showed rank bounds imply
black-box testers: if $R(\mathbb{F},k,d)$ is a rank bound for
simple minimal $\Sigma\Pi\Sigma(k,d,n)$ identities over field $\mathbb{F}$, then there is a deterministic
black-box identity tester for such circuits, that runs in $\text{poly}(n,d^{R(\mathbb{F},k,d)})$
$\mathbb{F}$-operations. (For the time complexity over $\mathbb{Q}$, we actually count the {\em bit
operations}.)
Dvir \& Shpilka~\cite{DS06} proved that the
rank of a simple, minimal $\Sigma\Pi\Sigma(k,d)$ identity is bounded by $2^{O(k^2)}(\log d)^{k-2}$.
This rank bound was improved to $O(k^3\log d)$ by Saxena \& Seshadhri \cite{SS09}.
Fairly basic identity constructions show that
the rank is $\Omega(k)$ over the reals and $\Omega(k \log d)$ for
finite fields~\cite{DS06,KS07,SS09}.
Dvir \& Shpilka \cite{DS06} conjectured that $\text{rk}(C)$ should be some $\text{poly}(k)$ over the reals.
Through a very insightful
use of Sylvester-Gallai theorems, Kayal \& Saraf \cite{KSar09} subsequently bounded
the rank of identities, over reals, by $k^{O(k)}$. This means that for a constant top fanin
circuit,
the rank of identities is constant, independent of the degree. This also
leads to the first truly polynomial-time deterministic black-box identity testers for this case.
Unfortunately, as soon as $k$ becomes even $\Omega(\log n)$, this bound
becomes trivial. We improve this rank bound
exponentially, to $O(k^2)$, which is almost optimal.
This gives a major improvement
in the running time of the black-box testers.
We also improve the rank bounds for general fields
from $O(k^3\log d)$ to $O(k^2\log d)$. We emphasize that
we give a unified framework to prove all these results.
Table~\ref{tab} should make it easier to compare the various
bounds.
\begin{table}[h]
\caption{Known rank bounds and black-box PIT}
\centering
\begin{tabular}{l l l l l}
\hline\hline
Paper & Result & Asymptotics & over field \\ [0.5ex]
\hline
& rank bound & $2^{k^2}\log^{k-2}d$ & any \\[-0.6ex]
\raisebox{1.5ex}{\cite{DS06, KSh08}} & time complexity & $n d^{(2^{k^2}\log^{k-2}d)}$ & any \\[1ex]
& rank bound & $k^3\log d$ & any \\[-1ex]
\raisebox{1.5ex}{\cite{SS09}} & time complexity & $n d^{k^3\log d}$ & any \\[1ex]
& rank bound & $k^k$ & $\mathbb{R}$ \\[-1ex]
\raisebox{1.5ex}{\cite{KSar09}} & time complexity & $n d^{(k^k)}$ & $\mathbb{Q}$ \\[1ex]
& rank bound & $k^2$ & $\mathbb{R}$ \\[-1ex]
\raisebox{1.5ex}{Ours} & time complexity & $n d^{k^2}$ & $\mathbb{Q}$ \\[1ex]
& rank bound & $k^2\log d$ & any \\
& time complexity & $n d^{k^2\log d}$ & any \\[1ex]
\hline
\end{tabular}
\label{tab}
\end{table}
Kayal \& Saraf~\cite{KSar09} connect Sylvester-Gallai theorems to rank bounds. They need
advanced versions of these theorems that deal with colored points and have to prove
certain \emph{hyperplane decomposition theorems}. We make the connection much more
transparent (at the loss of some color from the theorems).
We reiterate that our techniques are completely different, and employ
a very powerful algebraic framework to dissect identities.
This allows us to use as a ``black-box'' the most basic form
of the higher dimensional Sylvester-Gallai theorems.
\subsection{Our Results}
Before we state our results, it will be helpful to understand
Sylvester-Gallai configurations.
A set of points $S$ with the property
that every line through two points of $S$ passes through
a third point in $S$ is called a \emph{Sylvester-Gallai configuration}.
The famous Sylvester-Gallai theorem states: for a set $S$ of points
in ${\mathbb{R}}^2$, not all collinear, there exists
a line passing through exactly two points of $S$.
In other words, the only Sylvester-Gallai configuration in ${\mathbb{R}}^2$
is a set of collinear points.
This basic
theorem about point-line incidences was extended to
higher dimensions~\cite{Han65,BE67}.
We introduce the notion of \emph{Sylvester-Gallai rank bounds}.
This is a clean and convenient way of expressing these theorems.
\begin{definition}
Let $S$ be a finite subset of the {\em projective space} $\mathbb{F}\mathbb{P}^{n}$.
Alternately, $S$ is a subset of vectors in $\mathbb{F}^{n+1}$
without \emph{multiples}: no two vectors in $S$ are scalar
multiples of each other\footnote{When $|\mathbb{F}|>|S|$, such an $S$ is, wlog, a subset of
distinct vectors with first coordinate $1$.}.
Suppose, for every set $V \subset S$ of $k$ linearly independent vectors, the linear
span of $V$ contains at least $k+1$ vectors of $S$.
Then, the set $S$ is said to be \emph{$\text{SG}_k$-closed}.
The largest possible rank of an $\text{SG}_k$-closed set of at most $m$ vectors in $\mathbb{F}^n$ (for any $n$)
is denoted by $\text{SG}_k(\mathbb{F},m)$.
\end{definition}
The classic Sylvester-Gallai theorem essentially
states\footnote{To see this, take an $SG_2$-closed set $S$ of vectors. Think of each
vector being represented by an infinite line through the origin, hence giving
a set $S$ in the projective space. Take a $2$-dimensional plane $P$ not passing through
the origin and take the set of intersection points $I$ of the lines in $S$ with $P$.
Observe that the coplanar points $I$ have the property that a line
passing through two points of $I$ passes through a third point of $I$.} that for all $m$,
$\text{SG}_2(\mathbb{R},m)\le2$.
Higher dimensional analogues~\cite{Han65,BE67} prove that $\text{SG}_k(\mathbb{R},m) \leq 2(k-1)$.
One of our auxiliary theorems is such a statement for \emph{all fields}.
\begin{theorem}[$\text{SG}_k$ for all fields] \label{thm-sgk} For any field $\mathbb{F}$ and
$k,m\in\mathbb{N}^{>1}$, $\text{SG}_k(\mathbb{F},m)\leq 9k\lg m$.
\end{theorem}
Our main theorem is a simple, clean expression of how Sylvester-Gallai
influences identities.
\begin{theorem}[From $\text{SG}_k$ to Rank]
Let $|\mathbb{F}|>d$. The rank of
a simple and minimal $\Sigma\Pi\Sigma(k,d)$ identity over $\mathbb{F}$ is at most $2k^2 + k\cdot\text{SG}_k(\mathbb{F},d)$.
\end{theorem}
\noindent {\bf Remark.}
If $\mathbb{F}$ is small, then we choose an extension $\mathbb{F}'\supset\mathbb{F}$ of size $> d$ and get a
rank bound with $\text{SG}_k(\mathbb{F}',d)$. \\
Plugging in $\text{SG}_k$-rank bounds gives us the desired theorem for depth-$3$ identities.
We have a slightly stronger version of the above theorem that we use to
get better constants (refer to Theorem~\ref{thm-rank}).
\begin{theorem}[Depth-$3$ Rank Bounds]\label{thm-main}
Let $C$ be a $\Sigma\Pi\Sigma(k,d)$ circuit, over field $\mathbb{F}$, that is simple, minimal and zero. Then,
\begin{itemize}
\item For $\mathbb{F} = \mathbb{R}$, $\text{rk}(C)<3k^2$.
\item For any $\mathbb{F}$, $\text{rk}(C)<3k^2(\lg 2d)$.
\end{itemize}
\end{theorem}
As discussed before, a direct application of this result to Lemma 4.10 of \cite{KSh08}
gives a deterministic black-box identity test for $\Sigma\Pi\Sigma(k,d,n)$ circuits (we
will only discuss $\mathbb{Q}$ here as the other statement is analogous).
Formally, we get the following {\em hitting set generator} for $\Sigma\Pi\Sigma$ circuits with real
coefficients.
\begin{corollary}[Black-box PIT over $\mathbb{Q}$]
There is a deterministic algorithm that takes as input a triple
$(k,d,n)$ of natural numbers and in time $\text{poly}(nd^{k^2})$,
outputs a hitting set $\mathcal{H}\subset\mathbb{Z}^n$ with the following properties:
\begin{itemize}
\item[1)] Any $\Sigma\Pi\Sigma(k,d,n)$ circuit $C$ over $\mathbb{R}$
computes the zero polynomial iff $\forall a\in\mathcal{H}$, $C(a) = 0$.
\item[2)] $\mathcal{H}$ has at most $\text{poly}(nd^{k^2})$ points.
\item[3)] The total bit-length of each point in $\mathcal{H}$ is $\text{poly}(kn\log d)$.
\end{itemize}
\end{corollary}
\noindent {\bf Remark.}
\begin{itemize}
\item[1)]
Our black-box test
has {\em quasi-polynomial} in $n$ time complexity (with polynomial-dependence on $d$) for top fanin as large as
$k=\text{poly}\log(n)$,
and {\em sub-exponential} in $n$ time complexity (with polynomial-dependence on $d$) even for
top fanin as large as $k=o(\sqrt{n})$. This is the first tester to achieve
such bounds.
\item[2)] The fact that the points in $\mathcal{H}$ are integral and have ``small'' bit-length
is important to estimate the time complexity of our algorithm in terms of {\em bit
operations}. Thus, the hitting set generator takes at most $\text{poly}(nd^{k^2})$ {\em bit}
operations to compute $\mathcal{H}$.
\end{itemize}
\section{Proof Outline, Ideas, and Organization}
Our proof of the rank bound comprises of several new ideas, both at the conceptual and the technical
levels. In this section we will give the basic intuition of the proof.
The three notions that are crucially used (or developed) in
the proof are: ideal Chinese remaindering, matchings and Sylvester-Gallai rank bounds. These have
appeared (in some form) before in the works of Kayal \& Saxena \cite{KS07}, Saxena \& Seshadhri
\cite{SS09} and Kayal \& Saraf \cite{KSar09} respectively, to prove different kinds of results.
Here we use all three of them together to show quite a strong structure in $\Sigma\Pi\Sigma$ identities.
We will talk about them one by one in
the following three subsections outlining the three steps of the proof. Each step proves a new
property of identities which is interesting in its own right. The first two steps set
up the algebraic framework and prove theorems that hold for all fields.
The third step is where the Sylvester-Gallai theorems are brought in.
Some (new and crucial) algebraic lemmas and their proofs
have been moved to the Appendix. The flow of the actual proof
will be identical to the overview that we now provide.
\subsection{Step 1: Matching the Gates in an Identity}
We will denote the set $\{1,\ldots,n\}$ by $[n]$. We fix
the base field to be $\mathbb{F}$, so the circuits
compute multivariate polynomials in the {\em polynomial ring} $R:=\mathbb{F}[x_1,\ldots,x_n]$.
A \emph{linear form} is a linear polynomial in $R$ with zero constant term. We will denote
the set of all linear forms by
$L(R):= \left\{\sum_{i=1}^n a_ix_i \mid a_1,\ldots,a_n\in\mathbb{F}\right\}$.
Clearly, $L(R)$ is a vector (or linear) space over $\mathbb{F}$ and that will be quite useful.
Much of what we do shall deal with {\em multi}-sets of linear forms (sometimes polynomials in $R$ too), equivalence
classes inside them, and various maps across them. A \emph{list} of linear
forms is a multi-set of forms with an arbitrary order
associated with them. The actual ordering is unimportant:
we will heavily use maps between lists, and the ordering allows us to
define these maps unambiguously. The object, list, comes with all the usual set operations naturally defined.
\begin{definition} We collect some important definitions from \cite{SS09}:
{\bf[Multiplication term, $L(\cdot)$ \& $M(\cdot)$]} A {\em multiplication term} $f$ is an
expression in $R$ given as (the product may have repeated $\ell$'s),
$f := c\cdot\prod_{\ell\in S}\ell$, where $c\in\mathbb{F}^*$
and $S$ is a list of nonzero linear forms.
The {\em list of linear forms in $f$}, $L(f)$, is just the list $S$ of forms
occurring in the product above.
For a list $S$ of linear forms we define the {\em multiplication
term of $S$}, $M(S)$, as $\prod_{\ell\in S}\ell$ or $1$ if $S=\phi$.
{\bf[Forms in a Circuit]} We will represent a $\Sigma\Pi\Sigma(k,d)$ circuit $C$ as a sum of $k$ multiplication terms
of degree $d$, $C = \sum_{i=1}^k T_i$. The list of {\em linear forms occurring in $C$}
is $L(C):=$ $\bigcup_{i\in[k]}L(T_i)$. Note that $L(C)$ is a list of size exactly $kd$.
The {\em rank of $C$}, $\text{rk}(C)$, is just the number of linearly independent
linear forms in $L(C)$. (Remark: for the purposes of this paper $T_i$'s are given in circuit
representation and thus the list $L(T_i)$ is unambiguously defined from $C$)
{\bf [Similar forms]} For any two polynomials $f,g\in R$ we call $f$ {\em similar to} $g$ if there
exists $c\in\mathbb{F}^*$ such that $f=cg$. We say $f$ {\em is similar to $g$ mod
$I$}, for some ideal $I$ of $R$, if there is some $c\in\mathbb{F}^*$ such that
$f \equiv cg (\mod I)$. Note that ``similarity mod $I$" is an equivalence relation (reflexive, symmetric
and transitive) and partitions any list of polynomials into equivalence classes.
{\bf [Span $\text{sp}(\cdot)$]} For any $S\subseteq L(R)$ we let $\text{sp}(S)\subseteq L(R)$ be the
{\em linear span} of the linear forms in $S$ over the field $\mathbb{F}$. (Conventionally,
$\text{sp}(\emptyset)=\{0\}$.)
{\bf [Matchings]}
Let $U, V$ be lists of linear forms and $I$ be a subspace of $L(R)$. An {\em $I$-matching
$\pi$ between $U, V$} is a bijection $\pi$ between lists $U, V$ such
that: for all $\ell\in U$, $\pi(\ell)\in \mathbb{F}^*\ell+I$.
When $f,g$ are multiplication terms, an {\em $I$-matching between $f,g$}
would mean an $I$-matching between $L(f), L(g)$.
\end{definition}
We will show that all the multiplication terms of a minimal $\Sigma\Pi\Sigma$ identity can be matched by
a ``low'' rank space.
\begin{theorem}[Matching-Nucleus]\label{thm-mat-nucleus}
Let $C=T_1+\cdots+T_k$ be a $\Sigma\Pi\Sigma(k,d)$ circuit that is minimal and zero.
Then there exists a linear subspace $K$ of $L(R)$ such that:
\\\indent
1) $\text{rk}(K)< k^2$.
\\\indent
2) $\forall i\in[k]$, there is a $K$-matching $\pi_i$ between $T_1, T_i$.
\end{theorem}
The idea of matchings within identities was first introduced in~\cite{SS09},
but nothing as powerful as this theorem has been proven.
This theorem gives us a space of small rank, \emph{independent of $d$},
that contains most of the ``complexity" of $C$. All forms in $C$ outside $K$
are just mirrored in the various terms.
This starts connecting the algebra
of depth-$3$ identities to a combinatorial structure. Indeed, the graphical
picture (explained in detail below) that this theorem provides, really
gives an intuitive grasp on these identities. The proof
of this involves some interesting generalizations of the Chinese Remainder
Theorem to some special ideals.
\begin{definition}[mat-nucleus]
Let $C$ be a minimal $\Sigma\Pi\Sigma(k,d)$ identity. The linear subspace $K$ given by Theorem
\ref{thm-mat-nucleus} is called {\em mat-nucleus of $C$}.
\end{definition}
The notion of mat-nucleus is easier to see in the following unusual representation of
the $\Sigma\Pi\Sigma(4,d)$ circuit $C=\sum_{i\in[4]}T_i$. The four bubbles refer to the four
multiplication terms of $C$ and the points inside the bubbles refer to the linear forms
in the terms. The proof of Theorem \ref{thm-mat-nucleus} gives mat-nucleus as the space
generated by the linear forms in the dotted box. The linear forms that are not in
mat-nucleus lie ``above" the mat-nucleus and are all (mat-nucleus)-matched, i.e.
$\forall \ell\in (L(T_1)\setminus\text{mat-nucleus})$, there is a form similar to $\ell$ modulo
$\text{mat-nucleus}$ in each $(L(T_i)\setminus\text{mat-nucleus})$. Thus the essence of Theorem \ref{thm-mat-nucleus}
is: the mat-nucleus part of the terms of $C$ has low rank $k^2$, while the part of the terms
above mat-nucleus all look ``similar".
\begin{center}\includegraphics[scale=0.43]{mat-core.png}\end{center}
\subsubsection*{Proof Idea for Theorem \ref{thm-mat-nucleus}}
The key insight in the construction of mat-nucleus is a reinterpretation of the identity test
of Kayal \& Saxena \cite{KS07} as a structural result for $\Sigma\Pi\Sigma$ identities. Again, refer to
the following figure depicting a $\Sigma\Pi\Sigma(4,d)$ circuit $C$ and think of each bubble having
$d$ linear forms. Roughly, \cite{KS07} showed that $C=0$ iff for every {\em path}
$(v_1,v_2,v_3)$ (where $v_i\in L(T_i)$): $T_4\equiv0 (\mod v_1,v_2,v_3)$ or in ideal
terms, $T_4\in\ideal{v_1,v_2,v_3}$. Thus, roughly, it is enough to go through all the $d^3$ paths
to certify the zeroness of $C$. This is why the time complexity of the identity test of
\cite{KS07} is dominated by $d^k$.
\begin{center}\includegraphics[scale=0.35]{paths.png}\end{center}
Now if we are given a $\Sigma\Pi\Sigma(4,d)$ identity $C$ which is {\em minimal}, then we know that
$T_1+T_2+T_3\ne0$. Thus, by applying the above interpretation of \cite{KS07} to $T_1+T_2+T_3$
we will get a path $(v_1,v_2)$ such that $T_3\notin\ideal{v_1,v_2}$. Since $C=0$ this means
that $T_3+T_4\equiv0 (\mod v_1,v_2)$ but $T_3,T_4\not\equiv0 (\mod v_1,v_2)$ (if
$T_4$ is in $\ideal{v_1,v_2}$ then so will be $T_3$). Thus, $T_3\equiv -T_4 (\mod v_1,v_2)$
is a nontrivial congruence and it immediately gives us a $\ideal{v_1,v_2}$-matching
between $T_3, T_4$ (see Lemma \ref{lem-I-match}).
By repeating this argument with a different permutation of the terms we
could match different terms (by a different ideal), and finally we expect to match all the
terms (by the union of the various ideals).
This fantastic argument has numerous technical problems, but they can all be taken care
of by suitable algebraic generalizations. The main stumbling block is the presence
of \emph{repeating} forms. It could happen that $(\mod v_1)$, $v_2$ occurs in many
terms, or in the same term with a higher power.
The most important tool developed is an ideal version of Chinese
remaindering that forces us to consider not just linear forms $v_1, v_2$, but
{\em multiplication terms} $v_1, v_2$ dividing $T_1, T_2$ respectively. We give the full
proof in Section \ref{sec-mat-nucleus}.
(Interestingly, the {\em non}-blackbox identity test of
\cite{KS07} guides in devising a blackbox test of ``similar'' complexity over rationals.)
\subsection{Step 2: Certificate for Linear Independence of Gates}
Theorem \ref{thm-mat-nucleus} gives us a space $K$, of rank $<k^2$, that matches $T_1$ to each
term $T_i$. In particular, this means that the list $L_K(T_i):=L(T_i)\cap K$ has the same cardinality
$d'$ for each $i\in[k]$. In fact, if we look at the corresponding
multiplication terms $K_i:=M(L_K(T_i))$, $i\in[k]$, then they again form a $\Sigma\Pi\Sigma(k,d')$
identity! Precisely, $C'=\sum_{i\in[k]}\alpha_iK_i$ for some $\alpha_i$'s in $\mathbb{F}^*$ (see Lemma
\ref{lem-nucleus}) is an identity.
We would like $C'$ to somehow mimic the structure of $C$.
Of course $C'$ is simple but is it again minimal? Unfortunately, it may not be.
For reasons that will be clear later,
minimality of $C'$ would have allowed us to go directly to Step 3.
Now step 2 will involve increasing the space $K$ (but not by too much)
that gives us a $C'$ that ``behaves" like $C$.
Specifically,
if $T_1,\ldots,T_{k'}$ are {\em linearly independent} (i.e. $\nexists$
$\vec{\beta}\in\mathbb{F}^{k'}\setminus\{\vec{0}\}$ s.t. $\sum_{i\in[k']}\beta_iT_i=0$), then so are
$K_1,\ldots,K_{k'}$.
\begin{theorem}[Nucleus]\label{thm-nucleus}
Let $C=\sum_{i\in[k]}T_i$ be a minimal $\Sigma\Pi\Sigma(k,d)$ identity and let $\{T_i|i\in\mathcal{I}\}$ be
a maximal set of linearly
independent terms ($1\leqslant k':=|\mathcal{I}|<k$). Then there exists a linear subspace $K$ of $L(R)$ such that:
\begin{itemize}
\item[1)] $\text{rk}(K)< 2k^2$.
\item[2)] $\forall i\in[k]$, there is a $K$-matching $\pi_i$ between $T_1, T_i$.
\item[3)] (Define $\forall i\in\mathcal{I}$, $K_i:=M(L_K(T_i))$.) The terms $\{K_i|i\in\mathcal{I}\}$
are linearly independent.
\end{itemize}
\end{theorem}
\begin{definition}[nucleus]
Let $C$ be a minimal $\Sigma\Pi\Sigma(k,d)$ identity. The linear subspace $K$ given by Theorem
\ref{thm-nucleus} is called the {\em nucleus of $C$}. By Lemma~\ref{lem-nucleus}, the subspace $K$
induces an identity $C'=\sum_{i\in[k]} \alpha_i K_i$ which we call the {\em nucleus identity}.
\end{definition}
The notion of the nucleus is easier to grasp when $C$ is a $\Sigma\Pi\Sigma(k,d)$ identity that is
{\em strongly minimal}, i.e. $T_1,\ldots,T_{k-1}$ are linearly independent. Clearly, such a
$C$ is also minimal\footnote{If for some proper $S\subset[k]$, $\sum_{i\in S}T_i=$
$\sum_{i\in \vec{S}}T_i=0$ then linear independence of $T_1,\ldots,T_{k-1}$ is violated.}.
For such a $C$, Theorem \ref{thm-nucleus} gives a nucleus $K$ such that the corresponding
nucleus identity is strongly minimal. The structure of $C$ is very strongly
represented by $C'$. As a bonus, we actually end
up greatly simplifying the polynomial-time PIT algorithm of Kayal \& Saxena~\cite{KS07}
(although we will not discuss this point in detail in this paper).
\subsubsection*{Proof Idea for Theorem \ref{thm-nucleus}}
The first two properties in the theorem statement are already satisfied by mat-nucleus of $C$.
So we incrementally add linear forms to the space mat-nucleus till it satisfies property (3)
and becomes the nucleus. The addition of linear forms is guided by the ideal version of Chinese
remaindering. For convenience assume $T_1,T_2,T_3$ to be linearly independent. Then, by
homogeneity and equal degree, we have an equivalent ideal statement:
$T_2\notin\ideal{T_1}$ and $T_3\notin\ideal{T_1,T_2}$ (see Lemma \ref{lem-homo-ideal}). Even in
this general setting the path analogy (used in the last subsection) works and we essentially
get linear forms $v_1\in L(T_1)$ and $v_2\in L(T_2)$ such that: $T_2\notin\ideal{v_1}$ and $T_3\notin\ideal{v_1,v_2}$.
We now add these forms $v_1, v_2$ to the space mat-nucleus, and
call the new space $K$. It is expected that the new $K_1, K_2, K_3$ are now linearly
independent.
Not surprisingly, the above argument has numerous technical problems. But it can be made
to work by careful applications of the ideal version of Chinese
remaindering. We give the full proof in Section \ref{sec-nucleus}.
\subsection{Step 3: Invoking Sylvester-Gallai Theorems}
\label{sec-step-3}
We make a slight, but hopefully interesting, detour and leave
depth-$3$ circuits behind. We rephrase the standard Sylvester-Gallai theorems in terms of
\emph{Sylvester-Gallai closure} (or configuration) and \emph{rank bounds}.
This is far more appropriate
for our application, and seems to be very natural in itself.
\begin{definition}[$\text{SG}_k$-closed]\label{def-sg-1}
Let $k \in {\mathbb{N}}^{>1}$.
Let $S$ be a subset of non-zero vectors in $\mathbb{F}^{n}$
without \emph{multiples}: no two vectors in $S$ are scalar
multiples of each other\footnote{This is just a set of
elements in the projective space $\mathbb{F}\mathbb{P}^{n-1}$, but this
formulation in terms of vectors is more convenient for our applications.}.
Suppose that for every set $V$ of $k$ linearly independent vectors in $S$,
the linear span of $V$ contains at least $(k+1)$ vectors of $S$.
Then, the set $S$ is said to be \emph{$\text{SG}_k$-closed}.
\end{definition}
We would expect that if $S$ is finite then it will get harder to keep $S$ $\text{SG}_k$-closed as
$\text{rk}(S)$ is gradually increased. This intuition holds up when $\mathbb{F}=\mathbb{R}$.
As we mentioned earlier, the famous Sylvester-Gallai Theorem states: if a finite $S\subset\mathbb{R}^n$ is
$\text{SG}_2$-closed, then $\text{rk}(S)\le2$. It is optimal as the line
$S:=\{(1,0), (1,1), (1,2)\}$ has rank $2$ and is $\text{SG}_2$-closed.
In fact, there is
also a generalization of the Sylvester-Gallai theorem known (as stated in Theorem 2.1 of
\cite{BE67}) : {\em Let $S$ be a finite set in $\mathbb{R}\mathbb{P}^{2t}$ spanning that projective space.
Then, there exists a $t$-flat $H$ such that
$|H\cap S|=t+1$, and $H$ is spanned by those points $H\cap S$.}
\comment{
(as stated in Theorem 6.2 of
\cite{KSar09}) : {\em Let $S$ be a finite set of points
spanning an affine space $V\subseteq\mathbb{R}^n$ such that $\text{rk}(V)>2t$.
Then, there exists a $t$ dimensional hyperplane $H$ such that
$|H\cap S|=t+1$, and $H$ is spanned by the points
of $S$, i.e $\text{affine-span}(H\cap S)=H$. }
}
Let $S\subset\mathbb{R}^n$ be a finite set of points with first coordinate being $1$ and let $k\geq2$.
We claim that if $S$ is $\text{SG}_k$-closed, then $\text{rk}(S)\le2(k-1)$. Otherwise
the above theorem guarantees $k$ vectors $V$ in $S$ whose $(k-1)$-flat $H$ has only
$k$ points of $S$. If $\text{sp}(V)$ has a point $s\in S\setminus V$ then as $S$ has first
coordinates $1$, it would mean that a {\em convex} linear combination of $V$ (i.e. sum of
coefficients in the combination is $1$) is $s$. In other words, $s\in H$, which contradicts
$H$ having only $k$ points of $S$. Thus, $\text{sp}(V)$ also
has no point in $S\setminus V$, but this contradicts $\text{SG}_k$-closure of $S$. This shows that
higher dimensional Sylvester-Gallai theorem implies that if $S$ is $\text{SG}_k$-closed then
$\text{rk}(S)\le2(k-1)$. We prefer using this
rephrasal of the higher dimensional Sylvester-Gallai Theorem.
This motivates the following definitions.
\begin{definition}[SG operator]\label{def-sg-2}
Let $k, m \in {\mathbb{N}}^{>1}$.
{\bf [$\text{SG}_k(\cdot,\cdot)$]}
The largest possible rank of an $\text{SG}_k$-closed set of at most $m$ points in $\mathbb{F}^n$ is denoted by
$\text{SG}_k(\mathbb{F},m)$. For example, the above discussion entails $\text{SG}_k(\mathbb{R},m)\le2(k-1)$ which
is, interestingly, independent of $m$.
(Also verify that $\text{SG}_{k}(\mathbb{F},m)\leq\text{SG}_{k'}(\mathbb{F}',m')$ for $k\leqslant k'$,
$m\leqslant m'$ and $\mathbb{F}\subseteq\mathbb{F}'$.)
{\bf [$\text{SG}_k(\cdot)$]}
Suppose a set $S\subseteq\mathbb{F}^n$ has rank greater than $\text{SG}_k(\mathbb{F},m)$ (where $\#S\leqslant m$).
Then, by definition,
$S$ is not $\text{SG}_k$-closed. In this situation we say the
\emph{$k$-dimensional Sylvester-Gallai operator} $\text{SG}_k(S)$ (applied on $S$) returns
a set of $k$ linearly independent vectors $V$ in $S$ whose span has no point in $S\setminus V$.
\end{definition}
The Sylvester-Gallai theorem in higher dimensions can now be expressed
succintly.
\begin{theorem}[High dimension Sylvester-Gallai for $\mathbb{R}$]\cite{Han65,BE67}\label{thm-sg}
$\text{SG}_k(\mathbb{R},m)\le2(k-1)$.
\end{theorem}
\noindent {\bf Remark.} This theorem is also optimal, for if we set $S$ to be a union of $(k-1)$
``skew lines'' then $S$ has rank $2(k-1)$ and is $\text{SG}_k$-closed. For example, when
$k=3$ define $S:=\{(1,1,0,0),(1,1,1,0),(1,1,2,0)\}$ $\cup$
$\{(1,0,1,0),(1,0,1,1),(1,0,1,2)\}$. It is
easy to verify that $\text{rk}(S)=4$ and the span of every three linearly independent vectors in
$S$ contains a fourth vector!
Using some linear algebra and combinatorial tricks, we prove the
first ever Sylvester-Gallai bound for all fields. The proof is in Section~\ref{sec-sg},
where there is a more detailed discussion of this (and the connection
with LDCs).
\smallskip\noindent
{\bf Theorem \ref{thm-sgk}} ($\text{SG}_k$ for all fields). For any field $\mathbb{F}$ and
$k,m\in\mathbb{N}^{>1}$, $\text{SG}_k(\mathbb{F},m)\leq 9k\lg m$.
\subsubsection{Back to identities}
Let $C$ be a simple and strongly minimal $\Sigma\Pi\Sigma(k,d)$ identity.
Theorem \ref{thm-nucleus} gives us a nucleus $K$, of rank $<2k^2$, that matches $T_1$ to each
term $T_i$. As seen in Step 2, if we look at the corresponding
multiplication terms $K_i:=M(L_K(T_i))$, $i\in[k]$, then they again form a $\Sigma\Pi\Sigma(k,d')$
``nucleus identity'' $C'=\sum_{i\in[k]}\alpha_iK_i$, for some $\alpha_i$'s in $\mathbb{F}^*$, which is
simple and strongly minimal. Define the {\em non-nucleus part} of $T_i$ as
$L_K^c(T_i):=L(T_i)\setminus K$, for all $i\in[k]$ ($c$ in the exponent annotates
``complement", since $L(T_i)=L_K(T_i)\sqcup L_K^c(T_i)$). What can we say about the rank of
$L_K^c(T_i)$ ?
Define the {\em non-nucleus part of $C$} as $L^c_K(C):=$ $\bigcup_{i\in[k]}L_K^c(T_i)$.
Our goal in Step 3 is to bound $\text{rk}(L^c_K(C)\ \mod K)$ by $2k$
when the field is $\mathbb{R}$. This will give us a rank bound of $\text{rk}(K)+$ $\text{rk}(L^c_K(C) \mod K)$
$<(2k^2+2k)$ for simple and strongly
minimal $\Sigma\Pi\Sigma(k,d)$ identities over $\mathbb{R}$. The proof is mainly combinatorial, based on higher
dimensional Sylvester-Gallai theorems and a property of set partitions, with a sprinkling
of algebra.
We will finally apply $\text{SG}_k$ operator not directly on the forms in $L(C)$ but on a suitable
truncation of those forms. So we need another definition.
\begin{definition}[Non-$K$ rank]\label{def-non-nucleus}
Let $K$ be a linear subspace of $L(R)$. Then $L(R)/K$ is again a linear space (the {\em quotient
space}). Let $S$ be a list of forms in $L(R)$. The {\em non-$K$ rank of $S$} is defined to
be $\text{rk}(S \mod K)$ (i.e. the rank of $S$ when viewed as a subset of $L(R)/K$).
Let $C$ be a $\Sigma\Pi\Sigma(k,d)$ identity with nucleus $K$. The non-$K$ rank of the non-nucleus part
$L^c_K(T_i)$ is called the {\em non-nucleus rank of $T_i$}.
Similarly, the non-$K$ rank of the non-nucleus part $L^c_K(C):=$
$\bigcup_{i\in[k]}L^c_K(T_i)$ is called the {\em non-nucleus rank of $C$}.
\end{definition}
We give an example to explain the non-$K$ rank.
Let $R = \mathbb{F}[z_1,\cdots,z_n,y_1,\cdots,y_m]$. Suppose
$K = \text{sp}(z_1,\cdots,z_n)$ and $S \subset L(R)$. We can
take any element $\ell$ in $S$ and simply drop all the $z_i$ terms, i.e. `truncate' $z$-part of
$\ell$. This
gives a set of linear forms over the $y$ variables. The rank
of these is the non-$K$ rank of $S$.
We are now ready to state the theorem that is proved in Step 3. It basically shows a
neat relationship between the non-nucleus part and Sylvester-Gallai.
\begin{theorem}[Bound for simple, strongly minimal identities]\label{thm-strong-rank}
Let $|\mathbb{F}|>d$. The non-nucleus rank of
a simple and strongly minimal $\Sigma\Pi\Sigma(k,d)$ identity over $\mathbb{F}$ is at most $\text{SG}_{k-1}(\mathbb{F},d)$.
\end{theorem}
Given a simple, minimal $\Sigma\Pi\Sigma(k,d)$ identity $C$ that is not strongly minimal.
Let $T_1,\ldots,T_{k'}$ be linearly independent and form a basis of $\{T_i|i\in[k]\}$. Then
it is clear that $\exists\vec{a}\in\mathbb{F}^{k'}\setminus\{\vec{0}\}$ such that
$\sum_{i\in[k']}a_iT_i+T_{k'+1}$ is a strongly minimal $\Sigma\Pi\Sigma(k'',d)$ identity (for some
$1<k''\leq k'+1$).
Hence, we could apply the above theorem on this identity and get a rank bound for the non-nucleus
part. The only problem is this fanin-$k''$ identity may not be simple. Our solution
for this is to replace $T_{k'+1}$ by the suitable linear combination of $\{T_i|i\in[k']\}$
in $C$ and repeat the above argument on the new identity. In Section \ref{sec-gen-case}
we show this takes care of the
whole non-nucleus part and bounds its rank by $k\cdot\text{SG}_k(\mathbb{F},d)$. To state the theorem formally,
we need a more refined notion than the fanin of a $\Sigma\Pi\Sigma$ circuit.
\begin{definition}[Independent-fanin]
Let $C=\sum_{i\in[k]}T_i$ be a $\Sigma\Pi\Sigma(k,d)$ circuit. The {\em independent-fanin} of $C$, $\text{ind-fanin}(C)$, is defined
to be the size of the maximal $\mathcal{I}\subseteq[k]$ such that $\{T_i|i\in\mathcal{I}\}$ are linearly independent
polynomials. (Remark: If $\text{ind-fanin}(C)=k$ then $C\ne0$. Also, for an identity $C$, $C$ is strongly minimal
iff $\text{ind-fanin}(C)=k-1$.)
\end{definition}
We now state the following stronger version of the main theorem.
\begin{theorem}[Final bound]\label{thm-rank} Let $|\mathbb{F}|>d$. The rank of a simple, minimal
$\Sigma\Pi\Sigma(k,d)$, independent-fanin $k'$, identity is at most $2k^2 + (k-k')\cdot\text{SG}_{k'}(\mathbb{F},d)$.
\end{theorem}
\noindent {\bf Remark:}
In particular, the rank of a simple, minimal $\Sigma\Pi\Sigma(k,d)$ identity over reals is at most
$2k^2 + (k-k')\cdot\text{SG}_{k'}(\mathbb{R},d)$ $\leq$ $2k^2 + (k-k')2(k'-1)$ $<3k^2$, proving the main theorem
over reals. Likewise, for any $\mathbb{F}$, we get the rank bound of
$2k^2 + (k-k')\cdot\text{SG}_{k'}(\mathbb{F},d)$ $\leq$ $2k^2 + (k-k')9k'\lg d\leqslant$ $2k^2 + \frac{9k^2}{4}\lg d$
$<3k^2\lg 2d$, proving the main theorem.
\subsubsection*{Proof Idea for Theorem \ref{thm-strong-rank}}
Basically, we apply the $\text{SG}_k(\cdot)$ operator on the non-nucleus part of the term $T_1$,
i.e. we treat a linear form $\sum_ia_ix_i$ as the point
$(1,\frac{a_2}{a_1},\ldots,\frac{a_n}{a_1})\in\mathbb{F}^n$ for the purposes of Sylvester-Gallai and then
we consider $\text{SG}_k(L_K^c(T_1))$ assuming that the non-nucleus rank of $T_1$ is more
than $\text{SG}_k(\mathbb{F},d)$.
This application of Sylvester-Gallai is much more direct compared
to the methods used in \cite{KSar09}. There, they needed versions
of Sylvester-Gallai that dealt with colored points and had to
prove a {\em hyperplane decomposition} property after applying essentially a
$\text{SG}_{k^{O(k)}}(\cdot)$ operator on $L(C)$.
Since, modulo the nucleus, all multiplication terms look essentially the same,
it suffices to focus attention on just one of them.
Hence, we apply the $\text{SG}_k$-operator on a single multiplication term.
To continue with the proof idea, assume $C$ is a simple, strongly minimal $\Sigma\Pi\Sigma(k,d)$
identity with terms $\{T_i|i\in[k]\}$ and let $K$ be its nucleus given by Step 2. It will
be convenient for us to fix a linear form $y_0\in L(R)^*$ and a subspace $U$ of $L(R)$
such that we have the following {\em orthogonal} vector space decomposition
$L(R)=\mathbb{F} y_0\oplus U\oplus K$ (i.e. $\ell\in \mathbb{F} y_0\cap U$ implies $\ell=0$ and
$\ell\in (\mathbb{F} y_0\oplus U)\cap K$ implies $\ell=0$). This means for any form $\ell\in L(R)$,
there is a unique way to express $\ell=\alpha y_0+u+v$, where $\alpha\in\mathbb{F}$, $u\in U$ and
$v\in K$.
Furthermore, we will assume wlog that for every form $\ell\in L^c_K(T_1)$ the corresponding
$\alpha$ is nonzero, i.e. each form in $L^c_K(T_1)$ is {\em monic} wrt $y_0$ (see Lemma
\ref{lem-monic-forms}).
\begin{definition}[$\text{trun}(\cdot)$]\label{def-trun}
Fix a decomposition $L(R)=\mathbb{F} y_0\oplus U\oplus K$. For any form $\ell\in L^c_K(T_1)$,
there is a unique way to express $\ell=\alpha y_0+u+v$, where $\alpha\in\mathbb{F}^*$, $u\in U$ and
$v\in K$.
The {\em truncated form $\text{trun}(\ell)$} is the linear form obtained by dropping the $K$ part and
normalizing, i.e. $\text{trun}(\ell):=y_0+\alpha^{-1}u$.
Given a list of forms $S$ we define $\text{trun}(S)$ to be the corresponding {\em set} (thus no
repetitions) of truncated forms.
\end{definition}
To be precise, we fix a basis $\{y_1,\ldots,y_{\text{rk}(U)}\}$ of $U$ so that each form in $\text{trun}(L_K^c(T_1))$
has representation $y_0+\sum_{i\geq1}a_iy_i$ ($a_i\text{'s}\in\mathbb{F}$). We view each such form as the
{\em point}
$(1,a_1,\ldots,a_{\text{rk}(U)})$ while applying Sylvester-Gallai on $\text{trun}(L_K^c(T_1))$.
Assume, for the sake of
contradiction, that the non-nucleus rank of $T_1$, $\text{rk}(\text{trun}(L_K^c(T_1)))>\text{SG}_k(\mathbb{F},d)$ then
(by definition)
$\text{SG}_k(\text{trun}(L_K^c(T_1)))$ gives $k$ linearly independent forms $\ell_1,\ldots,\ell_k\in(y_0+U)$
whose span contains no {\em other} linear form of $\text{trun}(L_K^c(T_1))$.
For simplicity of exposition, let us fix $k=4$, $K$ spanned by $z$'s, $U$ spanned by $y$'s
and $\ell_i=y_0+y_i$ $(i\in[4])$. Note that (by definition)
$\text{trun}(\alpha y_0+\sum_i \alpha_i z_i + \sum_i \beta_i y_i) = y_0+\sum_i\frac{\beta_i}{\alpha} y_i$.
We want to derive a
contradiction by using the $\text{SG}_4$-tuple $(y_0+y_1,y_0+y_2,y_0+y_3,y_0+y_4)$ {\em and} the fact
that $C$ is a simple, strongly minimal $\Sigma\Pi\Sigma(4,d)$ identity. The contradiction is easy to see
in the following configuration: Suppose the linear forms in $C$ that are similar to a form in
$\bigcup_{i\in[4]}(y_0+y_i+K)$ are {\em exactly} those depicted in the figure. Let us consider
$C$ modulo the ideal $I:=\ideal{y_0+y_1+z_1, y_0+y_2+z_2, -y_0-y_4+z_2}$. It is easy to see
that these forms (call them $\ell_1',\ell_2',\ell_4'$) ``kill" the first three gates, leaving
$C\equiv T_4 (\mod I)$. As $C$ is an identity this means $T_4\in I$, thus there is a form
$\ell\in L(T_4)$ such that $\ell\in\text{sp}(\ell_1',\ell_2',\ell_4')$. Now none of the forms
$\ell_1',\ell_2',\ell_4'$ divide $T_4$.
Also, their non-trivial combination, say
$\alpha\ell_1'+\beta\ell_2'$ for $\alpha\beta\ne0$, cannot occur in $L(T_4)$.
Otherwise, by the matching property $\text{trun}(\alpha\ell_1'+\beta\ell_2')=$
$(\alpha+\beta)^{-1}(\alpha\ell_1+\beta\ell_2)$ will be
in $\text{trun}(L^c_K(T_1))$. This contradicts the $\ell_i$'s being a
$\text{SG}_4$-tuple. Thus, $T_4$ cannot be in $I$, a contradiction. This means
that the non-nucleus rank of $T_1$ is $\leq\text{SG}_4(\mathbb{F},d)$, which by matching properties implies
the non-nucleus rank of $C$ is $\leq\text{SG}_4(\mathbb{F},d)$.
\begin{center}\includegraphics[scale=0.48]{sg4.png}\end{center}
We were able to force a contradiction because we used a set of forms in
an SG-tuple that killed three terms and ``preserved" the last term.
Can we always do this? This is not at all obvious, and that is because
of repeating forms. Suppose, after going modulo form $\ell$,
the circuit looks like $x^3y + 2x^2y^2 + xy^3 = 0$. This is not simple,
but \emph{it does not have to be}. We are only guaranteed that the original
circuit is simple. Once we go modulo $\ell$, that property is lost.
Now, the choice of \emph{any} form kills all terms. In the figure
above, $\<y_0+y_1+z_1, y_0+y_2+z_2$, $y_0+y_3+z_3\>$ does not yield
a contradiction. We will use our more powerful Chinese remaindering
tools and the nucleus properties to deal with this.
We have to prove
a special theorem about partitions of $[k]$ and use strong minimality (which we did not use in the
above sketch). The full proof is given in Section \ref{sec-strong}.
\section{Matching the Terms in an Identity: Construction of mat-nucleus}
\label{sec-mat-nucleus}
\subsection{Chinese Remaindering for Multiplication Terms}
Traditionally, Chinese remaindering is the fact: if two coprime polynomials (resp. integers) $f, g$
divide a polynomial (resp. integer) $h$ then $fg$ divides $h$. The key tool in constructing mat-nucleus
is a version of Chinese
remaindering specialized for multiplication terms but generalized to ideals. Similar methods appeared
first in \cite{KS07} but we turn those on their head and give a ``simpler'' proof. In particular, we avoid the
use of local rings and Hensel lifting.
\begin{definition}[Radical-span]
Let $S:=\{f_1,\ldots,f_m\}$ be multiplication terms generating an ideal $I$.
Define linear space $\text{radsp}(S):=$ $sp(L(f_1)\cup\ldots\cup L(f_m))$.
When the set of generators $S$ are clear from the context we will also use the notation $\text{radsp}(I)$.
Similarly, $\text{radsp}(I,f)$ would be a shorthand for $\text{radsp}(S\cup\{f\})$.
\end{definition}
\noindent {\bf Remark.} Radical-span is motivated by the {\em radical} of an ideal but it is not quite that, for example,
$\text{radical}(x_1^2,x_1x_2)=$ $\ideal{x_1}$ but $\text{radsp}(x_1^2,x_1x_2)=sp(x_1,x_2)$. It is easy to see that
the ideal generated by $\text{radsp}$ always contains the radical ideal.
Now we can neatly state Chinese remaindering as an {\em ideal decomposition} statement.
\begin{theorem}[Ideal Chinese remaindering] \label{thm-crt}
Let $f_1,\ldots,f_m, z,f,g$ be multiplication terms. Define the ideal $I:=\ideal{f_1,\ldots,f_m}$.
Assume $L(z)\subseteq \text{radsp}(I)$ while, $L(f)\cap\text{radsp}(I)=\emptyset$ and $L(g)\cap\text{radsp}(I,f)=\emptyset$. Then,
$\ideal{I, zfg}=\ideal{I, z} \cap \ideal{I, f} \cap \ideal{I, g}$.
\end{theorem}
\begin{proof}
If $h$ is a polynomial in $\ideal{I, zfg}$ then clearly it is in each of the ideals $\ideal{I, z}$, $\ideal{I, f}$ and
$\ideal{I,g}$.
Suppose $h$ is a polynomial in $\ideal{I, z} \cap \ideal{I, f} \cap \ideal{I, g}$. Then by definition there exist
$i_1,i_2,i_3\in I$ and $a,b,c\in R$ such that,
$$h=i_1+az=i_2+bf=i_3+cg.$$
The second equation gives $bf\in \ideal{I,z}$. Since $L(f)\cap\text{radsp}(I,z)=L(f)\cap\text{radsp}(I)=\emptyset$,
repeated applications of Lemma \ref{lem-non-zd} give us, $b\in\ideal{I,z}$.
Implying $bf \in \ideal{I,z} f \subseteq \ideal{I, zf}$, hence $h=i_2+bf\in \ideal{I,zf}$.
This ensures the existence of $i_2'\in I$ and a polynomial $b'$ such that,
$$h=i_2'+b'zf=i_3+cg.$$
Again this system says that $cg\in \ideal{I,zf}$. Since $L(g)\cap\text{radsp}(I,zf)=L(g)\cap\text{radsp}(I,f)=\emptyset$,
repeated applications of Lemma \ref{lem-non-zd} give us $c\in\ideal{I,zf}$.
Implying $cg \in \ideal{I,zf} g\subseteq \ideal{I, zfg}$,
hence $h=i_3+cg\in \ideal{I,zfg}$. This finishes the proof.
\end{proof}
The conditions in this theorem suggest that factoring a multiplication term $f$ into parts corresponding
to the equivalence classes of ``similarity mod $\text{radsp}(I)$" would be useful.
\begin{definition}[Nodes]
Let $f$ be a multiplication term and let $I$ be an ideal generated by some multiplication terms.
As the relation ``similarity mod $\text{radsp}(I)$" is an equivalence relation on $L(R)$, it partitions, in
particular, the list $L(f)$ into equivalence classes.
{\bf[$\text{rep}_I(f)$]}
For each such class pick a representative $\ell_i$
and define $\text{rep}_I(f):=$ $\{\ell_1,\ldots,\ell_r\}$.
(Note that form $0$ can
also appear in this set, it represents the class $L(f)\cap\text{radsp}(I)$.)
{\bf[$\text{nod}_I(f)$]}
For each $\ell_i \in \text{rep}_I(f)$, we multiply the forms in $f$ that are similar to
$\ell_i$
mod $\text{radsp}(I)$. We define {\em nodes of $f$ mod $I$} as the set of polynomials $\text{nod}_I(f):=$
$\{M(L(f)\cap(\mathbb{F}^*\ell+\text{radsp}(I)))\ |\ \ell\in\text{rep}_I(f)\}$. (Remark: When $I=\{0\}$, nodes of $f$
are just the coprime powers-of-forms dividing $f$.)
{\bf [...wrt a subspace]}
Let $K$ be a linear subspace of $L(R)$. Clearly, the relation
``similarity mod $K$" is an equivalence relation on $L(R)$. It will be convenient for us to also
use notations $\text{rep}_K(f)$ and $\text{nod}_K(f)$. They are defined by replacing $\text{radsp}(I)$ in the
above definitions by $K$.
\end{definition}
Observe that the product of polynomials in $\text{nod}_I(f)$ just gives $f$.
Also, modulo $\text{radsp}(I)$, each node is just a form-power $\ell^r$. In other words,
modulo $\text{radsp}(I)$, a node is rank-one term. The choice of the word ``node"
might seem a bit mysterious, but we will eventually construct paths through these.
To pictorially see what is going on, think of each term $T_i$ as a set of
its constituent nodes.
We prove a corollary of the ideal Chinese remaindering
theorem that will be very helpful in both Steps 1 and 2.
\begin{corollary}\label{cor-crt}
Let $h\in R$, $f$ be a multiplication term, and let $I$ be an ideal generated by some multiplication
terms. Then, $h\notin\ideal{I,f}$ iff $\exists g\in\text{nod}_I(f)$ such that $h\notin\ideal{I,g}$.
\end{corollary}
\begin{proof}
If $h\notin\ideal{I,g}$, for some $g\in\text{nod}_I(f)$, then clearly $h\notin\ideal{I,f}$.
Conversely, assume $h\notin\ideal{I,f}$.
Let $\text{rep}_I(f)=$ $\{\ell_1,\ldots,\ell_r\}$ and correspondingly, $\text{nod}_I(f)=$ $\{g_1,\ldots,g_r\}$.
If $r=1$ then $f$ is similar to $g_1$, hence $h\notin\ideal{I,g_1}$ and we are done.
So assume
$r\geq2$. Also, in case $L(f)$ has a form in $\text{radsp}(I)$, assume wlog $\ell_1$ is the representative of
the class $L(f)\cap\text{radsp}(I)$. Define $G_i:=\prod_{i<j\leqslant r}g_i$, for all $i\in[r-1]$.
We claim that for all $i\in[r-1]$, $L(G_i)\cap\text{radsp}(I,g_i)=\emptyset$. Otherwise
$\exists\ell\in L(G_i)$ such that either $\ell\in\text{radsp}(I)$ or $\ell\in(\mathbb{F}^*\ell_i+\text{radsp}(I))$.
Former case contradicts $\ell_1$ being the representative of the class $L(f)\cap\text{radsp}(I)$, while
the latter case contradicts $\ell_{i+1},\ldots,\ell_r$ being {\em non}-similar to $\ell_i$
mod $\text{radsp}(I)$. Thus, for all $i\in[r-1]$, $L(G_i)\cap\text{radsp}(I,g_i)=\emptyset$, and by
applying Theorem \ref{thm-crt} on $\ideal{I,g_iG_i}$ for each $i\in[r-1]$, we deduce:
$$\ideal{I,f}=\left\<I,\prod_{i\in[r]}g_i\right\>=\bigcap_{i\in[r]}\ideal{I,g_i}.$$
Thus, $h\notin\ideal{I,f}$ implies the existence of some $i\in[r]$ such that $h\notin\ideal{I,g_i}$.
\end{proof}
\subsection{Applying Chinese Remaindering to $\Sigma\Pi\Sigma$ Circuits}
We showed the effect of ideal Chinese remaindering on a single multiplication
term $f$ in Corollary \ref{cor-crt}. Now we show the effect on a {\em tuple}
of multiplication terms, for example, appearing in a $\Sigma\Pi\Sigma$ circuit. We then
need, quite naturally, a notion of {\em path} of nodes.
\begin{definition}[Paths]
Let $I$ be an ideal generated by some multiplication terms.
Let $C=\sum_{i\in[k]}T_i$ be a $\Sigma\Pi\Sigma(k,d)$ circuit. Let $v_i$ be a {\em sub-term} of
$T_i$ (i.e. $L(v_i)\subseteq L(T_i)$), for all $i\in[k]$.
We call the tuple $(I,v_1,\ldots,v_k)$ a {\em path of $C$ mod $I$} if, for all $i\in[k]$,
$v_i\in\text{nod}_{\ideal{I,v_1,\ldots,v_{i-1}}}(T_i)$. It is of {\em length} $k$.
(Remark: We have defined path $\vec{p}$ as a tuple but, for convenience, we will sometimes
treat it as a {\em set} of multiplication terms, eg. when operated upon by $\text{sp}(\cdot)$,
$\ideal{\cdot}$, $\text{radsp}(\cdot)$, etc.)
Conventionally, when $k=0$ the circuit $C$ has just ``one'' gate: $0$. In that case,
the only path $C$ has is $(I)$, which is of length $0$.
We also define, for any subset $S\subseteq[k]$, the {\em sub-circuit} $C_S:=\sum_{s\in S}T_s$.
For an $i\in\{0,\ldots,k-1\}$, define $[i]':=[k]\setminus[i]$.
We set $[0]:=\emptyset$ and $C_\emptyset:=0$.
\end{definition}
We now show that if $C$ is a nonzero $\Sigma\Pi\Sigma(k,d)$ circuit then $\exists i\in\{0,\ldots,k-1\}$,
such that $C_{[i]}$ has a path $\vec{p}$ for which, $C\ (\mod \ideal{\vec{p}})$ is nonzero {\em and}
similar to some multiplication term. This rather special path inside $C$ can be seen as a
certificate for the nonzeroness.
The rank of the linear forms
appearing in this path can be at most $i+\text{rk}(\text{radsp}(I))$, since
the rank of each node is one, modulo the radical-span of the previous nodes in the path.
Hence, it is a {\em low-rank certificate for the
nonzeroness of $C$}.
\begin{theorem}[Certificate for a Non-identity]\label{thm-cert-non0}
Let $I$ be an ideal generated by some multiplication terms. Let $C=\sum_{i\in[k]}T_i$ be a
$\Sigma\Pi\Sigma(k,d)$ circuit that is nonzero modulo $I$. Then $\exists i\in\{0,\ldots,k-1\}$ such that
$C_{[i]}$ mod $I$ has a path $\vec{p}$ satisfying:
$C_{[i]'}\equiv \alpha\cdot T_{i+1}\not\equiv0$ $(\mod \vec{p})$ for some
$\alpha\in\mathbb{F}^*$.
\end{theorem}
Before we prove the theorem, we make an aside observation.
If the reader has kept the mental picture of the
terms as consisting of rank-one (modulo $\text{radsp}(I)$) nodes, then the notion of a path
has some meaning. A path $\vec{p}$ kills the terms that is passed through, and
collapses remaining circuit into a single term.
This is very reminiscent of the poly-time algorithm of Kayal \& Saxena~\cite{KS07}.
Indeed, this theorem is a (shorter) \emph{proof} of the correctness
of the algorithm. Why? Consider the path $\vec{p}$ given by the theorem
when $I$ is the zero ideal.
The path $\vec{p}$ can be represented by a list of at most $k$ `forms' in $L(C)$.
This path comes from some $C_{[i]}$, which means that $C_{[i]} = 0 (\mod \vec{p})$.
So, we get that $C \equiv \alpha \cdot T_{i+1} \not\equiv 0 (\mod \vec{p})$.
Since $T_{i+1}$ is a product of linear
forms, it is easy to algorithmically check if $C \equiv 0 (\mod \vec{p})$.
If $C$ is identically zero, such a path cannot exist. Since there are at most
$d^k$ different paths, we can exhaustively check all of them. That yields
an alternative view of \cite{KS07}'s test.
\begin{proof}
Fix an $i\in\{0,\ldots,k-1\}$ and a path $\vec{p}$ of $C_{[i]}$ mod $I$ such that:
1) $C_{[i]'}\notin\ideal{\vec{p}}$ and,
2) the set $J_i:=\{j\in[i]'\ |\ T_j\notin\ideal{\vec{p}}\}\ne\emptyset$ is the smallest possible
(over all $i$).
\smallskip\noindent
Note that for values $i=0$, $\vec{p}=(I)$, the condition (1) is satisfied and the corresponding
$J_i\ne\emptyset$. Thus, there also exist $i$ and $\vec{p}$ satisfying both the conditions (1) and (2).
Let $j^*$ be the smallest element in $J_i$. This means that for every $m$, $i<m<j^*$,
$T_m\in\ideal{\vec{p}}$. This means, by repeated applications of Lemma \ref{lem-non-zd},
$v_m:=M(L_{\text{radsp}(\vec{p})}(T_m))$ $\in\ideal{\vec{p}}$. Thus, $\ideal{\vec{p}}=$
$\ideal{\vec{p}\cup\{v_m|i<m<j^*\}}$. This makes $\vec{q}:=(\vec{p},(v_m|i<m<j^*))$ also a path
of $C_{[j^*-1]}$ mod $I$. We now claim that $\vec{q}$ is the path promised in the theorem statement.
Note that $C_{[j^*-1]'}\equiv C_{[i]'}$ $(\mod \vec{p})$ and $C_{[i]'}\notin\ideal{\vec{p}}=$
$\ideal{\vec{q}}$, in other words,
path $\vec{q}$ also satisfies:
1) $C_{[j^*-1]'}\notin\ideal{\vec{q}}$ and,
2) the set $J_{j^*-1}$ $=\{j\in[j^*-1]'\ |\ T_j\notin\ideal{\vec{q}}\}$ $=J_i$ is still
the smallest possible.
\smallskip\noindent
If $C_{[j^*-1]'}$
$\notin\ideal{\vec{q},T_{j^*}}$ then, by Corollary \ref{cor-crt}, there exists
$v_{j^*}\in\text{nod}_{\ideal{\vec{q}}}(T_{j^*})$ such that $C_{[j^*-1]'}$ $\notin\ideal{\vec{q},v_{j^*}}$,
hence $C_{[j^*]'}=$ $C_{[j^*-1]'}-T_{j^*}$ $\notin\ideal{\vec{q},v_{j^*}}$.
Define $\vec{q}':=(\vec{q},v_{j^*})$, clearly it is a path of $C_{[j^*]}$ mod $I$. Wrt this path $\vec{q}'$,
$J_{j^*}\subseteq J_i\setminus\{j^*\}$ $\subsetneq J_i$ together with $C_{[j^*]'}$
$\notin\ideal{\vec{q}'}$, contradicting the minimality assumption on $i$.
Thus, we assume $C_{[j^*-1]'}$ $\in\ideal{\vec{q},T_{j^*}}$. By Lemma \ref{lem-homo-ideal}, this
guarantees the existence of an $\alpha\in\mathbb{F}$ such that,
$$(C_{[j^*-1]'}-\alpha T_{j^*})\in\ideal{\vec{q}}=\ideal{\vec{p}}.$$
Since $C_{[j^*-1]'}\equiv C_{[i]'}$ $\not\equiv0$ $(\mod \vec{p})$,
the above equation can be rewritten as:
$$C_{[j^*-1]'}\equiv\alpha T_{j^*}\not\equiv0\ (\mod \ideal{\vec{q}}).$$
Thus, finishing the proof ($\alpha$ nonzero is implied).
\end{proof}
\noindent {\bf Remark.} The above theorem is quite powerful, for instance, it only needs the {\em non}-zeroness of
$C$ mod $I$ without referring to any simplicity or minimality requirements.
\subsection{Using Minimality to get mat-nucleus}
If we are given a circuit that is zero \& minimal (may not be simple) then a repeated application
of Theorem \ref{thm-cert-non0} gives us a space {\em mat-nucleus} that matches all the
multiplication terms of $C$.
\smallskip\noindent
{\bf Theorem \ref{thm-mat-nucleus}} (Matching-Nucleus).
Let $C=T_1+\cdots+T_k$ be a $\Sigma\Pi\Sigma(k,d)$ circuit that is minimal and zero.
Then there exists a linear subspace $K$ of $L(R)$ such that:
\begin{enumerate}
\item[1)] $\text{rk}(K)< k^2$.
\item[2)] $\forall i\in[k]$, there is a $K$-matching $\pi_i$ between $T_1, T_i$.
\end{enumerate}
\begin{proof}
The proof is an iterative process with at most $k$ rounds. We maintain a set $\mathcal{P}$,
containing paths of some sub-circuits of $C$, and an undirected graph $G = ([k],E)$.
For convenience, define $U:=\text{radsp}(p|p\in\mathcal{P})$ (i.e. consider each path $p$ as a set of
multiplication terms, take the union of all these sets, and compute its radical-span).
The {\em invariant} at the end of each round is: $(i,j)\in E$ iff $T_i, T_j$ are
$U$-matched. At the end of round $0$ we assume, $\mathcal{P}:=\{(0)\}$ and $E:=$
$\{(i,j)\in[k]^2\ |\ T_i, T_j\text{ are similar}\}$. We want to eventually make $G$ a
connected graph (infact a {\em $k$-clique}) by keeping $\text{rk}(U)$ as small as possible.
Suppose the invariant holds till the end of some round $(r-1)\geq0$. If $G$ is connected
then the process stops at round $(r-1)$. Otherwise, we will show how
to {\em decrease} the number of connected components of $G$ in round $r$.
Say, $G$ has a maximal connected component on vertices $S\subsetneq[k]$.
Since $C_S\ne0$ (by minimality), we can apply Theorem \ref{thm-cert-non0} on $C_S$ mod $\ideal{0}$
to get a path $p_S$ inside $C_S$ mod $\ideal{0}$ such that $\exists i\in S$,
$C_S\equiv \alpha T_i \not\equiv0\ (\mod p_S)$ for some $\alpha\in\mathbb{F}^*$.
Define $S':=[k]\setminus S$.
Now,
\begin{equation}\label{eqn-pf-mat-nucleus}
C \equiv C_{S'}+\alpha T_i \equiv 0\ (\mod p_S).
\end{equation}
This means $C_{S'}\notin\ideal{p_S}$ (otherwise $\alpha T_i\in\ideal{p_S}$, a contradiction).
Thus, we can apply Theorem \ref{thm-cert-non0} on $C_{S'}$ mod $\ideal{p_S}$ to get a path
$p_{S'}$ inside $C_{S'}$ mod $\ideal{p_S}$ such that, $C_{S'}\equiv\beta T_j\not\equiv0 (\mod p_{S'})$,
for some $\beta\in\mathbb{F}^*$. This allows us to rewrite Equation (\ref{eqn-pf-mat-nucleus}) as:
$$\alpha T_i\equiv -\beta T_j\not\equiv0\ (\mod p_{S'})$$
Define $K':=\text{radsp}(p_{S'})$.
As $p_{S'}$ is, after all, a path of some sub-circuit of $C\ \mod\ideal{0}$, of
length at most $|S|-1+|S'|-1=k-2$, we deduce that
$\text{rk}(K')<(k-1)$. Also, by Lemma \ref{lem-I-match}, the above congruence implies a $K'$-matching
between $T_i$ and $T_j$. We append the path $p_{S'}$ to $\mathcal{P}$ and update $U$. Note that for
any edge $(i,i')$ in the connected component $S$, and for any edge $(j,j')$ in the connected
component $\tilde{S}$ (of vertex $j$): since $T_i, T_{i'}$ are still $U$-matched;
$T_j, T_{j'}$ are still $U$-matched; $T_i,T_j$ are newly $K'$-matched; gives us that
$T_{i'}, T_{j'}$ are newly $U$-matched. In other words, the two different connected components
$S$ and $\tilde{S}$ of $G$ will now form a bigger connected component (infact a {\em clique})
when we update the graph as,
$E:=\{(a,b)\in[k]^2\ |\ T_a, T_b\text{ are }U\text{-matched}\}$.
So in every round we are increasing $\text{rk}(U)$ by at most $(k-1)$, maintaining the invariant, and
decreasing the number of connected components in $G$ by at least one. Thus, after at most
$(k-1)$ repetitions we get a $U$ that matches $T_1, T_i$, for all $i\in[k]$, and $\text{rk}(U)<k^2$.
We define this $U$ as $K$, finishing the proof.
\end{proof}
\section{Certificate for Linear Independence of terms: Constructing nucleus}
\label{sec-nucleus}
Suppose we have multiplication gates $T_1,\ldots,T_{k'}$ and a space $K'$ of $L(R)$
such that $T_1, T_i$ is $K'$-matched, for all $i\in[k']$. We show in this section that
if $T_1,\ldots,T_{k'}$ are linearly independent (i.e. $\nexists$
$\vec{\beta}\in\mathbb{F}^{k'}\setminus\{\vec{0}\}$ s.t. $\sum_{i\in[k']}\beta_iT_i=0$) then
$K'$ can be extended to a linear space $K$ of rank at most $(\text{rk}(K')+k'^2)$ such that:
$M(L_K(T_1)),\ldots,M(L_K(T_{k'}))$ are also linearly independent. This will prove
Theorem \ref{thm-nucleus}.
\smallskip\noindent
{\bf Theorem \ref{thm-nucleus}} (Nucleus).
Let $C=\sum_{i\in[k]}T_i$ be a minimal $\Sigma\Pi\Sigma(k,d)$ identity and let $\{T_i|i\in\mathcal{I}\}$ be
a maximal set of linearly
independent terms ($1\leqslant k':=|\mathcal{I}|<k$). Then there exists a linear subspace $K$ of $L(R)$ such that:
\begin{itemize}
\item[1)] $\text{rk}(K)< 2k^2$.
\item[2)] $\forall i\in[k]$, there is a $K$-matching $\pi_i$ between $T_1, T_i$.
\item[3)] (Define $\forall i\in\mathcal{I}$, $K_i:=M(L_K(T_i))$.) The terms $\{K_i|i\in\mathcal{I}\}$
are linearly independent.
\end{itemize}
\begin{proof}
For convenience, and wlog, assume $\mathcal{I}=[k']$.
The proof is an iterative process with at most $k'^2$ iterations, and gradually builds the
promised space $K$. Each iteration of the process maintains a space $U$ of $L(R)$ which
is intended to grow at each step and bring us closer to $K$. For convenience, define
$U_i:=M(L_U(T_i))$, for all $i\in[k']$. Also for each $i\in\{2,\ldots,k'\}$, define
ideal $\mathcal{I}_i:=\ideal{U_1,\ldots,U_{i-1}}$.
The process has two nested iterations, or phrased
differently,
a double induction. We will call the outer ``loop" a {\em phase}, and the inner loop a
{\em round}.
In each round the rank of $U$ increases by at most $1$, and the $i$-th phase has at most
$i$ rounds.
At the end of the $i$-th phase ($i\geq2$), we will ensure $T_i\notin\mathcal{I}_i$.
(Remark: By Lemma \ref{lem-non-zd} this is equivalent to ensuring $U_i\notin\mathcal{I}_i$,
which by Lemma \ref{lem-homo-ideal} means that $U_i$ is linearly independent of
$U_1,\ldots,U_{i-1}$.)
In the first phase we set $U:=K'$, where $K'$ is the matching-nucleus obtained by applying
Theorem \ref{thm-mat-nucleus} on $C$. This immediately gives us property (2) promised
in the theorem statement, i.e. the matching property. Also, $\text{rk}(U)<k^2$ at the end of
the first phase.
Now the second phase. As $T_1, T_2$ are linearly independent, we get, by Lemma
\ref{lem-homo-ideal}, that $T_2\notin\ideal{ T_1 }$. By an application of Corollary
\ref{cor-crt}, $\exists v\in\text{nod}_{\ideal{0}}(T_1)$ such that $T_2\notin\ideal{v}$.
We update $U\leftarrow(U+\text{radsp}(v))$. Note that after updation
$T_2\notin\ideal{U_1}=\mathcal{I}_2$ (otherwise $T_2\in\ideal{U_1}\subseteq\ideal{v}$, since $v|U_1$).
Now, for the $i>2$ phase. Inductively, we assume that $\forall r<i$, $T_r\notin\mathcal{I}_r$
(remember that all these ideals are wrt the {\em current} $U$).
The phase consists of various rounds. At the end of the $j$-th round ($1\leqslant j<i$),
we just want to ensure $T_i\notin\ideal{U_1,\ldots,U_j,T_{j+1},\cdots,T_{i-1}}$.
So we do nothing in the $j$-th round unless this is violated. What do we do when it is violated?
\begin{claim}\label{clm-extend}
Let $i>2$ and $1\leqslant j<i$. Suppose $\forall r<i$,
$T_r\notin\ideal{U_1,\ldots,U_{r-1}}$.
Suppose $T_i\in\ideal{U_1,\cdots,U_j,T_{j+1},\cdots,T_{i-1}}$
but $T_i\notin \ideal{U_1,\cdots,U_{j-1},T_j,\cdots,T_{i-1}}$. There exists a
$v\in\text{nod}_{\ideal{U_1,\cdots,U_{j-1}}}(T_j)$
such that for the updated $U'\leftarrow(U+\text{radsp}(v))$ we have
$T_i\notin\ideal{U_1',\cdots,U_j',T_{j+1},\cdots,T_{i-1}}$.
\end{claim}
\claimproof{clm-extend}{
Since $T_i \in \ideal{U_1,\cdots,U_j,T_{j+1},\cdots,T_{i-1}}$, by Lemma \ref{lem-homo-ideal},
we get
$T_i + \sum_{r=j+1}^{i-1} \alpha_r T_r \in\ideal{U_1,\cdots,U_j}$ for some $\alpha_r$-s in $\mathbb{F}$.
Suppose there are two
distinct choices for $\alpha_r$-s (we will call them $\alpha_r$ and $\alpha'_r$).
Then,
$$\left(T_i + \sum_{r=j+1}^{i-1} \alpha_r T_r\right),\
\left(T_i + \sum_{r=j+1}^{i-1} \alpha'_r T_r\right) \in \ideal{U_1,\cdots,U_j}.$$
Subtracting,
we get $\sum_{r=j+1}^{i-1} (\alpha - \alpha'_r) T_r \in \ideal{U_1,\cdots,U_j}$.
Let $s$ be the largest index such that $\alpha_s - \alpha'_s \neq 0$.
(By the distinctness of the sequences, such an index exists.)
We get that
$T_{s} \in \ideal{U_1,\cdots,U_j,T_{j+1},\cdots,T_{s-1}}$ $\subseteq \ideal{U_1,\cdots,U_{s-1}}$.
Since $s \leq i-1$, this contradicts the hypothesis. Hence, the sequence $\{\alpha_r\}$ is unique.
The claim hypothesis says that $T_i\notin \ideal{U_1,\cdots,U_{j-1},T_j,\cdots,T_{i-1}}$.
That implies
$T_i + \sum_{r=j+1}^{i-1} \alpha_r T_r$ $\notin \ideal{U_1,\cdots,U_{j-1},T_j}$.
Thus, by Corollary \ref{cor-crt}, $\exists v\in\text{nod}_{\ideal{U_1,\cdots,U_{j-1}}}(T_j)$
such that
$T_i + \sum_{r=j+1}^{i-1} \alpha_r T_r$ $\notin \ideal{U_1,\cdots,U_{j-1},v}$.
Let us update $U$ to $U'\leftarrow(U+\text{radsp}(v))$. (This updates $U_r$-s to $U_r'$-s.)
We now argue that $T_i \notin \ideal{U'_1,\ldots,U'_j,T_{j+1},\cdots,T_{i-1}}$.
Suppose not. Then, by Lemma \ref{lem-homo-ideal}, for some sequence $\beta_r$,
$T_i + \sum_{r=j+1}^{i-1} \beta_r T_r \in \ideal{U'_1,\ldots,U'_j}$
$\subseteq\ideal{U_1,\ldots,U_j}$ (since for all $r$, $U_r|U'_r$).
By the uniqueness of $\{\alpha_r\}$, we have $\beta_r = \alpha_r$, for all $r$.
But that implies $T_i + \sum_{r=j+1}^{i-1} \alpha_r T_r$ $\in \ideal{U'_1,\ldots,U'_j}$
$\subseteq \ideal{U_1,\cdots,U_{j-1},v}$.
This is a contradiction and hence completes the proof.
Let us look at the first round (i.e. $j=1$).
Suppose $T_i \notin \ideal{U_1,T_2,\cdots,T_{i-1}}$. Then, we move directly to the
second round, since we have already satisfied the round invariant.
Otherwise, $T_i\in\ideal{U_1,T_2,\cdots,T_{i-1}}$. Furthermore, by linear independence
and Lemma \ref{lem-homo-ideal}, we have $T_i \notin \ideal{T_1,\cdots,T_{i-1}}$, so
we can invoke Claim \ref{clm-extend} to get a $v\in\text{nod}_{\ideal{0}}(T_1)$. This allows us
to update $U\leftarrow(U+\text{radsp}(v))$ such that
$T_i \notin \ideal{U_1,T_2,\cdots,T_{i-1}}$.
Now for the induction step.
We assume that, by the end of the $(j-1)$th round,
$T_i \notin \ideal{U_1,\cdots,U_{j-1},T_{j},\cdots,T_{i-1}}$.
For the $j$-th round, either we would have to do nothing or have to apply Claim
\ref{clm-extend} and update $U$. In either case, $\text{rk}(U)$ increases by at most $1$.
At the end of the round, $T_i \notin \ideal{U_1,\cdots,U_{j},T_{j+1},\cdots,T_{i-1}}$.
This continues till $j = i-1$. We finally
have $T_i \notin \ideal{U_1,\cdots,U_{i-1}}$ $=\mathcal{I}_i$,
giving us the required invariant for the $i$-th phase.
This completes the proof.
\end{proof}
\section{Invoking Sylvester-Gallai Theorems: The Final Rank Bound}
\label{sec-final-rank}
In this section we will bound the non-nucleus rank of a simple, minimal $\Sigma\Pi\Sigma(k,d)$, independent-fanin $k'$,
identity $C$ by $(k-k')\cdot\text{SG}_{k'}(\mathbb{F},d)$. Thus, proving Theorem \ref{thm-rank}. We
divide the proof into two subsections. First, we bound the non-nucleus rank of a
simple, {\em strongly} minimal $\Sigma\Pi\Sigma(k,d)$ identity $C$ by $\text{SG}_{k-1}(\mathbb{F},d)$, finishing
the proof of Theorem \ref{thm-strong-rank}. Second, we show how to repeatedly use this
result on a simple, minimal but not strongly-minimal identity.
\subsection{The strongly minimal case}\label{sec-strong}
Assume that $C:=\sum_{i\in[k]}T_i$ is a simple, strongly minimal $\Sigma\Pi\Sigma(k,d)$ identity
(recall: then $T_1,\ldots,T_{k-1}$ are linearly independent polynomials).
Let $K$ be its nucleus given by Theorem \ref{thm-nucleus}.
There are two important properties of this nucleus that we restate (and elaborate upon) for emphasis.
The first is the \emph{matching property}.
For any $i\in[k]$, $L^c_K(T_1)$ $(= L(T_1)\setminus K)$ is $K$-matched to $L^c_K(T_i)$
$( = L(T_i)\setminus K)$.
In other words for any $\ell\in L^c_K(T_1)$, the degrees of
$M(L^c_K(T_1)\cap(\mathbb{F}^*\ell+K))$
and $M(L^c_K(T_i)\cap(\mathbb{F}^*\ell+K))$ are equal (remark: they are polynomials in $\text{nod}_K(T_1)$ and
$\text{nod}_K(T_i)$ respectively). This observation motivates the following definition.
\begin{definition}[Family]
Let $C$ be a $\Sigma\Pi\Sigma(k,d)$ identity and $K$ be its nucleus. Let $\ell\in L^c_K(C)$.
The {\em family of $\ell$} is defined to be the list, $\text{fam}(\ell):=$
$\{M(L^c_K(T_i)\cap(\mathbb{F}^*\ell+K))\ |\ i\in[k]\}$. Note that $\text{fam}(\ell)$ is a multiset of size
exactly $k$, having equal degree polynomials corresponding to each term $T_i$, we fix this
ordering on the list (i.e. $i$-th element in $\text{fam}(\ell)$ corresponds \& divides the multiplication
term $T_i$).
\end{definition}
Verify that any two forms in $L^c_K(C)$ that are ``similar mod $K$" have the same families.
\smallskip\noindent{\bf [Partition, Class, Split \& Preserve]}
Let us focus on a list $\text{fam}(\ell)$. The equivalence relation of similarity (i.e. mod
$\ideal{0}$) on $\text{fam}(\ell)$, induces a {\em partition} of $[k]$ (i.e. if $f_i,f_j\in\text{fam}(\ell)$
are similar then place
$i$ and $j$ in the same partition-class). Denote this partition induced on $[k]$, by $\text{Part}(\ell)$.
Observe that $\text{Part}(\ell)$ must contain at least $2$ classes (otherwise simplicity of $C$ is violated).
Each set in this partition is called a \emph{class}, and we naturally have
a class $\text{cl}(f)$ associated with each member of $f \in \text{fam}(\ell)$.
We say that $\text{Part}(\ell)$ \emph{splits} a subset $S \subseteq [k]$
if there is some class $X\in\text{Part}(\ell)$ such that $X \cap S \neq \emptyset, S$.
Otherwise, we say that $\text{Part}(\ell)$ {\em preserves} $S$. Note that a singleton is always
preserved.
For classes $A_1\in\text{Part}(\ell_1)$ and $A_2\in\text{Part}(\ell_2)$, the \emph{complement} $\vec{A_1\cup A_2}$ is just
the set
$[k] \setminus(A_1\cup A_2)$. We will be later interested in the properties of this complement set wrt the two
partitions.
\smallskip\noindent
The second property of the nucleus, the \emph{linear independence}, says something technical
about the nucleus identity. By definition $K_i=M(L_K(T_i))$, for all $i\in[k]$, and by Lemma
\ref{lem-nucleus} : $\sum_{i\in[k]}\alpha_iK_i=0$ for some $\alpha_i\text{-s}\in\mathbb{F}^*$.
Furthermore,
\begin{claim} \label{clm-kmin} For $1<r<k$, let $\{s_1,\cdots,s_r\}$ be
a subset $S \subsetneq [k]$, where $s_1 < s_2 < \cdots < s_r$. Then $K_{s_r}\notin\ideal{K_{s_1},\cdots,K_{s_{r-1}}}$.
\end{claim}
\begin{proof} If $s_r < k$, then this just holds from the linear independence of
$\{K_1,\ldots,K_{k-1}\}$ and Lemma \ref{lem-homo-ideal}.
So, we can assume $s_r = k$ and $K_k\in\ideal{K_{s_1},\cdots,K_{s_{r-1}}}$.
By Lemma \ref{lem-homo-ideal}, this
means $K_k = \sum_{i\in[r-1]} \beta_{s_i} K_{s_i}$ for some $\beta\text{-s}\in\mathbb{F}$.
The nucleus identity gives us $K_k = -\sum_{i\in[k-1]}\frac{\alpha_i}{\alpha_k}K_i$
$=\sum_{i\in[r-1]}\beta_{s_i}K_{s_i}$.
Since $r<k$, this implies that for some $\gamma$-s in $\mathbb{F}$, not all
zero, $\sum_{i\in[k-1]}\gamma_i K_i = 0$. This contradicts the linear independence of
$\{K_1,\ldots,K_{k-1}\}$, finishing the proof.
\end{proof}
Before applying Sylvester-Gallai-type theorems (i.e. the $\text{SG}_{k-1}$ operator) we emphasize
that, as discussed in Section \ref{sec-step-3}, there is a distinguished linear form
$y_0\in L(R)^*$ and a subspace $U$ of $L(R)$ such that $L(R)=\mathbb{F} y_0\oplus U\oplus K$
and every form in $L^c_K(C)$ is monic wrt $y_0$. Thus, for every $\ell\in L^c_K(C)$
there exists a unique way to express : $\ell=\alpha y_0+u+v$ ($\alpha\in\mathbb{F}^*$, $u\in U$
and $v\in K$). This allows us to define the truncation operator : $\text{trun}(\ell)=y_0+$
$\alpha^{-1}u$.
\begin{lemma}[Partitions from $\text{SG}_{k-1}$-tuple]\label{lem-split}
Suppose $\text{rk}(\text{trun}(L^c_K(T_1)))$ $>\text{SG}_{k-1}(\mathbb{F},d)$, and
$\text{SG}_{k-1}(\text{trun}(L^c_K(T_1)))$ gives the set $\{\ell_1, \ell_2,\cdots,\ell_{k-1}\}$.
For all $i\in[k-1]$, let $\ell_i'\in L^c_K(T_1)$ be a form satisfying
$\text{trun}(\ell_i')=\ell_i$.
Let $\mathcal{I}\subseteq[k-1]$ be nonempty, and $A_i$ be any class in $\text{Part}(\ell_i' )$ for all $i\in\mathcal{I}$. Suppose
$S:=\overline{\bigcup_{i\in\mathcal{I}}A_i}$ $\ne\emptyset$. Then $S$ is split
by $\text{Part}(\ell_c')$, for {\em some} $c\in\mathcal{I}$.
\end{lemma}
\begin{proof}
We prove by contradiction.
Suppose $S$ is preserved by $\text{Part}(\ell_i')$, for all $i\in\mathcal{I}$.
Since for all $i\in\mathcal{I}$, $A_i\in\text{Part}(\ell_i')$, by definition there exists an
$f_i\in\text{fam}(\ell_i')$ such that $A_i=\text{cl}(f_i)$. Similarly, for all $i\in\mathcal{I}$,
there exists a $g_i\in\text{fam}(\ell_i')$ such that $S\subseteq\text{cl}(g_i)$. Note that,
by definition, sets $A_i$ and $S$ are disjoint, hence the classes $\text{cl}(f_i)$ and
$\text{cl}(g_i)$ are different, implying $f_i, g_i$ are {\em not} similar, for all $i\in\mathcal{I}$.
Define ideal $I:=\ideal{f_i | i\in\mathcal{I}}$.
Let us focus on the sub-circuit $C_S = \sum_{j \in S} T_j$.
Since $C=0$ and $S=\overline{\bigcup_{i\in\mathcal{I}}\text{cl}(f_i)}$, we deduce $C_S\in I$ (as
$f_i$ ``kills" the term $T_r$ for all $r\in\text{cl}(f_i)$, and ``spares" the other terms).
For all $i\in\mathcal{I}$, $S\subseteq\text{cl}(g_i)$ we deduce that :
$\prod_{i\in\mathcal{I}}g_i$ divides $T_j$, for all $j\in S$. So $T_j':=T_j/(\prod_{i\in\mathcal{I}}$ $g_i)$
is again a multiplication term
with none of its form in $\bigcup_{i\in\mathcal{I}}(\mathbb{F}^*\ell_i'+K)=$ $\bigcup_{i\in\mathcal{I}}(\mathbb{F}^*\ell_i+K)$.
Thus, we get an important equation:
$$C_S\ =\ \(\prod_{i\in\mathcal{I}}g_i\)\cdot\(\sum_{j\in S}T_j'\)\ \in\ \ideal{f_i\ |\ i\in\mathcal{I}}. $$
By a repeated application of Lemma \ref{lem-cancel} on the above system, we get :
\begin{equation}\label{eqn-lem-split}
\sum_{j\in S}T_j' \in \ideal{f_i' \ |\ i\in\mathcal{I}}=:I',
\text{ where, } f_i':=\frac{f_i}{\gcd(f_i,g_i)}, \forall i\in\mathcal{I}.
\end{equation}
Since $f_i, g_i$ are not similar, $f_i'$ has degree $\geq1$, for all $i\in\mathcal{I}$.
Let the elements of $S$ be $s_1 < s_2 < \cdots < s_r$, for some $r\in[k-1]$. Since we have only
changed the non-nucleus part of $T_j$ to get $T_j'$, we deduce $K_{s_i}|T_{s_i}'$, for all $i\in[r]$.
Thus, modulo the ideal $I'':=\ideal{I',K_{s_1},\cdots,K_{s_{r-1}}}$, Equation (\ref{eqn-lem-split})
becomes : $T_{s_r}'\in I''$.
We have $\text{radsp}(I'')\subseteq\text{sp}(\ell_i\ |\ i\in\mathcal{I})+K$.
Let us factor $T_{s_r}' = B_0B_1$, where $B_0$ is the product of all forms in $\text{radsp}(I'')$
and $B_1$ is the remaining product. Thus, $B_0B_1\in I''$. By Lemma \ref{lem-non-zd}, $B_1$ can be
cancelled out and we get $B_0\in I''$.
Suppose all forms of $B_0$ are in $K$, so $B_0 = K_{s_r}$. This means $K_{s_r} \in I''$ implying,
\begin{equation}\label{eqn-lem-split-2}
K_{s_r}\in\ideal{K_{s_1},\cdots,K_{s_{r-1}}, \{f_i' \ |\ i\in\mathcal{I}\}}.
\end{equation}
Recall that each form in $f_i'$ is similar to some form in $(\mathbb{F}^*\ell_i+K)$, for all $i\in\mathcal{I}$.
Suppose form $(\beta_i\ell_i+u_i)|f_i'$, for all $i\in\mathcal{I}$, for some $\beta$-s in $\mathbb{F}^*$ and
$u$-s in $K$. In Equation (\ref{eqn-lem-split-2}) make the {\em evaluation} : $\ell_i\leftarrow$
$-\beta_i^{-1}u_i$, for all $i\in\mathcal{I}$. This is a valid evaluation since
$\{\ell_i \ |\ i\in\mathcal{I}\}$ are linearly independent mod $K$, and values substituted are from $K$.
Clearly, this evaluation leaves the polynomial $K_s$ ($s\in S$) unchanged. Thus, we get
$K_{s_r}\in\ideal{K_{s_1},\cdots,K_{s_{r-1}}}$, contradicting Claim \ref{clm-kmin}.
As a result, we have a form $\ell | B_0$ such that $\ell \notin K$.
We have $\ell \in \text{radsp}(I'')$ $\subseteq\text{sp}(\ell_i \ |\ i\in\mathcal{I})+K$, and
by the way $T_{s_r}'$ was defined,
$\ell \notin \bigcup_{i\in\mathcal{I}}(\mathbb{F}^*\ell_i + K)$.
By the matching property of the nucleus, this gives us an $\ell' \in L^c_K(T_1)$ such that :
$\ell'\in(\text{sp}(\ell_i | i\in\mathcal{I})+K)\setminus K$ and
$\ell'\notin \bigcup_{i\in\mathcal{I}}(\mathbb{F}^*\ell_i + K)$.
This means that there exist constants $\beta_i$-s in $\mathbb{F}$, not all zero, such that
$\ell'\in \sum_{i\in\mathcal{I}}\beta_i\ell_i+K$. As the coefficient of $y_0$ in $\ell'$ is nonzero while
that in $\ell_i$ ($i\in\mathcal{I}$) is $1$, we deduce :
$\text{trun}(\ell') =(\sum_{i\in\mathcal{I}} \beta_i)^{-1}(\sum_{i\in\mathcal{I}}\beta_i\ell_i)$.
If exactly one $\beta_i$ is nonzero, then $\ell'\in(\mathbb{F}^*\ell_i + K)$, which is a contradiction.
So at least two $\beta_i$-s are nonzero, implying that $\text{trun}(\ell')\in\text{trun}(L^c_K(T_1))$
is a non-trivial combination of the $\ell_i$-s, contradicting the
fact that $\{\ell_1,\cdots,\ell_{k-1}\}$ were obtained from $\text{SG}_{k-1}(\text{trun}(L^c_K(T_1)))$.
This contradiction proves that $S$ is split by $\text{Part}(\ell_i')$, for some $i\in\mathcal{I}$.
\end{proof}
To prove Theorem \ref{thm-strong-rank}, we need a combinatorial lemma
about general partitions. It is helpful to abstract out some of the
details specific to identities and frame this as a purely combinatorial problem.
Since the proof is fairly involved, we present
that in the next subsection. For now, we give the necessary
definitions and claims.
We have a universe $\mathcal{U} := [k]$ of elements. We deal with set systems
with special properties.
\begin{definition}[Unbroken chain]\label{def-unbroken}
A partition of $\mathcal{U}$ is \emph{trivial} if it contains the single set $\mathcal{U}$.
Let $\mathfrak{P}$ be a collection of non-trivial partitions of $\mathcal{U}$ (here a collection refers to a {\em multiset},
i.e. $\mathfrak{P}$ can have partitions repeated). A {\em chain in $\mathfrak{P}$} is
a sequence of sets $A_1,A_2,\cdots,A_s$ (for some $s$) such that each set comes
from a different element of $\mathfrak{P}$ (say $A_i \in \mathcal{P}_i \in \mathfrak{P}$).
The chain $A_1,A_2,\cdots,A_s$ is an \emph{unbroken chain}, if
$\overline{\bigcup_{i \in[s]} A_i}$ is non-empty and preserved in $\mathcal{P}_i$, for each $i\in[s]$.
\end{definition}
Note that if $\overline{\bigcup_{i \leq s} A_i}$ is a singleton then
it is trivially preserved in any partition, therefore, such a chain would be unbroken.
By Lemma~\ref{lem-split}, the collection $\{\text{Part}(\ell_i')|i\in[k-1]\}$ has
no unbroken chain. By purely studying partitions, we will show that such a phenomenon is absurd.
The following combinatorial lemma
implies Theorem~\ref{thm-strong-rank}.
\begin{lemma}[Partitions have unbroken chain]\label{lem-unbroken}
Let $\mathfrak{P}$ be a collection of
non-trivial partitions of $\mathcal{U}$. If $\mathfrak{P}$ contains at least $|\mathcal{U}|-1$ partitions then $\mathfrak{P}$
contains an unbroken chain.
\end{lemma}
\smallskip\noindent
{\bf Theorem \ref{thm-strong-rank}.} (Bound for simple, strongly minimal identities).
Let $|\mathbb{F}|>d$. The non-nucleus rank of
a simple and strongly minimal $\Sigma\Pi\Sigma(k,d)$ identity over $\mathbb{F}$ is at most $\text{SG}_{k-1}(\mathbb{F},d)$.
\begin{proof} (of Theorem~\ref{thm-strong-rank})
Let $C=\sum_{i\in[k]}T_i$ be a simple and strongly minimal $\Sigma\Pi\Sigma(k,d)$ identity over $\mathbb{F}$,
and let $K$ be the nucleus provided by Theorem \ref{thm-nucleus}. As $|\mathbb{F}|>d$ we can assume (wlog by Lemma
\ref{lem-monic-forms}) the existence of a truncation operator on $L^c_K(T_1)$.
We will show that the rank of $\text{trun}(L^c_K(T_1))$ is at most $\text{SG}_{k-1}(\mathbb{F},d)$.
By the matching property of the nucleus, $\text{trun}(L^c_K(T_1))$ together
with $K$ span $L(C)$. Therefore, a non-nucleus rank bound of the former suffices
to bound the non-nucleus rank of $L(C)$.
Assuming that the rank of $\text{trun}(L^c_K(T_1))$
is greater than $\text{SG}_{k-1}(\mathbb{F},d)$, as in Lemma~\ref{lem-split},
we invoke $\text{SG}_{k-1}(\text{trun}(L^c_K(T_1)))$ to get $\{\ell_1,\ell_2,\cdots,\ell_{k-1}\}$.
Associated with each of these, we have the partition $\text{Part}(\ell_i')$.
There are $k-1$ partitions in the collection $\mathfrak{P}:=$ $\{\text{Part}(\ell_i')|i\in[k-1]\}$, which are all non-trivial by the
simplicity of $C$. Lemma~\ref{lem-unbroken} tells us that $\mathfrak{P}$ has an unbroken chain, while
Lemma~\ref{lem-split} says that $\mathfrak{P}$ has none. This contradiction implies the rank of
$\text{trun}(L^c_K(T_1))$ is at most $\text{SG}_{k-1}(\mathbb{F},d)$, thus finishing the proof.
\end{proof}
\subsubsection{The combinatorial proof of Lemma \ref{lem-unbroken}}
Intuitively, when the partitions in $\mathfrak{P}$ have many classes then an unbroken chain should be easy to
find, for example, when $(k-1)$ partitions in $\mathfrak{P}$ are all equal to $\{\{1\},\ldots,\{k\}\}$ then there is an easy
unbroken chain, namely $\{1\},\ldots,\{k-1\}$. On the other hand, when the partitions in $\mathfrak{P}$ contain
few classes then we can effectively decrease the universe and apply induction. Most of this subsection would deal
with the former case. Let us first define the splitting property.
\begin{definition}[Splitting property]
Let $\mathfrak{P}$ be a collection of partitions of $\mathcal{U}$.
Suppose for all non-empty $S \subset \mathcal{U}$, $S$ is split by at least $(|S|-1)$ partitions of $\mathfrak{P}$.
Then $\mathfrak{P}$ is said to have the \emph{splitting property}.
\end{definition}
\begin{claim} \label{clm-singleton}
Let $\mathfrak{P}$ be a collection of at least $(k-1)$ non-trivial partitions of $[k]$.
If $\mathfrak{P}$ has the splitting property then there is a chain $A_1,\cdots,A_{k-1}$ in $\mathfrak{P}$ such that
$\overline{\bigcup_{i \leq k-1} A_i} = \{k\}$. (In particular, $\mathfrak{P}$ has an unbroken chain.)
\end{claim}
We defer its proof and, instead, first show why this claim would suffice.
\begin{proof} (of Lemma~\ref{lem-unbroken})
We will prove this by induction
on the universe size $k$. For the base case, suppose $k=3$ and $\mathfrak{P}=$ $\{\mathcal{P}_1,\mathcal{P}_2,\ldots\}$.
So we have at least two partitions. If any partition (say $\mathcal{P}_1$) contains exactly $2$ sets,
it must be a pair and a singleton (say $\mathcal{P}_1=\{\{1,2\},\{3\}\}$).
But then $\{1,2\}$ is itself an unbroken chain in $\mathfrak{P}$. So, all the partitions
can be assumed to consist only of singletons. But then we can take the set, say, $\{1\}$
from $\mathcal{P}_1$ and, say, $\{2\}$ from $\mathcal{P}_2$ to get an unbroken chain.
Now for the induction step. Suppose $\mathfrak{P}$ has at least $(k-1)$ partitions.
We assume that the claim is true for
universes of size upto $(k-1)$. If $\mathfrak{P}$ has the splitting property,
then we are done by Claim~\ref{clm-singleton}. If not, then for some subset $S \subset\mathcal{U}$ of
size at least $2$, $S$ is split in at most $(|S|-2)$ partitions. Let the collection of partitions
in $\mathfrak{P}$ that preserve $S$ be ${\bf\mathfrak{P}'}$. So ${\bf \mathfrak{P}'}$ contains
at least $(k-1)-(|S|-2)$ $=(k-|S|+1)$ partitions. Merge the elements
of $S$ into a new element, to get a new universe $\mathcal{U}'$ of size $(k-|S|+1)$.
The partitions in ${\bf \mathfrak{P}'}$ are valid partitions of $\mathcal{U}'$, and still maintain their structure.
We now have a universe of size $k-|S|+1<k$, and at least $k-|S|+1$ partitions.
By the induction hypothesis, there is an unbroken chain in ${\bf \mathfrak{P}'}$.
Observe that it is (under the natural correspondence) still an unbroken chain in the original collection
$\mathfrak{P}$, and we are done.
\end{proof}
\begin{proof} (of Claim~\ref{clm-singleton})
We will label the partitions in $\mathfrak{P}$ in such a way that its first $(k-1)$ elements,
$\mathcal{P}_1,\cdots,\mathcal{P}_{k-1}$ satisfy : $\mathcal{P}_{i}$ splits $\{i,k\}$, for all $i\in[k-1]$. Thus,
there is a set $A_{i} \in \mathcal{P}_{i}$
that contains $i$ but not $k$. Naturally, $\overline{\bigcup_{i \leq k} A_i} = \{k\}$.
We will construct this labelling through an iterative process.
In the $i$th phase, we will
find $\mathcal{P}_i$. At the end of this phase, we will
have $\mathcal{P}_1,\cdots,\mathcal{P}_{i}$ with the desired property
and the remaining {\em pool} $\mathfrak{P}$ of remaining partitions.
We warn the reader that this labelling is very dynamic, so
during the $i$th phase, we may change the labels
of $\mathcal{P}_1,\cdots,\mathcal{P}_{i-1}$ by moving them to $\mathfrak{P}$
and labelling new partitions with older labels.
At any stage, we have the \emph{labelled}
partitions and the {\em unlabelled} partitions $\mathfrak{P}$.
Before the beginning of the first phase, $\mathfrak{P}$ is just the given collection
of all permutations.
{\bf [Phase $1$]}
The first phase is easy to understand.
By the splitting property,
there is some partition that splits $\{1,k\}$.
We set this to $\mathcal{P}_1$.
{\bf [Phase $i$]}
The $i$th phase, $i\geq2$, is a rather involved process. We describe the
various sets associated with it and explain them. By the beginning of this phase,
we have already determined $\mathcal{P}_1,\cdots,\mathcal{P}_{i-1}$. The \emph{covered}
elements are just $[i-1]$. We maintain
a partition $E_1,\cdots,E_{i-1}$ of the covered elements.
We set $E_0 = \{i,k\}$.
Corresponding to each set $E_j$, we have a set
of partitions $\mathcal{C}_j$ ($:=\{\mathcal{P}_b|b\in E_j\}$). We fix $\mathcal{C}_0=\emptyset$.
Note that the $\mathcal{C}_j$'s form a partition of the labelled partitions.
We set $E_{\leq j} = \bigcup_{0 \leq l \leq j} E_l$.
We get a similar set of partitions $\mathcal{C}_{\leq j} = \bigcup_{0 \leq l \leq j} \mathcal{C}_l$.
We will \emph{always} maintain that $\mathcal{P}_j$ splits
$\{j,k\}$.
There will be various rounds in a phase.
To aid understanding, we will describe the first
and second round in detail.
{\bf [Phase $i$, Round $1$]}
We now explain the first round.
Initially, we set $E_1 = [i-1]$.
This is because, as of now, we know nothing
about the elements in the set containing $k$ in
the various $\mathcal{P}_j$'s.
Note that \emph{at any stage}, if we have
a partition in $\mathfrak{P}$ that splits $\{i,k\}$,
we can set this to $\mathcal{P}_i$ and we are done.
Now, for every element $b$ in $E_1$ (currently, it is $[i-1]$)
check if there is a partition in $\mathfrak{P}$ that
splits $\{i,k,b\}$. If so, call this a {\em success} for $b$.
Note that $\{i,k\}$ is not split in any partition $\mathfrak{P}$.
Now, we can label \emph{this}
partition as $\mathcal{P}_b$ and move the old one to the pool $\mathfrak{P}$.
All the labelled partitions still have their desired property.
If the old partition splits $\{i,k\}$, then we
are done (since this is now in $\mathfrak{P}$). So, we can assume
that (even after this switching) that $\{i,k\}$ is preserved
in all of $\mathfrak{P}$.
For our (new) $\mathcal{P}_b$, we know that $\{i,k\}$
is preserved. So we have some extra information about it.
This is represented by ``promoting" $b$ from $E_1$ to $E_2$.
This just involves removing $b$ from $E_1$ and putting it in $E_2$.
Let us repeat this for all elements in $E_1$ until
we have a maximal set $E_2$, containing all successful
elements. Note that when $\mathfrak{P}$ changes because of the
switching, we check all elements in $E_1$
again for successes.
We are now at the end of this round and have the following information.
$\{E_1, E_2\}$ is a partition of $[i-1]$.
For any successful element $b \in E_2$, $\mathcal{P}_b$ preserves
$\{i,k\}$. So, the labelled partitions $\mathcal{C}_2$ preserve
$\{i,k\} = E_0$.
For every failure $b \in E_1$, \emph{every} partition in $\mathfrak{P}$
preserves $\{i,k,b\}$. In other words,
every partition in $\mathfrak{P}$ preserves $E_1 \cup \{i,k\} = E_{\leq 1}$.
Successes create the new $E_2$, while failures increase the size of the
set preserved by $\mathfrak{P}$.
Let us understand this a little more.
Suppose \emph{all} elements are eventually successful, so $E_2 = [i-1]$.
Therefore, all labelled partitions preserve $\{i,k\}$. But so
do all partitions in $\mathfrak{P}$.
So $\{i,k\}$ is preserved in \emph{all} partitions,
contradicting the splitting property. There must
be some failures.
Suppose everything is a failure, so $E_1$ is still $[i-1]$.
The set $E_{\leq 1}$ has size $i+1$. But the only partitions that split $E_{\leq 1}$
are the labelled ones since $\mathfrak{P}$ preserves $E_{\leq 1}$.
There are only $i-1$ labelled partitions so this contradicts
the splitting property. So there are some successes and some
failures and $E_1, E_2$ form a non-trivial partition. In some
sense, we made ``progress".
{\bf [Phase $i$, Round $2$]}
We move to the next round. For every $b \in E_2$, we check
if $E_{\leq 1} \cup \{b\}$ is split in any partition of $\mathfrak{P}$. If we get
a success, then we set this partition to be the new $\mathcal{P}_b$.
We ``promote" $b$ from $E_2$ to a new set $E_3$.
We need to shift this old partition (call it $\mathcal{P}$)
to our pool $\mathfrak{P}$. But we want to ensure that $E_{\leq 1}$ is preserved in all of $\mathfrak{P}$,
and this may not happen for $\mathcal{P}$. So, first we check if $\{i,k\}$
is preserved in $\mathcal{P}$. If not, we are done. Assume otherwise. We start
checking if $\{i,k,c\}$ is preserved, for all $c \in E_1$.
If it is so for all $c$, then we know that $E_{\leq 1}$ is preserved in
$\mathcal{P}$. So, we maintain our condition about $\mathfrak{P}$, and we continue
to the next $b$. If not (this is the interesting part!),
then we have found a partition that separates $c$ from $\{i,k\}$.
Note that the reason why $c$ belongs to $E_1$, is, because we were unable
(in the previous round) to find such a partition. So, we label
$\mathcal{P}$ as $\mathcal{P}_c$. We ``promote" $c$ from $E_1$ to $E_2$.
The old $\mathcal{P}_c$ is moved to the pool $\mathfrak{P}$, so we repeat the
above procedure for this partition as well. So, either we maintain
the invariant that $E_{\leq 1}$ is preserved in all of $\mathfrak{P}$, or we promote
elements from $E_1$ to $E_2$. If, at some stage, there are no elements
in $E_1$, then we are done. (Why? Because every labelled partition now
preserves $\{i,k\}$, by the splitting property, there must be a partition in $\mathfrak{P}$ splitting
this.) For all the failures $b \in E_1$, we know that $E_{\leq 1} \cup \{b\}$ is preserved
in $\mathfrak{P}$. All successes are promoted to $E_2$. So at the end we have
the partition $E_1, E_2, E_3$ of $[i-1]$. All of $\mathfrak{P}$
preserves $E_{\leq 2}$. The partitions $\mathcal{C}_3$ preserve $E_{\leq 1}$ and
those in $\mathcal{C}_2$ preserve $E_{\leq 0}$. If $E_3$ is empty,
then $E_{\leq 2}$ is of size $i+1$.
All of $\mathfrak{P}$ preserves $E_{\leq 2}$ so the splitting property
is violated. If $E_1$ is empty, then also
we are done.
Let us give a formal proof by describing the invariant at the end
of a round.
\begin{claim} \label{clm-round} By the end of the ($i$th phase-) $j$th round,
suppose we do not find the right $\mathcal{P}_i$.
Then we can construct a partition of $[i-1]$, $E_1,\cdots,E_{j+1}$,
where $E_{j+1}$ is non-empty and the following hold:
the partitions $\mathcal{C}_l$ preserve $E_{\leq l-2}$, for all $2\leqslant l\leqslant(j+1)$, and the unlabelled
partitions $\mathfrak{P}$ preserve $E_{\leq j}$.
\end{claim}
\begin{proof} We prove by induction on $j$. We have already proven this
for $j=1,2$. Assuming this is true upto $j$, we will show this for $j+1$.
The round repeatedly ``processes" elements of $E_{j+1}$.
Processing $b \in E_{j+1}$ involves checking if all
partitions in $\mathfrak{P}$ preserve $E_{\leq j} \cup \{b\}$. If they do, then $b$
is a failure. If $\mathcal{P} \in \mathfrak{P}$ splits $E_{\leq j} \cup \{b\}$, then
we ``swap" it with $\mathcal{P}_b$, i.e. $\mathcal{P}$ is now the old $\mathcal{P}_b$ and is
denoted a \emph{hanging partition}. The element $b$ is promoted
from $E_{j+1}$ to the new set $E_{j+2}$. How to deal with the hanging partition $\mathcal{P}$?
We first check if it splits $E_0$. If so, we have found $\mathcal{P}_i$. Otherwise,
we check if it preserves $E_0 \cup \{c\}$, for all $c \in E_1$.
If it splits $\{i,k,c\}$, then we swap $\mathcal{P}$ with $\mathcal{P}_c$. We promote
$c$ from $E_1$ to $E_2$. The old $\mathcal{P}_c$ becomes the new hanging partition $\mathcal{P}$.
If $\mathcal{P}$ preserves $E_{\leq 1}$, then we move on to $E_2$.
In general, if $\mathcal{P}$ preserves $E_{\leq l}$,
then we check if $\mathcal{P}$ preserves all $E_{\leq l} \cup \{c\}$, for $c \in E_{l+1}$.
If $\mathcal{P}$ splits $E_{\leq l} \cup \{c\}$, we swap $\mathcal{P}$ with $\mathcal{P}_c$ and
promote $c$ from $E_{l+1}$ to $E_{l+2}$. Note that the sets $E_{\leq p}$ (for any $p$)
can only decrease on such a promotion. So still the partitions
in $\mathcal{C}_p$ preserve $E_{\leq p-2}$.
The old $\mathcal{P}_c$ becomes the new hanging partition $\mathcal{P}$
and we repeat this process. If, on the other hand, $\mathcal{P}$ preserves all $E_{\leq l} \cup \{c\}$,
then $\mathcal{P}$ preserves $E_{\leq l+1}$. So we repeat this process
with $E_{l+2}$, and so on. If we end up with $\mathcal{P}$
preserving $E_{\leq j}$, then
we can safely move $\mathcal{P}$ into $\mathfrak{P}$. Otherwise, we have made a promotion and
we deal with a new hanging partition.
Note that when $\mathfrak{P}$ changes, we again process all elements in $E_{j+1}$.
There can only be a finite number of promotions,
so this round must end. We end up with $E_0,\cdots,E_{j+2}$,
with $\mathcal{C}_l$ preserving $E_{\leq l-2}$.
All the failures are still in $E_{j+1}$, and $\mathfrak{P}$
preserves all $E_{\leq j} \cup \{c\}$, $\forall c \in E_{j+1}$.
So $\mathfrak{P}$ now preserves $E_{\leq j} \cup E_{j+1} = E_{\leq j+1}$.
Note that if $E_{j+2}$ is empty,
we have a contradiction. This is because $E_{\leq j+1}$
is of size $i+1$ and there are at most $i-1$ partitions splitting it.
\end{proof}
Now we show that in this phase $i\geq2$ there can be at most $i$ rounds before
we get the desired $\mathcal{P}_i$.
\begin{claim} \label{clm-nonempty} Suppose $E_1,\cdots,E_j$
is a partition of $[i-1]$ such that $C_l$ preserves $E_{\leq l-2}$
and $\mathfrak{P}$ preserves $E_{\leq j}$. Then all $E_l$'s are non-empty.
\end{claim}
\begin{proof} Suppose $E_l$ is empty, for some $l\in[j]$. So $C_l$ is also
empty. Any partition that is not in $C_{\leq l-1}$
is either in $\mathfrak{P}$ or in $C_p$, for some $p \geqq l+1$ (if it exists). All these
partitions preserve $E_{\leq l-1}$. Thus, the only partitions splitting $E_{\leq l-1}$
are those in $C_{\leq l-1}$.
Since $|C_{\leq l-1}| = |E_{\leq l-1}| - 2$, we contradict the splitting property.
\end{proof}
The sets $E_1,\cdots,E_j$ form a partition of $[i-1]$. The above
claim tells us that we can run at most $i-1$ rounds to completion.
Hence, if we do not find $\mathcal{P}_i$ by $i-1$ rounds, then, by Claim~\ref{clm-round},
we will find it in the $i$th round. This completes the proof.
\end{proof}
\subsection{The general case}\label{sec-gen-case}
Now, we deal with simple, minimal identities and remove
the strong minimality condition. This will come at a cost
of an extra $k$ factor in the rank bound. First, we recall the definition of gcd and simple
parts of a general $\Sigma\Pi\Sigma$ circuit, as given in older works \cite{DS06,SS09}.
\begin{definition}[Gcd \& Simple part]
Let $C=\sum_{i\in[k]}T_i$ be a $\Sigma\Pi\Sigma(k,d)$ circuit over a field $\mathbb{F}$. The {\em gcd} of $C$ is
defined to be the usual gcd of the polynomials $T_i$-s, i.e. $\gcd(C):=\gcd(T_i|i\in[k])$.
The {\em simple} part of $C$ is the $\Sigma\Pi\Sigma(k,d')$ circuit, $\text{sim}(C):=C/\gcd(C)$, where
$d':=d-\text{deg}(\gcd(C))$.
\end{definition}
The following will be shown to be a consequence of Theorem~\ref{thm-strong-rank}.
\smallskip\noindent
{\bf Theorem \ref{thm-rank}.} (Final bound). Let $|\mathbb{F}|>d$. The rank of a simple, minimal
$\Sigma\Pi\Sigma(k,d)$, independent-fanin $k'$, identity is at most $2k^2 + (k-k')\cdot\text{SG}_{k'}(\mathbb{F},d)$.
\begin{proof}
Let circuit $C$ be $T_1+\cdots+T_k = 0$. Wlog let $T_1,\cdots,T_{k'}$ be a
linear basis for $T_1,\cdots,T_k$. Obviously, we have $1< k' <k$ (first by simplicity and second
by zeroness). By Theorem \ref{thm-nucleus}, there exists a nucleus $K$ wrt the set $\mathcal{I}:=[k']$. The rank
of $K$ is at most $2k^2$.
So, it remains to bound the non-nucleus rank of $C$ by $(k-k')\cdot\text{SG}_{k'}(\mathbb{F},d)$.
As $T_1,\cdots,T_{k'}$ form a basis, for each $i\in[k'+1,k]$, there exists $\alpha_{i,j}$-s in
$\mathbb{F}$ such that we have a zero circuit
$D_i := \sum_{j\in[k']}\alpha_{i,j}T_j + T_i = 0$. Define $N_i$ to be the set of $j$-s for which
$\alpha_{i,j}\ne0$. Thus,
\begin{equation}\label{eqn-thm-rank}
\forall i\in[k'+1,k],\ D_i = \sum_{j\in N_i}\alpha_{i,j}T_j + T_i = 0
\end{equation}
Since $\{\alpha_{i,j}T_j\ |\ j\in N_i\}$ are $|N_i|$ linearly independent terms, we get
that $D_i$ is a strongly minimal $\Sigma\Pi\Sigma(|N_i|+1,d)$ identity, for all $i\in[k'+1,k]$.
By nucleus properties, $\{K_j|j\in N_i\}$ are linearly independent polynomials, implying that
the polynomials $\{K_j/g_i|j\in N_i\}$ are also linearly independent,
where $g_i:=M(L_K(\gcd(D_i)))$. Thus, the linear space $K$ remains a nucleus of the new identity
$\text{sim}(D_i)$, showing at the same time that it is strongly minimal. We conclude that $\text{sim}(D_i)$
is a simple, strongly minimal $\Sigma\Pi\Sigma(k_i,d_i)$ identity with nucleus $K$ (although of $\text{rk}<2k^2$),
$k_i\leqslant(k'+1), d_i\leqslant d$,
for all $i\in[k'+1,k]$. Theorem \ref{thm-strong-rank} bounds the non-nucleus (non-$K$ to be precise)
rank of each of these identities by $\text{SG}_{k'}(\mathbb{F},d)$.
Suppose a linear form $\ell|\gcd(D_i)$ for all $i\in[k'+1,k]$. Then $\ell$ divides $T_j$
for all $j\in\bigcup_{i\in[k'+1,k]}N_i$ $\cup[k'+1,k]$. Consider the case
$\bigcup_{i\in[k'+1,k]}N_i=[k']$, it means that $\ell$ divides every term in $C$,
contradicting simplicity. Thus, in that case every linear form $\ell$ of $C$ appears in at least
one of the circuits $\{\text{sim}(D_i)|i\in[k'+1,k]\}$, whose total non-nucleus rank we have already bounded
by $(k-k')\cdot\text{SG}_{k'}(\mathbb{F},d)$, so we will be done.
The case, left to handle, is when : $S:=\bigcup_{i\in[k'+1,k]}N_i\subsetneq[k']$. This means,
by summing over $i$ in Equation (\ref{eqn-thm-rank}), $\sum_{i\in[k'+1,k]}T_i=$
$\sum_{s\in S}\beta_sT_s$, for some $\beta$-s in $\mathbb{F}$.
Substituting this in the equation $C=0$ we get,
$$C=C_{[k']}+C_{[k'+1,k]}\ =\ \sum_{i\in[k']}T_i+\sum_{s\in S}\beta_sT_s\ =\ 0.$$
As $S$ is a proper subset of $[k']$, the above equation could only mean that a nontrivial
combination of $T_i$ $(i\in[k'])$ is vanishing, contradicting the linear independence of those
polynomials. Thus, $S=[k']$. This completes the proof.
\end{proof}
\section{Sylvester-Gallai Rank Bounds for any $\mathbb{F}$} \label{sec-sg}
We wish to bound $\sgkb{\mathbb{F},m}$, for \emph{any} field $\mathbb{F}$. We will prove the
following theorem, which can be seen as the first attempt ever to give a Sylvester-Gallai Theorem for
{\em all} fields. It is convenient to think of a set of vectors $S$ in $\mathbb{F}^n$ as
\emph{multiple-free}: this means that no two vectors
in $S$ are scalar multiples of each other.
\begin{theorem}[High dimension Sylvester-Gallai for any field]\label{thm-sgkbd}
Suppose $k \in {\mathbb{N}}^{>1}$ and $S$ is an $SG_k$-closed set of vectors in $\mathbb{F}^n$
of rank $r \geqslant 9k$. Then, $|S| \geqslant 2^{r/9k}$. In other words, for every $m\in\mathbb{N}^{>1}$,
$SG_k(\mathbb{F},m)\leq 9k\lg m$.
\end{theorem}
\noindent {\bf Remark.} This bound is not tight. Over $\mathbb{F}_p$ , the best construction
we can come up with is an $SG_k$-closed set
with rank around $k+\log_p m$. Consider $\mathbb{F}_p^{k-1+r}$, for some $r > 0$ and $p\nmid(k-1)$.
Take the set $S_1\subset\mathbb{F}_p^{k-1+r}$ of vectors $e_1,e_2,\cdots,e_{k-1}$,
$\frac{1}{k-1}\cdot\sum_{i\in[k-1]} e_i$,
where $e_i$ is just the unit vector in the $i$th direction. Then take
the set $S_2\subset\mathbb{F}_p^{k-1+r}$ of all (non-zero) vectors which have zeroes in the first
$(k-1)$ coordinates. Extend this to $\mathbb{F}_p^{k+r}$ by putting a $1$
in the new coordinate. This gives a set of non-multiple
distinct vectors in $\mathbb{F}_p^{k+r}$.
Observe that $S_1$ is $SG_{k-1}$-closed
and $S_2$ is $SG_2$-closed. Hence $S_1 \cup S_2$ is
$SG_k$-closed. The size of $S$ is $(k + p^r-1)$ and the rank
is just $k+r$.\\
In some sense, bounds for $\text{SG}_2(\mathbb{F},m)$ are already implicit in known theorems
(used to prove lower bounds for LDCs). Concretely, Corollary 2.9 of~\cite{DS06}
can be interpreted as a proof that $\text{SG}_2(\mathbb{F},m) = O(\log m)$. This is an extension
of theorems in~\cite{GKST02} that prove this for $\mathbb{F}_2$. In the context
of $\text{SG}_2$, these proofs
can be interpreted as a ``doubling trick". In essence, each time we
want to increase the rank of an $\text{SG}_2$-closed set by $1$, we are forced
to double the number of vectors. A naive attempt to implement this
for $\text{SG}_k$ does not work.
Roughly speaking, we want to argue that if we want to increase the rank
of an $\text{SG}_k$-closed set by $(k-1)$, then the size of the set must double.
But, when $k\geq3$, this is not true! It is possible to increase the rank by $(k-1)$
by adding a very small number of vectors. So we have a sort of two-pronged approach.
If the size does not increase much, even though the rank increases,
then we show that the set has some very special $\text{SG}$ properties. Namely, {\em many} small
subspaces of the set are $\text{SG}_{k-1}$-closed. Even though these subspaces
can intersect very heavily, we are still able to argue that the set
must now be very large.
We will require two auxiliary claims. The first claim is probably of independent interest,
but the second is tied to our current approach.
\begin{claim} \label{clm-coord} Suppose $k \in {\mathbb{N}}^{>1}$ and $S$ is $SG_k$-closed.
Let vectors $e_1,e_2,\cdots,e_r$ be elements of $S$ that form a basis for $S$. Hence, every element in $S$
is represented by an $r$-tuple of coordinates in $\mathbb{F}$. There exists some element
in $S$ whose representation has at least $r/(k-1)$ non-zero coordinates.
\end{claim}
\begin{proof} Consider any vector $v \in S$. Let $v = \sum_{i=1}^r \alpha_i e_i$, for $\alpha_i \in \mathbb{F}$.
We denote by $N(v)$ the index set $\{i | \alpha_i \neq 0\}$. In other words,
$N(v)$ is the set of indices for which the corresponding coordinates of $v$
are non-zero. Let $v_1 \in S$ be some vector that maximizes $|N(v_1)|$. Choose some $v_2 \in S$
such that $N(v_2) \cap N(v_1) = \emptyset$ and $|N(v_2)|$ is maximized. Such a $v_2$ exists as long as
$|N(v_1)|<r$, as in this case we can get $v_2$ from $\text{sp}(e_i|i\notin N(v_1))$.
Iteratively, choose $v_j \in S$ such that $N(v_j)$ is disjoint to $\bigcup_{l < j} N(v_l)$
and $|N(v_j)|$ is maximized. As long as $\bigcup_{l \leq j} N(v_l) \neq [r]$,
we can always choose the next $v_{j+1}$, again, from $\text{sp}(e_i|i\notin\bigcup_{l \leq j} N(v_l))$.
We keep choosing $v_j$'s until we cover all
coordinates. At the end, $\bigcup_{l \leq j} N(v_j) = [r]$. Note that the
sets $N(v_1),\cdots,N(v_j)$ form a partition of $[r]$.
Suppose $j \geq k$. Then take the vectors $v_1,\cdots,v_k$. They are certainly
linearly independent, since they are defined on a disjoint set of coordinates. By the $SG_k$-closure
of $S$, some non-trivial linear combination of these vectors exists in $S$. Suppose some
non-zero combination of $v_{i_1},v_{i_2},\cdots$, denoted by $v$, is in $S$ (where $i_1 < i_2 \cdots$).
Note that $N(v) \supset N(v_{i_1})$ and $N(v)$ is disjoint to $\bigcup_{l < i_1} N(v_l)$.
That contradicts that choice of $v_{i_1}$. Hence $j < k$. Because the sets
$N(v_l)$ form a partition of $[r]$, $|N(v_1)|\geq r/(k-1)$. That completes the proof.
\end{proof}
\begin{claim} \label{clm-subsets} Suppose $k \in {\mathbb{N}}^{>2}$.
Consider a set of linearly independent vectors
$e_1,\cdots,e_{r'}$ in $S$. Let $I \subseteq [r']$ and $|I|\geq r'/4$. Let the set $E_I$ be $\{e_i | i \in I\}$
and $S_I$ be the set $S \cap \text{sp}(E_I)$.
If for all such $I$, $S_I$ is $SG_{k-1}$-closed, then $|S_{[r']}| \geq 2^{r'/8k}$.
\end{claim}
\begin{proof}
By Claim~\ref{clm-coord}, for every such $I$,
there is a $v_I \in S$
such that $v_I$ has at least $|I|/(k-2)$ non-zero coordinates (wrt basis $E_I$). Fix these $v_I$, for
each such $I$. As before, we denote the set of indices corresponding to non-zero coordinates of $v$
by $N(v)$. We describe a random
process to generate a subset of $[r']$. Simply pick each element in $[r']$
independently with probability $1/2$. Let $I$ and $J$ be two sets
generated independently this way. A ``good" event occurs if $|I|\geq r'/4$ and
$v_I \notin S_J$. We will call this the {\em good event for $(I,J)$}.
How can a good event not happen? Either
$|I|< r'/4$, or, if $|I|\geq r'/4$, then $v_I \in S_J$.
What is the probability of a ``bad" event (call this $\mathcal{E}(I,J)$) happening?
This is:
$$ {\rm Pr}[|I|< r'/4]\ +\ {\rm Pr}[|I|\geq r'/4]\cdot{\rm Pr}[v_I \in S_J \ \big| \ |I|\geq r'/4] $$
Since $\hbox{\bf E}[|I|] = r'/2$, the probability that $|I| < r'/4$ is at most
$e^{-r'/8}$ (by a Chernoff bound, refer to notes \cite{oDCher} for the exact form used).
Obviously, ${\rm Pr}[|I|\geq r'/4] < 1$.
Now assume that $|I|\geq r'/4$. So $|N(v_I)| \geq r'/4(k-2)$. For $v_I$ to
be in $S_J$, $J$ must contain $N(v_I)$. By the random construction of $J$,
the probability of this is at most $2^{-r'/4(k-2)}$. Thus, the probability
of a bad event is at most $e^{-r'/8} + 2^{-r'/4(k-2)} \leq 2\cdot2^{-r'/4(k-1)}$.
Now, let us choose $q = 2^{r'/8k}$ subsets of $[r']$ independently
through this random distribution. Call these $I_1,I_2,\cdots,I_q$.
For indices $1\leqslant a< b \leq q$, let $X_{ab}$ be the indicator random variable
for the event $\mathcal{E}(I_a,I_b)$. What is the expected total number
of bad events?
$$ \hbox{\bf E}\left[\sum_{1\leqslant a< b \leq q} X_{ab}\right] = \sum_{1\leqslant a< b \leq q} \hbox{\bf E}[X_{ab}]
< \frac{q^2}{2} {\rm Pr}[\mathcal{E}(I_1,I_2)] \leq 2^{r'/4k} \cdot 2^{-r'/4(k-1)} < 1 $$
By the probabilistic method, there exist subsets of $[r']$,
$I_1,I_2,\cdots,I_q$ such that no event $\mathcal{E}(I_a,I_b)$ happens.
This means all sets $I_a$ are of size at least $r'/4$.
Furthermore, for $1\leqslant a< b \leq q$, $v_{I_a} \notin S_{I_b}$. This
means that $v_{I_a} \neq v_{I_b}$. Therefore, there must
be at least $q$ distinct vectors in $S_{[r']}$.
\end{proof}
\begin{proof} (of Theorem~\ref{thm-sgkbd}) It will be convenient to assume that $|S| < 2^{r/9k}$,
and arrive at a contradiction. Let $T \subset S$ be a subset of rank $t\leq r/9$.
We can construct a basis of $S$, $e_1,\cdots,e_{r'},e_{r'+1},\cdots,e_r$, using elements of $S$
such that :
$r' \geq 8r/9$, $\{e_1,\cdots,e_{r'}\}$ are orthogonal to $T$, and $\{e_{r'+1},\cdots,e_r\}$
spans $T$. If $k>2$ then by applying
Claim~\ref{clm-subsets} to $e_1,\cdots,e_{r'}$ we get some
$I\subseteq[r']$ such that $|I|\geq r'/4\geq$ $2r/9\geq 2k$, and $S_I$ is \emph{not} $SG_{k-1}$-closed
(recall, $S_I:=S\cap\text{sp}(e_i|i\in I)$).
There exist linearly independent $v_1,v_2,\cdots,v_{k-1} \in S_I$ such that no non-trivial
combination of these are in $S_I$ (and hence in $S$). On the other hand, if $k=2$ then define $v_1:=e_1$.
Note that $v_1,v_2,\cdots,v_{k-1}$ are all orthogonal to $T$.
Consider any $v \in T$. By the $SG_k$-closure of $S$,
there exists some linear combination of $\{v,v_1,\cdots,v_{k-1}\}$ in $S$.
Call this the \emph{image} of $v$. Take two different $v,v' \in T$.
We argue that their images are distinct. Let the image of $v$
be $\alpha v + \sum_{i \leq k-1} \alpha_i v_i$ and the image
of $v'$ be $\beta v' + \sum_{i \leq k-1} \beta_i v_i$.
Note that $\alpha\beta\ne0$ by the way we have chosen $v_1,v_2,\cdots,v_{k-1}$.
If these images are equal, then $\alpha v - \beta v' = \sum_{i \leq k-1} (\beta_i - \alpha_i)v_i$.
Since $v$ and $v'$ are not multiples of each other, the right hand side
is non-zero. The left hand side is a vector that is spanned by $e_{r'+1},\cdots,e_r$
but the right hand side is spanned by $e_1,\cdots,e_{r'}$. Contradiction.
Hence, all images are distinct.
Starting from any subset $T \subset S$ of rank $t\leq r/9$,
there exist linearly independent $v_1,\cdots,v_{k-1}$ $\in S$ such that
$S \cap \text{sp}\{v_1,\cdots,v_{k-1},e_{r'+1},\cdots,e_r\}$
contains at least $2|T|$ vectors. The rank of this intersection is $t + (k-1)$.
Starting from $T$ being just one vector, we keep repeating this process.
This can go on for $h$ iterations, where $h$ is the smallest integer
such that $1 + (k-1)h \geq r/9$. So, $h = \lceil (r/9 - 1)/(k-1) \rceil$ which is at least $r/9k$, since
$r\geq 9k$.
We have $|S| \geq 2^h \geq 2^{r/9k}$. Contradiction.
\end{proof}
\section{Conclusion}
In this work we developed the strongest methods, to date, to study depth-$3$ identities.
The ideal methods hinge on a classification of zerodivisors of the ideals
generated by gates of a $\Sigma\Pi\Sigma$ circuit (eg. Lemmas \ref{lem-non-zd}, \ref{lem-I-match} and
\ref{lem-cancel}). That is useful in proving an ideal version of Chinese remaindering tailor-made
for $\Sigma\Pi\Sigma$ circuits, which is in turn useful to show a connection between all the gates involved
in an identity. As a byproduct, it shows the existence of a low rank {\em nucleus identity} $C'$
sitting inside {\em any} given $\Sigma\Pi\Sigma(k,d)$ identity $C$ (when $C$ is not minimal, $C'$ can still
be defined but it might not be homogeneous). This nucleus identity is quite mysterious and it
might be useful for PIT to understand (or classify) it further. For example, {\bf can the
rank bound for the nucleus identity be improved to $O(k)$?}
We generalize the notion of Sylvester-Gallai configurations to {\em any} field
and define a parameter $\text{SG}_k(\mathbb{F},m)$ associated with field $\mathbb{F}$. This number seems to be a
fundamental property of a field, and as we show, is very closely related to $\Sigma\Pi\Sigma$ identities.
It would be interesting to obtain bounds for $\text{SG}_k(\mathbb{F},m)$ for different $\mathbb{F}$. For example,
as also asked by \cite{KSar09}, {\bf can we nontrivially bound the number $\text{SG}_k(\mathbb{F},m)$ for interesting
fields: $\mathbb{C}$,
finite fields with large characteristic, or even $p$-adic fields?}
Other than the bounds for $\mathbb{R}$, all that was known before
is $\text{SG}_2(\mathbb{C},m)\leq 3$ \cite{EPS03}.
We shed (a little) light on high dimension $\text{SG}$ rank bounds by showing $\text{SG}_k(\mathbb{F},m) = O(k\log m)$.
We conjecture: $\text{SG}_k(\mathbb{F},m)$
is $O(k)$ for zero characteristic fields, while $O(k+\log_pm)$ for fields of characteristic $p>1$.
We also prove a property of a general collection $\mathfrak{P}$ of partitions of a universe $\mathcal{U}$, namely,
if $\mathfrak{P}$ has at least $|\mathcal{U}|-1$ partitions then it has an unbroken chain. It is tight and gives an idea
of how a Sylvester-Gallai configuration in the non-nucleus part of a $\Sigma\Pi\Sigma(k,d)$ identity ``spreads'' around.
Finally, we ask: {\bf Can the rank bound for simple minimal real $\Sigma\Pi\Sigma(k,d)$ identities be improved to
$O(k)$?} The best constructions known, since \cite{DS06}, have rank $4(k-2)$. Likewise, over other fields,
our upper bound of $O(k^2\log d)$ still leaves some gap in understanding the exact dependence on $k$.
\section*{Acknowledgements}
We are grateful to Hausdorff Center for Mathematics, Bonn
for the kind support, especially, hosting the second author
when part of the work was done. The first author thanks
Nils Frohberg for several detailed discussions that clarified
the topic of incidence geometry and Sylvester-Gallai theorems.
We also thank Malte Beecken, Johannes Mittmann and Thomas Thierauf
for several interesting discussions.
\bibliographystyle{alpha}
|
\section{Introduction}
Recent neutron scattering experiments performed on the heavy-fermion superconductor CeCoIn$_5$
have revealed an unexpected behaviour of the vortex lattice (VL) form factor\cite{Bianchi, White,Esk} defined as the Fourier transform of the local magnetic field in the vortex lattice. The VL form factor of the type II superconductors is usually a decreasing function of the magnetic field.\cite{Clem}
On the contrary, the VL form factor in CeCoIn$_5$ was found to increase with increasing magnetic field and then to fall down at the approach of $H_{c2}$. \cite{Bianchi, White}
CeCoIn$_5$ is a tetragonal, d-wave pairing superconductor with a large GL parameter, and with the highest critical temperature
($T_c=2.3\,K$) among all the heavy fermion compounds.\cite{Petrovic,Izawa,Aoki}
It has already generated great interest caused by the signs of the existence of the FFLO phase for a magnetic field parallel to the ab plane (and possibly to the c-axis),\cite{Bianchi2} and by the presence of an interval where the superconducting/normal phase transition is first order.\cite{Bianchi3,Miclea}
An explanation of the form factor behaviour has been proposed,\cite{Machida} which was based on a numerical processing of the quasi-classical Eilenberger equations in type II superconductors
with strong paramagnetic contribution.
In this Letter we present an analytic derivation of the magnetic field distribution and VL form factor taking into account the paramagnetic effects. Our analysis is based on the electrodynamic theory of the Abrikosov lattice in the superconducting state where the diamagnetic superfluid currents are mainly determined by the Zeeman interaction of the electron spins with the local magnetic field as developed in the paper\cite{Hou2}.
The orbital and the Zeeman respective contribution is quantified by the Maki parameter $\alpha_M=\sqrt{2}H_{c20}/H_p$, where $H_{c20}=\phi_0/2\pi\xi_0^2$ is the orbital critical field while $H_p=\Delta_0/\sqrt{2}\mu$ is the Pauli limiting field, $\phi_0\simeq 2.07\times 10^{-7}\,G\cdot cm^2$ is the flux quantum. Unlike
the majority of superconductors, the Pauli limiting field in CeCoIn$_5$ is smaller \cite{Bianchi3} than the orbital critical field by a factor of $> 3$. Hence, the Zeeman interaction plays an important role in the mixed state
field and current distributions.
We demonstrate analytically how the VL form factor, which decreases with increasing magnetic field in the high temperatures region of the phase diagram, at lower temperatures turns to the behaviour increasing with increasing field.
We shall consider a square VL with wave vector $q=2\pi\sqrt{B/\phi_0}$, formed in a tetragonal
type II superconductor under magnetic field directed along the c-axis. The magnetic induction ${\bf B}=\overline{{\bf h}}$ is determined as the spatial average of the local magnetic field
$\mathbf{h}=\nabla\times\mathbf{A}$. The GL theory for the VL form factor, valid in the limit $\kappa\gg1$ and for the external field not too close to $H_{c2}$, was developed by J. Clem. \cite{Clem}
Starting from the general form of the order parameter for an isolated vortex
\begin{equation}
\Delta(\mathbf{r})=\Delta_\infty f(r) e^{-i\varphi}
\label{OP}
\end{equation}
($\varphi$ is the angle measured from one of the axes in the ab plane), he proposed to model $f(r)$ by the trial function
\begin{equation}
f(r)=\frac{r}{R},
\label{f}
\end{equation}
with $R=\sqrt{r^2+\xi_v^2}$. The variational parameter $\xi_v$ was constrained to minimize the vortex total energy and was found to be in the large $\kappa$ limit $\xi_v=\sqrt{2}\,\xi$, where $\xi$ is the coherence length defined below. He has calculated the field distribution due to the orbital current and obtained the form factor
\begin{equation}
F_{orb}=B\,\frac{K_1(Q \xi_v)}{Q\lambda K_1(\xi_v/\lambda)},
\label{CFF}
\end{equation}
where $Q=\sqrt{q^2+\lambda^{-2}}$, $K_1(x)$ is the modified Bessel function of first order,\cite{Abr} and $\lambda$ is the London penetration depth.
One can write an approximative form of it in the conditions $\kappa\gg1$, $q\gg\lambda^{-1}$, and $q\xi_v\ll1$,
\begin{equation}
F_{orb}\simeq \frac{\phi_0}{(2\pi\lambda)^2}-\frac{\xi^2B}{2\lambda^2}\ln\left(\frac{2e\phi_0}{\gamma^2(2\pi\xi)^2B}\right),
\label{Fgamma}
\end{equation}
$\ln\gamma=C\simeq0.577$ is the Euler constant.
The found form factor, that quite slowly decreases with magnetic field, is reliable if one neglects the paramagnetic interaction of the electron spins with the magnetic field.
The latter leads to two extra features that are important in the case of a large enough Maki parameter.
First, the two characteristic lengths $\xi$ and $\lambda$ in the above expression proves to be magnetic field dependent. Second, a new mechanism originating from the Zeeman interaction gives rise to the main contribution to the diamagnetic screening \cite{Hou2}
in the high magnetic field region of the phase diagram. To find it we start with the Ginzburg-Landau formulation including the paramagnetic effects.
The superconductor CeCoIn$_5$ has pairing symmetry $d_{x^2-y^2}$,\cite{Vorontsov} with order parameter
\begin{equation}
\Delta_{\mathbf{k}}(\mathbf{r})=\psi(\mathbf{\hat{k}})\Delta(\mathbf{r}),~~~~\psi(\mathbf{\hat{k}})=\sqrt{2}\cos(2\varphi).
\end{equation}
The free energy of the system is given by the Ginzburg-Landau (GL) functional
\begin{eqnarray}
\mathcal{F}=\int\mathrm{d}^2\mathbf{r}\,\Big(\,\frac{\mathbf{h}^2}{8\pi}+\big(\alpha+\varepsilon(h_z-B)\big)|\Delta|^2\nonumber\\+\beta|\Delta|^4+\gamma|\mathbf{D}\Delta|^2\Big),
\label{GL}
\end{eqnarray}
where $\mathbf{D}=-i\nabla+2e\mathbf{A}$ is the gauge-invariant derivative (from here we put $\hbar=c=1$), and
the coefficients are functions of temperature and induction. In the clean limit they are\cite{Hou2,Hou1}
\begin{eqnarray}
\alpha&=&N_0\big(\ln(T/T_c)+\Re\mathfrak{e}\Psi(w)- \Psi (1/2)\big),\nonumber\\
\varepsilon&=&\frac{N_0\mu}{2\pi T}\Im\mathfrak{m}\Psi'(w),\nonumber\\
\beta&=&-\frac{N_0}{8(2\pi T)^2}\langle|\psi(\mathbf{\hat{k}})|^4\rangle\Re\mathfrak{e}\Psi^{(2)}(w),\nonumber\\
\gamma&=&-\frac{N_0 v_F^2}{16(2\pi T)^2}\Re\mathfrak{e}\Psi^{(2)}(w),\nonumber
\end{eqnarray}
where $\Psi(w)$
is the digamma function, $\Psi^{(m)}(w)$ are its derivatives called by the polygamma functions, \cite{Abr} and
$$
w=\frac{1}{2}-\frac{i\mu B}{2\pi T}\cdot
$$
The coefficient $\varepsilon$ is proportional to $B$. Hence, the corresponding term in the functional is negligibly small in the ordinary GL region near $T_c$ ($B\to 0$).
In a quasi-two-dimensional case we deal in the first approximation with a cylindrical Fermi surface.
For the order parameter given above the average over the Fermi surface is $\langle|\psi(\mathbf{\hat{k}})|^4\rangle=3/2$.
The stationary condition for the functional in respect of the order parameter gives the non-linear GL equation
\begin{equation}
\alpha_1 \Delta+ 2\beta|\Delta|^2\Delta+\gamma\mathbf{D}^2\Delta=0.
\label{NonLinear}
\end{equation}
The coherence length is defined as
$
\xi=\sqrt{\gamma/|\alpha_1|},
$
with $\alpha_1=\alpha+\varepsilon(h_z-B)$.
This expression is meaningful away from the region where $\gamma=0$ (close to which the FFLO phase may exist).
The variational solution of the GL equation for an isolated vortex \cite{Clem} is given by eqns (\ref{OP}), (\ref{f}) with
the order parameter amplitude
$
\Delta_\infty=\sqrt{|\alpha_1|/2\beta}.
$
In the high $\kappa$ limit, we have $h\simeq B$ almost everywhere. Therefore as far as the configuration of the order parameter is concerned, we will take $\alpha_1\simeq\alpha$.
The stationary condition with respect to the vector potential gives the Maxwell equation
\begin{equation}
\frac{1}{4\pi}\nabla\times\mathbf{h}={\bf j}_{orb}+{\bf j}_Z
\label{Maxwell}.
\end{equation}
For an isolated vortex with order parameter given by eqns. (\ref{OP}), (\ref{f}) the vector potential is of the form
$\mathbf{A}_v(\mathbf{r})=A_v(r)\hat{\varphi}$,
the \emph{orbital} density of current is
\begin{equation}
{\bf j}_{orb}=-8e^2\gamma\big(A_v(r)-\frac{\phi_0}{2\pi r}\big)|\Delta|^2\hat{\varphi}
\label{jgamma},
\end{equation}
while the \emph{Zeeman} current \cite{Hou2} found from (\ref{GL}) is
\begin{equation}
{\bf j}_{Z}=\varepsilon\frac{d}{dr}|\Delta|^2\hat{\varphi}.
\end{equation}
Hence, we come to the equation that determines the vector potential $A_v(r)$
\begin{equation}
\frac{d}{dr}\big(\frac{1}{r}\frac{d}{dr}(rA_v)\big)-\frac{f^2}{\lambda^2}A_v=-\frac{\phi_0f^2}{2\pi\lambda^2r}-4\pi\varepsilon\Delta_\infty^2\frac{df^2}{dr},
\label{DE}
\end{equation}
where $\lambda=\sqrt{\beta/16\pi e^2\gamma|\alpha|}$ is the penetration depth. Writing (\ref{DE}) for the shifted potential $a(r)=A_v(r)-\phi_0/2\pi r$, we obtain the differential equation with an inhomogeneous term of Zeeman origin
\begin{equation}
\frac{d}{dr}\big(\frac{1}{r}\frac{d}{dr}(ra)\big)-\frac{f^2}{\lambda^2}a=-4\pi\varepsilon\Delta_\infty^2\frac{df^2}{dr}.
\end{equation}
The general solution of this equation $a(r)=a_h(r)+a_i(r)$ consists of the sum of the solutions of the homogeneous and the inhomogeneous equations.
The former determines the orbital part $A_{orb}=\phi_0/2\pi r+a_h$ of the vector potential $A_v=A_{orb}+A_Z$,
\begin{equation}
A_{orb}(r)=\frac{\phi_0}{2\pi r}\left (1-\frac{R}{\xi_v}\frac{K_1(R/\lambda}{K_1(\xi_v/\lambda}\right ),
\end{equation}
such that the corresponding magnetic field ${\bf h}_{orb}=h_{orb}\hat z$ is \cite{Clem}
\begin{equation}
h_{orb}=\frac{\phi_0}{2\pi\lambda\xi_v}\frac{K_0(R/\lambda)}{K_1(\xi_v/\lambda)},
\end{equation}
and the form factor is determined by eqn.(\ref{CFF}).
Taking into account the expressions for $\alpha, \beta$ and $\gamma$ we obtain field dependence
for the orbital part of the form factor determined at $\kappa\gg1$ by the first term in (\ref{Fgamma})
\begin{equation}
F_{orb}\simeq\frac{4}{3\pi}\phi_0e^2v_F^2|\alpha|.
\label{Fgammaeq}
\end{equation}
The Zeeman part of the vector potential is given by
\begin{equation}
A_Z(r)=a_i(r)=\frac{R}{r}K_1(R/\lambda) C(R/\lambda)
\end{equation}
with
\begin{eqnarray}
C(z)&=&\int_{\xi_v/\lambda}^z \mathrm{d} x\,\phi(x), ~~~~~\phi(z)=t(z)e^{-s(z)},\nonumber\\
t(z)&=&-\frac{8\pi\varepsilon\Delta_\infty^2\xi_v^2}{\lambda}\int \mathrm{d} z\,\frac{1}{z^3 K_1(z)}e^{s(z)},\nonumber\\
s(z)&=&-\int \mathrm{d} z\,\left(\frac{1}{z}+\frac{2K_0(z)}{K_1(z)}\right).\nonumber
\end{eqnarray}
The corresponding magnetic field ${\bf h}_Z=h_Z\hat z$ takes the form
\begin{equation}
h_Z=\frac{1}{\lambda}\left[K_0(R/\lambda)C(R/\lambda)+K_1(R/\lambda)C^\prime(R/\lambda)\right].
\label{heps}
\end{equation}
In order to find a simpler expression for this new term, we can look at its behaviour in the region not too far from the vortex core i.e. for $R\ll\lambda$, corresponding to the region between the vortices of the lattice. By using the Bessel functions asymptotic expressions in that limit,\cite{Abr} we find the dominating term in the eqn. (\ref{heps})
\begin{equation}
h_{Z}\simeq4\pi\varepsilon\Delta_\infty^2\frac{\xi_v^2}{R^2}.
\label{hepseq}
\end{equation}
We see that $h_Z(r)$ is concentrated near the core of a vortex, the characteristic length associated to it being of the order of $\xi_v$.
As a remark, this term may also be derived by considering the equation
$
\nabla\times\mathbf{h}_Z=4\pi{\bf j}_Z,
$
that is valid in the absence of the orbital current. From (\ref{heps}) one can find the correction to (\ref{hepseq}),
$
\delta h_Z=4\pi \varepsilon\Delta_\infty^2K_0(R/\lambda)\ln(R/\xi_v)/\kappa^2$,
which is small in the high $\kappa$ limit. Let us therefore consider expression (\ref{hepseq}) for deriving the new form factor.
The Fourier transform of the magnetic field around a single vortex is
\begin{equation}
{h}_Z(q)= \int \mathrm{d}^2\mathbf{r}\,h_Z(r)e^{-i{\bf q}\cdot{\bf r}}=8\pi^2\varepsilon \Delta_\infty^2
\xi_v^2K_0(q\xi_v).
\label{Fourier}
\end{equation}
Hence, the contribution to the form factor that originates from the interaction of the electron spins with the local magnetic field for an array of $B/\phi_0$ vortices per $cm^2$ is
\begin{equation}
F_{Z}=4\pi\frac{N_0{\mu B}}{\phi_0{T}}\Delta_\infty^2 \Im\mathfrak{m}\Psi'(w)\xi_v^2K_0(q\xi_v)
\end{equation}
One can evaluate the total energy of the magnetic field (quantities integrated over the whole 2-D plane) and deduce that the energy of a single vortex is (in the high $\kappa$ limit) dominated by the term $h_{orb}$.
Therefore the minimization of the energy of a single vortex gives the same variational parameter $\xi_v=\sqrt{2}\,\xi$ as exposed in the Clem paper.\cite{Clem} As a result we obtain
\begin{equation}
F_{Z}=\frac{4\pi}{3}\frac{N_0{\mu B}v_F^2}{\phi_0{T}} \Im\mathfrak{m}\Psi'(w)K_0(\sqrt{2}q\xi)
\label{ZFF}
\end{equation}
\begin{figure}[b]
\centering
\includegraphics[width=9cm]{Image7.png}
\caption{(Above) CeCoIn$_5$ phase diagram for $H ||$c-axis. The color lines represent the temperatures where we applied the model. (Below) Variations of the squared form factor $F^2$ at different temperatures including both the orbital and Zeeman contributions. The dashed lines represent the variations of the orbital part alone.}
\label{FormFactor}
\end{figure}
For $q\xi\ll 1$ and $\mu B<2\pi T$
the ratio of two form factors (\ref{ZFF}) and (\ref{Fgammaeq}) is
\begin{equation}
\frac{F_Z}{F_{orb}}\simeq\frac{7\zeta(3)\mu^2 B^2}{\pi T^2}\frac{\ln(\sqrt{2}/q\xi)}
{\ln(T_c/T)-7\zeta(3)(\mu B/2\pi T)^2},
\end{equation}
here $\zeta(x)$ is the Riemann zeta function.
We observe that the Zeeman part of the form factor prevails over its orbital part in the phase diagram region where $\mu B \simeq T$.
We are now able to analyze the behaviour of the total form factor
\begin{equation}
F=F_{orb}+F_Z,
\end{equation}
where $F_{orb}$ and $F_Z$ are given by equations (\ref{CFF}) and (\ref{ZFF}) correspondingly.
It is drawn in Fig. \ref{FormFactor}. In numerical calculations we assumed the values $\mu=g\mu_B/2=\mu_B$ for the electron magnetic moment in the material, and $v_F=5\times 10^5\,cm/s$ for the Fermi velocity inside the superconducting phase. A value for $v_F$ slightly bigger was given in\cite{Miclea} as a result of measurements of the upper critical field $H_{c2}$ near $T_c$. The 2-D density of states on the Fermi surface is independent of $v_F$ and is given by $N_0=m^*/2\pi\ell_c$, where we considered $m^*=100\,m_e$ for the electron effective mass, and $\ell_c=7.6\times10^{-8}\,cm$ is the lattice c-axis spacing.
In the form factor variations, there is first a domination of the orbital part in the low magnetic field region ($F_Z$ vanishes at $B=0$). We observe next a crossover to a region where the paramagnetic term is dominant. The regime where (\ref{CFF}) goes exactly against (\ref{ZFF}) is likely to explain the observed constant logarithm of the squared form factor\cite{Esk} in the interval $B=0.5-2T$. In addition, these features were observed in an experiment realised on the s-wave superconductor TmNi$_2$B$_2$C.\cite{Esk2} At larger fields and smaller temperatures region, the form factor increases with field.
The observed form factor\cite{Bianchi,White} falling towards zero near the phase transition line,
where the field variation amplitude decreases with the order parameter,
is out of the region of applicability of the developed theory where isolated vortices are supposed. To describe the form factor behaviour reaching the maximum and decreasing at the approach of the upper critical field, one must match the present theory to the Ginzburg-Landau description valid in the vicinity of $H_{c2}$
at temperatures above and also slightly below the tricritical point where the order parameter $\Delta_\infty^2$ takes a finite value at the critical field.
This will be published elsewhere.
In conclusion, making use of the generalized Clem approach
we have calculated the magnetic field dependence of the vortex lattice form factor.
Some results found previously in the framework of the Ginzburg-Landau-Abrikosov theory\cite{Hou2} lead us to find a new term originating from the interaction of the electron spins with the magnetic field existing inside the sample. The latter has a dominant contribution in the expression of the vortex lattice form factor at high magnetic field in the superconductors with a small enough Fermi velocity.\cite{footnote}
The magnetic field caused by the diamagnetic currents originating from the Zeeman interaction was found to be concentrated in the core of vortices and gives rise to new features of the form factor that accounts for the measurements that are currently being made on CeCoIn$_5$.\cite{Bianchi,White,Esk}
|
\section{Introduction}
Recently, the technical advances in cold atom trapping allowed the
experimental realization of the Anderson localization
\cite{Anderson} in quantum matter waves \cite{Billy,Roati}. The good
tunability and controllability of optical lattices offer myriad
opportunities for studying the disorder effects in ultracold atom
system \cite{Lewenstein}. So far, different techniques have been
devised for the introduction of disorder in the ultracold atom
system, such as speckle filed patterns in the optical lattice
\cite{Billy,Lye}, random localized impurities by loading a mixture
of two kinds of atoms with heavy and light masses \cite{Gavish}, and
incommensurate bichromatic optical lattices by superimposing two
one-dimensional (1D) optical lattices with incommensurate frequency
\cite{Roati,Fallani}. In particular, the experiment \cite{Fallani}
has provided evidences of the existence of a Bose-glass (BG) phase
\cite{Fisher89}.
Interesting phenomenons are expected to appear in disordered systems
when the interplay of disorder and interactions is taken into
account. Interactions between atoms can be controllably tuned by
Feshbach resonances in ultra-cold atom systems. While disorder can
lead to localization of the wave function of a particle,
delocalization can arise as the consequence of interactions in some
many-body systems. Theoretically, for a repulsive Bose gas it has
been predicted that there is a quantum phase transition from a
superfluid phase to an insulating BG phase with localized
single-particle states as disorder is increased
\cite{Giamarchi,Fisher,Delande,Fontanesi,Egger,Egger2,Cai}, however
unambiguous observation of the superfluid-Bose-glass transition is
still under debate \cite{Damski,Delande}. A lot of attention
\cite{Gurarie,Roux,Deng,Roscilde,Orso} has been payed to
investigating the combined role of disorder and interactions in the
strongly interacting ultra-cold atomic system. Except numerical or
approximate approaches, the exact solution for the many-body systems
with the interplay of disorder and interaction are rarely known. In
this paper, we study the interacting bosons on the incommensurate
optical lattices with harmonic trap in the limit case with
infinitely repulsive interaction which can be solved exactly. The 1D
Bose gas with infinitely repulsive interaction is known as the
hard-core boson (HCB) or Tonks-Girardeau (TG) gas \cite{Girardeau},
which can be exactly solved via the Bose-Fermi mapping
\cite{Girardeau} and has attracted intensive theoretical attention
\cite{Girardeau1,Minguzzi,Gangardt}. Experimental access to the
required parameter regime has made the TG gas a physical reality
\cite{Paredes,Kinoshita}. For the HCB in 1D optical lattices, it is
convenient to use the exact numerical approach proposed by Rigol and
Muramatsu \cite{Rigol}. Following the exact numerical approach, we
calculate the static properties of the hard-core bosons, such as
one-particle density matrices, density profiles, momentum
distributions, natural orbitals and their occupations, to exploit
the superfluid-to-BG phase transition for the systems in
incommensurate optical lattices with harmonic confining trap.
Furthermore, we study the nonequilibrium dynamical properties of
expanding clouds of hard-core bosons on the 1D incommensurate
lattices after turning off the harmonic trap suddenly. We find that
the expansion dynamics for the superfluid phase and BG phase exhibit
quite different behaviors, which may serve as a signature for
experimentally detecting the transition from superfluid-to-BG phase.
The paper is organized as follows. In Section II, we present the
model and the exact approach used in this paper. In Section III, we
show properties of the ground-state for hard-core bosons on the
incommensurate optical lattice with a harmonic trapping potential.
Section IV is devoted to studying the nonequilibrium dynamics of the
system after the harmonic trap is suddenly switched off. Finally, a
summary is presented in Section V.
\section{Model And Method}
In the present section we describe the exact approach which we used
to study 1D hard-core bosons on the incommensurate lattice with an
additional harmonic trap. Under the single-band tight-binding
approximation, the system of $N$ hard-core bosons in the 1D optical
lattice can be described by the following Hamiltonian:
\begin{equation}
\label{eqn2}
H=-t\sum_i(b^\dagger_ib_{i+1}+\mathrm{H.c.})+\sum_iV_in^b_i,
\end{equation}
where $b^\dagger_i(b_i)$ is the creation (annihilation) operator of
the boson which fulfills the hard-core constraints \cite{Rigol},
{\it i.e.,} the on-site anticommutation $(\{b_i,b^\dagger_i\}=1)$
and $[b_i,b^\dagger_j]=0$ for $i\neq j$; $n^b_i$ is the bosonic
particle number operator; $t$ is the hopping amplitude set to be
the unit of the energy $(t=1)$; $V_i$ is given by
\begin{equation}
V_i=V_I\mathrm{cos}(\alpha2\pi i+\delta)+V_H(i-i_0)^2.
\end{equation}
Here $V_I$ is the strength of incommensurate potential with $\alpha$
as an irrational number characterizing the degree of the
incommensurability and $\delta$ an arbitrary phase (in our
calculation it is chosen to be zero for convenience, without loss of
generality), $V_H$ is the strength of harmonic trap and $i_0$ is the
position of the vale of the harmonic trap.
In order to calculate the properties of hard-core bosons, it is
convenient to use the Jordan-Wigner transformation \cite{Jordan}
(JWT) or Bose-Fermi mapping for the lattice model
\begin{equation}
b^\dagger_j=f^\dagger_j\prod^{j-1}_{\beta=1}e^{-i\pi f^\dagger_\beta
f_\beta},b_j=\prod^{j-1}_{\beta=1}e^{+i\pi f^\dagger_\beta
f_\beta}f_j,
\end{equation}
which maps the Hamiltonian of hard-core bosons into the Hamiltonian
of noninteracting spinless fermions
\begin{equation}
\label{eqn1}
H_F=-\sum_i(f^\dagger_if_{i+1}+\mathrm{H.c.})+\sum_iV_in^f_i ,
\end{equation}
where $f^\dagger_i(f_i)$ is the creation (annihilation) operator of
the spinless fermion and $n^f_i$ is the particle number operator.
The ground-state wave function of the system with $N$ spinless free
fermions can be obtained by diagonalizing Eq.(\ref{eqn1}) and can be
represented as
\begin{equation}
\label{eqn3}
|\Psi^G_F\rangle=\prod^N_{n=1}\sum^L_{i=1}P_{in}f^\dagger_i|0\rangle
,
\end{equation}
where $L$ is the number of the lattice sites, $N$ is the number of
fermions (same as bosons), and the coefficients $P_{in}$ are the
amplitude of the $n$-th single-particle eigenfunction at the $i$-th
site which can form an $L \times N$ matrix $P$ \cite{Rigol}.
In order to get the static properties of the ground-state, we
calculate the one-particle Green function for the hard-core bosons
defined by
\begin{equation}
G_{ij}=\langle\Psi^G_{HCB}|b_ib^\dagger_j|\Psi^G_{HCB}\rangle\\
= \langle\Psi^A|\Psi^B\rangle ,
\end{equation}
where $|\Psi^G_{HCB}\rangle$ is the ground-state of hard-core
bosons, and $\langle\Psi^A| =
\left(f^\dagger_i\prod_{\beta=1}^{i-1}e^{-i\pi f^\dagger_\beta
f_\beta}|\Psi^G_F\rangle\right)^\dagger$, $ |\Psi^B\rangle
=f^\dagger_j\prod_{\gamma=1}^{j-1}e^{-i\pi f^\dagger_\gamma
f_\gamma}|\Psi^G_F\rangle $. Explicitly the state $\left| \Psi
^A\right\rangle $ can be represented as $ \left| \Psi^A\right\rangle
=\prod_{n=1}^{N+1}\sum_{l=1}^LP_{ln}^{A}f_l^{\dagger }\left|
0\right\rangle $ with $ P_{ln}^{A} = - P_{ln} $ for $l\leq i-1$,
$P_{ln}^{A} =P_{ln}$ for $l\geq i$ with $n\leq N$, $P_{iN+1}^{A}=1$
and $P_{lN+1}^{A}=0$ $\left( l\neq i\right) $. Similarly we can get
$P^{B}$ for the state $|\Psi^B\rangle$ with the replacement of $i$
by $j$. The Green function is a determinant dependent on the
$L\times \left( N+1\right) $ matrices $P^{A}$ and $P^{B}$
\cite{Rigol}
\begin{eqnarray}
G_{ij}=\left\langle \Psi^A|\Psi^B\right\rangle =\det \left[ \left(
P^{A}\right)^\dagger P^{B}\right] .
\end{eqnarray}
The one-particle density matrix can be evaluated from the relation
\begin{equation}
\rho_{ij}=\langle
b^\dagger_ib_j\rangle=G_{ij}+\delta_{ij}(1-2G_{ii}).
\end{equation}
Alternatively one can use the method proposed by Pareder {\it et
al.} \cite{Paredes} to calculate the one-particle density matrix
directly. The momentum distribution is defined by the Fourier
transform with respect to $i-j$ of the one-particle density matrix
with the form
\begin{equation}
n(k)=\tfrac{1}{L}\sum^L_{i,j=1}e^{-ik(i-j)}\rho_{ij},
\end{equation}
where $k$ denotes the momentum. The eigenfunctions or natural
orbitals $(\phi^\eta_i)$ of the one-particle density matrix
\cite{Penrose} can be obtained by solving
\begin{equation}
\sum^L_{j=1}\rho_{ij}\phi^\eta_j=\lambda_\eta\phi^\eta_i,
\end{equation}
which can be understood as being effective single particle states
with occupations $\lambda_\eta$. For noninteracting bosons, all the
particles occupy in the lowest natural orbital and bosons are in the
BEC phase at zero temperature, however only the quasi condensation
exists for the 1D hard-core bosons. For hard-core bosons in the
incommensurate optical lattice, the strong incommensurate potential
can destroy the quasi-BEC.
The nonequilibrium dynamical properties of expanding clouds of
hard-core bosons on 1D incommensurate lattice after turning off the
harmonic trap suddenly can be also calculated through the equal time
Green function. With the JWT, we can express the equal time Green
function for the expanding hard-core bosons as
\begin{eqnarray}
G_{ij}(t)&=&\langle\Psi^G_{HCB}(t)|b_ib^\dagger_j|\Psi^G_{HCB}(t)\rangle\\
&=& \langle\Psi^G_F(t)|\prod_{\beta=1}^{i-1}e^{i\pi f^\dagger_\beta
f_\beta}f_if^\dagger_j\prod_{\gamma=1}^{j-1}e^{-i\pi
f^\dagger_\gamma f_\gamma}|\Psi^G_F(t)\rangle,\notag
\end{eqnarray}
where $|\Psi^G_{HCB}(t)\rangle$ is the wave function of the
hard-core bosons at $t$ after turning off the harmonic trap and
$|\Psi^G_F(t)\rangle$ is the corresponding one for the
noninteracting fermions. The wave function $|\Psi^G_F(t)\rangle$ can
be easily calculated with the initial wave function
$|\Psi^G_F\rangle$
\begin{eqnarray}
|\Psi^G_F(t)\rangle=e^{-iH'_Ft}|\Psi^G_F\rangle=\prod_{n=1}^{N}\sum_{l=1}^LP_{ln}(t)f_l^{\dagger
}|0\rangle,
\end{eqnarray}
which is still a product of time-dependent single-particle states,
where $\hbar$ has been set to be unit of time in the evolution
operator, $H'_F$ is the $H_F$ in Eq.(\ref{eqn1}) with $V_H=0$, and
$P(t)$ is the matrix of $|\Psi^G_F(t)\rangle$ in the same way as
$|\Psi^G_F\rangle$. Then we can use the method described above to
calculate the equal time Green function, yet we can get the
one-particle density matrix, density profile, momentum distribution,
natural orbitals and their occupations.
\begin{figure}[tbp]
\includegraphics[width=\linewidth, height=10cm, bb=25 20 303 235]{Fig1.eps}
\caption{(Color online) The density profiles(a,c,e) and the local
average density distributions(b,d,f) for the systems with 1000
sites; 200 bosons; $V_H=3\times10^{-5}$; $\alpha=(\sqrt{5}-1)/2$;
and (a,b)$V_I=0$, (c,d)$V_I=1$, and (e,f)$V_I=2.4$. (g): The local
average density distributions for the systems with 1000 sites, 600
bosons, $V_H=3\times10^{-5}$, $\alpha=(\sqrt{5}-1)/2$ and different
$V_I$. (h): The mean particle fluctuation vs $V_I$ for the systems
with 1000 sites, 200 bosons, $V_H=3\times10^{-5}$,
$\alpha=(\sqrt{5}-1)/2$.} \label{Fig1}
\end{figure}
\section{STATIC PROPERTIES OF HARD-CORE BOSONS CONFINED IN HARMONIC TRAP}
Hard-core bosons in harmonic trap without the incommensurate
potential have been studied by Rigol and Muramatsu \cite{Rigol} in
detail. Quantitatively they characterize the system in a harmonic
trap with the length scale set by combination lattice-confining
potential $\zeta=(V_H/t)^{1/2}$, and the associated characteristic
density $\widetilde{\rho}=N/\zeta$ \cite{Rigol1}. They found that
there is a critical characteristic density
($\widetilde{\rho}_c\sim2.6-2.7$) in the system, for
$\widetilde{\rho}<\widetilde{\rho}_c$ the whole system is in the
superfluid phase at zero temperature, and for
$\widetilde{\rho}>\widetilde{\rho}_c$ there is a phase separation in
the system with a Mott insulating plateau in the middle of the trap
with filling factor equaling one surrounding by the superfluid phase
on the two sides. In this paper we mainly study the influence of the
incommensurate potential which acts as the quasi-random potential
and leads to localization of particles. We focus our study on
systems with low characteristic density
($\widetilde{\rho}<\widetilde{\rho}_c$). For system with high
characteristic density the properties are similar to the one with
low characteristic density except that there is a Mott insulator
plateau in the system. The Mott insulator plateau is not
dramatically influenced by the incommensurate potential because the
particles in the Mott phase are already localized with one-particle
density matrix $\rho_{ij}=\delta_{ij}$. Without especially
illustration, properties studied in this section are for systems
with low characteristic density.
Following the method described in the above section, we can get the
one-particle density matrix. The density profile is given by the
diagonal elements of the one-particle density matrix, i.e.,
$n_i=\rho_{ii}$. In the low characteristic density region, the
density profiles for three different strength of the incommensurate
potential are shown in Fig.\ref{Fig1}. In the absence of the
incommensurate potential ($V_I=0$), the density profile is shown in
Fig.1(a) with all the bosons being in the superfluid phase at the
vale of the harmonic trap. As $V_I$ increases but is still small,
the density profiles basically have the arc shape, but there are a
lot of drastic oscillations in the profiles induced by the
incommensurate potential. The amplitude of the oscillations becomes
more and more large with the increase in $V_I$. When $V_I$ becomes
large enough, the density almost oscillates in the range of $0$ to
$1$. The density profile looks like a belt at the center of harmonic
trap. Additionally, the width of the density profile almost does not
change with the change of $V_I$. For system in high characteristic
density region, the Mott plateau in the density profile basically
does not change with the increase in $V_I$ because the particles in
this phase are already localized. However the superfluid regions
exhibit similar behavior as the system in low characteristic density
region.
Since there are many drastic oscillations in the density
distributions, it is hard to tell that in which area the density is
high in a density profile. In order to reduce the drastic
oscillations, we define the local average density $\overline{n}_i$
as
\begin{eqnarray}
\overline{n}_i=\sum_{j=-M}^Mn_{i+j}/(2M+1),
\end{eqnarray}
where $2M+1$ is the length to count the local average density with
$M\ll L$. In this work, we set $M=10$. The local average density
distributions for three different values of $V_I$ are also shown in
Fig.\ref{Fig1}. When $V_I=0$, the local average density distribution
is almost the same as the density profile with no drastic
oscillations. As $V_I$ increases, plateaus appear at the shoulders
of the arc (see Fig.\ref{Fig1}(d)), then become wider and wider.
Such kind of plateau is induced by the disorder, and thus we call it
Anderson plateau hereafter. When $V_I>2$, the shape of the local
average density does not change with the increase in $V_I$. In
Fig.\ref{Fig1} (e) and (f), the density distribution and the local
average density distribution for $V_I=2.4$ are displayed.
The number of the plateaus and their locations are related to the
particle number and the strength of the harmonic trap. In
Fig.\ref{Fig1} there happened to be only two plateaus. For system
with high characteristic density, the local average density
distributions for different $V_I$ are shown in Fig.\ref{Fig1}(g).
The Mott insulator plateau always exists in the trap center as $V_I$
increases, except that the edges of the plateau are more and more
dissolved into superfluid phase when $V_I<2$. Meanwhile new plateaus
appear at the two sides of the distribution, which is similar to the
one with low characteristic density as $V_I$ increases. The Mott
plateau in the trap center is characterized by $n_i=1$ which does
not oscillate against the particle density in the Anderson plateaus
due to the existence of the incommensurate potential.
The mean particle fluctuation for various value of $V_I$ is shown in
Fig.\ref{Fig1} (h). It is defined by $\kappa=\sum_{i=1}^L\kappa_i/L$
with $\kappa_i$ being the local particle fluctuation
$\kappa_i\equiv\langle\hat{n}^2_i\rangle-n^2_i=n_i-n^2_i$ for
hard-core bosons. We can see that $\kappa$ decreases as $V_I$
increases because of the incommensurate lattice making the bosons
difficult to hop. For system in high characteristic density region,
the local particle fluctuation for particles in the Mott insulator
plateau always equals zero as $n_i=1$ for any $V_I$, whereas the
particle fluctuation is not equal to zero for particles in Anderson
plateaus. So we can distinguish the Mott region from the superfluid
region or Bose-glass region by the local particle fluctuation.
\begin{figure}[tbp]
\includegraphics[width=\linewidth, height=4.5cm, bb=25 20 303 235]{Fig2.eps}
\includegraphics[width=5cm, height=4.5cm, bb=25 20 303 235]{Fig2c.eps}
\caption{(Color online) One-particle density matrices for systems
with 1000 sites; 200 bosons; $V_H=3\times10^{-5}$;
$\alpha=(\sqrt{5}-1)/2$; and (a)$i=501$, (b)$i=601$. (c): The
momentum distributions for systems with 1000 sites, 200 bosons,
$V_H=3\times10^{-5}$, $\alpha=(\sqrt{5}-1)/2$.} \label{Fig2}
\end{figure}
In Fig.\ref{Fig2}(a,b), we show the one-particle density matrices
for different value of $V_I$ and site. In the absence of the
incommensurate potential ($V_I=0$), the density matrix has a
power-law decay with exponent $-1/2$ \cite{Rigol}. As $V_I$
increases but is still small, the density matrices still have the
power-law decay, but the exponents are smaller than $-1/2$ and have
lots of oscillations induced by the incommensurate potential.
Power-law decay of the density matrix is the character of the system
in superfluid phase, so when $V_I$ is small, the system in the
superfluid phase. But things change when the Anderson plateaus
appear in the local average density distribution as $V_I$ increases.
The one-particle density matrices for the sites in Anderson plateaus
have an exponential-law decay, while the ones for the sites out of
the Anderson plateaus still have the power-law decay. Different from
the case without the harmonic confining trap \cite{Cai}, the systems
are not in a uniform phase. When $V_I>2$, the local average density
distribution only consists of several Anderson plateaus, and the
density matrix has the exponential-law decay which is the character
of the system in Bose-glass phase.
For system with $\widetilde{\rho}>\widetilde{\rho}_c$, the
one-particle density matrix is given by $\rho_{ij}=\delta_{ij}$ for
$i,j$ in the Mott insulator plateau, and the one for $i,j$ on the
shoulders in the density profile behaves like the density matrix of
the system in low characteristic region. In Fig.\ref{Fig2}(c), we
show the momentum distributions for different value of $V_I$. When
$V_I$ is small, the momentum distribution is similar to the case of
$V_I=0$ and has a sharp peak at $k=0$ reflecting the coherence of
the system. As $V_I$ increases further, the peak becomes more and
more shallow and the distribution spreads wider. When $V_I>2$, there
is almost no obvious peak. The system is in the Bose-glass phase
with all the effective single-particle states being localized
states, and the coherence of the system becomes small.
\begin{figure}[tbp]
\includegraphics[width=\linewidth, height=9cm, bb=25 20 303 235]{Fig3.eps}
\caption{Occupations of the natural orbitals for systems with 1000
lattice sites; 200 bosons; $\alpha=(\sqrt{5}-1)/2$;
$V_H=3\times10^{-5}$; and (a)$V=0$, (b)$V=1.2$, and (c)$V=2.4$. (d):
The amplitude of the discontinuation ($Z$) which we define as
$Z=\lambda_N-\lambda_{N+1}$ relates to $V_I$ for systems with 1000
lattice sites, 200 bosons $\alpha=(\sqrt{5}-1)/2$ and
$V_H=3\times10^{-5}$. (e,f): Profiles of the two lowest natural
orbitals for systems with 1000 lattice sites, 200 bosons,
$\alpha=(\sqrt{5}-1)/2$ and $V_H=3\times10^{-5}$.} \label{Fig3}
\end{figure}
Now we study the properties of the natural orbitals and their
occupations for the systems with the low characteristic density. The
occupations of the natural orbitals for systems with different $V_I$
are shown in Fig.\ref{Fig3}. The occupations are plotted versus the
orbital numbers $\eta$, and ordered starting from the highest
occupied one. When $V_I$ is small, the occupation distribution has a
sharp single peak at $\eta=1$ which is the feature of bosons in
contrast to the step function for fermions. With the increase in
$V_I$, the occupation of the lowest natural orbital ($\lambda_1$)
decreases. When the system is in the Bose-glass phase for $V_I>2$,
no an obvious peak appears in the lowest natural orbital.
Additionally we can find that there is a discontinuation in the
occupation distribution at $\eta=N$ when $V_I>2$. We use
$Z=\lambda_N-\lambda_{N+1}$ \cite{Cai} to characterize the
discontinuation for a Bose gas with $N$ particles. In
Fig.\ref{Fig3}(d) $Z$ versus $V_I$ is plotted. There is an obvious
change around $V_I=2$. The difference of the occupations between
systems with the low and high characteristic density is that there
is a plateau in the occupation for system in high $\widetilde{\rho}$
region due to the existence of a Mott insulator plateau. Such a
plateau is basically not changed as $V_I$ increases. In
Fig.\ref{Fig3} we also show the effect of the incommensurate
potential on the natural orbitals. We plot profiles of two lowest
natural orbitals for different value of $V_I$. Without the
incommensurate lattice, the natural orbitals are similar to the ones
of hard-core bosons in the harmonic trap without lattice. As $V_I$
increases but is still small, the natural orbitals have similar
shapes as the case of $V_I=0$ ones, but the widths of the wave
functions become smaller and smaller due to the particles becoming
more localized.
When $V_I>2$, all the natural orbitals spread only over a few
lattice sites, and they become typical Anderson localized state
\cite{Cai,Sarma}. The system is in the Bose-glass phase with
exponential-law decay one-particle density matrix
when $V_I>2$. For system with high $\widetilde{\rho}$, the amplitude
of the natural orbitals in the Mott insulator region has to be zero
(see Ref.\cite{Rigol}).
\section{DYNAMICAL PROPERTIES OF HARD-CORE BOSONS}
\begin{figure}[tbp]
\includegraphics[width=\linewidth, height=7cm, bb=25 20 303 235]{Fig4.eps}
\caption{(Color online) The evolution of the density profile(a),
momentum distribution(b), one-particle density matrix(c) with
$i=501$, occupations(d) of the natural orbitals for system with 1000
lattice sites, 100 bosons, $\alpha=(\sqrt{5}-1)/2$,
$V_H=3\times10^{-5}$, and $V_I=0$. $t$ is the time after turning off
the trap. The "fermions" curve is the momentum distribution for
system with 1000 lattice sites, 100 free fermions,
$\alpha=(\sqrt{5}-1)/2$, $V_H=3\times10^{-5}$, and $V_I=0$. Insert
of (d): The occupation of the lowest natural orbital vs $t$ for
system with 1000 lattice sites, 100 bosons, $\alpha=(\sqrt{5}-1)/2$,
$V_H=3\times10^{-5}$, and $V_I=0$.} \label{Fig4}
\end{figure}
In this section, we study the nonequilibrium dynamical properties of
expanding clouds of hard-core bosons on 1D incommensurate lattice
after turning off the harmonic trap suddenly. We first consider the
situations that the systems are in the superfluid phase with small
$V_I$ before turning off the trap. The system with $V_I=0$ has been
studied by Rigol and Muramatsu \cite{Rigol1}. In order to see
clearly the effect of a nonzero value of $V_I$ and provide an
example for comparison, we first present results for the case of
$V_I=0$ following \cite{Rigol1}. As shown in Fig.\ref{Fig4}, the
evolution of the density profile is ordinary, i.e., after turning
off the harmonic trap the density profile spreads more and more
wider as the time increases. As for the evolution of the momentum
distribution, shortly after turning off the trap, the peak at $k=0$
disappears. After a long time, the momentum distribution has similar
shape as the distribution of the noninteracting fermions which does
not change during the expansion \cite{Rigol1}. Although the momentum
distribution displays the behavior of fermionization \cite{Rigol1},
the modulus of one-particle density matrix still has the power-law
decay which is the character of the bosons in superfluid phase. Here
we note that the elements of the density matrix are complex numbers
after turning off trap. The occupations of natural orbitals
basically remain unchanged, but the occupations of the lowest
natural orbitals slightly increase during the expansion. This can be
understood as a result due to the increase in the size of the
expanding atomic cloud. The evolution of the natural orbitals has
similar behavior as the density profile which becomes more and more
wider.
\begin{figure}[tbp]
\includegraphics[width=\linewidth, height=10cm, bb=25 20 303 235]{Fig5.eps}
\caption{(Color online) The evolution of the local average density
distribution(a), momentum distribution(b), one-particle density
matrix with (c)$i=501$, (d)$i=650$, occupations(f) of the natural
orbitals for system with 1000 lattice sites, 100 bosons,
$\alpha=(\sqrt{5}-1)/2$, $V_H=3\times10^{-5}$, and $V_I=1$. (e): The
lowest natural orbitals at different time during expansion for
systems with 1000 lattice sites, 100 bosons,
$\alpha=(\sqrt{5}-1)/2$, $V_H=3\times10^{-5}$, and $V_I=1$.}
\label{Fig5}
\end{figure}
Now we consider the evolution of the system in the presence of the
incommensurate potential. For the case that the system in a weak
incommensurate potential is initially in the superfluid phase, the
evolution of the density profile is similar to the one with $V_I=0$
after turning off the harmonic trap. To give a concrete example, in
Fig.\ref{Fig5} we present the evolution of the local average density
for the system with $V_I=1$. The density profile becomes more and
more wider and the plateaus vanish during expansion. Comparing to
the density profile in Fig.\ref{Fig4}, we can see that the expansion
is much more slow as the incommensurate potential is added in which
acts as a pseudo-random potential and prevents movement of
particles. The evolution of the momentum distribution is shown in
Fig.\ref{Fig5}(b). Shortly after turning off the harmonic trap, the
peak at $k=0$ disappears. In the presence of the incommensurate
potential, there appear some small peaks on the distribution. As
time increases, the peaks of the momentum distribution become small,
and the momentum distribution evolves to the one similar to the
corresponding noninteracting fermions. The evolution of the
one-particle density matrix is shown in Fig.\ref{Fig5}(c,d). The
density matrix $|\rho_{ij}|$ has the exponential-law decay at $t=0$
for $i$ in the Anderson plateaus. For a finite $t$, it still has the
exponential-law decay despite the disappearance of Anderson plateau.
For $i$ out of the Anderson plateaus, $|\rho_{ij}|$ has the
power-law decay at $t=0$ , and it also has the exponential-law decay
after turning off the harmonic trap. In Fig.\ref{Fig5}(f) we show
the lowest natural orbitals for different time during expansion. The
natural orbitals for system at different time have different shapes,
which can be attributed to the scattering of initial wave function
by the incommensurate potential. Correspondingly the occupations of
the natural orbitals are different at different time during
expansion.
\begin{figure}[tbp]
\includegraphics[width=\linewidth, height=4cm, bb=25 20 303 235]{Fig6.eps}
\caption{(Color online) The evolution of the local average density
distribution(a) and momentum distribution(b) for system with 1000
lattice sites, 100 bosons, $\alpha=(\sqrt{5}-1)/2$,
$V_H=3\times10^{-5}$, and $V_I=2.5$. } \label{Fig6}
\end{figure}
Finally, we study the evolution of system in the BG phase with
$V_I>2$. When $V_I>2$, the system is in the BG phase with all the
effective single-particle states being Anderson-localized states.
After turning off the harmonic trap, the evolution of the system is
very slow, and basically remains unchanged even after a long time. As
an example, we show the evolutions of the local average density and
momentum distribution in Fig.\ref{Fig6}, in which we cannot find the
difference between the initial one and the one after a long time.
This clearly indicates that the system is completely pinned down by
the strong incommensurate potential. The asymmetry of the local average
density is caused by the asymmetry of the incommensurate lattice.
\section{Summary}
In summary, we have studied the properties of hard-core bosons on
incommensurate optical lattices with harmonic confining trap. Using
the Bose-Fermi mapping and the exact numerical method proposed by
Rigol and Muramatsu \cite{Rigol,Rigol1}, we calculate the
one-particle density matrices, momentum distributions, natural
orbitals and their occupations for both the static system and the
dynamic system at different time. Particularly, we exploit the phase
transition from superfluid to the localized BG phase as the strength
of the incommensurate potential increases from weak to strong, and
the nonequilibrium dynamical properties of expanding clouds of
hard-core bosons on the incommensurate lattice after turning off the
harmonic trap suddenly. For the optical lattice with harmonic
confining trap, the density profiles show obvious different
characters for systems with weak, intermediate and strong
incommensurate potential. When the strength of the incommensurate
potential becomes strong enough, Anderson plateaus are founded in
the density distribution. The shape of the local average density
profile shall not change with the increase in the strength of
incommensurate potential if it exceeds a critical value of $V_I=2$.
When the harmonic trap is suddenly switched off, the expansion
dynamics for the systems with $V_I<2$ and with $V_I>2$ also exhibits
quite different behaviors. All of these quantities give clear
signature that there exists a superfluid-to-Bose glass phase
transition in the system when the strength of incommensurate
potential exceeds $V_I=2$. Our study provides an exact static and
dynamic example which unambiguously exhibits the transition from
superfluid-to-Anderson-insulator in the incommensurate optical
lattice.
\begin{acknowledgments}
This work has been supported by NSF of China under Grants
No.10821403 and No.10974234, programs of Chinese Academy of Science,
973 grant No.2010CB922904 and National Program for Basic Research of
MOST.
\end{acknowledgments}
|
\section{INTRODUCTION}
A wealth of information on the partonic structure of the nucleon is encoded in the generalized parton
distributions (GPD's) and extensive research programs are being pursued to gain information on nucleon
GPD's.
Our strategy to this end is to determine the quark-nucleon vertex function from an investigation of nucleon
electromagnetic (em) form factors within the light-front dynamics, and then to use the obtained vertex
function to evaluate the nucleon GPD's. Indeed, light-front dynamics opens a unique possibility to study
the hadronic state in both the valence and the nonvalence sector \cite{Bro}, since within a light-front
framework no spontaneous pair production occurs and a meaningful Fock state
expansion is possible :
\begin{eqnarray} &&
| baryon \rangle = {|qqq \rangle} +
{|qqq~q \bar{q} \rangle +|qqq~g \rangle
.....}
\label{Fock}
\end{eqnarray}
As a first step, in this contribution we present our preliminary results for the unpolarized longitudinal
and transverse parton momentum distributions in the nucleon.
\section{NUCLEON VERTEX FUNCTION AND NUCLEON ELECTROMAGNETIC FORM FACTORS}
We describe the quark-nucleon vertex function through a Bethe-Salpeter amplitude (BSA),
whose Dirac structure is suggested
by an effective Lagrangian \cite{de}.
Then the symmetrized BSA for the nucleon is approximated as follows
\begin{eqnarray} &&
\hspace{-0.6cm}\Phi^\sigma_N(k_1,k_2,k_3,P)
=\nonumber \\ &&
\imath
\left [~S(k_1)~\tau_y ~ \gamma^5 ~ S_C(k_2)C ~\otimes~ S(k_3)~+
~\right. \nonumber \\ &&
\hspace{-0.0cm}S(k_3)~ \tau_y ~ \gamma^5 ~S_C(k_1)C ~\otimes~ S(k_2)~+
\nonumber \\ &&
\left.~S(k_3)~ \tau_y ~ \gamma^5 ~S_C(k_2)C~\otimes~S(k_1)
\right ] \nonumber \\ &&
\times ~~ \Lambda(k_1,k_2,k_3) ~ \chi_{\tau_N} ~ U_N(P,\sigma)
\end{eqnarray}
where $S(k)$ is the constituent quark (CQ) free propagator,
$S_C(k)$ the charge conjugated propagator,
$\Lambda(k_1,k_2,k_3)$ describes the symmetric momentum dependence of the
vertex function upon the quark momentum variables, $k_i$,
$U_N(P,\sigma)$ is the nucleon spinor and $\chi_{\tau_N}$ the isospin eigenstate.
The matrix elements of the {\em{macroscopic}} current in the spacelike (SL) region
are approximated {\em{microscopically}} by the Mandelstam formula \cite{mandel}
\vspace{0.0cm}
\begin{eqnarray} &&
\hspace{-0.7 cm}\langle \sigma',P'|j^\mu~|P,\sigma\rangle
=~3~N_c ~ \times
\nonumber \\ && \int {d^4k_1 \over (2\pi)^4}\int {d^4k_2 \over (2\pi)^4} ~
{\Large{\Sigma}} \left \{ \bar \Phi^{\sigma'}_N(k_1,k_2,k'_3,P')
\right. ~ \times
\nonumber \\ && \hspace{-0.7 cm} \left.
S^{-1}(k_1) ~ S^{-1}(k_2)~{\cal I}^\mu(k_3,q)~
\Phi^\sigma_N(k_1,k_2,k_3,P)\right \}
\end{eqnarray}
where $N_c$ is the number of colors, $k'_{3} = k_{3} + q$, and ${\cal I}^\mu(k_3,q)$
is the quark-photon vertex.
An analogous expression holds
in the timelike (TL) region.
We adopt a Breit reference frame where~$~~{\bf q}_{\perp}=0$ and
$q^+=q^0 + q^3=|q^2|^{1/2}$. Our CQ mass is ~$m=m_u = m_d = 200 ~ MeV$.
The
Mandelstam Formula is projected out by an analytic integration on $k_1^-$ and $k_2^-$,
taking into account only the poles of the propagators.
Then the current becomes the sum of a purely valence contribution (diagram (a) in Fig. 1)
and a nonvalence (NV), pair-production contribution (diagram (b) in Fig. 1).
Clearly, after the $k^-$ integrations,
the vertex functions depend only upon the light-front three-momentum.
\begin{figure}
\includegraphics[width=7.5cm]{nucfig1.eps}
\vspace{-0.6cm}
\caption{Diagrams for the SL nucleon current: (a) valence (triangle) contribution with
$0 < k_{i}^+ < P^+$, and $~0 \le k_{3}^+ + q^+ < {P'}^{+}$;
(b) nonvalence, pair-production contribution with $0 > k_{3}^+ \ge -q^+$.
A cross indicates a quark on the $k^-$-shell, i.e. $k^- = k^-_{on} = (m^2 + k^2_{\perp})/k^+$.
Solid circles and solid square represent valence and NV vertex functions, respectively
(after Ref. \cite{nucleon}). }
\end{figure}
The quark-photon vertex has isoscalar and isovector contributions
\begin{eqnarray} &&
{\cal I}^\mu=~{\cal I}^\mu_{IS} +\tau_z
{\cal I}^\mu_{IV}
\label{curr}
\end{eqnarray}
and each term in Eq. (\ref{curr}) contains a purely valence contribution
(in the SL region only) and a contribution corresponding to the pair production (Z-diagram).
In turn the Z-diagram contribution can be decomposed in a bare term $+$ a
Vector
Meson Dominance (VMD) term (according to the decomposition of the photon
state in bare, hadronic [and leptonic] contributions), viz
\begin{eqnarray} &&
\hspace{-0.7 cm} {\cal I}^\mu_{i}(k,q) = {\cal N}_{i} ~ \theta(P^+-k^+)\theta(k^+)
\gamma^\mu+\theta({q}^+ + k^+)
\nonumber \\ &&\hspace{-0.7 cm} \times ~
\theta(-k^+)~\left \{ Z_B~{\cal N}_{i} ~ \gamma^\mu+
Z^i_{VM}~\Gamma^\mu[k,q,i]\right\}
\end{eqnarray}
with $i = IS, IV$, ${\cal N}_{IS}=1/6$ and ${\cal N}_{IV}=1/2$. The constants
$Z_B$ (bare term) and $Z^i_{VM}$ (VMD term) are unknown weights
to be extracted from
the phenomenological analysis of the data.
According to $i$, the VMD term $\Gamma^\mu[k,q,i]$
includes isovector or isoscalar mesons. Indeed in \cite{nucleon} we extended to isoscalar mesons
the microscopic model for the VMD successfully used in
\cite{DFPS} for the pion form factor in the SL and in the TL region
and based on the meson mass operator of Ref. \cite{Fred}.
As explained in \cite{nucleon}, $\Gamma^\mu[k,q,i]$
does not involve free parameters. We consider up to 20 mesons for achieving convergence at high $|q^2|$.
In the valence vertexes (solid circles in Fig. 1) the spectator quarks are on the
$k^-$-shell, and the BSA
momentum dependence
is approximated through a
nucleon wave function a la Brodsky (PQCD inspired), namely
\begin{eqnarray} &&
\hspace{-0.6cm} \Psi_N(\tilde{k}_1,\tilde{k}_2,P) = P^+ ~
{ {\Lambda}(k_1,k_2,k_3)|_{(k_{1on}^-,k_{2on}^-)} \over [m_N^2 - M^2_0(1,2,3)]^{~}} =
\nonumber \\ &&
\nonumber \\ && = ~ P^+ ~ {\cal {N}}~
{~(9~m^2)^{7/2}
\over (\xi_1\xi_2\xi_3)^{p}~ \left[\beta^2 + M^2_0(1,2,3)\right]^{7/2}}~
\end{eqnarray}
where $\tilde{k}_i \equiv (k_i^+,{\bf k}_{i\perp})$, $M_0(1,2,3)$
is the free mass of the three-quark system,
$\xi_i = k^+_i/P^+$ \quad ($i=1,2,3$)
and ${\cal N}$ a normalization constant.
The power $ 7/2 $ and the parameter $ p = 0.13 $ are chosen to have
an asymptotic decrease of the triangle contribution faster than the dipole.
Only the triangle diagram determines
the magnetic moments, which are weakly dependent on $p$. Then
$\beta = 0.645$ can be fixed by the magnetic moments and we obtain $\mu_p = 2.87\pm0.02$ ~
($\mu_p^{exp}$ = 2.793)
and $\mu_n = -1.85\pm0.02$ ~ ($\mu_n^{exp}$ = -1.913).
\begin{figure}[htb]
\vspace{-0.0cm}
\hspace{-.0cm}{\includegraphics[width=7.5cm]{Ratio_prl.eps}}
\vspace{-1.3cm}
\caption{Ratio between electric and magnetic form factors for
the proton vs $-q^2$. Solid line: full calculation, sum of triangle and
pair-production terms; dotted line: triangle contribution only (after Ref. \cite{nucleon}). }
\end{figure}
\begin{figure}[htb]
\hspace{+.0cm}
{\includegraphics[width=7.5cm]{Gmp_prl.eps} }
\vspace{-1.3cm}
\caption{Magnetic proton form factor vs $-q^2$. Solid and dotted lines as
in Fig. 2. $G_D(|q^2|) = [1 + |q^2|/(0.71$ (GeV/c)$^2]^{-2}$ (after Ref. \cite{nucleon}). }
\end{figure}
For the Z-diagram contribution, the NV vertex (solid square in Fig. 1)
is needed.
It can depend on the available invariants, i.e.
on the free mass, $M_0(1,2)$, of the (1,2) quark pair
and on the free mass, $M_0(N,\bar {3})$, of the ( nucleon - quark $\bar {3}$ ) system
entering the NV vertex.
Then in the SL region we approximate the
momentum dependence of the NV vertex
$ {\Lambda}_{NV}^{SL} = {\Lambda}(k_1,k_2,k_3)|_{k^-_{1on},k^{'-}_{3on}}$ by
\begin{eqnarray}
&&\hspace{-.9cm} {\Lambda}_{NV}^{SL}= [g_{12}]^{2}[g_{N\bar {3}}]^{3/2}
\left [{k_{12}^+ \over P^{\prime +} } \right ]
\left [ P^{\prime +} \over k_{\overline {3}}^+ \right ]^r
\left [P^{+} \over k_{\overline {3}}^+ \right ]^{r}
\end{eqnarray}
with
\begin{eqnarray}
k_{12}^+ = k_1^+ + k_2^+ ~~ \quad g_{AB} = {(m_A ~ m_B) \over \left
[\beta^2+M^2_0(A,B)\right]}
\end{eqnarray}
The power 2 of $[g_{12}]^{2}$ is suggested from counting rules.
The power 3/2 of $[g_{N\bar {3}}]^{3/2}$ and the parameter $r=0.17$ are chosen
to have an asymptotic dipole behavior for the NV contribution.
\begin{figure}[htb]
\vspace{-0.0cm}
\hspace{-.0cm}
{\includegraphics[width=7.cm]{Gen_prl.eps}}
\vspace{-0.9cm}
\caption{Electric neutron form factor vs $-q^2$.
Solid and dotted lines as in Fig. 2 (after Ref. \cite{nucleon}). }
\end{figure}
\begin{figure}[htb]
\hspace{+.0cm}
{\includegraphics[width=7.cm]{Gmn_prl.eps} }
\vspace{-0.9cm}
\caption{The same as in Fig. 3, but for the neutron (after Ref. \cite{nucleon}). }
\end{figure}
Analogous expressions with the same parameters are used for the nonvalence vertexes in the TL region
(see Ref. \cite{nucleon}
for the explicit expressions).
We perform a fit of our free parameters, $Z_B$, $Z^i_{VM}$, $p$, $r$ in the SL region
and obtain the form factors shown in Figs. 2 - 5,
with a $\chi ^2$/datum = 1.7. As a result
the weights for the pair-production terms
are $Z_B = Z_{VM}^{IV} = 2.283$ and
$Z_{VM}^{IS} / Z_{VM}^{IV} = 1.12$, remarkably close to one.
The same values of our weight parameters are adopted to evaluate the TL form factors shown in Figs. 6 and 7.
Preliminary results of our model for the nucleon form factors were presented in \cite{MFPPS}.
The Z-diagram turns out to be essential for the description of the form factors in our reference frame
with $q^+ > 0$.
In particular the possible zero in $G_E^p/G_M^p$ is strongly related to the pair-production contribution.
In the TL region our parameter free calculations give a fair description of the proton data,
apart for some missing strength at $q^2 = 4.5$ (GeV/c)$^2$ ~ and ~ $q^2 = 8$ (GeV/c)$^2$ (as
occurs for the pion
case \cite{DFPS}), which one could argue to be due to possible unknown vector mesons, missing
in the spectrum of Ref. \cite{Fred}.
\begin{figure}[htb]
\centering
\includegraphics[width=7.cm]{TLP_4a.eps}
\vspace{-.9cm}
\caption{Proton effective form factor $G^p_{eff}(q^2)/G_D(q^2)$ in the timelike region vs $q^2$.
Solid line: bare +
VMD; dotted line: bare term.
$G_{eff}(q^2) = [(|G_M(q^2)|^2 ~ - ~ \eta ~|G_E(q^2)|^2 ) /
(1 ~ - ~ \eta)]^{0.5}$ with $\eta = - 2 m_N^2 / q^2$ (after Ref. \cite{nucleon}). }
\end{figure}
\begin{figure}[htb]
\includegraphics[width=7.cm]{TLN_4b.eps}
\vspace{-.9cm}
\caption{The same as in Fig. 5, but for
the neutron. Dashed line: solid line arbitrarily multiplied by 2 (after Ref. \cite{nucleon}). }
\end{figure}
\section{LONGITUDINAL AND TRANSVERSE QUARK MOMENTUM DISTRIBUTIONS IN THE NUCLEON}
The longitudinal distribution $q(x)$ is the limit in
the forward case, $P' = P$, of the unpolarized
generalized parton distribution ${H}^q(x,\xi,t)$.
Indeed one can define the distributions ${H}^q(x,\xi,t)$ and $E^q(x,\xi,t)$ for the quark $q$
through the following relation
\begin{eqnarray}
\label{gpdisos}
\frac{1}{2P^+} \bar{u}(P',\lambda')\left [ {H}^q~ \gamma^+ +
{E}^q ~\imath \frac{\sigma^{+\alpha}q_{\alpha}}{2 M} \right ] {u}(P,\lambda)= \nonumber
\\ \int \frac{dz^-}{4\pi} e^{i x {\cal{P}}^+ z^-}
\left . \langle P',\lambda' | \bar \psi_q
(-\frac{z}{2} ) \gamma ^+ \, \psi_q (\frac{z}{2} ) | P,\lambda
\rangle \right |_{\tilde{z}=0} \nonumber
\end{eqnarray}
where $\lambda$ is the nucleon helicity,
$\tilde z\equiv \{z^+, {\bf z}_{\perp}\}$ ,
$\psi_q(z)$ is the quark field isodoublet, while
\begin{eqnarray} &&
{\cal{P}}=\frac12(P'+P)
\;\; \quad \quad t=(P' - P)^2
\label{kin1}
\end{eqnarray}
\begin{eqnarray} &&
\xi=-\frac{P'^+ - P^+}{2 {\cal{P}}^+} \quad \quad
x=\frac{k^+}{{\cal{P}}^+}
\label{kin2}
\end{eqnarray}
and $k$ is the average
momentum of the quark that interacts with the photon, i.e.
\begin{eqnarray} &&
k = \frac{k_3 + (k_3 +q)}{2}
\end{eqnarray}
For $P' = P$, both $q^+$ and $\xi$ are vanishing and $x = k_3^+ / P^+ = \xi_3$
coincides with the fraction of the longitudinal momentum carried by the active quark, i.e., with the
Bjorken variable. As a consequence the function ${H}^q(x,\xi,t)$ reduces to the longitudinal parton
distribution function $q(x)$ :
\begin{eqnarray}
\hspace{-2cm} {H}^q(x,0,0) = q(x) = \int d{\bf k}_{\perp} ~~ f_1^q(x,k_{\perp}) =&&
\label{struc} \\
\int \frac{dz^-}{4\pi} e^{i x P^+ z^-}
\left . \langle P | \bar \psi_q
(-{z\over 2}) \gamma ^+ \, \psi_q ( {z\over 2}) | P \rangle \right |_{\tilde{z}=0} \nonumber
\end{eqnarray}
where
an average on the nucleon helicities is understood.
Once all the parameters of the nucleon light-front wave function
$\Psi_N(\tilde k_1,\tilde k_2,P)$ have been determined, one can easily define
the transverse-momentum-dependent distributions
of the active quark
in terms of the nucleon light-front wave function :
\begin{eqnarray} &&
\hspace{-0.6cm} f_1^u(x,k_{\perp})=-{9~N_c\over 32~(2 \pi)^6}~{1\over P^{+2}} ~ \times ~\nonumber \\ &&
\hspace{-0.7cm}\int^{1 - x}_0 d\xi_2 {1 \over (1 - x -\xi_2) \xi_2} {1 \over
x^2}
\int d{\bf k}_{2 \perp} \left. {\cal H}_{u}
\right|_{(k^-_{1on},k^-_{2on})}
\nonumber \\ &&
\nonumber \\ &&
\hspace{-0.3cm} \times ~ |\Psi_N(P^+ ~ \xi_1,{\bf
k}_{1\perp},P^+ ~ \xi_2,{\bf k}_{2\perp},P)|^2
\label{partutr}
\end{eqnarray}
\begin{eqnarray} &&
\hspace{-0.6cm} f_1^d(x,k_{\perp})={9~N_c\over 8~(2 \pi)^6}~ {1\over P^{+2}} ~ \times ~\nonumber \\ &&
\hspace{-0.7cm}\int^{1 - x}_0 d\xi_2 {1 \over (1 - x-\xi_2) \xi_2} {1 \over
x^2}
\int d{\bf k}_{2 \perp} \left. {\cal H}_{d}
\right|_{(k^-_{1on},k^-_{2on})}
\nonumber \\ &&
\nonumber \\ &&
\hspace{-0.3cm}\times ~ |\Psi_N(P^+ ~ \xi_1,{\bf
k}_{1\perp},P^+ ~ \xi_2,{\bf k}_{2\perp},P)|^2~~
\quad
\label{partdtr}
\end{eqnarray}
where ${\cal H}_{u}$ and ${\cal H}_{d}$ are
proper traces of propagators and of the
currents ${\cal {I}}^+_u$ and ${\cal {I}}^+_d$, respectively.
From the nucleon light-front wave function
$\Psi_N(\tilde k_1,\tilde k_2,P)$ one can easily define through Eq. (\ref{struc})
also the longitudinal distribution of the stuck quark and from the isospin symmetry one has
\begin{eqnarray} &&
\hspace{-0.7cm}u_p(x)=d_n(x)=u(x); ~
d_p(x)=u_n(x)= d(x)
\end{eqnarray}
Our preliminary results for $f_1^{u(d)}(x,k_{\perp})$ in the proton and for $u(x)$ and $d(x)$ are
shown in Figs. 8, 9 and in Figs. 10, 11, respectively.
\begin{figure}[htb]
\vspace{-.8cm}
\hspace{-.0cm}{\includegraphics[width=7.5cm]{f1pnew.eps}}
\vspace{-2.5cm}
\caption{Valence transverse-momentum distributions for a $u$ quark inside
the proton. $G(k_{\perp}) = (1 ~ + ~ k_{\perp}^2/m_{\rho}^2)^{-5.5}$,
$m_{\rho}$ = 770 MeV and $k_{perp} = | k_{\perp} |$. }
\end{figure}
\begin{figure}[htb]
\vspace{-.8cm}
{\includegraphics[width=7.5cm]{f1nnew.eps}}
\vspace{-2.5cm}
\caption{The same as in Fig. 8, but for a $d$ quark inside
the proton. }
\end{figure}
It can be observed that the decay of our $f_1(x,k_{\perp})$ vs $k_{\perp}$
is faster than in diquark models of nucleon
\cite{Jacob}, while it is slower than in factorization
models for the transverse momentum distributions \cite{Anselm}.
As far as the longitudinal momentum distributions are concerned,
a reasonable agreement of our $u(x)$ with the CTEQ4 fit to the experimental data \cite{Lai}
can be seen in Fig. 10.
\begin{figure}[htb]
\centering
\hspace{.0cm}{\includegraphics[width=7.cm]{PaceE-fig3.eps}}
\vspace{-.9cm}
\caption{Longitudinal momentum distribution for a $u$ quark inside
the proton. Dashed lines: our preliminary results; thick solid lines: our results after
evolution to $Q^2$ = 1.6 (GeV/c)$^2$; thin solid lines: CTEQ4 fit to the experimental data \cite{Lai}.}
\end{figure}
\begin{figure}[htb]
\hspace{.0cm}{\includegraphics[width=7.cm]{PaceE-fig4.eps}}
\vspace{-.9cm}
\caption{The same as in Fig. 10, but for a $d$ quark inside
the proton. }
\end{figure}
\section{CONCLUSIONS}
A microscopical model for hadron em form
factors in both SL and TL region has been proposed. In our model
the quark-photon vertex for the process
where a
virtual photon materializes in a $q \bar{q}$ pair
is approximated by a microscopic VMD model plus a bare term.
Both for the pion
and the nucleon good results are obtained in the SL region.
The Z-diagram (i.e. higher Fock state component) has been shown to be essential,
in the adopted reference
frame ($q^+ \ne 0$).
The possible zero in $G^p_E
\mu_p/G^p_M$ turns out to be related to the pair-production contribution.
In the TL region our calculations give a fair description of the proton and pion data,
although some strength is lacking for
$q^2 = 4.5$ ~ (GeV/c)$^2$ ~ and ~ $q^2 = 8$ ~ (GeV/c)$^2$.
The analysis of nucleon form factors allows us to get
a phenomenological Ansatz for the nucleon LF wave function,
which reflects the asymptotic behaviour suggested by the one-gluon-exchange dominance.
This LF wave function is then
used to evaluate the unpolarized transverse momentum distributions
and the longitudinal momentum distributions of the quarks in the nucleon.
Our next step will be the calculation of polarized transverse and longitudinal momentum distributions.
|
\section{Introduction}
The spectrum of the radiation field emitted by a system of atoms interacting with an electromagnetic field is known to provide fundamental insight into the energy structure and properties of the combined (entangled) atom-field system~\cite{sz97}. The spectral distribution of light emitted by coherently driven atoms simultaneously coupled to a multi-mode field has been the subject of numerous investigations over the years. The subject received its initial stimulus in the famous paper by Mollow~\cite{m69}, in which he showed that the fluorescence spectrum of a strongly driven two-level atom has a three-peaked structure, quite different from the spectrum expected for weak-field resonance fluorescence~\cite{km76}. The radiation properties of atoms interacting with a single rather than multi-mode electromagnetic field were also investigated in the context of a search for signatures of quantum effects in the matter-radiation interaction~\cite{be94}. The single-mode interaction, also known as the Jaynes-Cummings model, can be realized by confining the atoms within an optical cavity. The atom-cavity field coupling leads to "dressed" states, the eigenstates of the Hamiltonian of the atoms plus the cavity field, that serve as the basis energy states of the system~\cite{ct92}. The physical origin of the spectral features observed in the radiation field emitted by the system is clearly explained the dressed-atom model which provides the positions, intensities and widths of the spectral features in analytic form. In the dressed-atom approach, the radiation field is quantized which, in addition to the convenient studies of the spectral features, offers a potential setting for investigations of statistical and quantum features of the system. Many novel quantum effects have been predicted in observed, such as atomic collapses and revivals~\cite{sm83}, vacuum Rabi splitting~\cite{ag84,cb89}, photon antibunching and squeezing~\cite{rt91,th95}. The spectral properties of strongly driven atoms inside a single-mode cavity were also investigated in context of searching for signatures of two-photon lasing~\cite{gw92}, the photon-number distribution of the cavity field~\cite{qf93} and sub-natural linewidths~\cite{lg97,fq93}.
The possibility of modifying linewidths of the spectral lines of the fluorescence field emitted by a three-level $V$-type atom using intense laser fields was first noted by Lorenzo Narducci and his colleagues~\cite{nso90}. They have found that the linewidth of the spectral lines on a given atomic transition can be switched into the linewidth of the other, much narrow atomic transition.
The phenomenon, termed as a dynamical narrowing of the spectral lines, can also be best understood in terms of the dressed-states of the system. The laser fields driving the atomic transitions lead to a coherent mixing of the atomic energy levels resulting in dressed states of the system. The transition rate between any two dressed states is proportional to the absolute square of the dipole transition moment connecting these two states. Since the dressed states depend on the Rabi frequencies of the driving fields, this results in the transition rates dependent on the Rabi frequencies and the spontaneous emission rates of the atomic transitions. Thus, the transition rates can be changed by changing the ratio between the Rabi frequencies.
Basically, the coherent mixing of the atomic energy levels modifies quantum fluctuations on the driven atomic transitions. Thus, the observed subnatural linewidth of the spectral line is the signature of the fluctuations stabilization in a driven atom by a coherent mixing of the atomic energy levels. In this context, the phenomenon of the dynamical narrowing of the spectral lines, predicted by Narducci et al. offers further interesting prospects for study of practical schemes for sub-natural spectral resolution and stabilization of the quantum fluctuations~\cite{gzm91}.
Narrowing of the spectral lines can also be achieved by placing the driven system inside a frequency dependent reservoir, the so called, tailored vacuum~\cite{kk95,lm88}. Transition rates between dressed states of the driven system can be modified with a suitable choosing of the Rabi frequency of the driving field. If the Rabi frequency is chosen to be much larger than the bandwidth of the reservoir field, some of the dressed-atom frequencies can then be found outside the bandwidth of the reservoir field. As a result, transitions at these frequencies are not affected by spontaneous emission. The selective decoupling of the dressed-atom transition frequencies may lead to narrowing and even suppression of the spectral lines of the fluorescence.
There have also been a considerable interest in the generation of lasing action between dressed states of strongly driven two-level atoms~\cite{zl91,lz90,lb91}. The concept of the dressed-atom laser has been introduced, where the lasing action can occur even in the absence of population inversion between the bare states of the atoms. This kind of lasing is often referred to as lasing without population inversion. In fact, the lasing occurs with population inversion which exists between the dressed states involved in the lasing transition.
Recently, the spectral and statistical properties of the single-atom dressed-state laser in a photonic band-gap material were shown to be qualitatively different from that arising from the ordinary frequency independent reservoir~\cite{fj04,fl06,tl08}. The band-gap material serves as a frequency dependent reservoir, where the density of the electromagnetic field modes varies sharply with frequency. Furthermore, it has been shown that in the presence of the photonic material, the dressed-atom system can operate as a thresholdless laser~\cite{ll09}.
In this paper, we examine the atomic fluorescence and the cavity field output spectra for features indicative of the thresholdless lasing which results in the single-atom dressed-state laser from suppression of spontaneous emission on the lasing transition~\cite{ll09}. Our interest in this particular problem stems from the fact that the thresholdless behavior of the system has been predicted by analyzing the photon statistics of the cavity field~\cite{rc94,tp94,dm88,bk94,pd99,sh99}. It is interesting to consider if a thresholdless behavior could also be observed in the spectra.
The predicted suppression of the threshold behavior in the statistics of the cavity photons occurs if a photonic band-gap material is employed to reduce the density of the electromagnetic field modes in the lasing region of the transition. As is well known, a strongly driven two-level atom fluorescence at three distinct frequencies, the driving laser frequency and at two sideband frequencies shifted symmetrically from the laser frequency by the Rabi frequency.
We consider an off-resonance driving field and the cavity mode tuned to resonance with the lower frequency Rabi sideband of the dressed atom. Our treatment is based on the master equation of the reduced density operator of the driven atom plus the cavity field, which we derive by combining two previously used approaches, the Keitel-Knight-Narducci-Scully~\cite{kk95} approach to the dynamics of a driven atom in a frequency dependent reservoir, and Freedhoff-Quang~\cite{fq93} approach to the dynamics of a driven atom coupled simultaneously to a single-mode cavity field and a frequency independent reservoir. In the absence of the band-gap material, we distinguish three regimes of the pumping rate, below, near and above threshold, at which the spectra exhibit significantly different structures. Below the threshold, the spectrum exhibits two peaks characteristic of the vacuum Rabi splitting. Near the threshold, a multi-peak structure is observed indicating a successive excitation of the discrete energy states of the system. Above the threshold, the cavity field spectrum is composed of a single very narrow peak, whereas the atomic fluorescence spectrum exhibits a Mollow-like triplet.
In the presence of the band-gap material, the spectra display interesting modifications. First of all, we can distinguish only two ranges of the pumping rate at which the spectra have different character. A multi-peak structure emerges at relatively low pumping rates. The spectra exhibit a multi-peak structures even in the limit of vanishing pumping. Secondly, we find that the spectral lines are significantly narrower than that predicted in the absence of the band-gap material. We observe an interesting feature that in the case of the cavity field spectrum, the multi-peak structure builds inside the vacuum Rabi doublet, whereas the fluorescence spectrum the structure builds inside as well as outside the doublet. Finally, we demonstarte that the cavity field spectrum exhibits a threshold behavior that at a certain pumping rate, the multi-peak spectrum converts into a single very narrow peak located at the cavity mode frequency. The threshold behavior exists despite the fact that there is no threshold behavior in the photon statistics of the cavity field.
The presence of the threshold raises a question of the meaning of a thresholdless laser, whether the thresholdless operation of the system should be determined by either, the photon statistics or the spectrum of the cavity field, or both.
\section{The dressed-atom system}
We consider a system consisting of an atom located within a single-mode cavity engineered inside a photonic crystal. The atom is modeled as a two-level system with lower level $\ket 1$ and upper level $\ket 2$ separated by energy $\hbar\omega_{0}$, and driven by an intense single-mode laser field of the resonant Rabi frequency $2\Omega_{0}$. The laser field propagates transverse to the cavity axis and is treated classically, while the cavity mode is treated as a quantum field in a state $\ket n$ containing $n$ photons of frequency $\omega_{c}$.
The reason for engineering the cavity inside a photonic crystal is twofold. On the one hand, the photonic crystal provides a system for a strong coupling of the atom to a previlage mode. On the other hand, the crystal provides a reservoir of a multi-mode electromagnetic field whose modal density varies sharply on a frequency scale much smaller than the Rabi frequency of the driving field~\cite{wj04,ya87,jo87,jq94}. Theoretically, the variation of the density of the field modes with frequency is expressed in terms of a step function that varies sharply at the cut-off frequency $\omega_{b}$.
The dynamics of the total system; the driven atom plus the cavity plus the reservoir, is determined by the density operator $\rho_{T}$, which in the interaction picture satisfies the master equation
\begin{eqnarray}
\frac{\partial }{\partial t}\tilde{\rho}_{T} = -\frac{i}{\hbar}\left[\tilde{H},\tilde{\rho}_{T} \right] +\frac{1}{2}\kappa{\cal L}_{c}\tilde{\rho}_{T} ,\label{e1}
\end{eqnarray}
where
\begin{eqnarray}
\tilde{\rho}_{T} = \exp\left(-\frac{i}{\hbar}H^{\prime}t\right)\rho_{T}\exp\left(\frac{i}{\hbar}H^{\prime}t\right) \label{e2}
\end{eqnarray}
is the density operator of the system in the interaction picture, with
\begin{eqnarray}
H^{\prime} = \hbar\omega_{L} a^\dagger a +\frac{1}{2}\hbar\omega_{L}\!\left(\sigma_{22} -\sigma_{11}\right) + \hbar\sum_{k}\omega_{L}a_{k}^{\dag}a_{k} ,\label{e3}
\end{eqnarray}
and
\begin{eqnarray}
{\cal L}_{c}\tilde{\rho}_{T} = \left(2a\tilde{\rho}_{T}{a}^\dag-{a}^\dag{a}\tilde{\rho}_{T} -\tilde{\rho}_{T}{a}^\dag{a}\right) \label{e4}
\end{eqnarray}
is an operator representing the damping of the cavity field with the rate $\kappa$.
The Hamiltonian $\tilde{H}$, appearing in Eq.~(\ref{e1}), is composed of three terms
\begin{eqnarray}
\tilde{H} = \tilde{H}_{0}+\tilde{H}_{L}+\tilde{H}_{I} ,\label{e5}
\end{eqnarray}
where
\begin{eqnarray}
\tilde{H}_{0} = \frac{1}{2}\hbar\Delta_{a}\left(\sigma_{22} -\sigma_{11}\right) + \hbar\Delta_{c}a^{\dag}a +\hbar\sum_{k}\Delta_{k}a_{k}^{\dag}a_{k} \label{e6}
\end{eqnarray}
is the free Hamiltonian of the atom, the cavity field, and the multi-mode reservoir, respectively,
\begin{eqnarray}
\tilde{H}_{L}= \hbar\Omega_{0}\left(\sigma_{12}+\sigma_{21}\right) \label{e7}
\end{eqnarray}
is the interaction Hamiltonian of the driving laser field with the atom, and
\begin{eqnarray}
\tilde{H}_{I} = i\hbar g\left(a^{\dag}\sigma_{12}-\sigma_{21}a\right)
+i\hbar\sum_{k}g_{k}\left(a^{\dag}_{k}\sigma_{12}-\sigma_{21}a_{k}\right) \label{e8}
\end{eqnarray}
is the interaction Hamiltonian of the privilege cavity mode and the remaining reservoir modes with the atom.
In Eqs.~(\ref{e1})-(\ref{e8}), the parameters $g$ and $g_{k}$ describe the coupling constants between the atom and the cavity mode and the reservoir modes, respectively, $a\ (a^{\dag})$ and $a_{k}\ (a^{\dag}_{k})$ are, respectively, the cavity-mode and the radiation reservoir annihilation (creation) operators, $\sigma_{ij} =\ket i \bra j \ (i,j=1,2)$ are the bare atomic operators, and $2\Omega_{0}$ is the resonant Rabi frequency of the laser field which, for simplicity, we have assumed is real and positive. The remaining parameters $\Delta_{a}=\omega_{a}-\omega_{L}$, $\Delta_{c}=\omega_{c}-\omega_{L}$, and $\Delta_{k}=\omega_{k}-\omega_{L}$ are the detunings of the atomic resonance frequency $\omega_{0}$, the cavity-mode frequency $\omega_{c}$, and the frequency $\omega_{k}$ of the $k$th mode of the radiation reservoir from the frequency $\omega_{L}$ of the driving laser field.
There are three fields interacting with the atom, the laser field whose the coupling strength to the atom is given by the Rabi frequency $\Omega_{0}$, the cavity field coupled to the atom with the strength $g$ and the multi-mode vacuum reservoir whose the $k$th mode is coupled to the atom with the strength $g_{k}$. We shall assume that the coupling $\Omega_{0}$ dominates over $g$ that, on the other hand, is much stronger than $g_{k}$, so that $\Omega_{0}\gg g\gg g_{k}$. We shall follow the following procedure, first "dressing" the atom in the laser field and then couple the resulting dressed-atom system to the cavity and the reservoir fields.
As we shall be interested in the spectral properties of the radiation field produced by a dressed-state laser, we make the customary transformation of representing the master equation in terms of dressed states of the driven atom~\cite{ct92}. The dressed states are the eigenstates of the Hamiltonian $\tilde{H}_{AL}= \frac{1}{2}\hbar\Delta_{a}\left(\sigma_{22} -\sigma_{11}\right)+\tilde{H}_{L}$. Since the driving field is treated classically in our calculations, the diagonalization of the Hamiltonian $\tilde{H}_{AL}$ produces the so-called semiclassical dressed states
\begin{eqnarray}
\ket{\tilde{1}} = \cos\phi\ket 1 -\sin\phi\ket 2 ,\quad \ket{\tilde{2}} = \sin\phi\ket 1 +\cos\phi\ket 2 ,\label{e9}
\end{eqnarray}
where $ \cos^{2}\phi = \frac{1}{2}\left[1+\Delta_{a}/(2\Omega)\right]$ and $2\Omega=(4\Omega_{0}^{2}+\Delta_{a}^{2})^{1/2}$ is the Rabi frequency of the detuned field.
We now introduce dressed-state operators, $R_{ij}=\ket{\tilde{i}}\bra{\tilde{j}}$, and find that the Hamiltonian $\tilde{H}_{0}+\tilde{H}_{L}$ becomes
\begin{eqnarray}
\tilde{H}_{0} +\tilde{H}_{L} \equiv \tilde{H}_{d} = \hbar\Delta_{c}a^{\dag}a +\hbar\Omega R_{3}+\sum_{\lambda}\Delta_{\lambda}a_{\lambda}^{\dag}a_{\lambda} ,\label{e10}
\end{eqnarray}
where $R_{3}=R_{22}-R_{11}$. Note that the Hamiltonian (\ref{e10}) is diagonal in the basis of the product states $\ket{\tilde{i}}\otimes\ket n \otimes\ket{n_{k}}$, which prompts us to make the unitary transformation
\begin{eqnarray}
\bar{\rho}_{T} = \exp\left(-\frac{i}{\hbar}H_{d}\, t\right)\tilde{\rho}_{T}\exp\left(\frac{i}{\hbar}H_{d}\, t\right) ,\label{e11}
\end{eqnarray}
which results in the density operator in the dressed-atom picture. It follows from Eqs.~(\ref{e1}) and~(\ref{e11}) that the maser equation for the density operator in the dressed-atom picture becomes
\begin{eqnarray}
\frac{\partial }{\partial t}\bar{\rho}_{T}\left( t\right) = -\frac{i}{\hbar}\left[\bar{H}_{I},\bar{\rho}_{T}\left( t\right) \right] ,\label{e12}
\end{eqnarray}
where
\begin{align}
\bar{H}_{I} &= g\, a^{\dag}\!\left[\frac{1}{2}\sin(2\phi)R_{3}\,{\rm e}^{i\Delta_{c}t}+\cos^{2}\!\phi R_{12}\,{\rm e}^{i(\Delta_{c}-2\Omega)t}\right.\nonumber\\
&\left. -\sin^{2}\!\phi R_{21}\,{\rm e}^{i(\Delta_{c}+2\Omega)t}\right]
+\sum_{k}g_{k}\, a_{k}^{\dag}\!\left[\frac{1}{2}\sin(2\phi)R_{3}\,{\rm e}^{i\Delta_{k}t}\right.\nonumber\\
&\left. +\cos^{2}\!\phi R_{12}\,{\rm e}^{i(\Delta_{k}-2\Omega)t}-\sin^{2}\!\phi R_{21}\,{\rm e}^{i(\Delta_{k}+2\Omega)t}\right] + {\rm H.c.} \label{e13}
\end{align}
Note that in the dressed-atom picture the evolution of the density operator is determined only by the interaction Hamiltonian (\ref{e8}).
We now proceed to eliminate the reservoir field by tracing the density operator of the total system over the space of the reservoir modes to obtain the reduced density operator of the dressed atom and the cavity mode only. Since the multi-mode field appears as a frequency dependent reservoir, we assume that the Rabi frequency is much larger than the width of the band edge of the band-gap material is much smaller than the spontaneous emission rate of the atomic bare transition and the dressed-atom transition frequencies $\omega_{L}$ and $\omega_{L}\pm 2\Omega$ are far away from the the band-edge frequency $\omega_{b}$. In this case, it is possible to obtain a master equation for the reduced density operator consistent with the Born-Markov approximation. In addition, we assume that the cavity mode is tuned to exact resonance with the low-frequency Rabi sideband, i.e. $\Delta_{c}=-2\Omega$. With this frequency tuning, we can distinguish that certain terms in the master equation are stationary in time and the other oscillating at frequencies $2\Omega$ and $4\Omega$. Since the Rabi frequency is large, these oscillatory terms make a negligible contribution to the dynamics of the system and we may ignore them. This simplification is a form of the rotating-wave approximation. With such approximations, and after using the standard procedure of eliminating of the reservoir modes, we find that the reduced density operator $\rho$ of the dressed atom plus the cavity mode has the form
\begin{align}
\frac{\partial{\rho}}{\partial{t}} &= g_{1}\!\left[\left(R_{12}a-{a}^\dag{R_{21}}\right),\rho\right] +\frac{1}{2}\kappa{\cal L}_{c}\rho \nonumber\\
&+\frac{1}{8}\gamma_{0}\!\left(2R_3\rho{R_3}-\rho{R_3^2}-R_3^2\rho\right) \nonumber\\
&+\frac{1}{2}\gamma_{-}\!\left(2R_{21}\rho{R_{12}}-R_{12}R_{21}\rho-\rho{R_{12}R_{21}}\right) \nonumber\\
&+\frac{1}{2}\gamma_{+}\!\left(2R_{12}\rho{R_{21}}-R_{21}R_{12}\rho-\rho{R_{21}R_{12}}\right) ,\label{e14}
\end{align}
where $g_1 = g\sin^{2}\phi$ is the "effective" coupling constant of the dressed atom to the cavity field,
\begin{eqnarray}
\gamma_{0} &=& \gamma \sin^{2}(2\phi) u(\omega_{L}-\omega_{b}) ,\ \
\gamma_{-} = \gamma \sin^{4}(\phi) u(\omega_{-}-\omega_{b}) ,\nonumber \\
\gamma_{+} &=& \gamma \cos^{4}(\phi) u(\omega_{+} -\omega_{b}) ,\label{e15}
\end{eqnarray}
are the damping rates between the dressed states of the system, and $u(\omega_{i} -\omega_{b}),\, \omega_{i}=\omega_{L},\omega_{\pm}$, is the unit step function describing the transfer function of the band-gap material. The function is evaluated at three frequencies corresponding to the three transition frequencies between the dressed states, $\omega_{i} =\omega_{L}$ and $\omega_{i} =\omega_{\pm}=\omega_{L}\pm 2\Omega$. Since the frequencies $\omega <\omega_{b}$ are forbidden in the band-gap material, this opens a possibility to eliminate spontaneous transitions at selected frequencies of the dressed-atom system.
The master equation (\ref{e14}) is very similar in form to that one found for the the frequency independent reservoir. In particular, the damping rates $\gamma_{0}, \gamma_{+}$ and $\gamma_{-}$ differ only in that the damping rate $\gamma$ of the bare atomic transition is replaced by $\gamma u(\omega_{i} -\omega_{b})$, the frequency dependent rates. Thus, modifications to the master equation brought by the band-gap material are essentially reflected in the appearance of three, frequency dependent, damping rates. As we shall see, the modifications, although look trivial at the first glance, will in fact significantly affect the dynamics of the system and the spectral distribution of the emitted light.
The damping rate $\gamma_{+}$ corresponds to spontaneous emission from the upper dressed
state $\ket{\tilde{2}}$ to the lower dressed state $\ket{\tilde{1}}$ of the manifold below and occurs at frequency $\omega_{+}=\omega_{L}+2\Omega$, whereas the damping rate $\gamma_{-}$ corresponds to spontaneous emission from the lower dressed state to the upper dressed state of the manifold below and occurs at frequency $\omega_{-}=\omega_{L}-2\Omega$.
It should be noted that the rate $\gamma_{-}$ appears as the decay rate on the transition resonant to the cavity frequency, whereas the rate~$\gamma_{+}$, that appears as a damping rate plays, in fact, the role of an incoherent pumping of the dressed system from $|\tilde{2}\rangle$ to $|\tilde{1}\rangle$. In other words, it is a pure incoherent pumping process that transfers the population to the upper state of the lasing transition.
Thus, by a suitable tuning of the dressed-atom frequency $\omega_{-}$ to the edge frequency $\omega_{b}$, one can eliminate spontaneous emission at the frequency of the cavity field.
We now introduce density-matrix elements with respect to the dressed states, denoting $\rho_{ij}= \bra{\tilde{i}}\rho\ket{\tilde{j}}$, and using the master equation (\ref{e14}), we obtain the following simple equations of motion for the populations and coherence
\begin{align}
\frac{\partial}{\partial{t}}{\rho_{11}} &= g_{1} \left(a\rho_{21}+\rho_{12}a^{\dagger}\right) -\gamma_{-}\rho_{11} +\gamma_{+}\rho_{22} +\frac{1}{2}\kappa{\cal L}_{c}\rho_{11} ,\nonumber\\
\frac{\partial}{\partial{t}}{\rho_{22}} &= -g_{1}\left(a^{\dagger}\rho_{12}+\rho_{21}a\right) -\gamma_{+}\rho_{22} +\gamma_{-}\rho_{11} +\frac{1}{2}\kappa{\cal L}_{c}\rho_{22} ,\nonumber\\
\frac{\partial}{\partial{t}}{\rho_{12}} &= g_{1}\left(a\rho_{22}-\rho_{11}a\right)
-\Gamma_{c}\,\rho_{12} +\frac{1}{2}\kappa{\cal L}_{c}\rho_{12} ,\nonumber\\
\frac{\partial}{\partial{t}}{\rho_{21}} &= g_{1}\left(\rho_{22}a^{\dagger}-a^{\dagger}\rho_{11}\right)
-\Gamma_{c}\,\rho_{21} +\frac{1}{2}\kappa{\cal L}_{c}\rho_{21} ,\label{e16}
\end{align}
where $\Gamma_{c} = (\gamma_{0}+\gamma_{+} +\gamma_{-})/2$ is the decay rate associated with the $\rho_{12}$ coherence.
Equations (\ref{e16}) are the basic equations for calculating the spectra of the atomic fluorescence and the cavity field. In the following we always assume that the Rabi frequency of the laser field is much larger than the coupling constant of the cavity mode and the damping rates of the atom and the cavity mode, $2\Omega \gg g,\gamma,\kappa$. Under such conditions, we can selectively tune the dressed-atom frequencies to the transmissive or to the forbidden band of the band-gap material.
\section{Atomic fluorescence and cavity field spectra}
We now follow the procedure of Freedhoff and Quang~\cite{fq93} to calculate the atomic fluorescence and cavity field spectra. The evaluation of the spectra simplifies by introducing
the following Hermitian combination of the density matrix elements
\begin{eqnarray}
\rho^{(1)}&=& \rho_{22}+\rho_{11} ,\ \rho^{(2)} = \rho_{22}-\rho_{11} ,\nonumber\\
\rho^{(3)}&=& \frac{1}{2}\left(\rho_{21}a+a^{\dagger}\rho_{12}\right) ,\nonumber\\
\rho^{(4)}&=& \frac{1}{2}\left(a\rho_{21}+\rho_{12}a^{\dagger}\right) .\label{e17}
\end{eqnarray}
The advantage of working with the Hermitian combinations is the most evident if one considers the photon-number representation, ${\rho^{(i)}_{n,n+m}}\equiv\langle{n}|\rho^{(i)}|{n+m}\rangle$, where it can be easily found that the equations of motion for the diagonal matrix elements decouple from the equations of motion for the off-diagonal elements.
Using Eq.~(\ref{e16}), we find that the equations of motion for the density matrix elements ${\rho^{(i)}_{n,n+m}}$ satisfy a set of coupled differential equations, which can be written in a matrix form as
\begin{eqnarray}
\frac{\partial}{\partial{t}}Z^{(m)}_{n} = {\bf A}^{(m)}_{n}Z^{(m)}_{n-1}+{\bf B}^{(m)}_{n}Z^{(m)}_{n}+{\bf C}^{(m)}_{n}Z^{(m)}_{n+1} ,\label{e18}
\end{eqnarray}
where $Z^{(m)}_{n}$ is the column vector composed of the density matrix elements, and ${\bf A}^{(m)}_{n}$, ${\bf B}^{(m)}_{n}$ and ${\bf C}^{(m)}_{n}$ are $4\times 4$ matrices composed of the coefficients of the equation. The explicit forms of the vector $Z^{(m)}_{n}$ and the matrices ${\bf A}^{(m)}_{n}$, ${\bf B}^{(m)}_{n}$ and ${\bf C}^{(m)}_{n}$ are given in Appendix A.
Equation (\ref{e18}) is in the form of a vector recurrence relation which is solved by a continued-fraction method, and the solution can be written in terms of Green's function matrices $G^{(m)}_{n,j}(t)$~as
\begin{align}
Z^{(m)}_{n}(t) = \sum_{j=0}^{\infty}Z^{(m)}_{n,j}(t) = \sum_{j=0}^{\infty}Z^{(m)}_{j}(0)G^{(m)}_{n,j}(t) ,\label{e19}
\end{align}
where $Z^{(m)}_{j}(0)$ are the initial values of the density matrix elements.
The general solution (\ref{e19}) can be applied to evaluate the spectrum of light emitted by the system. Two kinds of radiation are produced by the system, the atomic fluorescence emitted out of the sides of the cavity and the cavity-mode field, leaking the cavity through the cavity mirrors.
\subsection{Spectral distribution of the atomic fluorescence}
In the dressed-atom picture and under the condition of well separated dressed-atom transition frequencies, the spectrum of the atomic fluorescence can be evaluated separately at each of the dressed-atom frequency as
\begin{eqnarray}
S_{a}(\omega) = S^{(-)}(\omega)+S^{(0)}(\omega) + S^{(+)}(\omega) ,\label{e20}
\end{eqnarray}
where
\begin{align}
S^{(-)}(\omega) &= \gamma_{-}\,{\rm Re}\!\int_{0}^{\infty}\!d\tau {\rm e}^{-i(\omega-\omega_{-})\tau}\langle{R_{12}(t+\tau)R_{21}(t)\rangle}_{s} ,\nonumber \\
S^{(0)}(\omega) &= \frac{1}{4}\gamma_{0}\,{\rm Re}\!\int_{0}^{\infty}\!d\tau {\rm e}^{-i(\omega-\omega_{L})\tau}
\langle{R_{3}(t+\tau)R_{3}(t)\rangle}_{s} ,\nonumber\\
S^{(+)}(\omega) &= \gamma_{+}\,{\rm Re}\!\int_{0}^{\infty}\!d\tau{\rm e}^{-i(\omega-\omega_{+})\tau} \langle{R_{21}(t+\tau)R_{12}(t)\rangle}_{s} ,\label{e21}
\end{align}
and the subscript $s$ denotes the steady-state value of the correlation function. The required spectrum therefore follows directly from the dressed-state two-time correlation functions, which can be evaluated from the general solution (\ref{e19}).
Since we are interested in the spectral distribution of the field emitted at the cavity frequency, which is tuned to the frequency of the lower Rabi sideband, we will therefore focus our attention on the spectrum at the frequency $\omega_{-}=\omega_{L}-2\Omega$. The two-time correlation functions appearing in Eq.~(\ref{e21}) are found by applying the quantum regression theorem to the equation of motion (\ref{e16}). The general solution for the spectrum, given in terms of the Green function matrix is of the form
\begin{align}
&S^{(-)}(\omega) = \gamma_{-}\left\{\frac{\Gamma_{c}\langle{R_{12}R_{21}}\rangle_{s}}{\Gamma_{c}^{2} +\nu^{2}}\right. \nonumber\\
&\left. +{\rm Re}\!\sum^{\infty}_{n,j=0}\frac{g_1\sqrt{n+1}}{\Gamma_{c} +i\nu}\left[G^{(1)}_{n,j}(i\nu)Z^{(1)}_{j}(0)\right]_{2}\right\} ,
\end{align}
where $[G^{(1)}_{n,j}(i\nu)Z^{(1)}_{j}(0)]_{2} = \rho^{(2)}_{n,n+1}(i\nu)$ is the Laplace transformation of $\rho^{(2)}_{n,n+1}(\tau)$, and $\nu =\omega -\omega_{-}$.
\subsection{Spectral distribution of the cavity field}
The incoherent part of the spectrum of the cavity field evaluated at the frequency $\omega_{-}=\omega_{L}-2\Omega$ is defined as the Fourier transformation of the two-time correlation function of the fluctuation operators
\begin{align}
S_{c}(\omega) = 2{\rm Re}\int^{\infty}_{0}\!d\tau\, {\rm e}^{i(\omega-\omega_{-})\tau}
\langle\delta a^{\dag}(t+\tau) \delta a(t) \rangle_{s} ,
\end{align}
where $\delta a(t) = a(t) -\langle a(t)\rangle$.
Using the quantum regression theorem, one can show that the two-time correlation function of the field fluctuation operators can be expressed in terms of the density matrix elements as
\begin{eqnarray}
\langle\delta a^{\dag}(t+\tau) \delta a(t) \rangle_{s} = \sum^{\infty}_{n=0}\sqrt{n+1}\rho^{(1)}_{n,n+1}(\tau) ,\label{e24}
\end{eqnarray}
where $\rho^{(1)}_{n,n+1}(\tau)$ is the solution of Eq.~(\ref{e16}) with the initial condition given by the steady-state values of the density matrix elements.
The spectrum is obtained by taking the Fourier transform of Eq.~(\ref{e24}), or equivalently it can be done replacing $\tau$ by $\tau = i\nu$. Hence, we readily find the following formula for the cavity field spectrum
\begin{align}
S_{c}(\omega) = 2{\rm Re}\sum^{\infty}_{n,j=0}\sqrt{n+1}\left[G^{(1)}_{n,j}(-i\nu) Z^{(1)}_{j}(0)\right]_{1} .\label{e25}
\end{align}
In the following section, we present numerical results for the atomic fluorescence and the cavity field spectra evaluated for different values of the pumping rate of the lasing transition.
\section{Numerical results}
We now evaluate the atomic fluorescence and the cavity field spectra in a good cavity limit of $g\gg \kappa,\gamma$. We concentrate on the lower Rabi sideband and consider the variation in the spectra with changing $\gamma_{+}$, the pumping rate of the dressed-atom laser for the case of a frequency independent reservoir, $\gamma_{-}\neq 0$ and compare it with the result for the band-gap, $\gamma_{-}=0$ situation. The results are shown in~Figs.~\ref{fig1}-\ref{fig4}.
\begin{figure}[t]
\includegraphics[width=5cm,keepaspectratio,clip]{1a}\\
\includegraphics[width=5cm,keepaspectratio,clip]{1b}\\
\includegraphics[width=5cm,keepaspectratio,clip]{1c}\\
\includegraphics[width=5cm,keepaspectratio,clip]{1d}\\
\caption{The atomic fluorescence spectrum $S^{(-)}(\omega)$ as a function of frequency $\omega -\omega_{L}+2\Omega$ for $\kappa = 0.05$, $g=5$, $\gamma_{-}\neq 0$ and different $\delta_{a}$: (a) $\delta_{a}=-5$, (b) $\delta_{a}= 0$, (c) $\delta_{a}= 1$, (d) $\delta_{a}= 5$. All the parameters are normalized to $\gamma=1$.} \label{fig1}
\end{figure}
Let us first examine the spectra in the situation when the atom is coupled to a frequency independent reservoir that the spontaneous transitions are allowed at all three dressed-atom frequencies, i.e. $\gamma_{0}, \gamma_{+}$ and $\gamma_{-}$ are all different from zero. Figure~\ref{fig1} shows the atomic fluorescence spectrum for fixed $\kappa = 0.05\gamma$, and the cavity coupling constant $g=5\gamma$, and for progressively increasing pumping rate $\gamma_{+}$. At low pumping rate, the spectrum consists of two well-separated narrow peaks, centered at frequencies $\omega = \omega_{-}\pm g_{1}$. This is well known as the vacuum Rabi splitting. As the pumping rate increases, i.e. $\cos^{4}\phi$ increases, more peaks emerge at multiples of~$\pm g_{1}$. At moderate pumping rates, a multi-peaked spectrum is well visible with peaks inside as well as outside the vacuum Rabi doublet, but the most intense are still those corresponding to the vacuum Rabi splitting. Note that there is no fluorescence at $\omega_{-}$, the cavity field frequency. Moreover, the peaks group into three separate sets of multiplets. If we further increase the pumping rate, the multi-peak structures merge into the characteristic Mollow type triplet with the central component at the cavity field frequency $\omega_{-}$ and sidebands located at $\pm 2\sqrt{\langle n\rangle}g_{1}$, where $\langle n\rangle$ is the average number of photons in the cavity mode. This spectrum is very similar to that found before for a bichromatic driving field~\cite{ff96}, with one strong and one weak component. The strong component was resonant with the atomic transition whereas the weaker component was matched to the Rabi sideband induced by the strong component. In the case considered here, the role of the weaker component is played by the cavity field.
\begin{figure}[hbp]
\includegraphics[width=5cm,keepaspectratio,clip]{2a}\\
\includegraphics[width=5cm,keepaspectratio,clip]{2b}\\
\includegraphics[width=5cm,keepaspectratio,clip]{2c}\\
\includegraphics[width=5cm,keepaspectratio,clip]{2d}\\
\caption{The cavity field spectrum $S_{c}(\omega)$ as a function of frequency $\omega -\omega_{L}+2\Omega$ for $\kappa = 0.05$, $g=5$, $\gamma_{-}\neq 0$ and different $\delta_{a}$: (a) $\delta_{a}=-5$, (b) $\delta_{a}= 0$, (c) $\delta_{a}= 1$, (d) $\delta_{a}= 5$. All the parameters are normalized to $\gamma=1$.} \label{fig2}
\end{figure}
Somewhat similar effects appear in the cavity field spectrum. Figure~\ref{fig2} shows the cavity field spectrum for the same parameters as in Fig.~\ref{fig1}. We observe that the spectrum is quantitatively similar to the fluorescence spectrum, but there is a qualitative difference that the multi-peak structure builds up {\it only} inside the the vacuum Rabi doublet. When the pumping rate increases further, the multi-peak structure merges into a very marrow peak centered at the cavity frequency $\omega_{-}$. The appearance of the very narrow peak the cavity-field frequency is a clear evidence of a lasing action. Thus, there is a threshold value of $\gamma_{+}$, defined to be that value of the pumping rate for which the cavity field spectrum merges into a very narrow, almost monochromatic peak located at the frequency of the cavity mode.
It is interesting to note from Figs.~\ref{fig1} and~\ref{fig2} that we can distinguish three regions of the pumping rate at which the spectra have completely different structures. The low pumping region at which the spectra exhibit the vacuum Rabi splitting, the moderate region where multi-peak structure is seen, and the region of large pumping rates where only a single very narrow peak is observed.
\begin{figure}[hbp]
\includegraphics[width=5cm,keepaspectratio,clip]{3a}\\
\includegraphics[width=5cm,keepaspectratio,clip]{3b}\\
\includegraphics[width=5cm,keepaspectratio,clip]{3c}\\
\includegraphics[width=5cm,keepaspectratio,clip]{3d}\\
\caption{The atomic fluorescence spectrum $S^{(-)}(\omega)$ as a function of frequency $\omega -\omega_{L}+2\Omega$ for $\kappa = 0.05$, $g=5$, and the band-gap spontaneous emission cancelation $\gamma_{-}= 0$. In frames: (a) $\delta_{a}=-5$, (b) $\delta_{a}= 0$, (c) $\delta_{a}= 1$, (d) $\delta_{a}= 5$. All the parameters are normalized to $\gamma=1$.} \label{fig3}
\end{figure}
The behavior of the system and resulting spectral properties can be very different when the atom is coupled to a frequency-dependent reservoir exhibited by a band-gap material.
Figure~\ref{fig3} illustrates the filtering effect of the band-gap material on the atomic fluorescence spectrum by choosing $\gamma_{-}=0$. It corresponds to the situation of no spontaneous emission on the dressed-atom transition resonant to the cavity-mode frequency. For low pumping rates the behavior of the spectrum is seen to be qualitatively different from the previous case of the frequency independent reservoir. We start to see the multi-peaked structure at very low values of the pumping rate. In addition to a general appearance of the multi-peaked structure at low and moderate pumping rates, an enhancement of the fluorescence and the cavity field appear at frequencies symmetrically displaced from the cavity-mode frequency by the effective coupling constant $g_{1}$. It is notable that in both spectra the resonances occurring at $\pm g_{1}$ are the dominant features in the case of weak and moderate pumping rates.
\begin{figure}[hbp]
\includegraphics[width=5cm,keepaspectratio,clip]{4a}\\
\includegraphics[width=5cm,keepaspectratio,clip]{4b}\\
\includegraphics[width=5cm,keepaspectratio,clip]{4c}\\
\includegraphics[width=5cm,keepaspectratio,clip]{4d}\\
\caption{The cavity field spectrum $S_{c}(\omega)$ as a function of frequency $\omega -\omega_{L}+2\Omega$ for $\kappa = 0.05$, $g=5$, the band-gap spontaneous emission cancelation $\gamma_{-}= 0$. In frames: (a) $\delta_{a}=-5$, (b) $\delta_{a}= 0$, (c) $\delta_{a}= 1$, (d) $\delta_{a}= 5$. All the parameters are normalized to $\gamma=1$.} \label{fig4}
\end{figure}
Figure~\ref{fig4} shows the behavior of the cavity field spectrum in the band-gap situation with $\gamma_{-}=0$. Again, the multi-peaked structure, which is caused by the population of the low energy entangled states of the combined dressed-atom plus the cavity mode system, is seen to emerge at very low pumping rates. As the pumping rate increases, a threshold behavior occurs that the multi-peak spectrum converts into a single very narrow peak. Simultaneously, the atomic fluorescence spectrum converts into a triplet, similar to the Mollow triplet, with the central component at frequency $\omega_{-}$. It is interesting that the magnitude and the width of the narrow peak is the same as in the case of the frequency independent reservoir, i.e. is independent of whether $\gamma_{-}=0$ or $\gamma_{-}\neq 0$. Thus, the dominant effects for the filtering properties of the band-gap material are see at low and moderate pumping rates.
\section{Interpretation of the results}
The physics underlying the modifications of the spectra in the presence of the band-gap material can be understood more intuitively by an analytical examination of the energy structure of the combined dressed-atom plus the cavity mode system. The combined states are entangled states between the atom and the cavity field. The knowledge of the states allows us to calculate the populations of these states and transition rates between them. Our treatment is an extension of the Freedhoff and Quang approach to the case of a frequency dependent reservoir. In these calculations, we use the fully quantized dressed-atom approach which gives a complementary view of the cavity effects.
We use the quantum dressed-atom model, in which we quantize both the driving laser and the cavity-mode fields. In our calculations, we first couple the bare atom to the laser field, then we couple the resulting dressed atom to the cavity mode. In the fully quantum description we replace the semiclassical dressed states by their quantum counterparts
\begin{eqnarray}
|\tilde{1}, N\rangle = \cos\phi |1,N\rangle- \sin\phi |2,N-1\rangle ,\nonumber\\
|\tilde{2}, N\rangle= \sin\phi |1,N\rangle+ \cos\phi |2,N-1\rangle ,\label{e30}
\end{eqnarray}
where $|i,N\rangle$ is a state in which the atom is in the $i$th state and $N$ photons are present in the driving field.
The eigenstates of the noninteracting dressed-atom system plus the cavity mode are product states $|\tilde{i},N,n\rangle=|\tilde{i},N\rangle\otimes|n\rangle$, where $\ket n$ is the eigenstate of the cavity mode, with energies
\begin{eqnarray}
E_{i,N,n} = N\omega_{L}+(-1)^{i}\Omega+n(\omega_{L}-2\Omega) ,\label{e31}
\end{eqnarray}
where $n$ is the number of photons in the cavity mode.
It is easily to show that the state $|2,N,0\rangle$ is a single state, whereas the state $|2,N-n,n\rangle$ and $|1,N-n+1,n-1\rangle$ from doubly degenerate pairs.
When we include the interaction between the dressed atom and the cavity mode
\begin{eqnarray}
V_{ac} = \hbar g\left(a^{\dagger}\sigma_{12}+\sigma_{21}a\right) ,\label{e32}
\end{eqnarray}
the degeneracy is lifted, resulting in doublets with eigenstates, the atom-cavity-mode entangled states of the form
\begin{widetext}
\begin{eqnarray}
\ket{\Psi_{N,\pm n}} = \left\{\begin{array}{ll}
|\tilde{2},N,0\rangle ,&\textrm{for $n={0}$}\\
\\
\frac{1}{\sqrt{2}}\left(|\tilde{2},N-n,n\rangle\pm |\tilde{1},N-n+1,n-1\rangle\right) , &\textrm{for $n\neq{0}$}
\end{array}\right. \label{e33}
\end{eqnarray}
\end{widetext}
corresponding to energies
\begin{align}
E_{N,0} &= \hbar\left(N\omega_{L}+\Omega\right) ,\nonumber\\
E_{N,\pm{n}} &= \hbar\left[N\omega_{L} -(2n-1)\Omega\pm{g_1}\sqrt{n}\,\right] .\label{e34}
\end{align}
The entangled states (\ref{e33}) are the energy states of the system. They group into manifolds, each containing a singlet and an infinite number of doublets. Neighboring doublets are separated by~$2\Omega$, while the intra-doublet states are separated by $2\sqrt{n}g_{1}$. The spectral properties of the atomic fluorescence and the cavity field are a signature of the entangled system, its energy levels, their populations and probabilities of spontaneous transitions between them. Thus, one would expect multiple structures in the spectra.
\subsection{Transition rates}
The energy states of the system interact with the vacuum reservoir modes. This results in a spontaneous emission cascade by the system down its energy manifold ladder. The probability of a transition between any two energy states is proportional to the absolute square of the dipole transition moment between them. Since the interaction between the atom and the field is determined by a single-photon Hamiltonian, the transition dipole moments have selection rules that nonzero transitions can occur only between states $\ket{\Psi_{N,\pm n}}$ and $\ket{\Psi_{N-1,\pm n^{\prime}}}$ with $n^{\prime} = n, n\pm 1$. The probabilities (spontaneous emission rates), corresponding to transitions at frequencies
\begin{eqnarray}
\omega_{\pm n,n^{\prime}} = \hbar^{-1}\left(E_{N,\pm n} - E_{N-1,n^{\prime}}\right) ,\label{e35}
\end{eqnarray}
are given by
\begin{align}
\Upsilon_{\pm{n},\pm{n}^{\prime}} &= \frac{1}{4}\gamma_{0}|\bra{\Psi_{N,\pm n}}R_{3}\ket{\Psi_{N-1,\pm n^{\prime}}}|^{2} \nonumber\\
&+\gamma_{+}|\bra{\Psi_{N,\pm n}}R_{21}\ket{\Psi_{N-1,\pm n^{\prime}}}|^{2} \nonumber\\
&+\gamma_{-}|\bra{\Psi_{N,\pm n}}R_{12}\ket{\Psi_{N-1,\pm n^{\prime}}}|^{2} .\label{e36}
\end{align}
Similarly, cavity damping cause transition to occur from states
$|N,\pm{n}\rangle$ to states $|N-1,\pm{n}^{\prime}\rangle$ with probabilities
\begin{eqnarray}
\kappa_{\pm{n},\pm{n}^{\prime}} = \kappa|\bra{\Psi_{N,\pm n}}a^{\dagger}\ket{\Psi_{N-1,\pm n^{\prime}}}|^{2} .\label{e37}
\end{eqnarray}
Since our interest is principally on the spectral distribution at the frequency of the lower Rabi sideband, $\omega_{-}=\omega_{L}-2\Omega$, we therefore focus on spontaneous emission rates between states $\ket{\Psi_{N,\pm{n}}}\rightarrow\ket{\Psi_{N-1,\pm{(n+1)}}}$ and $\ket{\Psi_{N,\pm{(n+1)}}}\rightarrow\ket{\Psi_{N-1,\pm{n}}}$.
Using Eqs.~(\ref{e33}), (\ref{e34}) and (\ref{e36}), we readily find that the spontaneous emission rates corresponding to several frequencies centered at $\omega_{-}$ are given by the expressions
\begin{align}
\Upsilon_{\pm{n},\pm{(n+1)}} &= \frac{1}{4}\gamma_{+} (1+\delta_{n,0}) ,\ \omega_{\pm{n},\pm{(n+1)}} = \omega_{-}\mp \nu_{n}^{(-)} ,\nonumber\\
\Upsilon_{\mp{n},\pm{(n+1)}} &= \frac{1}{4} \gamma_{+} (1+\delta_{n,0}) ,\ \omega_{\mp{n},\pm{(n+1)}} = \omega_{-}\mp \nu_{n}^{(+)} ,\nonumber\\
\Upsilon_{\pm{(n+1)},\pm{n}} &=\frac{1}{4}\gamma_{-}(1+\delta_{n,0}) ,\ \omega_{\pm
(n+1),\pm{n}} = \omega_{-}\pm \nu_{n}^{(-)} ,\nonumber\\
\Upsilon_{\pm{(n+1)},\mp{n}} &= \frac{1}{4}\gamma_{-}(1+\delta_{n,0}) ,\ \omega_{\pm (n+1),\mp{n}} = \omega_{-}\pm \nu_{n}^{(+)} ,\label{e38}
\end{align}
where $\nu_{n}^{(\pm)} = \left(\sqrt{n+1}\pm\sqrt{n}\,\right)\!g_{1}$.
Similarly, we find from Eqs.~(\ref{e33}), (\ref{e34}) and (\ref{e37}) that the cavity damping causes transitions with the rates
\begin{align}
\kappa_{\pm{(n+1)},\pm{n}} &=\frac{1}{4}\kappa \left(\sqrt{n+1}+\sqrt{n}\, \right)^2 , \ \omega_{\pm (n+1),\pm{n}} = \omega_{-}\pm \nu_{n}^{(-)} \nonumber\\
\kappa_{\pm{(n+1)},\mp{n}} &=\frac{1}{4}\kappa \left(\sqrt{n+1}-\sqrt{n}\, \right)^2 ,\ \omega_{\pm (n+1),\mp{n}} = \omega_{-}\pm \nu_{n}^{(+)} .\label{e38a}
\end{align}
From the transition rates and the corresponding transition frequencies, it is apparent that the spectra consist of multiple lines located symmetrically about $\omega_{-}$. The multiples split into two sets of lines, $\nu_{n}^{(-)}$ and $\nu_{n}^{(+)}$, which we call the inner and outer sidebands, respectively. The spontaneous emission rates are dependent on the density of the vacuum reservoir modes at both the higher and lower Rabi sideband. However, only the rates that depend on the density of the modes at the lower Rabi sideband are, in addition, affected by the cavity damping. Moreover, the effect of the cavity damping on the transition rates corresponding to the inner sidebands increase with increasing number of photons $n$, whereas the transition rates corresponding to the outer sideband decreases with increasing $n$. This results in the enhancement (reduction) the inner (outer) sidebands amplitudes with the increasing pumping rate, as seen in Figs.~\ref{fig2} and \ref{fig4}.
It is interesting to note that there are no spontaneous transitions allowed at the cavity field frequency $\omega_{-}$ unless $n\gg 1$ for which $\sqrt{n+1}-\sqrt{n}\approx 0$. Consequently, there is no spectral line at $\omega_{-}$, the cavity mode frequency, until a large number of photons accumulate in the cavity mode, $n\gg 1$. Then, the inner sidebands at $\pm\nu_{n}^{(-)}$ merge into a single peak at $\omega_{-}$, while the outer sidebands merge into single lines located at frequencies $\pm 2\sqrt{n}g_{1}$. In this limit, the atomic fluorescence spectrum exhibits three lines. The merging of the fluorescence spectrum into the familiar Mollow triplet is a clear evidence that in this limit the cavity field appears as a monochromatic coherent field. We could call the point at which the cavity field merges into a single narrow peak as a threshold for the cavity field to operate as a conventional laser. On the other hand, the cavity field spectrum converts into a single peak, that the outer sidebands decreases with increasing $n$ with the inner sidebands exhibiting a strong enhancement. This is consistent with the variation with $n$ of the cavity induced damping rates~(\ref{e38a}).
\subsection{Populations}
To evaluate the populations of the energy states of the system, we project the master
equation~(\ref{e14}) onto $\ket{\Psi_{N,\pm n}}$ on the right and $\bra{\Psi_{N,\pm n}}$ on the left, and make the secular approximation that ignores couplings between populations and coherences. We then introduce the reduced populations
\begin{eqnarray}
\Pi_{\pm n} = \sum_{N} \bra{\Psi_{N,\pm n}} \rho \ket{\Psi_{N,\pm n}} ,\label{e39}
\end{eqnarray}
and find that in the steady state, $\Pi_{\pm n}$ satisfy the recurrence relation
\begin{eqnarray}
&&\gamma_{+}{\Pi}_{n-1} - \left[\gamma_{+}+\gamma_{-}+(2n-1)\kappa\right]\Pi_{n}\nonumber\\
&&+\left[\gamma_{-}+(2n+1)\kappa\right]\Pi_{n+1} = 0 ,\label{e40}
\end{eqnarray}
with $(\gamma_{-}+\kappa)\Pi_{1} = \gamma_{+}{\Pi}_{0}$ and ${\Pi}_{-n}={\Pi}_{n}$.
The solution of Eq.~(\ref{e36}) gives
\begin{align}
{\Pi}_{\pm{n}}=\Pi_{0}\prod^{n}_{m=1}\frac{\gamma_{+}}{\gamma_{-}+(2m-1)\kappa} ,\label{e41}
\end{align}
where $\Pi_{0}$ is determined by the normalization condition.
It is evident from Eq.~(\ref{e41}) that the population of the energy states depends solely on the density of the vacuum modes at the Rabi sideband frequencies. The population is unequally distributed among the states, but the redistribution depends strongly on whether $\gamma_{-}=0$ or~$\gamma_{-}\neq 0$. For $\gamma_{-}\neq 0$ and $\kappa\ll\gamma$, the population distribution depends weakly on $n$ leaving the state $n>1$ practically unpopulated. This explains why in the limit of a low pumping, $\gamma_{+}\ll 1$, only the characteristic vacuum Rabi splitting is seen in the spectra, as shown in Figs.~\ref{fig1} and~\ref{fig2}. However, in the band-gap situation of $\gamma_{-}=0$ the population distribution is strongly sensitive to $n$ and the population can be redistributed over the $n>1$ state even for relatively weak pumping rate. This clearly explains why in the band-gap situation, shown in Figs.~\ref{fig3} and~\ref{fig4}, a multi-peak structure is observed in the spectra even for very low pumping rate.
\section{Conclusions}
We have examined the atomic fluorescence and cavity field spectra of a strongly driven two-level atom coupled to a single-mode cavity and a frequency-dependent reservoir exhibited by a band-gap material. The band-gap spectral density of the electromagnetic vacuum modes changes sharply at the band-edge frequency $\omega_{b}$, such that $\omega_{b}-\omega_{-} \gg \Gamma$, where $\omega_{-}$ is the frequency of the lower Rabi sideband and $\Gamma$ is its bandwidth.
We have found that the frequency dependent reservoir affects the spectra for low and moderate pumping rates, while for large pumping rates the spectra do not differ substantially from that predicted for the case of a frequency independent reservoir.
The effect of the frequency dependent reservoir is to narrow the spectral lines and to suppress the below threshold regime characterized by the vacuum Rabi splitting. Multi-peaked spectra are observed for essentially very low pumping rates. This feature can be regarded as a signature of the thresholdless lasing. However, the multi-peaked structure of the cavity field spectrum indicates that the system does not operate as a source of a monochromatic light.
A conventional laser is usually described as a source of a highly monochromatic coherent light.
We have found that the cavity field spectrum exhibits a threshold behavior that at a certain pumping rate, the multi-peak spectrum converts into a single very narrow peak located at the cavity mode frequency. The threshold behavior exists despite the fact that there is no threshold behavior in the photon statistics of the cavity field.
\section*{Acknowledgments}
We acknowledge financial support from the National Natural Science Foundation of China (Grant Nos. 10674052 and 60878004), the Ministry of Education under project NCET (grant no NCET-06-0671), SRFDP (under grant no. 200805110002), and the National Basic Research Project of China (grant no. 2005 CB724508).
|
\section{Introduction}
Light is a natural carrier of information and since the advent of quantum information and metrology it solidified as a practical tool for quantum communication and quantum sensing. However, light is not sufficient on its own. For example, in a long distance quantum communication it has to be supplemented by quantum repeaters \cite{repeaters} based on, for example, atomic memory \cite{memory}. For the quantum sensing of magnetic fields, a cloud of atoms is used as a sensor, which is prepared and probed by the light beam \cite{sensing}. Light is also used as an efficient probe and resource for the preparation and monitoring of quantum vibrations of a mechanical oscillator \cite{mirrors}. It is therefore apparent that important elements of the quantum toolbox should be interface protocols, which allow for a flawless transfer from one physical system (light) to another (matter) \cite{interface}. One feature, this protocols should keep in regard, is the often limited capability of these matter systems to interact with light, be measured, and be controlled. The interfaces should therefore rely neither on pre-processing, nor on extensive post-processing of the matter system. Furthermore, the perfect interface should be achievable even for a weak coupling between the two physical systems and for a very noisy input state of the matter system.
Previous study of this problematic revealed that for almost any Gaussian coupling, with Hamiltonian quadratic in quadrature operators $x$ and $p$, a suitable pre-processing of the well-controllable system, measurement, and feed-forward can be used to implement a perfect state transfer without any regards for the initial state of the target system \cite{filip09}. This approach unfortunately does not work for the asymmetrical quantum-non-demolition (QND) coupling used to couple light to the macroscopic ensemble of atoms \cite{atoms}, or between light and macroscopic vibrations of the mechanical oscillator \cite{mirrors}. At its core, the QND interaction allows to transmit only a single quadrature from light to the matter system. A method of working around this limiting factor and implementing a state transfer through a single QND interaction has been studied in Ref.~\cite{filip08}, but it was shown impossible for weak coupling strengths. Hence we will focus on the problem of building, in principle, a perfect interface for this kind of weak-strength QND coupling, for a moment neglecting other technical details of the light-matter interaction.
In this paper, we propose a deterministic universal quantum interface based on two sequential weak QND couplings between two light beams and a noisy matter system. The interface is based on all-optical (finite gain) pre-processing, a homodyne detection of light beams, and a conditional feed-forward correction displacing the matter system. By a construction we prove that the interface can perfectly transfer any (even unknown) quantum state to the initially noisy matter system, with no regards to strength of the QND couplings and the initial noise of the matter system. This interface method is then extended to a parallel QND coupling, represented by a sum of two QND couplings jointly probing the matter system.
We also present a probabilistic version of the interface, which is also able to transfer any state under the same conditions. In this scenario, there is no need for the feed-forward operation, a simplification, which is paid for by a reduced probability of success. We analyze the probabilistic scenario with regards to transfer of a single-photon state and demonstrate that weak coupling strengths and high initial noise can be compensated by a more severe post-selection.
\section{Deterministic interface}
The class of QND couplings between two harmonic oscillators is characterized by interaction Hamiltonian proportional to a product of two quadratures, one corresponding to a continuous-variable of light and the other to a continuous variable of matter. All the QND couplings (of the same strength) are theoretically equivalent, differing just by a simple local unitary transformation (phase-space rotation) of the relevant light and matter modes. Therefore, all the results presented below can be easily adapted for any kind of QND coupling. A single particular QND coupling can be characterized by the interaction Hamiltonian of the form $H = \chi x_A p_L$, where $\chi$ stands for the interaction strength, $x$ and $p$ denote the quadrature operators with $[x_{.},p_{.}] = i$, and the subscripts $L$ and $A$ mark the participating light mode and the matter mode (atomic ensemble mode or mechanical oscillator mode), respectively. In Heisenberg picture, the quadrature operators transform as
\begin{eqnarray}
x_L^{out} = x_L^{in} + \kappa x_A^{in},~p_L^{out} = p_L^{in},\nonumber \\
x_A^{out} = x_A^{in},~p_A^{out} = p_A^{in} -\kappa p_L^{in}.
\end{eqnarray}
The interaction transfers information about quadratures $x_A$ and $p_L$ to quadratures $x_L$ and $p_A$, while leaving the original quadratures undisturbed. This behavior is where the quantum-non-demolition coupling got its name. In following, we shall employ parameter $\kappa = \chi T$ (where $T$ is the effective lenght of the interaction) as the gain of the QND interaction. This simple unitary model ignores all the decoherence effects typical for the matter system, therefore it can only describe a coupling, which is very fast, faster than any relevant decoherence time. As a natural consequence, the range of values of $\kappa$ is very limited, typically corresponding to a weak coupling regime ($\kappa<1$).
As was mentioned previously, it is possible to conceive a noise free transfer using a single QND coupling, measurement and feed-forward, and a suitable squeezing of the initial state of light \cite{filip08}. However, there are several drawbacks to approach. The transfer can be noiseless, but it is still affected by loss and squeezing. The amount of loss depends on the strength of the coupling, with transmission parameter $\eta=\kappa^2/(1+\kappa^2)$, so unless the coupling gain is quite large, $(\kappa\gg 1)$, the loss will reduce the quality of the transfer. The squeezing can be compensated by another QND interaction with ancillary mode of light followed by a measurement and a feed-forward, but this incurs further loss and the overall transmission ends up only $\eta'=\kappa^2/(1+\kappa^2)^2$. All in all, there is no set of finite parameters for which lossless transfer is possible. Furthermore, the transfer is noiseless only when the mode of the matter system is prepaired in the ground state, which, despite the significant progress done by vacuum state preparation of the atomic ensemble or the mechanical oscillator, can still be a problematic task on its own.
\begin{figure}
\centerline{\psfig{figure=fig_2QND_setup2.eps,width=0.9\linewidth}}
\caption{(color online) Setup for the universal quantum interface based on the sequential QND couplings: $L$, $M$ stands for the light modes and $A$ denotes the matter mode. QND$_i$ - QND interaction with gain $\kappa_i$, HD$_q$, $q=X,P$ - balanced homodyne detection measuring a quadrature $q$, $\gamma_q$ - gain in the feed-forward loop for quadrature $q$; D - displacement operation; S - optional squeezing operation. }\label{fig_2QND_setup}
\end{figure}
The need for the second QND coupling with an ancillary mode of light to compensate for the squeezing may be seen as an indication that a different scheme, which employs two sequential couplings with independent modes of light, may achieve better results. Clearly, these two couplings have to address both the complementary matter variables $x_A$ and $p_A$. Since the matter state is typically encoded up to a slowly rotating frame in the phase space, it is just a question of precise timing to correctly implement a sequence of the two couplings. Let us consider a system with two modes of light, $L$ and $M$ and one matter mode $A$. After the interaction mediated by the two sequential QND couplings characterized by interaction Hamiltonians $H_1=\chi_1 x_A p_M$, $H_2 = \chi_2 x_L p_A$ and gains $\kappa_1$, $\kappa_2$, the quadrature operators of the participating modes are transformed to
\begin{eqnarray}
x'_M &=& x_M + \kappa_1 x_A, \nonumber \\
p'_M &=& p_M, \nonumber \\
x'_L &=& x_L, \nonumber \\
p'_L &=& p_L - \kappa_2 p_A + \kappa_1 \kappa_2 p_M, \nonumber \\
x'_A &=& x_A + \kappa_2 x_L , \nonumber \\
p'_A &=& p_A - \kappa_1 p_M.
\end{eqnarray}
To complete the transfer we need to employ a pair of balanced homodyne detectors to measure values of the operators $x'_M$ and $p'_L$. This differs from the method of Ref.~\cite{filip09}, as there are no collective measurements required here, due to the specific nature of the QND interaction. The measured data are used in the feed-forward loop to drive the displacement of the mode $A$, $x_A^{out}= x'_A+\gamma_x x'_M$ and $p_A^{out}= p'_A + \gamma_p p'_L$, where $\gamma_x$ and $\gamma_p$ are the adjustable electronic gains. If these gains are chosen so $\gamma_x = -1/\kappa_1$ and $\gamma_p = 1/\kappa_2$, the influence of the initial state of the matter mode completely vanishes and its final state quadratures read
\begin{eqnarray}
x_A^{out}&=& \kappa_2 x_L - \frac{1}{\kappa_1} x_M , \nonumber \\
p_A^{out}&=& \frac{1}{\kappa_2} p_L.
\end{eqnarray}
As in \cite{filip09}, this dependance on collective quadratures of the initial modes of light can be removed by a suitable pre-processing. Namely, a squeezing of the mode $L$ with gain $g = \kappa_2$, followed by a QND coupling with interaction hamiltonian $H_3 = \chi_3 x_M p_L$ and gain $\kappa_3 = 1/(\kappa_1 \kappa_2)$, will achieve realization of a perfect transfer $x_A^{out} = x_L^{in}$ and $p_A^{out} = p_L^{in}$. Such the pre-processing can be built using an interferometric scheme based on either in-line squeezers or a measurement-induced QND (with the off-line squeezing), both of which are experimentally available \cite{QNDinline,QNDonline}. However, as of yet, the quality and possible gains of the pre-processing are limited by strengths of the in-line interactions or the available off-line squeezing.
A very interesting aspect of this protocol, a full depiction of which can be found in Fig.~\ref{fig_2QND_setup}, is that it can principally achieve a perfect transfer for an arbitrarily weak coupling between the light and the matter modes. Furthermore, although the demand on resources during the pre-processing phase can be substantial, as the gain $\kappa_3$ increases rapidly as $\kappa_1$ and $\kappa_2$ are getting smaller, the gains of both the QND transformation and the squeezing are always finite. Furthermore, the strength of interaction between light and matter is compensated for by a stronger all-optical coupling. Consequently, improvement in our ability to produce a more intensive coupling between optical modes would allow for a construction of better quantum interfaces between light and matter.
A possible inconvenience of the presented method lies in its sequential nature. The light modes $L$ and $M$ are supposed to interact with the non-light mode $A$ one at the time and therefore the second one needs to be delayed while the first one interacts. This can be problematic if the light modes are defined over a substantially long time, which is a typical situation in experiments, where it allows to effectively increase the interaction strength.
A possible remedy lies in using a joint same-time QND interaction of the three modes, $M$, $L$ and $A$. A similar strategy was already proposed for a readout from the atomic memory \cite{Fiurasek06}. This coupling, consisting of the simultaneous addressing of the matter system by both the light systems and having the interaction Hamiltonian $H = \chi ( x_A p_L + p_A x_M)$, transforms the quadrature operators as:
\begin{eqnarray}
x'_M &=& x_M, \nonumber \\
p'_M &=& p_M - \kappa p_A + \frac{\kappa^2}{2} p_L, \nonumber \\
x'_L &=& x_L + \kappa x_A + \frac{\kappa^2}{2}x_M, \nonumber \\
p'_L &=& p_L , \nonumber \\
x'_A &=& x_A + \kappa x_M , \nonumber \\
p'_A &=& p_A - \kappa p_L,
\end{eqnarray}
where $\kappa$ is the gain of the joint QND coupling. After homodyne measurement of quadratures $x'_L$ and $x_M$, the state of the mode $A$ can be suitably displaced, $x_A^{out}= x'_A +\gamma_x x'_M$ and $p_A^{out}= p'_A +\gamma_p x'_L$, and if the gains are tuned properly, $\gamma_x = -\gamma_p = -1/\kappa$, the output state operators are
\begin{eqnarray}
x_A^{out} &=& \frac{\kappa}{2} x_M - \frac{1}{\kappa}x_L ,\nonumber \\
p_A^{out} &=& \frac{1}{\kappa} p_M - \frac{\kappa}{2} p_L.
\end{eqnarray}
After a suitable pre-processing, which is depicted in Fig.~\ref{fig_JQND_setup}, and which consists of a balanced beam splitter and a pair of squeezing operations with gain $g = \sqrt{2}/\kappa$,
\begin{eqnarray}
x_M &=& \frac{1}{\kappa}(x^{in}_M+x^{in}_L),\nonumber \\
p_M &=& \frac{\kappa}{2}(p^{in}_M + p^{in}_L),\nonumber \\
x_L &=& \frac{\kappa}{2}(x^{in}_M - x^{in}_L),\nonumber \\
p_L &=& \frac{1}{\kappa}(p^{in}_M - p^{in}_L),
\end{eqnarray}
the quadratures of the input mode $L$ are perfectly transcribed onto the output mode $A$, $x_A^{out} = x_L^{in}$, $p_A^{out} = p_L^{in}$.
\begin{figure}
\centerline{\psfig{figure=fig_JQND_setup.eps,width=0.9\linewidth}}
\caption{(color online) Setup for the perfect state transfer. $L$, $M$ and $A$ denote the participating modes. HD$_q$ - balanced homodyne detection measuring a quadrature $q$, $\gamma_q$ - gain in the feed-forward loop for quadrature $q$; S - squeezing operation; PS - phase shift operation, BS balanced beam splitter. }\label{fig_JQND_setup}
\end{figure}
\section{Probabilistic interface}
The both already mentioned sequences of interactions, measurements, and conditional operations are capable of a deterministic transfer of an unknown quantum state from one physical system to another. However, for the state preparation it is often enough if the known quantum state can be transferred probabilistically, using a post-selection instead of the feed-forward. Specifically, instead of using the measured values of the operators $x'_M$ and $p'_L$ (or $x'_L$ and $p'_M$ for the simultaneous interaction) to drive the correcting displacement, we post-select the state of the target system if the detected values are zero and discard the state otherwise. To accurately describe the projection nature of the measurement procedure we need to abandon the Heisenberg picture employed so far and look at the evolution of the actual physical states. Since the two scenarios are very similar, we shall focus only on the sequential QND interface illustrated in Fig.~\ref{fig_2QND_setup}.
First, let us check whether the probabilistic interface can be implemented in the limiting case of perfect transfer and pure states. These states can be described in coordinate representation by their wave functions, $|\psi\rangle = \int \psi(x)|x\rangle dx$, and any unitary evolution can be characterized by its action on the basis states. In our scenario, the global basis state is composed of three states $|x_L\rangle_L|x_M\rangle_M|x_A\rangle_A$ and our aim is to transform it to $|x_L\rangle_A$. In this representation, the separate parts of the evolution illustrated in Fig.~\ref{fig_2QND_setup} act as follows:
\begin{eqnarray}
\mbox{squeezing}:&\,\,\,& |x\rangle_L \rightarrow |x/g \rangle, \nonumber \\
\mbox{QND}_1:&\,\,\,&|y\rangle_M|z\rangle_A\rightarrow |y+\kappa_1 z\rangle_M|z\rangle_A,\nonumber\\
\mbox{QND}_2:&\,\,\,&|x\rangle_L|z\rangle_A\rightarrow |x\rangle_L|z+\kappa_2x\rangle_A,\nonumber\\
\mbox{QND}_3:&\,\,\,&|x\rangle_L|y\rangle_M\rightarrow |x+\kappa_3y\rangle_L|y\rangle_M.
\end{eqnarray}
After the squeezing $S$, an all-optical coupling QND$_3$, and sequential QND$_1$ and QND$_2$ interactions with the atomic system $A$, the states
$|x_L\rangle_L|x_M\rangle_M|x_A\rangle_A$ are transformed to
\begin{equation}
|x_L/g+\kappa_3\rangle_L|x_M+\kappa_1x_A\rangle_M|x_A+\kappa_2(x_L/g +\kappa_3 x_M)\rangle_A.
\end{equation}
After the projective measurements, which are represented by $_L\langle p = 0|_M\langle y=0|$, where $|p=0\rangle_L \propto \int_{-\infty}^{\infty}|x\rangle_L dx$, are applied to the optical modes $L$ and $M$, the resulting state is proportional (up to a normalization) to \begin{equation}
\int_{-\infty}^{\infty}\hspace{-0.4cm}\delta(x-x_L/g-\kappa_3x_M)dx |-(\kappa_1^{-1}-\kappa_2\kappa_3)x_M+\kappa_2x_L /g\rangle_A.
\end{equation}
Now, the integral is equal to unity, and if we have chosen a suitable preprocessing with $g = \kappa_2$ and $\kappa_3 = (\kappa_1\kappa_2)^{-1}$ the final basis state is simply $|x_L\rangle_A$. As a result we have an atomic system corresponding perfectly to a wave function of the initial light system. The state of the optical mode $L$ has been perfectly transferred to the atomic mode $A$ without any regards for the initial state of any of the three participating modes.
The projective measurements can be approached by homodyne detectors followed by a post-selection if the measured values fell into a narrow interval around the desired values $p_L = 0$ and $x_M = 0$. However, the ideal situation, when the width of the interval approaches zero, also leads to a zero probability of success for the scheme. To realistically analyze the influence of the non-zero post-selection interval on the performance of the method and the probability of success, we will employ the formalism of Wigner functions \cite{wigner}. Wigner function of a single mode state is a real function of a pair of real variables $x$ and $p$, which directly correspond to the quadrature operators used before. The complete state of the three initial systems is represented by a joint Wigner function
\begin{equation}
W_{in}(\xi) = W_L(x_L,p_L)W_M(x_M,p_M)W_A(x_A,p_A),
\end{equation}
where the $\xi$ is a shorthand notation of the vector of the variables, $\xi = (x_L,p_L,x_M,p_M,x_A,p_A)^T$. The subscripts $L,M,$ and $A$ again denote the three participating systems. The unitary evolution of the complete system can be expressed by a symplectic matrix $U$, which transforms the variables as $\xi' = U\xi$. Note that the same matrix can be used to implement the evolution for a vector of quadrature operators. The sequence of interactions depicted in Fig.~\ref{fig_2QND_setup} with $\kappa_3 = -1/\kappa_1\kappa_2$ is represented by matrix
\begin{equation}
U = \left(\begin{array}{cccccc}
0 & 0 & \kappa_1^{-1} & 0 & -1 & 0 \\
0 & -\kappa_2^{-1} & 0 & 0 & 0 & -1 \\
\kappa_1\kappa_2 & 0 & 1 & 0 & -\kappa_1 & 0 \\
0 & (\kappa_1\kappa_2)^{-1} & 0 & 1 & 0 & \kappa_1^{-1} \\
-\kappa_2 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & \kappa_1 & 0 & 1
\end{array}
\right).
\end{equation}
After the transformation, the projective measurements are implemented as integration of the measured variables over interval $[-Q,Q]$, leading to Wigner function
\begin{equation}\label{Wig_out}
W_{out}(x_A,p_A) = \frac{1}{PS}\int_{-\infty}^{\infty}\hspace{-0.4cm} dp_M \int_{-\infty}^{\infty}\hspace{-0.4cm} dx_L \int_{-Q}^{Q}\hspace{-0.4cm}dx_M \int_{-Q}^{Q}\hspace{-0.4cm} dp_L W_{in}(U \xi),
\end{equation}
where $PS$ is a normalization factor which also stands for the probability of success. A general evaluation of the integral is not an easy task. Fortunately, under a realistic assumption that the modes $M$ and $A$ are initially in thermal states with Wigner functions
\begin{eqnarray}\label{Wig_thermal}
W_M(x,p) &=& \frac{\exp(-(x^2+p^2)/2 V_M)}{2\pi V_M} \nonumber \\
W_A(x,p) &=& \frac{\exp(-(x^2+p^2)/2 V_A)}{2\pi V_A},
\end{eqnarray}
the limit $Q\rightarrow 0$ is indeed $W_{out}(x_A,p_A) = W_L(x_A,p_A)$, as predicted by the simple method.
To discuss the convergence of the method with regards to the post-selection interval $Q$ we need to consider a particular quantum state. The chosen state should be quite vulnerable to loss and noise, so we can clearly see their influence. This translates to a nonclassical non-Gaussian state, a clear specimen of which is the single-photon state with its trademark negativity in the Wigner function. The Wigner function itself is
\begin{equation}\label{Wig_sp}
W_L(x,p) = \frac{\exp(-x^2-p^2)}{\pi} (2 x^2 + 2p^2 -1),
\end{equation}
and, under the assumption (\ref{Wig_thermal}), the final state can be analytically obtained from (\ref{Wig_out}). To evaluate the quality of the transfer we can calculate fidelity
\begin{equation}
F = 2\pi \int_{-\infty}^{\infty}\hspace{-0.4cm} dx \int_{-\infty}^{\infty}\hspace{-0.4cm} dp W_{out}(x,p)W_L(x,p),
\end{equation}
to gain a measurable overlap between the ideal and the existing state. We can also explicitly try to look at transfer of the non-classical property --
the highly negative value of the Wigner function at the point of origin. For our purpose, we can define the negativity of the Wigner function as $N = \min[ W(x,p)]$.
\begin{figure}
\centerline{\psfig{figure=Qdependance.eps,width=0.9\linewidth}}
\caption{(Color online) The fidelity \textbf{(a)}, the negativity of the Wigner function \textbf{(b)}, and the probability of success \textbf{(c)} for a transferred single-photon state as functions of the post-selection interval half-width $Q$. The parameters were $\kappa_1 = 0.3$, $\kappa_2 = 0.3$, $V_M = 0.5$, $V_A = 5$ for the blue solid line and $\kappa_1 = 0.5$, $\kappa_2 = 0.5$, $V_M = 0.5$, $V_A = 5$ for the red dashed line. }\label{fig_Qdependance}
\end{figure}
\begin{figure}
\centerline{\psfig{figure=kappadependance.eps,width=0.9\linewidth}}
\caption{(Color online) Values of the fidelity and the negativity of the Wigner function of a transferred
single-photon state relative to the coupling parameter $\kappa = \kappa_1
= \kappa_2$. The three curves correspond to different probabilities of
succes. Blue solid line: $PS = 10^{-2}$, green dashed line: $PS =
10^{-3}$, red dotted line: $PS = 10^{-4}$. Other parameters were: $V_M =
0.5$, $V_A = 5$.}\label{fig_kappadependance}
\end{figure}
\begin{figure}
\centerline{\psfig{figure=VAdependance.eps,width=0.9\linewidth}}
\caption{(Color online) Values of the fidelity and the negativity of the Wigner function a transferred single
photon state relative to the level of initial noise in the atomic system
as characterized by $V_A$. The three curves correspond to different
probabilities of succes. Blue solid line: $PS = 10^{-2}$, green dashed
line: $PS = 10^{-3}$, red dotted line: $PS = 10^{-4}$. Other parameters
were: $V_M = 0.5$, $\kappa_1 = \kappa_2 = 0.5$.}\label{fig_VAdependance}
\end{figure}
These indicators, together with the probability of success, are shown in Fig.~\ref{fig_Qdependance} relative to the logarithm of the threshold value $Q$. By comparing the three panels we see that for a success probability of $PS \approx 0.01$ we can achieve fidelity $F \approx 0.9$ for both considered values of the coupling parameters. This may not seem much, but observe that the transferred state is still highly nonclassical, as witnessed by the negativity $N\approx -0.3$, which is only slightly worse than the ideal negativity of $N = - 1/\pi$. Further analysis revealed that the initial states of the $M$ and $A$ modes, as well as the QND coupling constants $\kappa_1$ and $\kappa_2$, are only relevant as far as the probability of success is concerned.
Fixing the success rate, the fidelity and the negativity are depicted in Figs.~\ref{fig_kappadependance},\ref{fig_VAdependance} as functions of coupling strength $\kappa=\kappa_1=\kappa_2$ and the initial noise of the matter system $V_A$. Both the characteristics are robust against low coupling and large noise, if the success probability is sufficiently low. This clearly demostrates a basic feasibility of the probabilistic transfer for the state preparation of highly non-classical states of atomic ensembles or micro-mechanical oscillators.
\section{Conclusion}
We have proposed a quantum interface capable of transferring a quantum state from one quantum system to another, for example, from light to a continuous-variable matter system, such as a collective spin mode of an atomic cloud or a vibrational mode of a mechanical oscillator. The main building blocks of the interface are QND interactions, which naturally arise as a coupling between modes of light and the aforementioned matter systems. The main benefit of the proposed scheme lies in its deterministic nature and its flawless performance - unit fidelity of transfer can be, in principle, achieved for arbitrary states of all participating modes and for arbitrarily small values of the QND coupling parameters. Furthermore, the resources required for the perfect transfer are always finite.
The interface can be also implemented probabilistically, thus eliminating the necessity of the feed-forward correction performed on the matter system. This is especially interesting for the state preparation of non-classical states of atoms or mechanical oscillators, where non-classical states of light are used as a resource. For a comparison, a single QND interaction, employed by cite{filip08}, allows to transmit a single-photon state perfectly, but a superposition of coherent states is transmitted with a reduction of the amplitude proportional to the strength of the QND coupling. The presently proposed method does not suffer from this -- the state of light is perfectly transferred to the atomic or mechanical system, without any feed-forward correction required.
Finally, this paper demonstrates a principal theoretical capability of constructing a perfect interface with help of QND couplings. Actual experimental considerations strongly depend on the employed physical system and will be addressed in a separate publication.
\medskip
\noindent {\bf Acknowledgments}
This research has been supported by projects MSM 6198959213, LC06007 and Czech-Japan Project ME10156 (MIQIP) of the Czech Ministry
of Education. We also acknowledge grant 202/08/0224 and P205/10/P319 of GA CR and and EU grant FP7 212008 - COMPAS.
|
\section{Introduction}
\label{sec: intro}
Since many decades, a lot of attention has been devoted to
ferromagnetic semiconductors. More recently, the intense research
has been triggered by the synthesis of the (III,Mn)V diluted
magnetic semiconductor (Ga,Mn)As,\cite{Ohno:1996_APL} which has
become the canonical example of a dilute ferromagnetic
semiconductor.\cite{Matsukura:2002_B,Dietl:2008_B} It has been
demonstrated that a number of pertinent properties of this
material can be explained by the $p-d$ Zener
model.\cite{Dietl:2000_S,Dietl:2001_PRB,Jungwirth:2006_RMP,Dietl:2008_B}
Magnetic anisotropy of strained (Ga,Mn)As layers can be calculated
within this theory, and many experimental
studies\cite{Tang:2003_PRL,Welp:2003_PRL,Sawicki:2004_PRB,Sawicki:2005_PRB,Wang:2005_PRL,Liu:2006_g,Thevenard:2007_PRB,Gourdon:2007_PRB,Gould:2008_a}
were devoted to verify its predictions. However, despite these
intense studies, some important features of magnetic anisotropy in
this system are at present not completely understood.
An example of such a property is a rather strong in-plane uniaxial
magnetic anisotropy of epitaxial (Ga,Mn)As layers grown on GaAs
substrates of $(001)$ orientation. Owing to the presence of the
twofold symmetry axes $[100]$ and $[010]$, the in-plane
zinc-blende directions $[110]$ and $[\overline110]$ are expected
to be equivalent. Yet, as implied by the character of magnetic
anisotropy, the symmetry is lowered from $D_{2d}$ to $C_{2v}$,
possibly due to the growth-induced lack of symmetry between the
bottom and the top of the
layer,\cite{Welp:2004_APL,Sawicki:2004_PRB,Sawicki:2005_PRB} which
can be phenomenologically described by introducing a shear
strain.\cite{Sawicki:2005_PRB,Zemen:2009_PRB,Glunk:2009_PRB}
Since these symmetry considerations are limited to $(001)$ layers,
investigation of layers grown on substrates of other orientations
may not only allow to compare experimental observations with
predictions of the $p-d$ Zener model in a more general situation,
but also provide information from which conclusions on the nature
of the additional anisotropy can be drawn.
In this paper we present results of studies on
$\mathrm{Ga}_{1-x}\mathrm{Mn}_{x}\mathrm{As}$ layers grown by
low-temperature molecular beam epitaxy (MBE) on GaAs substrates
with the $(113)\mathrm{A}$ orientation. Previously, magnetic
anisotropy in such films was probed at low temperatures by
magnetoresistance\cite{Wang:2005_PRB_b,Limmer:2006_PRB,Limmer:2006_MJ},
scanning Hall probe microscopy\cite{Pross:2006_JAP} and
ferromagnetic resonance
measurements.\cite{Limmer:2006_MJ,Bihler:2006_APL,Liu:2007_JPCM}
Experimental techniques employed here include superconducting
quantum interference device (SQUID) magnetometry, ferromagnetic
resonance (FMR), and polar magnetooptical Kerr effect (PMOKE). Our
measurements are carried out over a wide temperature and magnetic
field range. We find that magnetic anisotropy can be consistently
described taking into account three contributions: a uniaxial
anisotropy with the hard axis tilted from $[113]$ toward the
$[001]$ direction, an in-plane uniaxial anisotropy with the easy axis
along the $[\overline110]$ direction, and a cubic anisotropy with
easy $\langle 100 \rangle$ directions. The general form of
anisotropy is, therefore, similar to the case of (001) films, but
the direction of the hard axis is found to be neither along [001]
nor perpendicular to the film plane in the (113) case. The
accumulated experimental results allow us to determine how the
three relevant magnetic anisotropy constants $K$ as well as the
tilt angle depend on the temperature. We find that the magnitudes
of energies corresponding to the competing cubic and uniaxial
anisotropies in the (001) plane depend, as could be expected, as
the fourth and second power of spontaneous magnetization $M(T)$,
respectively. In contrast, a complex dependence on $M(T)$ is
observed in the case of the energy characterizing the out-of-plane
uniaxial anisotropy. We assign this behavior to the
spin-splitting-induced and, hence, temperature dependent
redistribution of holes between the valence band subbands that are
characterized by different directions of the angular momentum and,
hence, of the easy axes.
\begin{figure}[bp]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig01}
\end{center}
\caption{(Color online) Crystallographic directions for a GaAs substrate of (113) orientation.}
\label{fig: 1}
\end{figure}
In the theoretical part, we present a theory of magnetic
anisotropy in epitaxially strained layers of (Ga,Mn)As and related
systems within the $p-d$ Zener model. Our approach generalizes
earlier theories developed for (001)
films\cite{Dietl:2000_S,Dietl:2001_PRB,Abolfath:2001_PRB,Glunk:2009_PRB,Zemen:2009_PRB}
by allowing for an arbitrary crystallographic orientation of the
substrate. Similarly to previous
studies,\cite{Sawicki:2005_PRB,Zemen:2009_PRB} in order to explain
the experimental findings, we introduce an additional shear strain
whose three components constitute adjustable
parameters. We also take into account the Hamiltonian terms linear in $k$ and find
that they give a minor contribution to the magnitude of magnetic
anisotropy constants.
\section{Samples and experiment}
\label{seq: samples}
We study a 50~nm thick
$\mathrm{Ga}_{1-x}\mathrm{Mn}_{x}\mathrm{As}$ layer which has been
grown on a (113)A GaAs substrate (see Fig.~\ref{fig: 1}) by low
temperature MBE.\cite{Reinwald:2005_JCG} The total Mn
concentration of $x = 6.4\%$ has been determined by means of
secondary ion mass spectrometry, however only more than twice
lower value of an effective Mn concentration $x_{\text{eff}}$ can
be inferred from low temperature experimental saturation
magnetization, $M_{\text{exp}}$. This reduction of $x$ is
primarily caused by a presence of Mn interstitials. These point
defects act as double donors and form strongly coupled spin
singlet pairs with neighbor substitutional Mn
cations.\cite{Jungwirth:2006_RMP,Yu:2002_PRB,Blinowski:2003_PRB}
These pairs neither participate in the ferromagnetic order, nor
they contribute to $M$. Thus, the effective concentration of Mn
ions which generates $M_{\text{exp}}$ gets reduced to
$x_{\text{eff}} = x - 2x_{\text{I}}$, where $x_{\text{I}}N_0$ is
the concentration of the Mn interstitials and $N_0$ is the cation
concentration. However, the experimentally measured
$M_{\text{exp}}$ is further reduced by holes magnetization, $M_h$,
which is oppositely oriented to magnetization of Mn spins,
$M_{\text{Mn}}$, and so $M_{\text{Mn}}$ = $M_{\text{exp}}$ +
$|M_h|$ should be used to calculate $x_{\text{eff}}$, with $M_h$
being computed in the framework of the the mean-field $p-d$ Zener
model.\cite{Dietl:2000_S,Dietl:2001_PRB} We perform these
calculations in a self-consistent way taking the hole
concentration as $p = N_0(x- 3x_{\text{I}}) = N_0(3x_{\text{eff}}
- x)/2$, that is neglecting other charge compensating defects.
The open air post growth annealing at temperatures below or
comparable to the growth
temperature\cite{Hayashi:2001_APL,Edmonds:2002_APL} is a
frequently used procedure for improving material parameters of
(Ga,Mn)As, since the corresponding out-diffusion and passivation
of Mn interstitials\cite{Edmonds:2004_PRL} increases
$x_{\text{eff}}$, $p$, and eventually the Curie temperature
$T_{\mathrm{C}}$. Therefore in order to widen the parameter space
employed here to study the magnetic anisotropy in this compound we
investigate both the as-grown material (sample S1) and the samples
annealed at $200^{\circ}$C for 1.5 hour (sample S2) and 5 hours
(sample S3). Taking the determined values of $M_{\text{exp}}$ data
we end up with $x_{\text{eff}}$ = 2.7, 3.1, and 3.3\% and $p=
2.0$, 3.3, and $3.8 \times 10^{20}$~cm$^{-3}$ for which calculated
values of $T_{\mathrm{C}}$ = 46, 73, and 85~K compares favorably
with the experimentally established values of 65, 77, and 79~K,
for samples S1, S2, and S3, respectively.
Magnetic properties referred to above and described further on
have been obtained by employing a Quantum Design MPMS XL-5
magnetometer. A special demagnetization procedure has been
employed to minimize the influence of parasite fields on
zero-field measurements. The temperature dependence of remnant
magnetization, TRM, serves to obtain an overview of magnetic
anisotropy as well as to determine $T_{\mathrm{C}}$
(Sec.~\ref{sec: rem}). After cooling the sample across
$T_{\mathrm{C}}$ down to $5\,\mathrm{K}$ in the external magnetic
field of 0.1~T, the field is removed, allowing magnetization to
assume the direction along the closest easy axis. The magnitude of
the magnetization component along the magnet axis, TMR$_i$, is then
measured while heating, where $i$ indicates one of the three mutually
orthogonal directions of the magnetizing field, corresponding to
the surface normal $\bm{n}_1 = [113]$ and the two edges $\bm{n}_2
= [33\overline2]$, $\bm{n}_3 = [\overline110]$, as depicted in
Fig.~\ref{fig: 1}. Since, except to the immediate vicinity of the
spin reorientation transition, magnetization of (Ga,Mn)As films
tend to align in a single domain state, the measurements performed
for the three orthogonal axes provide the temperature dependence
of the magnetization magnitude and direction.
To study magnetic anisotropy in a greater detail, magnetic
hysteresis loops $M_i(H)$ have been recorded in external magnetic
field in the range of $\pm 0.5$~T along the three directions $i$.
The measurements have been carried out at various temperatures,
and the parameters of the anisotropy model (Sec.~\ref{sec:
phenomodel}) have been fitted to reproduce the magnetization data.
To cross check magnetic anisotropy constants obtained from SQUID
studies, FMR measurements have been performed at $\omega/2\pi =
9.3 \, \mathrm{GHz}$ and $T = 10$~K. We have performed angle
dependent measurements of the resonance field in the four
different crystallographic planes $(\overline110)$,
$(33\overline2)$, $(113)$ and $\frac{1}{2\sqrt{11}}(3-\sqrt{11},
3+\sqrt{11}, -2)$. As discussed in Sec.~\ref{sec: fmr}, the FMR
data are in a good agreement with the anisotropy model, employing
parameters determined from the SQUID measurements.
\section{Experimental results}
\subsection{Overview of magnetic anisotropy}
\label{sec: rem}
The TRM studies of all three samples are summarized in
Fig.~\ref{fig:TRM}. We immediately see that the
TRM$_{[\overline110]}$ component of TRM is the strongest for all
of the samples and that at elevated temperatures its magnitude is
nearly equal to the saturation magnetization $M(T)$, established
by the measurement in $\mu_0H=0.1$~T. Since the magnitude of the
other two magnetization components is vanishingly small, we find
that in this temperature range the in-plane uniaxial anisotropy
with the easy axis along $[\overline110]$ direction dominates.
This perfectly uniaxial behavior at $T \rightarrow T_{\mathrm{C}}$
allows us to use TRM$_{[\overline110]}$ to precisely determine
$T_{\mathrm{C}}$ in the studied samples (already given in the
previous section). On the other hand, below a certain temperature
$T^*$ (marked by arrow for every sample in Fig.~\ref{fig:TRM})
TRM$_{[\overline110]}$ gets visibly smaller than $M(T)$, and the
other in-plane TRM component, TRM$_{[33\overline2]}$, acquires
sizable values, followed at still lower temperatures by the
out-of-plane component TRM$_{[113]}$. This clearly indicates a
departure of the easy direction from the $[\overline110]$
direction below these characteristic temperatures. Such a scheme
turns out to be fully equivalent to the general pattern of
magnetic anisotropy in (001) (Ga,Mn)As under compressive
strain.\cite{Wang:2005_PRL,Welp:2003_PRL,Tang:2003_PRL,Sawicki:2005_PRB,Wang:2005_PRB_a,Kato:2004_JJAP,Welp:2004_APL}
In such films uniaxial anisotropy between $[110]$ and
$[1\overline10]$ directions, dominating at elevated temperatures,
gives way at low $T$ to biaxial anisotropy with in-plane $\langle
100\rangle$ easy axes. This spin reorientation transition takes
place at a temperature, at which uniaxial and biaxial anisotropy
constants equilibrate,\cite{Wang:2005_PRL} and is corroborated
numerically in our samples from analysis of the magnetization
processes presented in Sec.~\ref{sec: phenomodel}.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig02TD}
\end{center}
\caption{(Color online) Temperature dependence of remnant
magnetization components in all three samples (points). Black
lines: $M(H)$ at 0.1~T along $[\overline110]$ - the in-plane
uniaxial easy axis. Solid lines of lighter shades: the same
magnetization components calculated according to a model of only
two: uniaxial and biaxial magnetic anisotropies operating at (001)
plane and undergoing a spin reorientation transition at
temperatures $T^*$ (marked by arrows), as in (001) (Ga,Mn)As.}
\label{fig:TRM}
\end{figure}
In an analogy to (001) (Ga,Mn)As, let us assume for a moment that
$\bm M$ of a (113) sample remains (without a magnetic field) in
(001) plane. Then, a similar description in terms of two in-plane
anisotropies (one biaxial and one uniaxial) is possible.
Furthermore, assuming that the uniaxial anisotropy constant is
proportional to $M(T)^2$, the biaxial anisotropy constant is
proportional to $M(T)^4$ (Ref.~\onlinecite{Wang:2005_PRL}), and
that both equilibrate at $T^*$ we are able to model qualitatively
temperature induced rotation of magnetization in the sample and
calculate all three components of magnetization that would be
measured by SQUID. The thick solid lines in Fig.~\ref{fig:TRM}
show the results for sample S1 and we find them reproducing the
experimental findings reasonably well. Therefore we identify $T^*$
as the temperature at which the spin reorientation transition from
a biaxial anisotropy along $\langle 100 \rangle$ to uniaxial one
along $[1\overline10]$ takes place in this system. On the other
hand, the discrepancies seen in Fig.~\ref{fig:TRM} indicate that a
more elaborated model is needed. In particular, we can infer from
the low temperature TRM data that the orientation of $\bm M$ at
5~K moves actually away from $\langle 100 \rangle$ on annealing.
The angle between $\bm M$ and $\langle 100 \rangle$ is increasing
from 9, through 19 to 26~deg for samples S1, S2 and S3
respectively. At the same time the angle between $\bm M$ and (113)
plane is dropping from 13 to 7~deg. This indicates that the plane
in which both easy orientations of $\bm M$ reside at low $T$ is
tilting away from (001) towards (113) plane. This observation is
fully confirmed from the comprehensive analysis of the magnetic
anisotropy presented in the next section.
We remark here that the origin of the symmetry breaking between
$[110]$ and $[1\overline10]$ in (001) (Ga,Mn)As is still unknown
and it is very stimulating to see favoring the in-plane
$[1\overline10]$ also in layers of different surface
reconstruction than (001) GaAs.
\subsection{Experimental determination of anisotropy constants}
\label{sec: phenomodel}
In order to build up a more complete anisotropy description we
analyze full magnetization curves $M(H)$. It was shown by Limmer
et al. (Ref.~\onlinecite{Limmer:2006_MJ,Limmer:2006_PRB}) that an
accurate description the magnetic anisotropy in (113)A (Ga,Mn)As
requires at least four components: a cubic magnetic anisotropy
with respect to the $\langle 001 \rangle$ axes, uniaxial in-plane
anisotropy along the $[\overline110]$ direction, and two uniaxial
out-of-plane anisotropies along the $[113]$ and $[001]$
directions. The first two anisotropy components are commonly
observed in (001)-oriented (Ga,Mn)As samples, and as shown in the
previous section they are sufficient to provide even
semi-quantitative description in (113) case. The other two arise
from the epitaxial strain and demagnetizing effect, both of which
depend on the orientation of the substrate. In our approach we
combine the two out-of-plane magnetic anisotropy contributions
into a single one, with its hard axis oriented between the $[001]$
and $[113]$ directions. Accordingly, we write the free energy in
the form,
\begin{eqnarray}
\label{eq: F}
F & = & -\mu_0\bm{H} \cdot \bm{M} + K_C (w_x^2 w_y^2 + w_y^2 w_z^2 + w_z^2 w_x^2 ) \nonumber\\
& & {} + K_{u\overline110} \sin^2 \Theta \sin^2 \Phi \nonumber\\
& & {} + K_{u1} (\cos\Theta_A \cos\Theta
- \sin\Theta_A \sin\Theta \cos\Phi)^2.
\end{eqnarray}
Here, $K_C$, $K_{u\overline110}$ and $K_{u1}$ are the lowest order
cubic, in-plane uniaxial and out-of-plane uniaxial anisotropy
energies, respectively; $\Theta_A$ describes the angle between
the $K_{u1}$ hard axis and $[113]$ direction; $w_x$, $w_y$, and $w_z$ denote
direction cosines of the magnetization vector with respect to the main
crystallographic directions $\langle 100 \rangle$; $\Theta$ is the
angle between $\bm M$ and the $[113]$ direction, and $\Phi$ is the
angle between the projection of the $\bm M$ onto the sample plane and
the $[33\overline2]$ direction.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=0.98\columnwidth]{fig03SI}
\end{center}
\caption{(Color online) Examples of magnetization curves for the
as-grown sample measured along three, mutually orthogonal, major
sample directions $[\overline110]$, $[33\overline2]$, and $[113]$:
(a) below spin reorientation transition at 10~K and (b) above, at
38~K. Symbols indicate measurement points; lines represent the
best fit of the model described by Eq.~(\ref{eq: F}).}
\label{fig:M-H}
\end{figure}
By numerical minimizing of the free energy with respect to
$\Theta$ and $\Phi$ we are able to trace the rotation of $\bm M$,
starting from the given orientation, while sweeping or rotating
external magnetic field. Adjusting the obtained ``trace'' to the
experimental data we get the values of the four parameters of the
model. We perform this procedure numerically for every sample for
all three orientations and for all temperatures the $M_i(H)$
curves have been recorded. Figure~\ref{fig:M-H} shows an example
of the measured and fitted $M_i(H)$ for the sample S1 at two
different temperatures.
\begin{figure*}[tbp]
\begin{center}
\includegraphics[width=0.98\textwidth]{fig04SI}
\end{center}
\caption{(Color online) (a) and (b) points: temperature dependence
of $K_C$, $K_{u\overline110}$ and $K_{u1}$ (a) and angle
$\Theta_A$ (b) obtained from numerical fitting of Eq.~(\ref{eq:
F}) to experimental magnetization curves for all three samples
considered in this study. Solid, dashed and dotted arrows in (a)
indicate the spin reorientation temperature in samples S1, S2 and
S3, respectively. (c), (d) and (e): $K_C$, $K_{u\overline110}$ and
$K_{u1}$ dependence on $M^4$, $M^2$ and $M^2$, respectively. In
all panels, the various lines are guides for eye only.}
\label{fig:magnetic anisotropyconstants}
\end{figure*}
\begin{figure}[tbp]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig05SI}
\end{center}
\caption{(Color online) Angular dependence of the ferromagnetic
resonance fields for external magnetic field rotating in four
different crystallographic planes: $(113)$, $(\overline{1}10)$,
$(33\overline{2})$ and $\frac{1}{2\sqrt{11}}(3-\sqrt{11},
3+\sqrt{11}, -2)$. Points show measured resonance field values.
Lines show resonant field values calculated using anisotropy
energy described by Eq.~(\ref{eq: F}) and magnetic anisotropy
constants obtained from hysteresis loops.}
\label{fig: 5}
\end{figure}
Temperature dependence of the three magnetic anisotropy constants
and the angle $\Theta_A$ are presented in Fig.~\ref{fig:magnetic
anisotropyconstants}a and b, respectively. All $K_i$'s
monotonically decrease with temperature, and, like in (001)
(Ga,Mn)As, the cubic anisotropy constant $K_C$ (with $\langle 100
\rangle$ easy axes, see Fig.~\ref{fig:magnetic
anisotropyconstants}c) and in-plane uniaxial constant
$K_{u\overline110}$ (with $[1\overline10]$ easy axis, see
Fig.~\ref{fig:magnetic anisotropyconstants}d) are proportional to
$M^4$ and $M^2$, respectively, so confirming the validity of the
single domain approach used to analyze the observed magnetization
rotations. The $K_C(T)$ and $K_{u\overline110}(T)$ data point to
the presence of the spin reorientation transition in (001) plane.
This already inferred from TRM data magnetic easy axis changeover
must take place as $K_C(T)$ and $K_{u\overline110}(T)$ swap their
intensities in our samples. The relevant temperatures are marked
in Fig.~\ref{fig:magnetic anisotropyconstants}a by arrows.
Importantly, we find these temperatures to agree within 2-3~K with
those indicated in Fig.~\ref{fig:TRM}, what strongly underlines
the correctness of the approach we employ here to describe the
magnetic anisotropy in our samples. We note that the SRT shifts
to lower temperatures on going from sample S1 to S3, since on
annealing the in-plane uniaxial anisotropy gets strongly enhanced
relative to the cubic one (compare Figs.~\ref{fig:magnetic
anisotropyconstants}c and d).
In contrast, $K_{u1}$ shows a more complex dependence on $M^2$,
see Fig.~\ref{fig:magnetic anisotropyconstants}e. A
proportionality of the out-of-plane anisotropy constant to $M^2$
is seen only at low $M$, that is at high $T$. On lowering
temperature $K_{u1}$ departures from this trend, and the effect is
the strongest for the sample~S1. We connect this behavior with
the proximity of this system to another spin reorientation
transition, the transition from the hard to easy out-of-plane axis
of the $K_{u1}$ uniaxial magnetic anisotropy. Such a SRT takes
place in compressively strained (001) (Ga,Mn)As on lowering $T$,
and was already observed in samples with moderate or high $x$ but
rather low hole density\cite{Sawicki:2004_PRB}. The effect depends
on the ratio of valence band spin splitting to the Fermi energy.
Therefore the S1 sample, the one with the lowest $p$ is expected
to show the strongest deviations from the expected functional
form. Then on annealing, along with the increase of $p$, we expect
the so called in-plane magnetic anisotropy (for the compressively
strained layers) to become more robust [less dependent on the
magnitude of the valence band splitting, that is on $M(T)$], as
experimentally observed.
Finally, we comment on $\Theta_A$, the parameter that can serve as
a measure of the angle between an `easy plane' with respect to
$K_{u1}$ (perpendicular) hard axis and the sample face. As
indicated in Fig.~\ref{fig:magnetic anisotropyconstants}b that
angle remains nearly constant at elevated temperatures and shows a
weak, but noticeable turn towards [001] below temperatures which
can be associated with $K_{u\overline110} \Leftrightarrow K_C$
SRT. This behavior again indicates the departure of easy direction
of $\bm{M}$ from $[1\overline10]$ direction in the (113) plane.
However, the maximum determined value of $\Theta_A \cong 15^o$
indicates, that the rotation of $\bm{M}$ actually neither takes
place in the (001) plane, nor is it directed exactly towards
$\langle 100 \rangle$ directions. $\bm{M}$ rather follows a
complex route in between (001) and (113) planes, a conclusion that
is a numerical confirmation of the results of the simple analysis
of the TRM data presented in the previous section.
\subsection{Ferromagnetic resonance}
\label{sec: fmr}
A tool widely used to study magnetic anisotropy is ferromagnetic
resonance spectroscopy. Magnetic anisotropy in thin (Ga,Mn)As
films on (113)A GaAs was recently studied by
Bihler\cite{Bihler:2006_APL} and
Limmer\cite{Limmer:2006_MJ,Limmer:2006_PRB}. In a ferromagnetic
resonance experiment the magnetization vector $\bm M$ of the
sample precesses around its equilibrium direction in given
external magnetic field $\bm H$ with Larmor frequency $\omega_L$.
The resonant condition at a fixed frequency of microwaves $\omega$
is given by,
\begin{equation}
\label{eq: 2}
\left( \frac{\omega}{\gamma} \right)^2 =
\frac{1}{\sin^2\Theta} \left[ \frac{\partial^2 F}{\partial
\Theta^2} \frac{\partial^2 F}{\partial \Phi^2}
- \left( \frac{\partial^2 F}{\partial \Theta \, \partial
\Phi} \right)^2 \right].
\end{equation}
Here, $\gamma = g \mu_B \hbar^{-1}$ is the gyromagnetic ratio,
$g$ is the $g$-factor, $\mu_B$ the Bohr magneton, and $\hbar$ is
the Planck constant. Resonance field is obtained by evaluating
Eq.~(\ref{eq: 2}) at the equilibrium position of $\bm M$
($\partial F/\partial \Theta = 0$ and $\partial F/\partial \Phi =
0$).
In Fig.~5 the dependence of the measured resonant fields on the
orientation of the applied magnetic field is shown for the sample
S2, along with the results of a calculation made according to
Eq.~(\ref{eq: 2}) with the magnetic anisotropy parameters obtained
from SQUID magnetization curves. The agreement between the
calculation and the measured data is very good, indicating that
Eq.~(\ref{eq: F}) captures main features of magnetic anisotropy
and that the numerical procedure employed to extract the
anisotropy constants is correct.
\subsection{Magnetization reversal}
\begin{figure}[tbp]
\begin{center}
\includegraphics[width=0.9\columnwidth]{fig06SI}
\end{center}
\caption{(Color online) The out-of-plane magnetization component
dependence on the in-plane magnetic field (swept along
$[33\overline2]$). The measurement is performed on the as-grown
sample S1 at $T = 15$~K. The cartoons inserted in the figure
illustrate the process.}
\label{fig: 06}
\end{figure}
We end the experimental part evidencing an interesting mechanism
of the reversal of the out-of-plane magnetization component by an
in-plane magnetic field.
Below the spin reorientation transition (about 20-30~K) the
magnetization easy axes are moving close to the $\langle 100
\rangle$ directions, that is they are tilted up from the sample
face, the (113) plane, and so $\bm M$ is acquiring a sizable
non-zero component $M_{[113]}$. These two axes define a plane
lying between (113) and (001) planes, which shares the common
$[\overline110]$ direction with those two planes. Therefore, any
sweep of an external field, except that along the $[\overline110]$
direction, will result in a magnetization rotation across the
$[\overline110]$ line from one half of that plane (say that one
`above' sample face) to the other one (say `below') resulting in
$M_{[113]}$ reversal.
Such a process is illustrated in Fig.~\ref{fig: 06} in the most
interesting case, when the field is swept in the sample plane,
along $[33\overline2]$. We record the $M_{[113]}$ magnetization
component using the polar MOKE technique and a clear change of
sign of the signal evidences the reversal of $\bm M$ by the
application of an in-plane magnetic field. The cartoons inserted
in this figure visualize the mechanism of this reversal.
\section{Theory of magnetic anisotropy}
\label{sec: theory}
\subsection{$\bm{k} \cdot \bm{p}$ Hamiltonian}
The current theory describing the properties of the (Ga,Mn)As
ferromagnetic semiconductor is the $p$-$d$ Zener
model.\cite{Dietl:2000_S} In this model, the thermodynamic
properties are determined by the valence band carriers
contribution to the free energy of the system, which is calculated
taking the spin-orbit interaction into account within the $\bm{k}
\cdot \bm{p}$
theory\cite{Dietl:2000_S,Dietl:2001_PRB,Jungwirth:2006_RMP,Zemen:2009_PRB}
or tight binding model\cite{Werpachowska:2009_arXiv} with the
$p$-$d$ exchange interaction between the carriers and the
localized Mn spins considered within the virtual-crystal and
molecular-field approximations. Within this approach, magnetic
anisotropy depends on the strain tensor components.
The 6-band Luttinger $\bm{k} \cdot \bm{p}$ Hamiltonian of
a valence band electron in a zinc-blende semiconductor is a block
matrix (cf.\ Ref.~\onlinecite{Sliwa:2006_PRB}):
\begin{equation}
\mathcal{H}_{\mathrm{6x6}} = \begin{pmatrix}
H^{vv}& H^{vs} \\
H^{sv}& H^{ss}
\end{pmatrix},
\end{equation}
where
\begin{eqnarray}
\mathcal{H}^{vv} & = & - \frac{\hbar^2}{m} \left\{
\frac12 \gamma_1 k^2
- \gamma_2 \left[\left(J_x^2 - \frac13 J^2\right) k_x^2
+ c.p. \right] \right. \\
&& \left. {} - 2 \gamma_3 \left[
\{J_x, J_y\} \{k_x, k_y\} + c.p. \right] \right\} \nonumber \\
\mathcal{H}^{vs} & = & - \frac{\hbar^2}{m} \left[
-3\gamma_2 (U_{xx} k_x^2 + c.p.) \right. \\
&& \left. {} - 6 \gamma_3 (U_{xy} \{k_x, k_y\} + c.p.) \right]
\nonumber \\
\mathcal{H}^{ss} & = &
- \left( \Delta_0 + \frac{\hbar^2}{2 m} \gamma_1 k^2 \right)
\end{eqnarray}
(we use the notation of Ref.~\onlinecite{Sliwa:2006_PRB}). Our basis is related to that of Ref.~\onlinecite{Dietl:2001_PRB} as follows: $u_1 = -\left| \frac32, \frac 32 \right>$, $u_2 = -i\left| \frac32, \frac 12 \right>$, $u_3 = \left| \frac32, -\frac 12 \right>$, $u_4 = i\left| \frac32, -\frac 32 \right>$, $u_5 = -\left| \frac12, \frac 12 \right>$, $u_6 = i\left| \frac12, -\frac 12 \right>$, \emph{i.e.} we use the standard basis of angular momentum eigenvectors (notice the change of sign in $\left| \frac12, \frac 12 \right>$ and $\left| \frac12, -\frac 12 \right>$ with respect to Ref.~\onlinecite{Sliwa:2006_PRB} that accounts for the difference in the sign of $H^{vs}$, $H^{sv}$). In this basis the $p-d$ exchange Hamiltonian is:
\begin{equation}
\mathcal{H}_{pd} = B_G \begin{pmatrix}
2 ( \bm{J} \cdot \bm{w} )&
-6 ( \bm{U} \cdot \bm{w} )\\
-6 ( \bm{T} \cdot \bm{w} )&
- ( \bm{\sigma} \cdot \bm{w} )
\end{pmatrix},
\end{equation}
where $\bm{w} = \bm{M}/M$ and $B_G$
is given by equation~2 of Ref.~\onlinecite{Dietl:2001_PRB},
\begin{equation}
B_G = A_F \beta M /(6 g \mu_B),
\label{eq:B_G}
\end{equation}
while the strain Hamiltonian is:
\begin{eqnarray}
\mathcal{H}^{vv}_{\epsilon} & = & -b \Bigl[ (J_x^2 - \frac13 J^2) \epsilon_{xx} +
c.p. \Bigr] \\ & & {} - \frac{d}{\sqrt{3}} \Bigl[ 2\{ J_x, J_y \}
\epsilon_{xy} + c.p. \Bigr] \\
\mathcal{H}^{vs}_{\epsilon} & = & -3 b (U_{xx} \epsilon_{xx} + c.p.) \\
& & {} - \sqrt{3} d (2 U_{xy} \epsilon_{xy} + c.p.) \\
\mathcal{H}^{ss}_{\epsilon} & = & 0.
\end{eqnarray}
Since the strain tensor for the $(113)$ substrate orientation features non-zero non-diagonal components, it is necessary to include in the $\bm{k} \cdot \bm{p}$ Hamiltonian the so called $k$-linear terms, {\em i.e.}, terms linear in $\bm{k}$ and $\bm{\epsilon}$ coming \emph{via} second-order perturbation (Ref.~\onlinecite{Bir:1974_B}, \S 15) from the terms in the $8\times 8$ Kane Hamiltonian\cite{Pikus:1984_B} that mix the conduction and the valence bands. The corresponding Hamiltonian is
\begin{equation}
\mathcal{H}_{k\epsilon} = C_4 \begin{pmatrix}
\bm{J} \cdot \bm{\varphi}&
\frac32 \left( 1 - \frac\eta 2 \right)
\bm{U} \cdot \bm{\varphi}\\
\frac32 \left( 1 - \frac\eta 2 \right)
\bm{T} \cdot \bm{\varphi}&
\left( 1 - \eta \right) \bm{\sigma} \cdot \bm{\varphi}
\end{pmatrix},
\end{equation}
where $\eta = \Delta_0/(E_g + \Delta_0)$ and the components of the vector $\bm{\varphi}$ are $\varphi_z = \epsilon_{zx} k_x - \epsilon_{zy} k_y$ (c.p.). The numerical value given in
Ref.~\onlinecite{Pikus:1984_B} is $C_3/\hbar = 8 \times 10^5 \, \mathrm{m} \, \mathrm{s}^{-1}$, hence for $C_4 = -C_3/(2 \eta)$ we obtain $C_4 / \hbar = -2.2 \times 10^6 \, \mathrm{m/s}$.
\subsection{Strain tensor}
Determining the components of the strain tensor for an unrelaxed epitaxial layer grown on a lattice mismatched substrate can be considered a classical topic. The two possible approaches to this problem are (i) to solve a system of linear equations for the strain and stress components assuming that some components of those tensors vanish (this is our approach) or (ii) to determine the strain of the layer by minimizing the elastic energy (this is the approach formulated in Ref.~\onlinecite{Yang:1994_APL}). Our approach involves a transformation of the coordinate system that is feasible in general only using a computer algebra system. Using one we arrive to the form of the symmetric strain tensor that is in a perfect agreement with that of Ref.~\onlinecite{Yang:1994_APL}.
The result for a $(11n)$ oriented substrate is,
\begin{equation}
\bm{\epsilon}_{\text{epi}} = \frac{f}{A+B+C}
\begin{pmatrix}
B+C& -A& -D/2\\
-A& B+C& -D/2\\
-D/2& -D/2& B-2C
\end{pmatrix},
\end{equation}
where
\begin{eqnarray}
A & = & 3\left[n^2(c_{11}-c_{12})-(n^2-2)c_{44}\right] (c_{11}+2c_{12}), \\
B & = & 2 \left[(n^4+4n^2+1) (c_{11}-c_{12})+6 c_{44}\right] c_{44}, \\
C & = & (n^2-1)(n^2+2) (c_{11}+2c_{12})c_{44}, \\
D & = & 3n \left[(n^2+1)(c_{11}-c_{12})+2 c_{44}\right] (c_{11}+2c_{12}),\quad
\end{eqnarray}
and
\begin{equation}
f = -\Delta a/a = (a_0-a)/a
\end{equation}
is the relative lattice constant misfit.
Here, the components of the epitaxial strain tensor are given with respect to the coordinates $(x, y, z)$ associated with the crystallographic axes, $x = [100]$, $y = [010]$, and $z = [001]$ (\emph{i.e.} the quantities given above are the strain components that enter the $\bm{k} \cdot \bm{p}$ Hamiltonian).
Using the values $c_{11} = 119. \, \mathrm{GPa}$, $c_{12} = 53.8 \, \mathrm{GPa}$, and $c_{44} = 59.5 \, \mathrm{GPa}$ (Ref.~\onlinecite{Madelung:1991_B}, p.~105) we obtain
\begin{equation}
\bm{\epsilon}_{\text{epi}} = -\frac{\Delta a}{a}
\begin{pmatrix}
\phantom{-}0.9488 & -0.0512 & -0.3478 \\
-0.0512 & \phantom{-}0.9488 & -0.3478 \\
-0.3478 & -0.3478 & -0.6260
\end{pmatrix}.
\end{equation}
For the purpose of determining the strain components from X-ray diffraction data, the components of the strain tensor in the coordinate system associated with the epitaxial film are needed. We take as the coordinate system: $x' = [3, 3, -2]$, $y' = [-1, 1, 0]$, $z' = [1, 1, 3]$. The relative difference of the lattice constants along the $[113]$ direction between that layer and the substrate is
\begin{widetext}
\begin{eqnarray}
\label{eq:Delta_d}
\Delta d/d &= & \frac{[27(c_{11}-c_{12})+67c_{44}](c_{11}+2c_{12})}{9 (c_{11} - c_{12})(c_{11} + 2 c_{12}) + (101 c_{11}-34c_{12}+4c_{44}) c_{44}} \Delta a/a = 1.7284 \, \Delta a/a.
\end{eqnarray}
For the sake of completeness we notice that there is also a shear strain component
\begin{eqnarray}
\epsilon_{x'z'} &=& \frac{12\sqrt{2}[(c_{11}-c_{12})-2c_{44}](c_{11}+2c_{12})}{9 (c_{11} - c_{12})(c_{11} + 2 c_{12})+(101c_{11}-34c_{12}+4c_{44}) c_{44}}\Delta a/a = -0.2746 \, \Delta a/a.
\end{eqnarray}
\end{widetext}
Following Ref.~\onlinecite{Sawicki:2005_PRB}, to account for the
mechanism which generates the in-plane uniaxial anisotropy in
$(001)$ samples, we incorporate in the $p-d$ Zener model an
additional Hamiltonian term corresponding to shear strain
$\bm{\epsilon}'$,
\begin{equation}
\bm{\epsilon} = \bm{\epsilon}_{\text{epi}} + \bm{\epsilon}'.
\end{equation}
In case of a $(001)$-oriented substrate the additional strain $\bm{\epsilon}'$ has a non-zero $xy$ component, $\epsilon'_{xy}$. The corresponding anisotropy is of the form $K_{xy} w_x w_y$ (as in Ref.~\onlinecite{Bowden:2008_JPCM}), hence it is a difference of uniaxial anisotropies on the $[110]$ and $[\overline110]$ directions, and the anisotropy field is $H_u = 2 K_{xy}/(\mu_0 M)$ (this is the field required to align magnetization along the hard axis, e.g. $[110]$; only $H_u/2$ is required to align magnetization along the $z$ direction). In the case of a $(113)$-oriented substrate the additional strain may have more non-zero components. We assume that the mirror symmetry with respect to the $(\overline110)$ plane is preserved, hence $\epsilon'_{xz} = \epsilon'_{yz}$.
\subsection{Numerical procedure}
The numerical procedure serving to determine the magnetic anisotropy from the Hamiltonian matrix is described in Ref.~\onlinecite{Dietl:2001_PRB}. Let us note that including the $k$-linear terms in the Hamiltonian leads to a tenfold increase of the processing time, although in specific cases it is possible to generate a symbolic expression for the characteristic polynomial of the $6\times 6$ Hamiltonian matrix. Moreover, since numerical interpolation of the dependence of the hole concentration on the Fermi energy may lead to uncontrollable inaccuracies, an alternative procedure that avoids those inaccuracies is to directly integrate the energy of the carriers in the momentum space. However, the integration has to be done separately for each hole concentration (this is an advantage if a single hole concentration is specified). Moreover, one still needs to solve the inverse eigenvalue problem to find the discontinuities of the integrand.
In a numerical calculation, it is possible to determine the full magnetic anisotropy by computing the free energy of the carriers for a number of directions of magnetization. In our case we choose a grid of directions that is rectangular in the spherical coordinates ($w_x = \sin\theta\cos\phi$, $w_y = \sin\theta\sin\phi$, $w_z = \cos\theta$), \emph{i.e.} $\theta = \theta_i$ and $\phi = \phi_j$, where $\cos\theta_i$, $i = 1, 2, \ldots, N_\theta$ are the nodes of a Gaussian quadrature and $\phi_j = 2 \pi j / N_\phi$, $j = 0, 1, \ldots, N_\phi-1$ are equally spaced. Then, following the method used in the software package SHTOOLS\cite{Wieczorek:2009_WWW} (routine \texttt{SHExpandGLQ}) we expand the magnetic anisotropy (free energy) into a sum of low-order spherical harmonics. Since the free energy is even, choosing even $N_\phi$ allows to restrict the grid to a half of the sphere. We use the standard quantum mechanics (orthonormalized) spherical harmonics $Y_{lm}(\theta, \phi)$, and denote the coefficients of this expansion $r_{lm}, m = 0, 1, \ldots, l$, and $s_{lm}, m = 1, 2, \ldots l$,
\begin{equation}
\label{eq:F_theory}
F_c = \sum_{l} \left[ r_{l0} Y_{l0} + \sum_{m = 1}^{l} (r_{lm} \Re Y_{lm} + s_{lm} \Im Y_{lm}) \right],
\end{equation}
where the outer sum is over $l = 0, 2, \ldots, N_{\theta}-1$. The scalar product in this representation is diagonal, with a weight of $1$ for $r_{l0}$ and $1/2$ for $r_{lm}$, $s_{lm}$, $m > 0$.
\subsection{Magnetic anisotropy}
There are a few sources of magnetic anisotropy in epitaxial $\mathrm{Ga}_{1-x}\mathrm{Mn}_{x}\mathrm{As}$: the cubic anisotropy of the valence band, epitaxial strain, the additional off-diagonal strain $\epsilon'$, and the shape anisotropy caused by the demagnetization effect. Since $\epsilon'$ is unknown, it is inevitable to parametrize the anisotropy in a manner that separates the components affected by $\epsilon'$ from what is predictable. If we measure $(\theta, \phi)$ in the spherical harmonic representation of the carriers' free energy with respect to the crystallographic axes, the above spherical harmonic representation allows one to separate the components brought about by a non-zero value of $\epsilon'$ from the remaining sources of magnetic anisotropy. Indeed, as far as $l = 2$ is concerned, $\epsilon'$ affects primarily only $K_{xy}$, $K_{yz}$, and $K_{xz}$, where $K_{xy} = \sqrt{\frac{15}{8\pi}} s_{22}$, $K_{yz} = -\sqrt{\frac{15}{8\pi}} s_{21}$, and $K_{xz} = -\sqrt{\frac{15}{8\pi}} r_{21}$. The remaining components are $r_{20}$ and $r_{22}$. As $r_{22} = 0$ due to the mirror symmetry, they can be collected into one term $K_{u001} (w_z^2 - \frac13)$, with $K_{u001} = \frac34 \sqrt{\frac{5}{\pi}} r_{20}$. Thus, we describe the $l = 2$ anisotropy (without the demagnetization contribution) by $K_{xy}$, $K_{xz} = K_{yz}$, and $K_{u001}$. Finally, the cubic anisotropy corresponds to $(r_{40}, r_{44}) = -\frac{2 \sqrt{\pi}}{15} K_C (1, \sqrt{10/7})$.
We have to relate now the components of the spherical harmonic
representation, $r_{lm}$ and $s_{lm}$, to the experimentally
determined magnetic anisotropy constants $K_C, K_{u1},
K_{u\overline110} $, and $\Theta_A$, as specified in Eq.~\ref{eq: F}
and presented in Fig.~\ref{fig:magnetic anisotropyconstants}.
Since the demagnetization effect adds to
$(r_{20}, r_{21}, s_{21}, r_{22}, s_{22})$ the contribution
$\frac{4\sqrt{10 \pi}}{165} K_d (4 \sqrt2, -3 \sqrt3, -3 \sqrt3, 0, \sqrt3)$,
with $K_d = \mu_0 M^2/2$, the constants are related
to those of Eq.~(\ref{eq: F}) as follows,
\begin{eqnarray}
K_{u001} & = & \frac{1 + 3 \cos 2\Theta_A'}4 K_{u1} - \frac12 K_{u\overline110} - \frac{8}{11} K_d, \label{eq:Ku001} \quad \\
K_{xy} & = & \frac{1 - \cos 2\Theta_A'}2 K_{u1} - K_{u\overline110} - \frac2{11} K_d, \label{eq:Kxy} \\
K_{xz} & = & -\frac{\sqrt2 \sin 2\Theta_A'}{2} K_{u1} - \frac{6}{11} K_d, \label{eq:Kxz}
\end{eqnarray}
where $\Theta_A' = \Theta_A - \arccos(3/\sqrt{11})$.
\begin{figure}[b]
\begin{center}
\includegraphics[width=0.95\columnwidth]{deltakl}
\end{center}
\caption{(Color online) The contribution of the $k$-linear terms to the $[001]$ uniaxial anisotropy for $C_4 / \hbar = -2.18 \times 10^6 \, \mathrm{m/s}$ and $\epsilon_{xy} = 0.05\%$.}
\label{fig: 7}
\end{figure}
\begin{figure}[b]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig_08}
\end{center}
\caption{(Color online) Hole concentration dependence of the
in-plane uniaxial anisotropy field due to shear strain
$\epsilon_{xy} = 0.05\%$ for various values of the valence-band
spin-splitting.}
\label{fig: 8}
\end{figure}
We carry out numerical calculations with band structure
parameters and deformation potentials specified
previously.\cite{Dietl:2001_PRB,Sawicki:2005_PRB} We include
hole-hole exchange interactions {\em via} the Landau parameter of
the susceptibility enhancement, $A_F = 1.2$
(Ref.~\onlinecite{Dietl:2001_PRB}). This parameter, assumed here
to be independent of the hole density and strain, enters into the
relation between $M$ and $B_G$, but also divides the anisotropy
constants. More specifically, we make the calculation with $B_G$
enhanced by the factor $A_F$, and divide the resulting anisotropy
constants by $A_F^{n-1}$, where $n$ is the power of
magnetization~$M$ to which a given anisotropy constant is
proportional. The result is proportional to $A_F$. We have $n = 2$
for the uniaxial anisotropies and $n = 4$ for the lowest order
cubic anisotropy (the proportionality holds for $B_G$ smaller than
a few meV). We note that the cubic anisotropy field shown in
Fig.~9 of Ref.~\onlinecite{Dietl:2001_PRB} was divided by $A_F$
rather than $A_F^3$.
To evaluate the effect of the $k$-linear terms, we use a non-zero
value of $C_4$ and calculate the difference of the resulting
anisotropy with respect to the $C_4 = 0$ case. This difference has
only one noticeable component, $\Delta r_{20} = r_{20}(C_4 \ne
0)-r_{20}(C_4 = 0)$, which corresponds to a uniaxial anisotropy
with a $[001]$ axis (or with $[100]$, $[010]$ axes for
$\epsilon_{yz} \ne 0$, $\epsilon_{xz} \ne 0$ respectively). A plot
of $\Delta K_{u001} \propto \Delta r_{20}$ is shown in
Fig.~\ref{fig: 7} for $\epsilon_{xy} = 0.05\%$ (as implied by the
symmetry, $\Delta K_{u001}$ is second order in $\epsilon_{xy}$).
The values are rather small. In fact, assuming $\Delta a/a =
0.5\%$, we have $\epsilon_{xz} = \epsilon_{yz} \approx 0.17\%$,
and the magnitude of $\Delta K_{u100} = \Delta K_{u010}$ is below
$10 \, \mathrm{J}/\mathrm{m}^3$ if we consider the epitaxial
strain only. This estimate appears to remain valid in case of a
general strain of a similar magnitude, although other anisotropy
components are affected as well and the dependence on strain
components is non-linear. However, we stress that this estimate
depends on the value of the parameter $C_4$, which is somewhat
uncertain and may be different for the ordinary strain and
$\bm{\epsilon}'$. Considered this, it is justified to set $C_4 =
0$ in the remaining part of this paper.
Before we proceed to the calculations specific to the particular
samples, we make a remark that the data originally shown in Fig.~6
of Ref.~\onlinecite{Sawicki:2005_PRB} were not correct due to a
numerical error in the form of the strain Hamiltonian. We show
corrected results for~$H_u$ in Fig.~\ref{fig: 8}. The present
results are in agreement with Fig.~17 of
Ref.~\onlinecite{Zemen:2009_PRB} (remember that our model includes
the parameter the Landau parameter $A_F$, neglected in
Ref.~\onlinecite{Zemen:2009_PRB}).
As discussed previously,\cite{Sawicki:2004_PRB,Sawicki:2005_PRB}
owing to sign oscillations of anisotropy constants, the direction
of magnetization can be changed by temperature ($B_{\text{G}}$) or
the hole concentration, particularly in the vicinity of $p =
6\times 10^{20}$ and $1\times 10^{20}$~cm$^{-3}$, according to the
results displayed in Fig.~8 The corresponding in-plane spin
reorientation transition has indeed been observed by some of us
either as a function of temperature\cite{Sawicki:2005_PRB} or the
gate voltage in metal-insulator semiconductor
structures\cite{Chiba:2008_N,Sawicki:2010_NP} in these two hole
concentration regions in (Ga,Mn)As, respectively.
\section{Comparison between experiment and theory}
We detailed above a microscopic model of the magnetic anisotropy
in a DMS. In order to assess the applicability of this model to an
arbitrarily oriented DMS we compare its predictions with
experimental findings for (113) (Ga,Mn)As.
First, we specify the magnitude of the lattice mismatch to
establish the components of the strain tensor. According to Fig.~2
of Ref.~\onlinecite{Daeubler:2006_APL}, for a $(113)$ sample
containing 6.4\% of Mn we expect $\Delta d/d = 5.6 \times 10^{-3}$
which, employing Eq.~\ref{eq:Delta_d}, translates into $\Delta a/a
= 0.323\%$. We assume this value throughout this section.
Then, using the values of $x_{\text{eff}}$ already established in
Sec.~\ref{seq: samples}, we calculate
\begin{equation}
M_{\text{Mn}} = x_{\text{eff}} N_0 S g \mu_B
\label{eq: M_Mn}
\end{equation}
and obtain $B_G$ for each of our samples from Eq.~\ref{eq:B_G}. It
is worth repeating here, that the established upon total $x$ and
$M_{\text{exp}}$ values of $x_{\text{eff}}$ and $p$ reproduce,
within the same model, experimental values of $T_{\text{C}}$
remarkably well. This boosts our confidence in the accuracy of the
material parameters used here for the computations of the magnetic
anisotropy and gives a solid ground to the presented conclusions.
The calculations are performed as a function of $p$, employing for
each sample the corresponding value of $B_G$: $-16.4$, $-18.7$, and
$-19.7$~meV for samples S1, S2, and S3, respectively. The results
are presented in Fig.~\ref{fig:All_K} as curves, while full
symbols represent experimentally established values of the
anisotropy constants (a square for the sample S1, a triangle for
S2, and a circle for S3). The experimental $K_{u001}$, $K_{xy}$,
$K_{xz}$, and $K_C$ are obtained from $K_{u\overline110}$,
$K_{u1}$, $K_C$ and $\Theta_A$ using Eqs.~\ref{eq:Ku001},
\ref{eq:Kxy} and \ref{eq:Kxz}. We take the $T = 5 \, \mathrm{K}$
experimental anisotropy data, as this is what is consistent with
the $T = 0$ limit, implicitly assumed in Eq.~\ref{eq: M_Mn}.
\begin{figure*}[htp]
\begin{center}
\includegraphics[width=0.95\textwidth]{FigAllK}
\end{center}
\caption{(Color online) Lines: theoretical dependence of the
anisotropy constants (a) $K_{u001}$, (b) $K_{xy}$, (c) $K_{xz}$,
and (d) $K_C$ on the hole concentration~$p$, calculated within
mean-field Zener model. The values of the exchange parameter $B_G$
and the lattice constant mismatch $\Delta a/a$ are specific to the
investigated samples and are specified in the text. Symbols
depict values determined from the experiment. $\bm{\epsilon}' = 0$
is assumed here except for (c), where also the case of
$\epsilon'_{xz} = -0.1\%$ is included.}
\label{fig:All_K}
\end{figure*}
We start by discussing the strongest component of the magnetic
anisotropy, the $K_{u001}$ term. The calculated curves are
presented in Fig.~\ref{fig:All_K}a. The calculations have been
performed without introducing the fictitious shear strain
$\bm{\epsilon}'$ (\emph{i.e.} $\epsilon' = 0$). However, since we know
that the magnitude of $K_{u001}$ is negligibly affected by
$\bm{\epsilon}'$, the results should already match the
experimental data, and indeed they do. Although the spread of the
experimental points in Fig.~\ref{fig:All_K}a is significantly
larger than that of the theoretical curves, the correspondence
between the computed and experimental values is good, and it has
been achieved without introducing into the model any adjustable
parameters. This has been only possible by including the hole
liquid magnetization in the calculation of $x_{\text{eff}}$. When
$M_h$ is disregarded, the experimental values of $K_{u001}$ are
systematically above the \emph{maxima} of the theoretical curves.
As already mentioned, in the case of the (001) (Ga,Mn)As films an
additional low-symmetry term has to be introduced into the
Hamiltonian in order to reproduce the experimentally observed
uniaxial in-plane magnetic anisotropy. In a general case of an
arbitrarily oriented substrate, there are three anisotropy
constants of this kind, $K_{xy}$, $K_{yz}$, and $K_{xz}$. In the case of
(113) (Ga,Mn)As, the symmetry requires that $K_{xz} = K_{yz}$, the
assumption confirmed by the description of the experimental
results. The two relevant anisotropy constants $K_{xy}$ and
$K_{xz}$ are presented in Figs.~\ref{fig:All_K}(b) and~(c), and
are seen to be non-zero even in the absence of a symmetry lowering
perturbation, $\epsilon'_{xy} = 0$. The computed
magnitude of $K_{xy}$ for $\epsilon'_{xy} = 0$ yields an
acceptable agreement with the experimental data. We note, moreover,
that an exact match is possible when allowing for non-zero values
of $\epsilon'_{xy} \cong 0.001$\% for sample S1 and 0.01\% for
samples S2 and S3.
In contrast, according to Fig.~\ref{fig:All_K}(c), the theoretical
description of the experimental values of the $K_{xz}$ anisotropy
constant requires a quite sizable value of the corresponding strain component
$\epsilon'_{xz} = -0.1\%$. Thus, remarkably and contrary to the case of
(001) (Ga,Mn)As, one barely needs any in-plane shear strain to
reproduce in-plane uniaxial anisotropy, whereas for the
out-of-plane component two times stronger shear strain is needed
comparing to the (001) case. This finding should be taken as a
strong evidence that the in-growth surface reconstruction, and a
related with it orientational preferences of Mn incorporation must
play a decisive role in the mechanism leading to the lowering of
magnetic symmetry.
Finally, we turn to the case of the cubic anisotropy constant
$K_C$, shown in Fig.~\ref{fig:All_K}d. Since $K_C$ shows only a
small sensitivity to $\bm{\epsilon}'$, we present the results of
computations only for $\bm{\epsilon}' = 0$. We find that similarly
to the (001) case,\cite{Sawicki:2004_PRB} the present theory
underestimates the magnitude of $K_C$, particularly in the low
hole concentration region, where the theoretically expected change
of sign of $K_C$ is not observed experimentally. The origin of
this discrepancy, and in particular its relation to the symmetry
lowering perturbation is presently unknown.
We have also examined theoretically how the particular anisotropy
constants depend on magnetization $M$. As could be expected, and
in a qualitative agreement with the experimental finding shown in
Fig.~\ref{fig:magnetic anisotropyconstants}, the uniaxial
anisotropy constants $K_{u001}$, $K_{xy}$, and $K_{xz}$ (or
equivalently $K_{u1}$ and $K_{u\overline110}$) are proportional to
$M^2$, whereas $K_C$ to $M^4$, except to the hole concentration
region in the immediate vicinity of the sign change.
\section{Conclusions}
We have investigated the magnetic properties of as-grown and
annealed (Ga,Mn)As layers grown by MBE on GaAs substrates of the
$(113)$ orientation and provided the most complete to date
description of the magnetic anisotropy in the whole temperature
range up to $T_{\mathrm{C}}$. At higher temperatures the
$[\overline110]$ direction is the easy magnetization axis before
and after annealing. At low temperature the spin reorientation
transition to a pair of easy axes near the $[100]$ and $[010]$
directions takes place and to a first approximation, magnetization
behavior as a function of temperature is similar to that observed
in $(001)$ (Ga,Mn)As in the absence of an external magnetic
field.\cite{Wang:2005_PRL} However, the magnetization vector
resides in a plane close to (001) plane only for a low value of
the hole concentration. When it increases, the plane rotates along
$[\overline110]$ towards the sample face (113), and the two cubic easy
directions move towards the $[33\overline2]$ direction.
We have estimated the values of magnetic anisotropy constants by
fitting our phenomenological model to the hysteresis loops
measured by SQUID. The comparison to results of FMR measurements
confirms the correctness of this approach. The obtained values of
the cubic and uniaxial in-plane magnetic anisotropy constants are
proportional to $M^4$ and $M^2$, respectively. Inflections from
the $M^2$ dependence of the out-of-plane uniaxial anisotropy
constant indicate a proximity to another spin reorientation
transition at which the out-of-plane axis becomes easy on lowering
temperature. It has been evidenced by MOKE that it is possible to
reverse the out-of-plane magnetization component by applying an
in-plane magnetic field.
For the hole and effective Mn concentrations determined from the
values of saturation magnetization and the total Mn concentration,
the $p-d$ Zener model explains, with no adjustable parameters, the
magnitude of the Curie temperature as well as the sign and
magnitude of the uniaxial $[001]$ anisotropy constant $K_{u001}$
caused by biaxial strain. At the same time, however, the predicted
values of the cubic anisotropy constant are smaller than those
found experimentally in the hole concentration range studied here.
For the substrate orientation in question there are two additional
non-zero second order ($l = 2$) components $K_{xy}$ and $K_{xz}$.
The comparison of their experimental and theoretical values points
to the presence of an additional shear strain. The non-vanishing
components of this additional strain are $\epsilon_{xz}' =
\epsilon_{yz}' \approx -0.1$\%, in contrast with $(001)$ samples,
for which a non-zero value of $\epsilon_{xy}'\approx = 0.05\%$ has
to be assumed in order to explain the experimental data. This
finding provides a hint that a preferential Mn incorporation
during the growth process accounts for the mysterious lowering of
the (Ga,Mn)As symmetry.
\section*{Acknowledgments}
The work was supported by EU FunDMS Advanced Grant of the European
Research Council within the "Ideas" 7th Framework Programme,
InTechFun (POIG.01.03.01-00-159/08), SemiSpinNet
(PITN-GA-2008-215368) and Polish MNiSW 2048/B/H03/2008/34 grant.
We thank M.~Kisielewski and M.~Maziewski for valuable discussions
on optical measurements.
|
\section{Introduction}
Quasinormal modes (QNMs) of black holes (BHs) have been studied for
nearly 40 years due to their importance in the context of
gravitational wave astronomy. In the last decade, there has been
moreover an increase of activity in BH QNM studies motivated by
potential applications in analog models of gravity, quantum gravity,
string theory and related topics (TeV-scale gravity, AdS/CFT
correspondence, alternative theories of gravity, BH area
quantization, phase transitions in BH systems,...). For excellent
reviews on the status of QNMs prior to 1999 and on their relevance
to gravitational wave astronomy, we refer to the articles by
Kokkotas and Schmidt \cite{KokkotasSchmidt99} and by Nollert
\cite{Nollert99}. For a more recent review on BH QNMs, we refer to
the article by Berti, Cardoso and Starinets
\cite{BertiCardosoStarinets2009}: it updates the two previously
cited articles and it also presents the aspects of QNM physics
linked to gauge-gravity duality; it includes, furthermore, an
interesting historical introduction on the subject as well as a
useful impressive bibliography on all the aspects of BH physics
linked to QNMs.
Immediately after the publication of one of the first papers on QNMs
by Press \cite{Press1971} where he identified the gravitational
ringing of the Schwarzschild BH as due to its ``free oscillations",
Goebel suggested a physically intuitive interpretation of the
associated QNMs \cite{Goebel}: they could be interpreted in terms of
gravitational waves in spiral orbits close to the unstable circular
photon/graviton orbit at $r=3M$ which decay by radiating away energy
(here $M$ denotes the mass of the BH). This appealing interpretation
has been developed by other authors for various field theories
defined on BH backgrounds using the eikonal approximation, i.e., in
a framework based on geodesics and bundle of geometrical rays (see
Refs.~\cite{FerrariMashhoon84,Mashhoon85,Stewart89,
AnderssonOnozawa96,ZerbiniVanzo2004,CardosoMirandaBertietal2009,DolanOttewill_2009,Hod2009}
as well as Ref.~\cite{BarretoZworski1997} for a more mathematical
approach). It has permitted them to obtain analytical approximations
for the leading-order terms of the characteristic complex
frequencies of various BH spectra from an interpretation in terms of
massless particles ``trapped" near unstable circular null geodesics
(see, more particularly, Ref.~\cite{CardosoMirandaBertietal2009}
where the relation with the Lyapunov exponent corresponding to
geodesic motion is clearly emphasized).
A potentially much richer implementation of the Goebel
interpretation of BH QNMs which is not limited to purely geometrical
considerations but based on wave/field theory and which goes beyond
the leading oder terms has also been formulated
\cite{DecaniniFJ_cam_bh,DecaniniFolacci2009,DecaniniFolacci2010a}
(see also Ref.~\cite{DolanOttewill_2009}). It uses complex angular
momentum (CAM) techniques (or, in other words, the Regge pole
machinery) which play a central role in scattering theory. Since, as
noted by Chandrasekhar and coworkers \cite{Chandrasekhar_1983} (see
also Ref.~\cite{Ferrari1992}), BH perturbation theory can be
formulated as a resonant scattering problem, CAM techniques arise
naturally in BH physics. For reviews of the CAM method, we refer to
the monographs of Newton \cite{New82}, Nussenzveig \cite{Nus92} and
Collins \cite{Collins77} as well as to references therein for
various applications in quantum mechanics, nuclear physics, high
energy physics, electromagnetism and seismology.
Some years ago, the CAM method was used in gravitational wave
physics by Chandrasekar and Ferrari \cite{Chandra} to express the
flow of energy due to nonradial oscillations of relativistic stars
and by Andersson and Thylwe to describe scattering from the
Schwarzschild BH \cite{Andersson1} as well as to interpret the
Schwarzschild BH glory \cite{Andersson2}. In this context, Andersson
established, for the Schwarzschild BH of mass $M$, the existence of
a family of ``surface waves" (each one associated with a Regge pole
of the $S$-matrix) orbiting close to the unstable photon orbit at
$r=3M$ (see also Ref.~\cite{DecaniniFJ_cam_bh} for a more rigorous
approach). Recently, from these ``surface waves", we have been able
to theoretically and numerically construct the spectrum of the
weakly damped complex frequencies of the Schwarzschild BH QNMs
\cite{DecaniniFJ_cam_bh,DecaniniFolacci2010a} and to interpret them
as Breit-Wigner resonances. This has been achieved by obtaining
analytically the nonlinear dispersion relation as well as the
damping of the ``surface waves" propagating close to the photon
sphere. Let us also note two related papers concerning analytical or
numerical determinations of the Regge poles of the Schwarzschild BH
\cite{GlampedakisAndersson2003,DolanOttewill_2009} and that, this
last year, Regge poles have also been used to understand the
resonant aspects of the BTZ BH \cite{DecaniniFolacci2009} as well as
to analyze some aspects of self-force calculations
\cite{Casals_et_al_2009}.
In the present paper, we extend the analysis developed for the
Schwarzschild BH to more general BHs and we establish, from Regge
pole considerations, a precise connection between the existence of a
photon sphere and the properties of the ``surface waves" propagating
close to it. More precisely, we consider a wide class of static
spherically symmetric BHs of arbitrary dimension with a photon
sphere, i.e., a hypersurface on which a massless particle can orbit
the BH on unstable circular null geodesics. For more rigorous
definitions of the photon sphere concept in static spherically
symmetric spacetime, we refer to the article by Claudel, Virbhadra
and Ellis \cite{ClaudelVirbhadraEllis2001}. This class of BHs
includes various spacetimes of physical interest such as
Schwarzschild, Schwarzschild-Tangherlini and Reissner-Nordstr\"om
BHs, the canonical acoustic BH or the Schwarzschild-de Sitter BH.
For this class of BHs, we provide general analytical expressions
beyond the leading-order terms for the Regge poles of the $S$-matrix
associated with a massless scalar field theory. These results permit
us to obtain analytically the nonlinear dispersion relation and the
damping of the ``surface waves" lying close to the photon sphere as
well as, from Bohr-Sommerfeld--type resonance conditions, the
complex frequencies corresponding to the weakly damped QNMs.
Our paper is organized as follows. In Sec.~II, we display our
general working assumptions and we justify them physically. We then
explain how to construct the $S$-matrix permitting us to analyze the
resonant aspects of a scalar field theory defined on an
asymptotically flat static spherically symmetric BH of arbitrary
dimension with a photon sphere and we finally define its Regge poles
as well as its complex quasinormal frequencies. In Sec.~III, we
provide a general analytical expression for the Regge poles. This is
achieved by using and extending the WKB approach developed in the
context of the determination of the QNMs by Schutz and Will
\cite{SchutzWill} and by Will and Iyer \cite{Iyer1,Iyer2} (see also
Ref.~\cite{BenderOrszag1978} for general aspects of WKB theory and
for particular aspects connected with eigenvalue problems). Our
result permits us to describe the Regge trajectories of a general
asymptotically flat static spherically symmetric BH of arbitrary
dimension with a photon sphere and to obtain, from semiclassical
formulas, analytical expressions for the QNM complex frequencies.
Our WKB analysis permits us moreover to show that (i) the dispersion
relation of the $n$th ``surface wave" is nonlinear and depends on
the index $n$ and that (ii) the damping of the $n$th ``surface wave"
is frequency dependent. In Sec.~IV, we apply the general theory
developed in Sec.~III to particular BHs (Schwarzschild,
Schwarzschild-Tangherlini, Reissner-Nordstr\"om and canonical
acoustic BHs). In a brief conclusion, we consider some consequences
of our work as well as possible extensions. In Appendix A, we
establish the semiclassical connection between the Regge poles of a
static spherically symmetric BH of arbitrary dimension with a photon
sphere and the complex frequencies of its weakly damped QNMs. In
Appendix B, we consider the particular case of the Schwarzschild-de
Sitter BH. Indeed, even if such a gravitational background is not
asymptotically flat, the formalism developed in Secs.~II and III
naturally applies to it.
In this paper, we shall use units with $\hbar=c=G=1$.
\section{Quasinormal frequencies and Regge poles of static
spherically symmetric black holes: General theory}
We consider a static spherically symmetric spacetime of arbitrary dimension $d \ge 4$
with metric
\begin{equation} \label{metric_BH}
ds^2=-f(r)dt^2+\frac{dr^2}{f(r)}+r^2d\sigma_{d-2}^{2}.
\end{equation}
Here $d\sigma_{d-2}^{2}$ denotes the line element on the unit sphere
$S^{d-2}$. On $S^{d-2}$, we introduce the usual angular coordinates
$\theta_i \in [0,\pi]$ with $i=1, \dots, d-3$ and $\varphi \in
[0,2\pi]$. We have
\begin{equation} \label{metric_Sphere}
d\sigma_{d-2}^{2}={d\theta_1}^2 + \sum_{k=2}^{d-3} \left(\prod_{i=1}^{k-1}
\sin^2\theta_i\right)d\theta_k^2 + \left(\prod_{i=1}^{d-3}
\sin^2\theta_i\right)d\varphi^2.
\end{equation}
Of course, a metric such as (\ref{metric_BH}) does not describe the
most general static spherically symmetric spacetime but it will
permit us to consider a wide class of BHs of physical interest.
In Eq.~(\ref{metric_BH}), we shall furthermore assume that $f(r)$
is a function of the usual radial coordinate
$r$ with the following properties:
\begin{itemize}
\item (i) There exists an interval $I=]r_h,+\infty[ \subset \bf{R}$
with $r_h>0$ such as $f(r)>0$ for $r \in I$.
\item (ii) $r_h$ is a simple root of $f(r)$, i.e.,
\begin{equation} \label{Assup_fr_1a}
f(r_h)=0 \qquad \text{and} \qquad f'(r_h)\neq 0,
\end{equation}
and $f(r)$ moreover satisfies
\begin{equation}
\underset{r \to +\infty}{\lim}f(r)=1.
\end{equation}
\item (iii) There exists a value $r_c \in I$ for which
\begin{equation} \label{Assup_fr_2a}
f'(r_c) - \frac{2}{r_c}f(r_c)=0
\end{equation}
and
\begin{equation} \label{Assup_fr_2b}
f''(r_c)-\frac{2}{r_c^2}f(r_c) <0.
\end{equation}
\end{itemize}
We shall now briefly discuss assumptions (i)-(iii) previously
introduced. Assumptions (i) and (ii) indicate that the spacetime
considered is an asymptotically flat BH with an event horizon at
$r_h$, its exterior corresponding to $r \in I$. Assumption (iii)
implies the existence of a photon sphere which is the support of
unstable circular null geodesics (see below for more details). It
should be noted that, as a consequence of (i) and (ii), the tortoise
coordinate $r_\ast=r_\ast(r)$ defined for $r \in I$ by the relation
$dr_\ast/dr=1 / f(r)$ and the condition $r_\ast(r_c)=0$ provides a
bijection $r_\ast=r_\ast(r)$ from $I$ to $]-\infty,+\infty[$.
Let us consider a free-falling massless particle orbiting the BH.
Without loss of generality, we can consider that its motion lies on
the equatorial hyperplane defined by $\theta_i=\pi/2$ for $i=1,
\dots, d-3$. Because it moves along a null geodesic, we have [cf.
Eqs.~(\ref{metric_BH}) and (\ref{metric_Sphere})]
\begin{equation}\label{nullgeod}
-f(r)\left(\frac{dt}{d\alpha}\right)^2
+\frac{1}{f(r)}\left(\frac{dr}{d\alpha}\right)^2+r^2\left(\frac{d\varphi}{d\alpha}\right)^2=0
\end{equation}
where $\alpha$ is an affine parameter and, of course, there exist
two integrals of motion respectively associated with the Killing
vectors $\partial / \partial t$ and $\partial / \partial \varphi$
and given by
\begin{subequations}
\begin{eqnarray}
&&f(r)\left(\frac{dt}{d\alpha}\right)=E,\label{motioncst1}\\
&&r^2\left(\frac{d\varphi}{d\alpha}\right)=L.\label{motioncst2}
\end{eqnarray}
\end{subequations}
Here $E$ and $L$ denote respectively the energy and the angular momentum of the massless particle.
Inserting Eqs.~(\ref{motioncst1}) and (\ref{motioncst2}) into (\ref{nullgeod}), we obtain
\begin{equation}\label{eqmotion}
\left(\frac{dr}{d\alpha}\right)^2+V_{\mathrm{eff}}(r)=E^2
\end{equation}
where
\begin{equation}\label{Veff}
V_{\mathrm{eff}}(r)=\frac{L^2}{r^2}f(r).
\end{equation}
From these last two equations and from assumption (iii), one can
easily remark that the massless particle can orbit the BH on an
unstable circular geodesic defined by $r=r_c$. Indeed, we have in
particular
\begin{subequations}
\begin{equation}\label{Veff_d1}
\left. \frac{d}{dr}V_{\mathrm{eff}}(r)\right|_{r=r_c}=0
\end{equation}
and
\begin{equation}\label{Veff_d2}
\left.
\frac{d^2}{dr^2}V_{\mathrm{eff}}(r)\right|_{r=r_c}=\frac{L^2}{r_c^2}\left(f''(r_c)-\frac{2}{r_c^2}f(r_c)
\right) <0.
\end{equation}
\end{subequations}
On this orbit, the massless particle takes the time
\begin{equation}\label{Torbit}
T=\frac{2\pi r_c}{\sqrt{f(r_c)}}
\end{equation}
to circle the BH. This result can be obtained by integrating Eq.~(\ref{nullgeod}).
The wave equation for a massless scalar field propagating on a
general gravitational background is given by
\begin{equation}\label{WaveEq}
\Box \Phi=g^{\mu\nu}\nabla_{\mu}\nabla_{\nu}\Phi=
\frac{1}{\sqrt{-g}}\partial_{\mu}\left(\sqrt{-g}g^{\mu\nu}\partial_{\mu}\Phi\right)=0.
\end{equation}
If the spacetime metric is given by (\ref{metric_BH}), after
separation of variables and the introduction of the radial partial
wave functions $\Phi_\ell(r)$ with $\ell=0,1,2, \dots$, this wave
equation reduces to the Regge-Wheeler equation
\begin{equation}\label{RW}
\frac{d^2 \Phi_\ell}{dr_*^2} + \left[ \omega^2 - V_\ell(r)\right]
\Phi_\ell=0.
\end{equation}
[Here we have assumed a harmonic time dependence $\exp(-i\omega t)$
for the massless scalar field.] In Eq.~(\ref{RW}), $V_\ell(r)$ is
the Regge-Wheeler potential given by
\begin{eqnarray}\label{EffectivePot}
V_\ell(r)=f(r) \left[ \frac{\ell(\ell+d-3)}{r^2} \right. &+& \left.
\frac{(d-2)(d-4)}{4r^2}f(r)\right. \nonumber \\
&+& \left. \left(\frac{d-2}{2r}\right)f'(r) \right].
\end{eqnarray}
It should be noted that
\begin{itemize}
\item $\lim_{r \to r_h} V_\ell(r)=0$ and $\lim_{r \to +\infty} V_\ell(r)=0$
and therefore the solutions of the radial equation (\ref{RW}) have a behavior
in $\exp(\pm i \omega r_{\ast})$ at the horizon and at infinity.
\item For $\ell \gg 1$, $V_{\ell}(r)$ has a local maximum at $r=r_c$ because,
in this limit, $V_{\ell}(r)$ and $V_{\mathrm{eff}}(r)$ are similar.
\item For any finite value of $\ell$, the local maximum of $V_{\ell}(r)$ is close to $r=r_c$.
\end{itemize}
For a given angular momentum index $\ell$, the $S$-matrix element
$S_\ell (\omega)$ is defined by seeking the solution of the
Regge-Wheeler equation (\ref{RW}) which has a purely ingoing
behavior at the event horizon $r=r_h$, i.e., which satisfies
\begin{equation}\label{bc1}
\Phi_\ell (r) \underset{r_* \to -\infty}{\sim}
T_\ell(\omega)e^{-i\omega r_* }
\end{equation}
and which, at spatial infinity $r \to +\infty$, presents an
asymptotic behavior of the form
\begin{eqnarray}\label{bc2}
&& \Phi_\ell(r) \underset{r_* \to +\infty}{\sim}
e^{-i\omega r_* + i\left(\ell+\frac{d-3}{2}\right)\frac{\pi}{2}-i\frac{\pi}{4}} \nonumber \\
&& \qquad \qquad\qquad -S_\ell(\omega)e^{+i\omega r_* -
i\left(\ell+\frac{d-3}{2}\right)\frac{\pi}{2}+i\frac{\pi}{4}}.
\end{eqnarray}
We recall that the $S$-matrix permits us to analyze the resonant
aspects of the considered BH as well as to construct the
form factor describing the scattering of a monochromatic scalar wave (see Appendix A).
To describe semiclassically resonance phenomena, the dual structure
of the $S$-matrix plays a crucial role. Indeed, the $S$-matrix is a
function of both the frequency $\omega$ and the angular momentum
index $\ell$. It can be analytically extended into the complex
$\omega$-plane as well as into the complex $\lambda$-plane (CAM
plane) with $\lambda=\ell+(d-3)/2$. From now on, we shall denote by
$S_{\lambda -(d-3)/2}(\omega)$ this double analytical extension. For
$\ell=\lambda -(d-3)/2 \in \bf{N}$, the simple poles lying in the
fourth quadrant of the complex $\omega$-plane [let us recall that
they are also simple poles of $T_{\lambda -(d-3)/2}(\omega)$] are
the complex frequencies of the QNMs. These modes are therefore
solutions of the radial wave equation (\ref{RW}) which are purely
outgoing at infinity and purely ingoing at the horizon. We shall
denote by $\omega_{\ell n}=\omega^{(o)}_{\ell n}- i\Gamma_{\ell
n}/2$ where $n\in {\bf N}^\ast$ the quasinormal frequencies. We
recall that $\omega^{(o)}_{\ell n}>0$ and $\Gamma_{\ell n}>0$
represent respectively the frequency of the oscillation and the
damping corresponding to the associated QNM. We assume that, in the
immediate neighborhood of $\omega_{\ell n}$, $S_\ell(\omega)$ has
the Breit-Wigner form, i.e.,
\begin{equation}\label{BW}
S_\ell(\omega ) \propto \frac{\Gamma _{\ell n}/2}{\omega
-\omega^{(o)}_{\ell n}+i\Gamma _{\ell n}/2}.
\end{equation}
For a given value $\omega >0$ of the frequency, the simple poles lying
in the first quadrant of the complex $\lambda$-plane are the so-called Regge poles.
It should be noted that they are also poles of $T_{\lambda -(d-3)/2}(\omega)$ and therefore
the associated modes (Regge modes) are purely outgoing at infinity and purely ingoing at
the horizon. We shall denote the Regge poles by $\lambda_n(\omega)$, the index $n=1,2,\dots$
permitting us to distinguish each pole.
The structure of the $S$-matrix in the complex $\lambda$-plane
allows us, by using integration contour deformations, Cauchy's
theorem and asymptotic analysis, to provide a semiclassical
description of scattering (see Appendix A for more precisions). The
curves traced out in the CAM plane by the Regge poles as a function
of the frequency $\omega$ are the so-called Regge trajectories. They
permit us to interpret Regge poles in terms of ``surface waves" (see
Appendix A): $\mathrm{Re} \, \lambda_n (\omega)$ provides the
dispersion relation for the $n$th ``surface wave" while $\mathrm{Im}
\, \lambda_n (\omega)$ corresponds to its damping. Furthermore, from
the Regge trajectories, we can semiclassically construct the
resonance spectrum [see, in Appendix A, formulas
(\ref{sc1_Appendix}), (\ref{sc2_Appendix}) and
(\ref{sc4_Appendix})]. The semiclassical formula (a
Bohr-Sommerfeld--type quantization condition)
\begin{equation}\label{sc1}
\mathrm{Re} \, \lambda_n \left(\omega^{(0)}_{\ell n} \right)= \ell
+ \frac{d-3}{2}, \qquad \ell \in {\bf N}
\end{equation}
provides the location of the excitation frequencies $\omega^{(0)}_{\ell n}$ of
the resonances generated by $n$th
``surface wave", while a second semiclassical formula gives the widths of these resonances
\begin{equation}\label{sc2} \frac{\Gamma _{\ell n}}{2}= \left. \frac{\mathrm{Im} \,
\lambda_n (\omega )[d/d\omega \, \mathrm{Re}\, \lambda_n (\omega )
]}{[d/d\omega \, \mathrm{Re} \, \lambda_n (\omega ) ]^2 + [d/d\omega
\, \mathrm{Im} \, \lambda_n (\omega ) ]^2 } \right|_{\omega
=\omega^{(0)}_{\ell n}}.
\end{equation}
It should be moreover noted that this formula reduces, in the
frequency range where the condition $| d/d\omega \,\mathrm{Re} \,
\lambda_n (\omega ) | \gg |d/d\omega \, \mathrm{Im}\, \lambda_n
(\omega ) |$ is satisfied, to
\begin{equation}\label{sc4}
\frac{\Gamma _{\ell n}}{2}= \left. \frac{\mathrm{Im} \, \lambda_n
(\omega )}{d/d\omega \, \mathrm{Re} \, \lambda_n (\omega ) }
\right|_{\omega =\omega^{(0)}_{\ell n}}.
\end{equation}
\section{WKB approximations for the Regge poles and
semiclassical expressions of the complex quasinormal frequencies}
In general, it is not possible to solve exactly the Regge-Wheeler
equation (\ref{RW}) and therefore we can obtain only analytical
approximations for the Regge poles and for the complex quasinormal
frequencies. For example, the WKB approach developed in the general
context of eigenvalue problems (for more details see
Ref.~\cite{BenderOrszag1978}) has been adapted for the determination
of the Schwarzschild BH QNMs by Schutz and Will \cite{SchutzWill}
and by Will and Iyer \cite{Iyer1,Iyer2} and, in
Ref.~\cite{DecaniniFolacci2010a}, for the determination of the
Schwarzschild BH Regge poles. It can be extended to the more general
case considered in this paper. By using third-order WKB
approximations \cite{Iyer1,Iyer2} for the Regge modes of
Eq.~(\ref{RW}), we find that the Regge poles $\lambda$ are the
complex solutions of the equation
\begin{eqnarray}\label{WKB_RP_1}
& &\omega^2=\left[V_0(\lambda) + {[-2V_0^{(2)}(\lambda)]}^{1/2} \,
{\overline \Lambda}(\lambda,n) \right] \nonumber \\ && \qquad - i \,
\alpha (n) \, {[-2V_0^{(2)}(\lambda)]}^{1/2}[1+ {\overline
\Omega}(\lambda ,n)]
\end{eqnarray}
with $\omega >0$ and $n=1,2,3, \dots $. Here
\begin{subequations}\label{WKB_RP_2}
\begin{eqnarray}\label{WKB_RP_2a}
& & {\overline
\Lambda}(\lambda,n)=\frac{1}{{[-2V_0^{(2)}(\lambda)]}^{1/2}}
\left[\frac{1}{8}
\frac{V_0^{(4)}(\lambda)}{V_0^{(2)}(\lambda)}\left( \frac{1}{4}+
\alpha(n)^2 \right) \right. \nonumber \\
& & \qquad \left. -\frac{1}{288}
\left(\frac{V_0^{(3)}(\lambda)}{V_0^{(2)}(\lambda)}\right)^2\left(
7+ 60 \, \alpha(n)^2 \right) \right]
\end{eqnarray}
and
\begin{eqnarray}\label{WKB_RP_2b}
& &{\overline \Omega}(\lambda,n)=\frac{1}{[-2V_0^{(2)}(\lambda)]} \times \nonumber \\
&& \qquad \left[\frac{5}{6912}
\left(\frac{V_0^{(3)}(\lambda)}{V_0^{(2)}(\lambda)}\right)^4\left(
77+ 188 \, \alpha(n)^2 \right) \right. \nonumber \\
&& \qquad \left. -\frac{1}{384}
\left(\frac{{[V_0^{(3)}(\lambda)]}^2V_0^{(4)}(\lambda)}{{[V_0^{(2)}(\lambda)]}^3}\right)
\left(
51+ 100 \, \alpha(n)^2 \right)\right. \nonumber \\
&& \qquad \left. + \frac{1}{2304}
\left(\frac{V_0^{(4)}(\lambda)}{V_0^{(2)}(\lambda)}\right)^2 \left(
67+ 68 \, \alpha(n)^2 \right) \right. \nonumber \\
&& \qquad \left. +\frac{1}{288} \left(\frac{ V_0^{(3)}(\lambda)
V_0^{(5)}(\lambda)}{{[V_0^{(2)}(\lambda)]}^2}\right) \left( 19+ 28
\, \alpha(n)^2 \right) \right. \nonumber \\
&& \qquad \left. -\frac{1}{288}
\left(\frac{V_0^{(6)}(\lambda)}{V_0^{(2)}(\lambda)}\right)\left( 5 +
4 \, \alpha(n)^2 \right) \right].
\end{eqnarray}
\end{subequations}
In Eqs.~(\ref{WKB_RP_1}) and (\ref{WKB_RP_2}), we have
introduced the notations
\begin{equation}\label{WKB_RP_3a}
\alpha(n)=n-1/2
\end{equation}
and, for $p \in {\bf N}$,
\begin{equation}\label{WKB_RP_3b}
V_0^{(p)} (\lambda) = \left. \frac{d^p}{{dr_*}^p }
V_{\lambda-(d-3)/2}(r_*)\right|_{r_*={(r_*)}_0}
\end{equation}
with ${(r_*)}_0$ which denotes the maximum of the function
$V_{\lambda-(d-3)/2}(r_*)$.
In order to solve Eq.~(\ref{WKB_RP_1}), we need to express the asymptotic
expansions for $|\lambda| \to +\infty$ of $V_0(\lambda)$ and
$\left[-2V_{0}^{(2)}(\lambda)\right]^{1/2}$ and of the various ratios
appearing in Eqs.~(\ref{WKB_RP_2a}) and (\ref{WKB_RP_2b}). In order to simplify
the results we have introduced the notations
\begin{equation}
f_c^{(p)}=f^{(p)}(r_c)
\end{equation}
and
\begin{equation}\label{Eta}
\eta_c=\frac{1}{2}\sqrt{4f_{c}-2r_{c}^{2}f_{c}^{(2)}}.
\end{equation}
It is worth noting that the $\eta_c$ parameter is directly linked to
the second derivative (\ref{Veff_d2}) of the effective potential
(\ref{Veff}) taken at $r=r_c$. As a consequence, it represents a
kind of measure of the instability of the circular orbits lying on
the photon sphere. In fact, it can be expressed in terms of the
Lyapunov exponent $\Lambda_c$ corresponding to these orbits
introduced in Ref.~\cite{CardosoMirandaBertietal2009} and which is
the inverse of the instability time scale associated with them: we
have
\begin{equation}\label{Eta_c_Lyapunov}
\eta_c= \frac{r_c}{\sqrt{f_c}}|\Lambda_c|.
\end{equation}
We will say no more about this connection because, as already
mentioned in Sec.~I, we intend to go beyond purely geometrical
considerations in our analysis of the resonant behavior of BHs.
After a tedious calculation, we obtain
\begin{widetext}
\begin{subequations}
\begin{equation}\label{DevEffPot}
V_{0}(\lambda)=\lambda^{2}\frac{f_{c}}{r_{c}^{2}}+\frac{f_{c}}{4r_{c}^{2}}
\left[d(d-2)f_{c}-(d-3)^{2}\right]+\underset{|\lambda| \to
+\infty}{\cal O}\left( \frac{1}{\lambda^2}\right)
\end{equation}
and
\begin{eqnarray}\label{DevDerSecPot}
\left[-2V_{0}^{(2)}(\lambda)\right]^{1/2}&=& \frac{2\eta_c
f_{c}}{r_{c}^{2}} \lambda -\frac{f_{c}}{8 \eta_c^{3}r_{c}^{2}}\left[
2f_c^2[(d-3)^{2}+(d-2)(d-4)f_c] \phantom{\left(f_{c}^{(2)}\right)^{2}}\right. \nonumber
\\ && \left. +r_{c}^{2}f_{c}f_{c}^{(2)}
[(d-2)(d+8)f_c-2(d-3)^2]+d(d-2)r_{c}^{3}f_{c}^{2}f_{c}^{(3)} \phantom{\left(f_{c}^{(2)}\right)^{2}}\right. \nonumber
\\ && \left. +r_{c}^{4}
\left(f_{c}^{(2)}\right)^{2}[(1/2)(d-3)^2-(d^2-4)f_c]\right]\frac{1}{\lambda}
+\underset{|\lambda| \to +\infty}{\cal O}\left(
\frac{1}{\lambda^3}\right)
\end{eqnarray}
\end{subequations}
as well as
\begin{subequations}
\begin{eqnarray}\label{RapportWKB_1}
\frac{V_{0}^{(4)}(\lambda)}{V_{0}^{(2)}(\lambda)}&=&-\frac{f_{c}}{2\eta_c^{2}r_{c}^{2}}
\left[16f_{c}^{2}-16r_{c}^{2}f_c
f_{c}^{(2)}+4r_{c}^{3}f_{c}f_{c}^{(3)}
\phantom{\left(f_{c}^{(2)}\right)^{2}}\right. \nonumber
\\ && \left.+r_{c}^{4}\left(4\left(f_{c}^{(2)}\right)^{2}+f_{c}f_{c}^{(4)}\right)\right]+\underset{|\lambda| \to +\infty}{\cal O}\left(
\frac{1}{\lambda^2}\right),
\end{eqnarray}
\begin{eqnarray}\label{RapportWKB_2}
\left(\frac{V_{0}^{(3)}(\lambda)}{V_{0}^{(2)}(\lambda)}\right)^{2}=
\frac{r_{c}^{4}f_{c}^{2}\left(f_{c}^{(3)}\right)^{2}}{4\eta_c^{4}}
+\underset{|\lambda| \to +\infty}{\cal O}\left( \frac{1}{\lambda^2}\right),
\end{eqnarray}
\begin{eqnarray}\label{RapportWKB_4}
\frac{{[V_{0}^{(3)}(\lambda)]}^{2}V_{0}^{(4)}(\lambda)}{{\left[V_{0}^{(2)}(\lambda)\right]}^{3}}
&=&-\frac{r_{c}^{2}f_{c}^{3}\left(f_{c}^{(3)}\right)^{2}}{8\eta_c^{6}}\left[16f_{c}^{2}
-16r_{c}^{2}f_cf_{c}^{(2)}+4r_{c}^{3}f_{c}f_{c}^{(3)}
\phantom{\left(f_{c}^{(2)}\right)^{2}}\right. \nonumber
\\ && \left.+r_{c}^{4}\left(4\left(f_{c}^{(2)}\right)^{2}
+f_{c}f_{c}^{(4)}\right)\right]+\underset{|\lambda| \to
+\infty}{\cal O}\left( \frac{1}{\lambda^2}\right),
\end{eqnarray}
\begin{eqnarray}\label{RapportWKB_6}
\frac{V_{0}^{(3)}(\lambda)V_{0}^{(5)}(\lambda)}{{\left[V_{0}^{(2)}(\lambda)\right]}^{2}}
&=&\frac{r_{c}^{2}f_{c}^{3}f_{c}^{(3)}}{4\eta_c^{4}}\left[-10f_{c}f_{c}^{(3)}+10r_{c}f_{c}f_{c}^{(4)}\right.
\nonumber
\\ && \left.+r_{c}^{2}\left(15f_{c}^{(2)}f_{c}^{(3)}+f_{c}f_{c}^{(5)}\right)\right]
+\underset{|\lambda| \to +\infty}{\cal O}\left( \frac{1}{\lambda^2}\right),
\end{eqnarray}
\begin{eqnarray}\label{RapportWKB_7}
& & \frac{V_{0}^{(6)}(\lambda)}{V_{0}^{(2)}(\lambda)}=
-\frac{f_{c}^{2}}{2\eta_c^{2}r_{c}^{4}}
\left[-272f_{c}^{3}+408r_{c}^{2}f_{c}^{2}f_{c}^{(2)}
-88r_{c}^{3}f_{c}^{2}f_{c}^{(3)}\phantom{\left(f_{c}^{(2)}\right)^{2}}\right.
\nonumber
\\ && \qquad \left.+r_{c}^{4}f_{c}\left(38f_{c}f_{c}^{(4)}
-204\left(f_{c}^{(2)}\right)^{2}\right)+r_{c}^{5}f_{c}\left(104f_{c}^{(2)}f_{c}^{(3)}
+18f_{c}f_{c}^{(5)}\right)
\phantom{\left(f_{c}^{(2)}\right)^{2}}\right. \nonumber
\\ && \qquad \left.+r_{c}^{6}\left(34\left(f_{c}^{(2)}\right)^{3}+15f_{c}\left(f_{c}^{(3)}\right)^{2}
+26f_{c}f_{c}^{(2)}f_{c}^{(4)}+f_{c}^{2}f_{c}^{(6)}\right)\right]
+\underset{|\lambda| \to +\infty}{\cal O}\left(
\frac{1}{\lambda^2}\right).
\end{eqnarray}
\end{subequations}
\end{widetext}
We can now solve Eq.~(\ref{WKB_RP_1}) by
assuming $|\lambda| \gg 1$ as well as $\mathrm{Re} \, \lambda \gg
\mathrm{Im} \, \lambda$. We obtain for the solutions a family
$\lambda_n(\omega)$ with $n=1,2,3, \dots $ given by the
approximation
\begin{widetext}
\begin{equation}\label{PRapproxWKB}
\lambda_{n}(\omega) \approx \left[\frac{r_{c}^{2}}{f_{c}} \,
\omega^{2}
+a_{n}+2\eta_c^2\alpha(n)^{2}\epsilon_{n}(\omega)\right]^{1/2}+i\eta_c\alpha(n)\left[1+\epsilon_{n}(\omega)\right]
\end{equation}
where
\begin{eqnarray}\label{an}
&& a_{n}=-~\frac{1}{1152\eta_c^{4}} \left\{
288f_{c}^{2}\left[(d^2-2d-1)f_{c}-(d-3)^{2}\right]\phantom{\left(f_{c}^{(2)}\right)^{3}}
\right. \nonumber \\ && \left. \quad +144r_{c}^{2}f_{c}f_{c}^{(2)}\left[2(d-3)^{2}-(2d^2-4d-3)f_{c}\right]
\phantom{\left(f_{c}^{(2)}\right)^{3}}\right.\nonumber \\
&&\left.\quad- 72r_{c}^{3}f_{c}^{2}f_{c}^{(3)} -18
r_{c}^{4}\left[4(d-3)^{2} \left(f_{c}^{(2)}\right)^{2} -4(d-3)(d+1)
f_{c}\left(f_{c}^{(2)}\right)^{2} + f_{c}^{2}f_{c}^{(4)}\right]\right.\nonumber \\
&&\left.\quad +
36r_{c}^{5}f_{c}f_{c}^{(2)}f_{c}^{(3)} + \,
r_{c}^{6}\left[36\left(f_{c}^{(2)}\right)^{3}-7f_{c}\left(f_{c}^{(3)}\right)^{2}
+9f_{c}f_{c}^{(2)}f_{c}^{(4)}\right] \right\} \nonumber \\
&& \quad +~\alpha(n)^{2}\frac{r_c^3f_c}{96\eta_c^{4}} \left\{24f_cf_c^{(3)}+6r_cf_cf_c^{(4)}-12r_c^2f_c^{(2)}f_c^{(3)}+r_c^3\left(5\left(f_c^{(3)}\right)^2-3f_c^{(2)}f_c^{(4)}\right)\right\} \nonumber \\
&&
\end{eqnarray}
and
\begin{equation}\label{Epsilon}
\epsilon_{n}(\omega)=\frac{b_{n}}{(r_{c}^{2}/f_{c}) \, \omega^{2}+a_{n}+\eta_c^2\alpha(n)^{2}}
\end{equation}
with
\begin{eqnarray}\label{bn}
&& b_{n}=\frac{1}{442368\eta_c^{10}} \left\{
-55296(d-3)^{2}f_{c}^{5}(1+f_{c})
+27648r_{c}^{2}f_{c}^{4}f_{c}^{(2)}\left[5(d-3)^{2}+2(d^2-12d+23)f_{c}\right]
\right. \nonumber \\ && \left. \quad - 9216 (3d^2-6d -2)
r_{c}^{3}f_{c}^{5}f_{c}^{(3)}
+6912r_{c}^{4}f_{c}^{3}\left[-20(d-3)^{2}\left(f_{c}^{(2)}\right)^{2}
+2(2d^2+36d-95)f_{c}\left(f_{c}^{(2)}\right)^{2}\right. \right. \nonumber \\ &&
\left. \left. \quad +9f_{c}^{2}f_{c}^{(4)} \right]
+3456r_{c}^{5}f_{c}^{4}\left[6(2d^2-4d+1)f_{c}^{(2)}f_{c}^{(3)}+5f_{c}f_{c}^{(5)}\right]
\right. \nonumber \\ && \left. \quad
-192r_{c}^{6}f_{c}^{2}\left[-360(d-3)^2\left(f_{c}^{(2)}\right)^{3}+144(2d^2+6d-25)
f_{c}\left(f_{c}^{(2)}\right)^{3}
+423f_{c}^{2}f_{c}^{(2)}f_{c}^{(4)}
-200f_{c}^{2}\left(f_{c}^{(3)}\right)^{2}\right. \right. \nonumber \\ &&\left. \left.\quad -5f_{c}^{3}f_{c}^{(6)}
\right]
-96r_{c}^{7}f_{c}^{3}\left[36(6d^2-12d+17)\left(f_{c}^{(2)}\right)^{2}f_{c}^{(3)}
-257f_{c}f_{c}^{(3)}f_{c}^{(4)}+270f_{c}f_{c}^{(2)}f_{c}^{(5)}\right]
\right. \nonumber \\ && \left. \quad
+12r_{c}^{8}f_{c}\left[-1440(d-3)^2\left(f_{c}^{(2)}\right)^{4}
+288(7d^2+6d-55)f_{c}\left(f_{c}^{(2)}\right)^{4}
+2376f_{c}^{2}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(4)}\right. \right.
\nonumber \\ && \left. \left. \qquad \qquad \quad +
67f_{c}^{3}\left(f_{c}^{(4)}\right)^{2}+152f_{c}^{3}
f_{c}^{(3)}f_{c}^{(5)}
-2744f_{c}^{2}f_{c}^{(2)}\left(f_{c}^{(3)}\right)^{2}-120 f_{c}^{3}f_{c}^{(2)}f_{c}^{(6)} \right]
\right. \nonumber \\ && \left. \quad
+48r_{c}^{9}f_{c}^{2}\left[12(6d^2-12d+59)\left(f_{c}^{(2)}\right)^{3}f_{c}^{(3)}
+153f_{c}\left(f_{c}^{(3)}\right)^{3}-514f_{c}f_{c}^{(2)}f_{c}^{(3)}f_{c}^{(4)}
+270f_{c}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(5)}\right] \right.
\nonumber \\ && \left. \quad + 12 r_{c}^{10}\left[
144(d-3)^{2}\left(f_{c}^{(2)}\right)^{5}-288(d^2-7)
f_{c}\left(f_{c}^{(2)}\right)^{5} -
67f_{c}^{3}f_{c}^{(2)}\left(f_{c}^{(4)}\right)^{2}-152f_{c}^{3}f_{c}^{(2)}f_{c}^{(3)}f_{c}^{(5)}
\right. \right. \nonumber \\ && \left. \left. \qquad \qquad +
344f_{c}^{2}\left(f_{c}^{(2)}\right)^{2}\left(f_{c}^{(3)}\right)^{2}
+60f_{c}^{3}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(6)}
+108f_{c}^{2}\left(f_{c}^{(2)}\right)^{3}f_{c}^{(4)}
+153f_{c}^{3}\left(f_{c}^{(3)}\right)^{2}f_{c}^{(4)}\right] \right.
\nonumber \\ && \left. \quad -24
r_{c}^{11}f_{c}f_{c}^{(2)}\left[252\left(f_{c}^{(2)}\right)^{3}f_{c}^{(3)}
+153f_{c}\left(f_{c}^{(3)}\right)^{3}-257f_{c}f_{c}^{(2)}f_{c}^{(3)}f_{c}^{(4)}
+90f_{c}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(5)}\right] \right.
\nonumber \\ && \left. \quad
+r_{c}^{12}\left[-864\left(f_{c}^{(2)}\right)^{6}+385f_{c}^{2}\left(f_{c}^{(3)}\right)^{4}
-1512f_{c}\left(f_{c}^{(2)}\right)^{4}f_{c}^{(4)}
-918f_{c}^{2}f_{c}^{(2)}\left(f_{c}^{(3)}\right)^{2}f_{c}^{(4)}
\right. \right. \nonumber \\ && \left. \left. \qquad \quad +
201f_{c}^{2}\left(f_{c}^{(2)}\right)^{2}\left(f_{c}^{(4)}\right)^{2}
+456f_{c}^{2}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(3)}f_{c}^{(5)} +
1368f_{c}\left(f_{c}^{(2)}\right)^{3}\left(f_{c}^{(3)}\right)^{2}
-120f_{c}^2 \left(f_{c}^{(2)}\right)^{3}f_{c}^{(6)} \right] \right\}
\nonumber \\ &&
\quad +\frac{\alpha(n)^{2}r_{c}^{3}f_{c}}{110592
\eta_c^{10}}\left\{
9216f_{c}^{4}f_{c}^{(3)}+13824r_{c}f_{c}^{4}f_{c}^{(4)}
+3456r_{c}^{2}f_{c}^{3}\left[-2f_{c}^{(2)}f_{c}^{(3)}+f_{c}f_{c}^{(5)}\right]
\phantom{\left(f_{c}^{(3)}\right)^{2}} \right. \nonumber \\ &&
\left. \quad
+192r_{c}^{3}f_{c}^{3}\left[72\left(f_{c}^{(3)}\right)^{2}
-99f_{c}^{(2)}f_{c}^{(4)}+f_{c}f_{c}^{(6)}\right]
-288r_{c}^{4}f_{c}^{2}\left[12\left(f_{c}^{(2)}\right)^{2}f_{c}^{(3)}
-29f_{c}f_{c}^{(3)}f_{c}^{(4)}+18f_{c}f_{c}^{(2)}f_{c}^{(5)}\right]
\right. \nonumber \\ && \left. \quad
+12r_{c}^{5}f_{c}^{2}\left[648\left(f_{c}^{(2)}\right)^{2}f_{c}^{(4)}
+ 17f_{c}\left(f_{c}^{(4)}\right)^{2}+56f_{c}f_{c}^{(3)}f_{c}^{(5)}
-
1032f_{c}^{(2)}\left(f_{c}^{(3)}\right)^{2}-24f_{c}f_{c}^{(2)}f_{c}^{(6)}
\right] \right. \nonumber \\ && \left. \quad
+144r_{c}^{6}f_{c}\left[28\left(f_{c}^{(2)}\right)^{3}f_{c}^{(3)}
+25f_{c}\left(f_{c}^{(3)}\right)^{3}-58f_{c}f_{c}^{(2)}f_{c}^{(3)}f_{c}^{(4)}
+18f_{c}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(5)}\right] \right.
\nonumber \\ && \left. \quad
+12r_{c}^{7}f_{c}\left[-36\left(f_{c}^{(2)}\right)^{3}f_{c}^{(4)}
+75f_{c}\left(f_{c}^{(3)}\right)^{2}f_{c}^{(4)}-
17f_{c}f_{c}^{(2)}\left(f_{c}^{(4)}\right)^{2}
-56f_{c}f_{c}^{(2)}f_{c}^{(3)}f_{c}^{(5)} \right. \right. \nonumber \\
&& \left. \left. \qquad \qquad \quad +
168\left(f_{c}^{(2)}\right)^{2}\left(f_{c}^{(3)}\right)^{2} +12
f_{c}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(6)} \right]\right.
\nonumber
\\ && \left. \quad
-72r_{c}^{8}f_{c}^{(2)}\left[12\left(f_{c}^{(2)}\right)^{3}f_{c}^{(3)}
+25f_{c}\left(f_{c}^{(3)}\right)^{3}-29f_{c}f_{c}^{(2)}f_{c}^{(3)}f_{c}^{(4)}
+6f_{c}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(5)}\right]\right.
\nonumber \\ && \left. \quad
+r_{c}^{9}\left[235f_{c}\left(f_{c}^{(3)}\right)^{4}-216\left(f_{c}^{(2)}\right)^{4}f_{c}^{(4)}
-450f_{c}f_{c}^{(2)}\left(f_{c}^{(3)}\right)^{2}f_{c}^{(4)} +
51f_{c}\left(f_{c}^{(2)}\right)^{2}\left(f_{c}^{(4)}\right)^{2}
\right. \right. \nonumber \\ && \left. \left. \qquad
+168f_{c}\left(f_{c}^{(2)}\right)^{2}f_{c}^{(3)}f_{c}^{(5)} +
360\left(f_{c}^{(2)}\right)^{3}\left(f_{c}^{(3)}\right)^{2}
-24f_{c}\left(f_{c}^{(2)}\right)^{3}f_{c}^{(6)} \right] \right\}.
\end{eqnarray}
\end{widetext}
Equation (\ref{PRapproxWKB}) is the main result of our paper. As we
shall see below (see also Appendix A), it also provides expressions
for the dispersion relation and the damping of the ``surface waves"
lying on (close to) the photon sphere of the considered BH.
It is moreover possible to simplify (\ref{PRapproxWKB}) and to
obtain a high frequency approximation for the Regge poles. We have
\begin{widetext}
\begin{equation}\label{RPHF}
\lambda_{n}(\omega)=\left[\frac{r_{c}}{\sqrt{f_{c}}}~\omega
+\frac{a_{n}}{(2r_{c}/\sqrt{f_{c}})}~\frac{1}{\omega}\right]
+i\eta_c\alpha(n)\left[1+\frac{b_n}{(r_{c}^{2}/f_{c})}~\frac{1}{\omega^{2}}\right]
+\underset{\omega \to +\infty}{\cal O}\left(
\frac{1}{\omega^3}\right).
\end{equation}
\end{widetext}
It is important to understand the difference between the
approximation (\ref{PRapproxWKB}) and its much more elegant version
(\ref{RPHF}). The approximation (\ref{RPHF}) is meaningful as a
$1/\omega$ expansion with $\omega \to +\infty$. By contrast, the
approximation (\ref{PRapproxWKB}) remains valid in a large range of
frequencies thanks to WKB theory. Indeed, in order to establish
Eq.~(\ref{WKB_RP_1}) we have considered as a perturbation parameter
of the WKB method the distance between the turning points of
Eq.~(\ref{RW}) [i.e., the roots of $Q_\ell(r^\ast)= \omega^2 -
V_{\ell}(r^\ast)$] and the location of the peak of $Q_\ell(r^\ast)$
(see also Refs.~\cite{BenderOrszag1978,SchutzWill,Iyer1,Iyer2})
instead of $1/\omega$. Of course, in order to solve
Eq.~(\ref{WKB_RP_1}) and to obtain the expression
(\ref{PRapproxWKB}), we have furthermore assumed that $|\lambda| \gg
1$ as well as $\mathrm{Re} \, \lambda \gg \mathrm{Im} \, \lambda$
and, as a consequence, we cannot expect from (\ref{PRapproxWKB}) a
very high accuracy for very low frequencies (see also
Refs.~\cite{DolanOttewill_2009} and \cite{DecaniniFolacci2010a} for
related numerical studies in the case of the Schwarzschild BH).
It is furthermore interesting to provide an interpretation of the
previous results in terms of ``surface waves" as it is customary in
the CAM approach. As we have noted in Appendix A, the resonant part
of the form factor is a superposition of terms like
$\exp[i(\lambda_n(\omega)\theta_1-\omega t)]$ (in this paragraph we
take into account the harmonic time dependence $\exp[-i\omega t]$).
By inserting the leading-order terms of Eq.~(\ref{RPHF}) into these
wavelike contributions, we can easily note that the contribution of
the $n$th ``surface wave" reduces to
\begin{equation}\label{surface_wave}
\exp[-\eta_c \alpha(n) \theta_1] \, \exp\left[i \left(r_c/\sqrt{f_c}
\,\omega \theta_1-\omega t \right)\right]
\end{equation}
The second term describes the propagation of this ``surface wave"
near the photon sphere at $r=r_c$. Indeed, it circles the BH in time
$T'=2\pi (r_c/\sqrt{f_c})$ which is exactly the time (\ref{Torbit})
needed for a massless particle to orbit the BH on an unstable
circular null geodesic. The first term corresponds to an exponential
decay of this ``surface wave" due to continual reradiation of
energy. It is very interesting to rewrite Eq.~(\ref{surface_wave})
in the form
\begin{equation}\label{Interp_SW}
\exp[-k''_n(\omega){\cal L}]\exp[i(k'_n(\omega){\cal L}-\omega t)]
\end{equation}
with ${\cal L}=r_c \, \theta_1$ which denotes the arc length taken
on the photon sphere. Now, $k'_n(\omega)=\omega/\sqrt{f_c}$
represents the wavenumber of the $n$th ``surface wave" or, in other
terms, its dispersion relation, while $k''_n(\omega)=\eta_c
\alpha(n)/r_c$ is its damping constant.
To the leading order, the dispersion relation is linear and
independent of the index of the ``surface wave" while the damping
constant depends only on the index $n$. Of course, if we go beyond
the leading-order terms, the dispersion relation reads [see
Eq.~(\ref{PRapproxWKB})]
\begin{subequations}\label{PR_WKB_disp}
\begin{equation}\label{PRapproxWKB_disp}
k'_{n}(\omega) \approx \left[\frac{1}{f_{c}} \,\omega^{2} +
\frac{a_{n}}{r_c^2}+2\left(\frac{\eta_c^2}{r_c^2
}\right)\alpha(n)^{2}\epsilon_{n}(\omega)\right]^{1/2}
\end{equation}
which implies [see Eq.~(\ref{RPHF})]
\begin{equation} \label{PRasympWKB_disp}
k'_{n}(\omega)=\left[\frac{1}{\sqrt{f_{c}}}~\omega
+\frac{a_{n}}{(2r_c^2/\sqrt{f_{c}})}~\frac{1}{\omega}\right]
+\underset{\omega \to +\infty}{\cal O}\left(
\frac{1}{\omega^3}\right)
\end{equation}
\end{subequations}
and is clearly nonlinear as well as dependent on the index $n$. This
result could have important consequences in strong gravitational
lensing (see also Ref.~\cite{DecaniniFolacci2010a} and the
discussion in Appendix B.1 of the present paper). With such a
potential application in mind, it is worth noting that, in
Eqs.~(\ref{Interp_SW}) and (\ref{PR_WKB_disp}), $\omega$ denotes the
frequency of the scalar photon observed by a static observer at
infinity. If we consider the gravitational redshift of this photon
and introduce its frequency ${\tilde \omega}=\omega/\sqrt{f_{c}}$
measured by a static observer lying on the photon sphere, the
dispersion relation (\ref{PR_WKB_disp}) reads
\begin{subequations}\label{PR_WKB_disp_GS}
\begin{equation}\label{PRapproxWKB_disp_GS}
k'_{n}({\tilde \omega}) \approx \left[{\tilde \omega}^{2}
+\frac{a_{n}}{r_c^2}+2\left(\frac{\eta_c^2}{r_c^2
}\right)\alpha(n)^{2}\epsilon_{n}({\tilde \omega})\right]^{1/2}
\end{equation}
and we have
\begin{equation}\label{RPHF_disp_GS} k'_{n}({\tilde
\omega})=\left[{\tilde \omega}
+\frac{a_{n}}{2r_c^2}~\frac{1}{{\tilde \omega}}\right]
+\underset{\omega \to +\infty}{\cal O}\left( \frac{1}{{\tilde
\omega}^3}\right).
\end{equation}
\end{subequations}
When $a_n>0$, Eq.~(\ref{PR_WKB_disp_GS}) provides a superluminal
dispersion relation. Indeed, it leads to a group velocity
$v_g({\tilde \omega})=d{\tilde \omega} /dk'_{n}({\tilde \omega})
>1$. The condition $a_n>0$ is satisfied $\forall n \in {\bf N}$ for
spins 1 and 2 propagating on the Schwarzschild BH [see Eq.~(13) of
Ref.~\cite{DecaniniFolacci2010a}]. It is also satisfied for the
massless scalar field when $n\not= 0$. As a consequence, it cannot
be considered as an exotic physical condition. It even seems to be
true in most cases encountered (see the examples considered in
Sec.~IV). At first sight, this result may seem a little bit
puzzling. But it is important to note that superluminal dispersion
relations for the ``surface waves" lying on the photon sphere do not
necessarily lead to a violation of the relativistic principle of
causality. Indeed, for the transfer of information between a source
and a receptor located outside a BH, various ``channels" are
involved. Of course, there are channels associated with diffraction
by the BH photon sphere [for a source and a receptor at infinity,
they correspond to the sum over the Regge poles in
Eq.~(\ref{ampliPR2})] but there are also channels associated with
geometrical rays [for a source and a receptor at infinity, they come
from the background integral (\ref{SWTransfo}) over the contour
$\Gamma$ of Fig.~\ref{fig:watson} after asymptotic evaluation]. In
order to study causality, it would be necessary to carefully take
into account interferences between all the monochromatic spectral
components of the signal carrying the information for all the
channels involved. We believe that such a study would not show a
violation of causality. The situation encountered here is similar to
that discussed
by many authors working in electromagnetism of dispersive media
(see, e.g., Ref.~\cite{Brillouin1960} and references therein). In such a context, it
has been observed that a group velocity greater than the light velocity does not violate
causality because it is not the velocity of information transmission or the energy velocity.
Finally, the WKB result (\ref{PRapproxWKB}) and the associated Regge
trajectories permit us to derive, from the semiclassical formulas
(\ref{sc1}) and (\ref{sc4}), useful analytical expressions for the
QNM complex frequencies. Indeed, by inserting (\ref{PRapproxWKB})
into Eqs.~(\ref{sc1}) and (\ref{sc4}), we obtain the large $\ell$
behaviors
\begin{widetext}
\begin{subequations}\label{asympQNM}
\begin{eqnarray}
& & \omega ^{(o)}_{\ell n} = \frac{\sqrt{f_{c}}}{r_{c}} \left[ \left(\ell +\frac{d-3}{2}\right)
- \frac{a_n}{2 \ell} + (d-3)\frac{a_n}{4 \ell^2} + \underset{\ell \to
+\infty}{\cal O}\left( \frac{1}{\ell^3} \right)
\right], \\
& & \frac{\Gamma _{\ell n}}{2} =
\eta_c\frac{\sqrt{f_{c}}}{r_{c}} \alpha(n)\left[1+ \frac{c_n}{2 \ell^2} +
\underset{\ell \to +\infty}{\cal O}\left( \frac{1}{\ell^3}
\right)\right],
\end{eqnarray}
\end{subequations}
with $\ell \in {\bf N}$ and $n=1, 2, \dots$ Here
\begin{equation}\label{c_n}
c_n=a_n+2 b_n.
\end{equation}
\end{widetext}
It should be noted that the leading-order terms of
Eq.~(\ref{asympQNM}) have already been obtained in
Refs.~\cite{ZerbiniVanzo2004,CardosoMirandaBertietal2009,DolanOttewill_2009}.
The higher-order terms in $1/\ell$ and $1/\ell^2$ are new (see,
however, Sec.~5.2 of Ref.~\cite{DolanOttewill_2009}). They are
directly linked to the nonlinear behavior of both the dispersion
relation and the damping of the ``surface waves" lying on (close to)
the photon sphere.
\section{Applications}
In this section we shall apply the previous formalism to various
spacetimes of physical interest.
\subsection{Schwarzschild-Tangherlini black holes}
For Schwarzschild-Tangherlini BHs \cite{Tangherlini63}, the function $f(r)$ reads
\begin{equation}
f(r)=1-\left(\frac{r_0}{r}\right)^{d-3}
\end{equation}
with
\begin{equation}\label{r0dimd}
r_0^{d-3}=\frac{16\pi M}{(d-2){\cal A}_{d-2}} \quad \text{and} \quad
{\cal A}_{d-2}=\frac{2\pi^{(d-1)/2}}{\Gamma[(d-1)/2]}.
\end{equation}
Here $M$ is the mass of the BH, ${\cal A}_{d-2}$ the area of the
unit sphere $S^{d-2}$ and the event horizon is located at $r_h=r_0$.
The unstable circular null geodesics are located at
\begin{subequations}
\begin{equation}\label{rcSchTang}
r_c=r_0\left(\frac{d-1}{2}\right)^{1/(d-3)}
\end{equation}
and the associated $\eta_c$ parameter is given by
\begin{equation}\label{etaSchTang}
\eta_c=\sqrt{d-3}.
\end{equation}
\end{subequations}
\subsubsection{d=4. The Schwarzschild black hole}
For $d=4$, the Schwarzschild-Tangherlini solution is nothing but the
ordinary Schwarzschild BH and we have
\begin{subequations}
\begin{eqnarray}
&&r_{c}=\frac{3}{2}r_{0}=3M,\\
&&\eta_c=1,
\end{eqnarray}
\end{subequations}
as well as
\begin{subequations}
\begin{eqnarray}
&&a_{n}=-\frac{29}{216}+\frac{5}{18}\alpha(n)^{2},\\
&&b_{n}=\frac{1357}{15552}-\frac{305}{3888}\alpha(n)^{2},\\
&&c_{n}=\frac{313}{7776}+\frac{235}{1944}\alpha(n)^{2},\\
&&\epsilon_{n}(\omega)=\frac{b_{n}}{27M^{2}\omega^{2}+a_{n}+\alpha(n)^{2}}.
\end{eqnarray}
\end{subequations}
The WKB approximation (\ref{PRapproxWKB}) for the Regge poles leads to
\begin{eqnarray}\label{Sch41}
&&\lambda_{n}(\omega) \approx
\left[27M^{2}\omega^{2}+a_{n}+2\alpha(n)^{2}\epsilon_{n}(\omega)\right]^{1/2}
\nonumber \\ && \qquad+i\alpha(n)\left[1+\epsilon_{n}(\omega)\right]
\end{eqnarray}
and their high frequency behavior (\ref{RPHF}) provides
\begin{eqnarray}\label{Sch42}
&&\lambda_{n}(\omega)=\left[3\sqrt{3}M\omega+\frac{\sqrt{3}a_n}{18M\omega}\right]\nonumber \\
&& \qquad+i\alpha(n)\left[1+\frac{b_{n}}{27M^{2}\omega^{2}}\right]
+\underset{\omega \to +\infty}{\cal O}\left(
\frac{1}{\omega^3}\right).
\end{eqnarray}
Formulas (\ref{Sch41}) and (\ref{Sch42}) are in agreement with the results
obtained in Ref.~\cite{DecaniniFolacci2010a}.
The resonance excitation frequencies and the damping of the QNMs given by
the general formulas (\ref{asympQNM}) reduce to
\begin{subequations}\label{QNMSch}
\begin{eqnarray}
&&\omega_{\ell
n}^{(0)}=\frac{1}{3\sqrt{3}M}\left[\left(\ell+\frac{1}{2}
\right)-\frac{a_{n}}{2\ell} +\frac{a_{n}}{4\ell^{2}}+\underset{\ell
\to +\infty}{\cal O}\left(\frac{1}{\ell^3}\right)\right],
\nonumber \\
&&\\
&&\frac{\Gamma_{\ell n}}{2}=\frac{\alpha(n)}{3\sqrt{3}M}\left[1+\frac{c_{n}}{2\ell^{2}}
+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right].
\end{eqnarray}
\end{subequations}
Formulas (\ref{QNMSch}) are in agreement with the results obtained in Ref.~\cite{Iyer2}
(see also Ref.~\cite{DecaniniFolacci2010a}).
\subsubsection{The five-dimensional Schwarzschild-Tangherlini black hole}
For $d=5$, we have
\begin{subequations}
\begin{eqnarray}
&&r_{c}=\sqrt{2}r_{0},\\
&&\eta_c=\sqrt{2},
\end{eqnarray}
\end{subequations}
as well as
\begin{subequations}
\begin{eqnarray}
&&a_{n}=-\frac{5}{16}+\frac{3}{4}\alpha(n)^{2},\\
&&b_{n}=\frac{115}{512}-\frac{31}{128}\alpha(n)^{2},\\
&&c_{n}=\frac{5}{48}+\frac{17}{64}\alpha(n)^{2},\\
&&\epsilon_{n}(\omega)=\frac{b_{n}}{2r_{c}^{2} \, \omega^{2}+a_{n}+2\alpha(n)^{2}}.
\end{eqnarray}
\end{subequations}
The WKB approximation (\ref{PRapproxWKB}) for the Regge poles leads to
\begin{eqnarray}
&&\lambda_{n}(\omega) \approx
\left[2r_{c}^{2}\omega^{2}+a_{n}+4\alpha(n)^{2}\epsilon_{n}(\omega)\right]^{1/2}\nonumber
\\ && \qquad+i\alpha(n)[1+\epsilon_{n}(\omega)]
\end{eqnarray}
and their high frequency behavior (\ref{RPHF}) provides
\begin{eqnarray}
&&\lambda_{n}(\omega)=\left[\sqrt{2}\,r_{c}\omega+\frac{a_{n}}{2\sqrt{2}\,r_{c}\omega}\right]\nonumber \\ &&\qquad+i\sqrt{2}\,\alpha(n)\left[1+\frac{b_n}{2r_{c}^{2}\omega^{2}}\right]+\underset{\omega \to +\infty}{\cal O}\left(
\frac{1}{\omega^3}\right).
\end{eqnarray}
The resonance excitation frequencies and the damping of the QNMs given
by the general formulas (\ref{asympQNM}) reduce to
\begin{subequations}
\begin{eqnarray}
&&\omega_{\ell n}^{(0)}=\frac{1}{\sqrt{2}\,r_c}\left[\left(\ell+1\right)-\frac{a_{n}}{2\ell}+\frac{a_{n}}{2\ell^{2}}+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right],\nonumber \\
&&\\
&&\frac{\Gamma_{\ell n}}{2}=\frac{\alpha(n)}{r_c}\left[1+\frac{c_{n}}{2\ell^{2}}+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right].
\end{eqnarray}
\end{subequations}
\subsubsection{The six-dimensional Schwarzschild-Tangherlini black hole}
For $d=6$, we have
\begin{subequations}
\begin{eqnarray}
&&r_{c}=\left(\frac{5}{2}\right)^{1/3}r_{0},\\
&&\eta_c=\sqrt{3},
\end{eqnarray}
\end{subequations}
as well as
\begin{subequations}
\begin{eqnarray}
&&a_{n}=-\frac{31}{60}+\frac{7}{5}\alpha(n)^{2},\\
&&b_{n}=\frac{1411}{3600}-\frac{91}{180}\alpha(n)^{2},\\
&&c_{n}=\frac{481}{1800}+\frac{7}{18}\alpha(n)^{2},\\
&&\epsilon_{n}(\omega)=\frac{b_{n}}{(5/3)r_{c}^{2} \, \omega^{2}+a_{n}
+3\alpha(n)^{2}}.
\end{eqnarray}
\end{subequations}
The WKB approximation (\ref{PRapproxWKB}) for the Regge poles leads to
\begin{eqnarray}
&&\lambda_{n}(\omega) \approx \left[\frac{5}{3}r_{c}^{2}\omega^{2}
+a_{n}+6\alpha(n)^{2}\epsilon_{n}(\omega)\right]^{1/2}\nonumber \\
&& \qquad+i\sqrt{3}\alpha(n)\left[1+\epsilon_{n}(\omega)\right]
\end{eqnarray}
and their high frequency behavior (\ref{RPHF}) provides
\begin{eqnarray}
&&\lambda_{n}(\omega)=\left[\left(\frac{5}{3}\right)^{1/2}r_{c}\omega+\frac{a_{n}}{2(5/3)^{1/2}r_{c}\omega}\right]\nonumber \\ && +i\sqrt{3}\,\alpha(n)\left[1+\frac{b_{n}}{(5/3)r_{c}^{2} \, \omega^{2}}\right]+\underset{\omega \to +\infty}{\cal O}\left(
\frac{1}{\omega^3}\right).\nonumber \\
&&
\end{eqnarray}
The resonance excitation frequencies and the damping of the QNMs given by the general formulas (\ref{asympQNM}) reduce to
\begin{subequations}
\begin{eqnarray}
&&\omega_{\ell
n}^{(0)}=\left(\frac{3}{5}\right)^{1/2}\left(\frac{1}{r_{c}}\right)
\left[\left(\ell+ \frac{3}{2} \right)\phantom{\underset{\ell \to
+\infty}{\cal O}\left( \frac{1}{\ell^3}\right)}\right.\nonumber\\ &&
\qquad \quad \left. -\frac{a_{n}}{2\ell}+3\frac{a_{n}}{4\ell^{2}}
+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right],\\
&&\frac{\Gamma_{\ell n}}{2}=
\sqrt{3}\left(\frac{3}{5}\right)^{1/2}\left(\frac{1}{r_{c}}\right)\alpha(n)\nonumber\\
&& \qquad \quad \times \left[1+\frac{c_{n}}{2\ell^{2}}+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right].
\end{eqnarray}
\end{subequations}
\subsubsection{Leading-order terms for the Schwarzschild-Tangherlini black hole of arbitrary dimension}
It should be noted that, in the $d$-dimensional case, we can easily
derive the leading-order terms of the Regge poles and quasinormal
complex frequencies. We have
\begin{eqnarray}
&&\lambda_{n}(\omega)=\sqrt{\frac{d-1}{d-3}}\,r_c\omega+i\sqrt{d-3}\,\alpha(n)+\underset{\omega \to +\infty}{\cal O}\left(\frac{1}{\omega}\right)\nonumber\\
&&
\end{eqnarray}
and
\begin{subequations}\label{QNMSchTang}
\begin{eqnarray}
&&\omega_{\ell n}^{(0)} = \sqrt{\frac{d-3}{d-1}}\left(\frac{1}{r_{c}}\right)\left(\ell+\frac{d-3}{2}\right)+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell}\right),\nonumber \\
&&\\
&&\frac{\Gamma_{\ell n}}{2} = \frac{d-3}{\sqrt{d-1}}\left(\frac{1}{r_{c}}\right)\alpha(n)+\underset{\ell \to +\infty}{\cal O}\left(\frac{1}{\ell}\right).
\end{eqnarray}
\end{subequations}
Formulas (\ref{QNMSchTang}) are in agreement with the results
obtained in Refs.~\cite{Konoplya2003} and \cite{ZerbiniVanzo2004}.
\subsection{The Reissner-Nordstr\"om black hole}
For the $d$-dimensional Reissner-Nordstr\"om BH \cite{Tangherlini63}, the function $f(r)$ reads
\begin{equation}
f(r)=1-\left(\frac{r_0}{r}\right)^{d-3}+\frac{\theta^{2}}{r^{2(d-3)}}
\end{equation}
with (see Ref.~\cite{MyersPerry1986} or Appendix A of Ref.~\cite{NatarioSchiappa2004})
\begin{equation}
r_{0}^{d-3}=\frac{16\pi M}{(d-2){\cal A}_{d-2}}\doteq 2\mu \quad \text{and}
\quad \theta^2=\frac{8\pi q^2}{(d-2)(d-3)}.
\end{equation}
Here $M$ is the mass of the BH and $q$ is its charge. We furthermore
assume that $\mu^2>\theta^2$. It should be noted that here we
consider electromagnetism in the Heaviside system of units. For this
background there are two horizons, the so-called inner and outer
horizons. They are respectively located at
\begin{subequations}
\begin{eqnarray}
&&r_{-}=\left(\mu-\sqrt{\mu^{2}-\theta^{2}}\right)^{1/(d-3)},\\
&&r_{+}=\left(\mu+\sqrt{\mu^{2}-\theta^{2}}\right)^{1/(d-3)}.
\end{eqnarray}
\end{subequations}
We are only interested in the outer horizon with radius at
$r_h=r_{+}$ because we have $f(r)>0$ for $r \in ]r_h,+\infty[$ [see
assumption (i) of Sec.~II].
The unstable circular null geodesics are located at $r_c>r_h$ given
by
\begin{subequations}
\begin{equation}
r_c=\left(\frac{(d-1)\mu+\sqrt{[(d-1)\mu]^{2}-4(d-2)\theta^{2}}}{2}\right)^{1/(d-3)},
\end{equation}
and we have for the associated parameter $\eta_c$
\begin{equation}
\eta_c=\sqrt{(d-3)-\frac{(d-2)(d-3)\theta^2}{r_c^{2(d-3)}}}.
\end{equation}
\end{subequations}
For $\theta=0$ we recover the Schwarzschild-Tangherlini BH results.
It should be noted that it is also possible to express $\eta_c$ in two other
equivalent forms since, as a consequence of Eq.~(\ref{Assup_fr_2a}),
the parameters $r_c$, $\mu$ and $\theta$ are linked by
\begin{equation}
\left(r_c^{d-3}\right)^{2}-(d-1)\mu r_c^{d-3} +(d-2)\theta^2=0.
\end{equation}
\subsubsection{The four-dimensional Reissner-Nordstr\"om black hole}
For $d=4$, we have
\begin{subequations}
\begin{eqnarray}
&&r_c=\frac{1}{2}(3M+\sqrt{9M^{2}-8Q^{2}}),\\
&&\eta_c=\sqrt{1-\frac{2Q^{2}}{r_c^{2}}},
\end{eqnarray}
\end{subequations}
as well as
\begin{widetext}
\begin{subequations}
\begin{eqnarray}
&&a_{n}=\frac{\left(1-Q^2/r_c^2\right)^2}{216(1-2Q^2/r_c^2)^2}
\left[-29+\frac{86Q^2}{r_c^2}\right]
+\frac{1-Q^2/r_c^2}{36(1-2Q^2/r_c^2)^2}\left[10-\frac{62Q^2}{r_c^2}
+\frac{124Q^4}{r_c^4}\right]\alpha(n)^{2},\\
&&b_{n}=\frac{\left(1-Q^2/r_c^2\right)^2}{15552(1-2Q^2/r_c^2)^5}\left[1357
-\frac{12478 Q^2}{r_c^2} + \frac{42333 Q^4}{r_c^4} -\frac{64324 Q^6}{r_c^6}
+ \frac{43588 Q^8}{r_c^8}\right]\nonumber\\
&& \quad +\frac{1-Q^2/r_c^2}{3888(1-2Q^2/r_c^2)^5}\left[-305 + \frac{3943 Q^2}{r_c^2} -
\frac{21335 Q^4}{r_c^4} + \frac{56357 Q^6}{r_c^6} - \frac{69544 Q^8}{r_c^8}
+ \frac{34772 Q^{10}}{r_c^{10}}\right]\alpha(n)^{2}, \\
&&c_{n}=\frac{\left(1-Q^2/r_c^2\right)^2}{7776(1-2Q^2/r_c^2)^5}\left[313
- \frac{3118 Q^2}{r_c^2}
+ \frac{11229 Q^4}{r_c^4} - \frac{18820 Q^6}{r_c^6} + \frac{18820 Q^8}{r_c^8}\right]\nonumber\\
&& \quad +\frac{1-Q^2/r_c^2}{1944(1-2Q^2/r_c^2)^5}\left[235 -
\frac{2645 Q^2}{r_c^2} + \frac{11929 Q^4}{r_c^4} - \frac{28315
Q^6}{r_c^6} + \frac{37592 Q^8}{r_c^8} - \frac{18796
Q^{10}}{r_c^{10}}\right]\alpha(n)^{2},
\end{eqnarray}
and
\begin{eqnarray}
&&\epsilon_{n}(\omega)=\frac{b_{n}}{3r_c^2/(1-Q^2/r_c^2)\,\omega^{2}
+a_{n}+(1-2Q^{2}/r_c^{2})\alpha(n)^{2}}.
\end{eqnarray}
\end{subequations}
Here we have noted $Q^2=4\pi q^2$. This permits us to compare our
results with those for which electromagnetism is expressed in the
Gaussian system of units. The WKB approximation (\ref{PRapproxWKB})
for the Regge poles leads to
\begin{eqnarray}
&&\lambda_{n}(\omega) \approx \left[\frac{3r_c^2}{1-Q^2/r_c^2}
\omega^{2}+a_{n}
+2\left(1-\frac{2Q^{2}}{r_c^{2}}\right)\alpha(n)^{2}\epsilon_{n}(\omega)\right]^{1/2}
+i\sqrt{1-\frac{2Q^{2}}{r_c^{2}}}\alpha(n)[1+\epsilon_{n}(\omega)]
\end{eqnarray}
\end{widetext}
and their high frequency behavior (\ref{RPHF}) provides
\begin{eqnarray}
&&\lambda_{n}(\omega)=\left[\frac{\sqrt{3}\,r_c}{\sqrt{1-Q^2/r_c^2}} \omega+\frac{a_{n}}{2\sqrt{3}\,r_c/\sqrt{1-Q^2/r_c^2} \,\omega}\right]\nonumber \\ && \quad +i\sqrt{1-\frac{2Q^{2}}{r_c^{2}}}\alpha(n)\left[1+\frac{b_n}{3r_c^2/(1-Q^2/r_c^2)\,\omega^{2}}\right]\nonumber \\ && \quad +\underset{\omega \to +\infty}{\cal O}\left(
\frac{1}{\omega^3}\right).
\end{eqnarray}
The resonance excitation frequencies and the damping of the QNMs given
by the general formulas (\ref{asympQNM}) reduce to
\begin{subequations}\label{QNMRN}
\begin{eqnarray}
&&\omega_{\ell
n}^{(0)}=\sqrt{1-\frac{Q^2}{r_c^2}}\left(\frac{1}{\sqrt{3}\,r_c}\right)
\left[\left(\ell+ \frac{1}{2} \right) \phantom{\left(
\frac{1}{\ell^3}\right)}\right. \nonumber \\ && \left.\qquad
-\frac{a_{n}}{2\ell}+\frac{a_{n}}{4\ell^{2}}+\underset{\ell \to
+\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right],\\
&&\frac{\Gamma_{\ell n}}{2}= \sqrt{1-\frac{2Q^{2}}{r_c^{2}}}\sqrt{1-\frac{Q^2}{r_c^2}}
\left(\frac{1}{\sqrt{3}\,r_c}\right)\,\alpha(n) \nonumber \\ && \qquad \times
\left[1+\frac{c_{n}}{2\ell^{2}}+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right].
\end{eqnarray}
\end{subequations}
The leading-order terms of (\ref{QNMRN}) are in agreement with
formulas (37) and (38) of Ref.~\cite{FerrariMashhoon84}.
Furthermore, it should be noted that, for $Q=0$, all the results
obtained for the Regge poles and the complex quasinormal frequencies
of the four-dimensional Reissner-Nordstr\"om BH reduce to the
Schwarzschild BH results of Sec.~IV.A.1.
\subsubsection{Leading-order terms for the Reissner-Nordstr\"om black hole of arbitrary dimension}
In the $d$-dimensional case, we can easily derive the leading-order
terms of the Regge poles and quasinormal complex frequencies. We
have
\begin{eqnarray}\label{PR_RNd}
&&\lambda_{n}(\omega)=\sqrt{\frac{d-1}{d-3}}\frac{r_{c}\omega}{\sqrt{1-\frac{\theta^2}{r_c^{2(d-3)}}}
} \nonumber \\ && \qquad +i\sqrt{d-3}\sqrt{1-\frac{(d-2)\theta^2}{r_c^{2(d-3)}}}\alpha(n)
+\underset{\omega \to +\infty}{\cal O}\left(\frac{1}{\omega}\right)\nonumber\\
&&
\end{eqnarray}
and
\begin{subequations}\label{QNMRNd}
\begin{eqnarray}
&&\omega_{\ell n}^{(0)} = \sqrt{\frac{d-3}{d-1}}\sqrt{1-\frac{\theta^2}{r_c^{2(d-3)}}}
\left(\frac{1}{r_{c}}\right)\left(\ell+\frac{d-3}{2}\right)\nonumber \\
&& \qquad +\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell}\right),\\
&&\frac{\Gamma_{\ell n}}{2} = \frac{d-3}{\sqrt{d-1}}\sqrt{1-\frac{\theta^2}{r_c^{2(d-3)}}}
\sqrt{1-\frac{(d-2)\theta^2}{r_c^{2(d-3)}}}\nonumber \\ && \qquad \times
\left(\frac{1}{r_{c}}\right)\alpha(n)+\underset{\ell \to +\infty}{\cal O}\left(\frac{1}{\ell}\right).
\end{eqnarray}
\end{subequations}
It should be noted that for $\theta=0$, from (\ref{PR_RNd}) and
(\ref{QNMRNd}), we recover the Schwarzschild-Tangherlini BH results
of Sec.~IV.A.4.
\subsection{The canonical acoustic black hole}
For the canonical acoustic BH (for more details see Sec.~8 of Ref.~\cite{Visser98}), the function $f(r)$ reads
\begin{equation}
f(r)=1-\left(\frac{r_0}{r}\right)^{4}.
\end{equation}
The sonic event horizon is located at $r_h=r_0$, the radius of the
unstable circular null geodesics is given by
\begin{subequations}
\begin{equation}
r_c=3^{1/4}r_0
\end{equation}
and we have for the corresponding $\eta_c$ parameter
\begin{equation}
\eta_c=2.
\end{equation}
\end{subequations}
Furthermore, we have
\begin{subequations}
\begin{eqnarray}
&&a_{n}=\frac{1}{108}+\frac{20}{9}\alpha(n)^{2},\\
&&b_{n}=-\frac{85}{1944}-\frac{215}{243}\alpha(n)^{2},\\
&&c_{n}=-\frac{19}{243}+\frac{110}{243}\alpha(n)^{2},\\
&&\epsilon_{n}(\omega)=\frac{b_{n}}{(3/2)r_{c}^{2} \, \omega^{2}+a_{n}+4\alpha(n)^{2}}.
\end{eqnarray}
\end{subequations}
The WKB approximation (\ref{PRapproxWKB}) for the Regge poles leads to
\begin{eqnarray}
&&\lambda_{n}(\omega) \approx
\left[\frac{3}{2}r_{c}^{2}\omega^{2}+a_{n}
+8\alpha(n)^{2}\epsilon_{n}(\omega)\right]^{1/2}\nonumber \\ &&
\qquad+2i\alpha(n)\left[1+\epsilon_{n}(\omega)\right]
\end{eqnarray}
and their high frequency behavior (\ref{RPHF}) provides
\begin{eqnarray}
&&\lambda_{n}(\omega)=\left[\frac{r_c}{\sqrt{2}}\omega+\frac{a_{n}}{\sqrt{2}\,r_{c}\omega}\right]\nonumber \\
&& \qquad+2i\alpha(n)\left[1+\frac{b_{n}}{(3/2)r_{c}^{2}\omega^{2}}\right]
+\underset{\omega \to +\infty}{\cal O}\left(
\frac{1}{\omega^3}\right).\nonumber \\
&&
\end{eqnarray}
The resonance excitation frequencies and the damping of the QNMs given by the general formulas (\ref{asympQNM})
reduce to
\begin{subequations}\label{QNMCABH}
\begin{eqnarray}
&&\omega_{\ell n}^{(0)}=\frac{\sqrt{2}}{r_{c}}\left[(\ell+1/2)-\frac{a_{n}}{2\ell}+\frac{a_{n}}{4\ell^{2}}+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right],\nonumber \\
&& \label{QNMCABHa}\\
&&\frac{\Gamma_{\ell n}}{2}=\frac{2\sqrt{2}}{r_{c}}\alpha(n)\left[1+\frac{c_{n}}{2\ell^{2}}+\underset{\ell \to +\infty}{\cal O}\left(
\frac{1}{\ell^3}\right)\right]\label{QNMCABHb}.
\end{eqnarray}
\end{subequations}
The leading-order terms of (\ref{QNMCABH}) are in agreement with
formula (53) of Ref.~\cite{BertiCardosoLemos2004}. In
Ref.~\cite{DolanOttewill_2009}, Dolan and Ottewill have obtained for
$n=0$ the expansions of $\omega_{\ell n}^{(0)}$ and $\Gamma_{\ell
n}/2$ up to order $1/(\ell+1/2)^{4}$. Our results (\ref{QNMCABH})
are consistent with their Eq.~(78).
\section{Conclusion}
As noted by Chandrasekhar in the mid-1970s, BH perturbation theory
can be formulated as a resonant scattering problem. As a
consequence, all the techniques developed in the framework of
scattering theory can be naturally introduced in the context of BH
physics. One of the central concepts of scattering theory is the
concept of a Regge pole which permits one to obtain semiclassical
interpretations of resonance phenomena. We have recently used it in
Refs.~\cite{DecaniniFJ_cam_bh} and \cite{DecaniniFolacci2010a} in
order to understand, from a new point of view, some aspects of the
resonant Schwarzschild BH. Our results have permitted us to
established more particularly, on a rigorous basis and in the
wave/field theory context, the appealing and intuitive
interpretation of Schwarzschild BH QNMs suggested by Goebel in 1972
\cite{Goebel}, i.e., that they could be interpreted in terms of
gravitational waves in spiral orbits close to the unstable circular
photon/graviton orbit at $r=3M$ which decay by radiating away
energy.
In the present paper, we have greatly extended the approach
initiated in Ref.~\cite{DecaniniFolacci2010a}. More precisely, we
have provided general analytical formulas for the Regge poles of the
$S$-matrix associated with a massless scalar field theory defined on
a static spherically symmetric BH of arbitrary dimension with a
photon sphere. This has been achieved by using third-order WKB
approximations to solve the associated radial wave equation and by
emphasizing more particularly the role of the photon sphere or, in
other words, of the unstable circular null geodesics on which a
massless particle can orbit the BH. But, it is important to note
that our results are only a first step to understand, in a
semiclassical framework, various aspects of static spherically
symmetric BH physics such as wave scattering, gravitational lensing,
Hawking radiation... Moreover, we think that it would now be very
interesting to introduce the CAM method in a less symmetric
situation, e.g., to study Kerr and Kerr-Newman BHs (see
Ref.~\cite{GlampedakisAndersson2003} for a first numerical step in
this direction) and to analyze the splitting of their complex
quasinormal frequencies or to understand, from a semiclassical point
of view, the superradiance phenomenon. We leave such a study for the
future.
|
\section{INTRODUCTION}
In Wilman et al.~(2008) (hereafter W08) we presented a semi-empirical simulation of the extragalactic radio continuum sky primarily intended to aid the design of
scientific programmes for the next generation of radio telescope facilities, culminating in the {\em Square Kilometer Array} (SKA).
The simulation covers a sky area of $20 \times 20$~deg$^{2}$ and contains radio-loud and radio-quiet active galactic nuclei (AGN) and star-forming galaxies out
to redshift $z=20$ within a framework for their large-scale clustering. The full source catalogue -- containing 320 million sources above 10~nJy at five frequencies ranging from 151~MHz to 18~GHz --
can be accessed at the SKADS Simulated Skies ($S^{3}$) database\footnote{http://s-cubed.physics.ox.ac.uk} under $S^{3}$--SEX (semi-empirical extragalactic simulation).
There are by necessity numerous uncertainties and limitations in the $S^{3}$--SEX simulation, including but not limited to issues such as the form of the high-redshift evolution of the AGN and galaxies,
the lack of star-forming/AGN hybrid galaxies, and the abundance of highly-obscured Compton-thick AGN. In so far as possible, flexibility was built into the simulations to allow post-processing to improve their accuracy as observations in the years ahead lead to improved constraints. Major advances in these areas are expected from the far-infrared surveys to be performed by the {\em Herschel} Space Observatory (Pilbratt 2008). To facilitate such comparisons we have post-processed the $S^{3}$--SEX simulation to cover these wavelengths. In this paper, we describe the recipes for assigning infrared spectral energy distributions (SEDs) to the radio sources, using existing mid- and far-infrared results from {\em Spitzer} and sub-mm survey data from {\em SCUBA} as the primary constraints. We then present our predictions for {\em Herschel} surveys, bolstered by a comparison with results from the
Balloon-borne Large Aperture Submillimetre Telescope ({\em BLAST}) which offer a foretaste of {\em Herschel}'s capabilities. In keeping with the philosophy of the S$^{3}$ project to maximise the degree of interaction between the user and the database, the infrared fluxes are also provided on the $S^{3}$ webpage from which users can generate simulation products for comparison with observations. The radio--infrared connection is of immense empirical
and theoretical interest for extragalactic surveys, for the identification and follow-up of {\em Herschel} sources, and for probing the physics of the far-infrared--radio correlation and its possible evolution. The simulation
can nevertheless also serve as a standalone {\em Herschel} simulation, for comparison with others such as the phenomenologically-inspired backward evolution models of Valiante et al.~(2009) and Pearson \& Khan~(2009), and the model of Lacey et al.~(2009) which is based on a semi-analytical galaxy formation model.
\section{THE ASSIGNMENT OF INFRARED SEDs TO THE RADIO SOURCE POPULATIONS}
The input radio source catalogue was obtained from the $S^{3}$-SEX online database and negative evolution at high
redshift was imposed using the `default post-processing options' described in W08. Each radio source was assigned an
infrared SED from various template libraries appropriate for star-forming galaxies and AGN, and the output flux density of
each infrared model component for each galaxy [in log(F$_{\rm{\nu}}$(Jy)] was computed at observed wavelengths of 24, 70, 100, 160, 250, 350, 450, 500, 850,
and 1200 {\rm\thinspace $\mu$m}. The wavelengths of 24, 70 and 160 {\rm\thinspace $\mu$m}~are those of the {\em Spitzer} MIPS instrument, for which an abundance of extragalactic survey
data is available to guide the construction of our model and to compare against its output (see the review of Soifer et al.~2008). {\em Herschel} will conduct surveys at 250, 350 and 500 {\rm\thinspace $\mu$m}~with the SPIRE instrument and at 70, 100 and 160{\rm\thinspace $\mu$m}~with PACS. Longer-wavelength constraints are available from submillimetre surveys by SCUBA at 450 and 850{\rm\thinspace $\mu$m}, and from the MAMBO instrument on the 30-metre IRAM telescope at 1200{\rm\thinspace $\mu$m}. The catalogued output flux densities are monochromatic values, with the exception of the {\em Spitzer} 24 and 70{\rm\thinspace $\mu$m}~MIPS bands for which the modelled spectra were convolved with the MIPS bandpass transmission curves, following the {\em Spitzer} Synthetic Photometry Recipe\footnote{http://ssc.spitzer.caltech.edu/postbcd/cookbooks/synthetic\underline{ }photometry.html}. This is due to the spectral complexity in the rest-frame 10{\rm\thinspace $\mu$m}~region, particularly for the star
formation component due to the presence of Polycyclic Aromatic Hydrocarbon (PAH) features.
The format of the output from the infrared extension, as it appears in the $S^{3}$-SEX online database, is described in Appendix A.
\subsection{Star-forming galaxies}
The population of star-forming galaxies in W08 comprises two populations, normal (or quiescent) galaxies and starbursts, identified respectively with the low and high-luminosity Schechter function components of the local 1.4~GHz luminosity function of Yun, Reddy \& Condon (2001). The entire population was evolved with pure luminosity evolution (1+z)$^{3.1}$ out to z=1.5 (defined in an Einstein-de Sitter cosmology but adapted to the flat-$\Lambda$ cosmology used for the simulation). The default post-processing option consists of negative space density evolution of the form $(1+z)^{-7.9}$ above $z=4.8$. We stress that the terms `normal galaxy' and `starburst' are merely convenient labels for these two components of the local luminosity function in our phenomenological model. Physically speaking, the terms may not necessarily offer an accurate description of the nature of the star formation in these populations, especially beyond the local Universe.
The first step in assigning an infrared SED is to use the far-infrared--radio correlation for
star-forming galaxies (eqn.~5 of Yun, Reddy \& Condon~2001) in order to derive the rest-frame far-infrared luminosity, L(FIR), given the rest-frame 1.4~GHz luminosity, L$_{\rm{1.4 GHz}}$. The relation is characterised by the ``q'' parameter:
\begin{equation}
q = \rm{ log [L(FIR,W)}/3.75 \times 10^{12} Hz] - log [\rm{L_{\rm{1.4 GHz}} (W Hz^{-1})}],
\end{equation}
for which we assume a gaussian distribution with $\mu = 2.34$ and $\sigma = 0.26$ (Yun, Reddy \& Condon). There is evidence that this relation holds out to redshift $z > 1$ (e.g. Appleton et al.~2004, Garn et al.~2009b),
although such conclusions are sensitive to the assumed infrared SEDs and associated ``k-corrections''. Indeed, using recent {\em BLAST} stacking measurements, Ivison et al.~(2009) reported evidence for a tentative decline in $q_{\rm{IR}}$ (defined as the logarithmic ratio of the the bolometric 8--1000{\rm\thinspace $\mu$m}~infrared to monochromatic 1.4~GHz radio fluxes), of the form $q_{\rm{IR}} \propto (1+z)^{-0.15 \pm 0.03}$ (see also Kovacs et al.~2006).
The far-infrared luminosity, L(FIR), is traditionally defined through a linear combination of IRAS 60 and 100{\rm\thinspace $\mu$m}~flux densities~(see Sanders \& Mirabel~1996), which for an SED
consisting of a superposition of modified blackbodies yields the 42.5--122.5{\rm\thinspace $\mu$m}~luminosity to sub-percent accuracy (Helou et al.~1988). Given L(FIR),
galaxies were assigned a model SED from the library of star-forming galaxy templates assembled by Rieke et al.~(2009). The latter consists of 14 SEDs uniformly spaced in total
infrared luminosity (8--1000{\rm\thinspace $\mu$m}), L(TIR), from log L(TIR)(\hbox{$\rm\thinspace L_{\odot}$}) = 9.75 to 13.00. The templates were taken from the online version of Table 4 in Rieke et al.~(2009) but they
terminate on the short wavelength side at 5.26{\rm\thinspace $\mu$m}~for those with log L(TIR) (\hbox{$\rm\thinspace L_{\odot}$}) $\leq 11$ and at 4.02{\rm\thinspace $\mu$m}~at higher luminosities. In order to predict
24{\rm\thinspace $\mu$m}~fluxes for galaxies at redshift $z \geq 3$, we need to extend these templates to shorter wavelengths. For this purpose we employed the spectra of the individual local luminous
(LIRGs) and ultraluminous infrared galaxies (ULIRGs) in Table 3 of Rieke et al.~(2009), which extend down to 0.4{\rm\thinspace $\mu$m}. From these, we computed mean LIRG and ULIRG spectra and matched
them onto the original Rieke templates below the cut-off wavelengths of 4.02 and 5.26{\rm\thinspace $\mu$m}. The ULIRG spectrum was used for the templates with log L(TIR)(\hbox{$\rm\thinspace L_{\odot}$}) $>$ 12, and the
LIRG spectrum for the remaining templates. The resulting templates should be considered as schematic below $\sim 3${\rm\thinspace $\mu$m}, as we did not attempt to model the stellar population and rest-frame optical
extinction in a self-consistent fashion.
The template SEDs were integrated over the appropriate wavelengths ranges to yield the mapping between
L(TIR) to L(FIR) shown in Fig.~\ref{fig:LTIRtoFIR_Rieke}; the adopted relation between the two quantities is mildly non-linear: log L(FIR) = 1.1log L(TIR) -- 1.42.
Based on the IRAS Bright Galaxy Sample (Soifer et al.~2007), Marcillac et al.~(2006) assumed a linear relation: L(FIR) = 1.91 L(TIR). In Fig.~\ref{fig:RiekesameFIR} we show the Rieke
templates scaled to a common L(FIR).
At a given L(FIR), the local galaxy population exhibits a distribution in 60--100{\rm\thinspace $\mu$m}~colour or, equivalently, dust temperature, as characterised by Chapin et al.~(2009). We use this
distribution (Chapin et al.~2009, equations 8 and 9) to assign each star-forming galaxy a 60--100{\rm\thinspace $\mu$m}~colour, $C$, defined as the ratio of the rest-frame monochromatic 60 and 100{\rm\thinspace $\mu$m}~flux densities:
$C = \rm{log} L_{\rm{60}}/L_{\rm{100}}$. The template nearest in colour is then selected from the Rieke library using the empirical $C$--L(TIR) conversion for these templates shown in Fig.~\ref{fig:RiekeCOLOR}, for
which a functional fit is:
\begin{equation}
C = 0.33 \rm{tanh}[1.1(\rm{log L(TIR)} - 11.0)] - 0.23.
\end{equation}
Having selected the template shape from the Rieke library by this process, the final step is to normalise it to actual value of L(FIR) originally specified by the far-infrared--radio relation. This is done using the non-linear L(FIR)--L(TIR) relation given in Fig.~\ref{fig:LTIRtoFIR_Rieke}. Chaplin et al.~(2009) presented evidence for evolution
in the colour-luminosity relation, but we initially assume that the local relation applies at all redshifts.
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{RIEKE_FIRvsTIR.ps}
\caption{\normalsize The symbols show the far-infrared luminosity (42.5--122.5~{\rm\thinspace $\mu$m}), L(FIR)(\hbox{$\rm\thinspace L_{\odot}$}), of the star-forming galaxy templates of Rieke et al.~(2009) as a function of
total infrared luminosity, L(TIR). To assign SED templates to the simulated galaxies, we assume the non-linear relation given by the solid line: log L(FIR) = 1.1log L(TIR) -- 1.42. The
dashed line shows the linear relation L(FIR) = 1.91 L(TIR) assumed by Marcillac et al.~(2006).}
\label{fig:LTIRtoFIR_Rieke}
\end{figure}
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{RIEKEtemplates_NORMsameFIR_tempex.ps}
\caption{\normalsize The star-forming galaxy template SEDs scaled to a common far-infrared luminosity using the L(TIR)--L(FIR) scaling in Fig.~\ref{fig:LTIRtoFIR_Rieke}. With increasing
L(TIR), the SEDs peak at shorter wavelengths and exhibit weaker spectral features in the $\lambda <20${\rm\thinspace $\mu$m}~region.}
\label{fig:RiekesameFIR}
\end{figure}
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{RIEKE_COLORvsTIR.ps}
\caption{\normalsize Colour of the Rieke templates [$C = \rm{log} L_{\rm{60}}/L_{\rm{100}}$] as a function of log L(TIR)(\hbox{$\rm\thinspace L_{\odot}$}), with the fit given by equation (2) of section 2.1.}
\label{fig:RiekeCOLOR}
\end{figure}
\subsection{Radio-quiet AGN}
In W08 a population of radio-quiet AGN was incorporated using the Ueda et al.~(2003) hard X-ray luminosity function and a relation between 2--10{\rm\thinspace keV}~and 1.4~GHz
luminosities. To assign infrared SEDs the first task is to split the population into subgroups of unobscured AGN, and Compton-thin and Compton-thick obscured AGN. This was not done in
W08 but is now necessary because the infrared SEDs are sensitive to this classification. We start from the findings of Hasinger~(2008), who used a compilation of hard X-ray surveys to parameterise
the fraction of obscured Compton-thin AGN (as a proportion of a total population comprising unobscured AGN and obscured Compton-thin AGN, but excluding Compton-thick obscured AGN)
as a function of $L_{\rm{X}}$ (2--10{\rm\thinspace keV}~luminosity; \hbox{$\erg\s^{-1}\,$}):
\begin{equation}
f_{\rm{2}} = [0.27 + \beta(log L_{\rm{X}} - 43.75)]g(z),
\end{equation}
where $\beta = -0.281$ and $g(z)$ is evolution of the form (1+z)$^{0.62}$ out to $z=2$ and flat thereafter. The number of Compton-thick obscured AGN is not as well
constrained, and in W08 we simply boosted the space density of the Ueda et al.~(2003) luminosity function by 50~per cent in a notional allowance for them, but left the
issue open for subsequent refinement in post-processing. Ueda et al. showed that a reasonable match to the X-ray background results if the number of Compton-thick AGN simply equals the
number of Compton-thin obscured AGN. We thus assume that the abundance of Compton-thick AGN population is a factor $f_{\rm{Cthick}}$ times that of the combined population of unobscured and
Compton-thin obscured AGN, where $f_{\rm{Cthick}}$ = min[0.5,$f_{\rm{2}}$], with the upper bound of 0.50 hard-wired into the W08 simulation. After the removal of any excess Compton-thick AGN,
the remaining sources in the W08 catalogue are probabilistically identified as unobscured (type I), Compton-thin or Compton-thick obscured (type II) AGN. Finally, 10~per cent of all
sources are flagged for removal because W08 included radio-loud AGN with a separate radio luminosity function but this population is also implicitly present in the Ueda et al. luminosity function.
The fractions of retained radio-quiet AGN which are unobscured, Compton-thin and Compton-thick obscured are shown as functions of $L_{\rm{X}}$ and redshift in Fig.~\ref{fig:RQAGNfrac}.
\begin{figure*}
\includegraphics[width=0.35\textwidth,angle=0]{RQAGN_hist1_2deg.ps}
\includegraphics[width=0.35\textwidth,angle=0]{RQAGN_hist2_2deg.ps}
\includegraphics[width=0.35\textwidth,angle=0]{RQAGN_hist3_2deg.ps}
\includegraphics[width=0.35\textwidth,angle=0]{RQAGN_hist4_2deg.ps}
\caption{\normalsize The fractions of retained radio-quiet AGN identified as unobscured
AGN (solid lines), Compton-thin obscured (dashed lines) and Compton-thick obscured AGN (dotted lines), for intrinsic luminosity thresholds of $L_{\rm{X}}> 10^{40}$, $10^{42}$, $10^{43}$ and $10^{44}$\hbox{$\erg\s^{-1}\,$}.
Evaluated from a 4~deg$^{2}$ simulation area.}
\label{fig:RQAGNfrac}
\end{figure*}
\subsubsection{AGN emission}
For the retained sources, $L_{\rm{X}}$ is inferred from the radio luminosity using equation (2) of W08. Sources with $L_{\rm{X}}$ above and below a threshold
of $3 \times 10^{44}$\hbox{$\erg\s^{-1}\,$}~are classified as quasars and Seyferts, respectively. Unobscured (type I) quasars are assigned at random one of the three templates for optically-selected
QSOs compiled by Polletta et al.~(2007) (termed QSO1, BQSO1 and TQSO1, which differ in their relative optical/infrared flux ratios)\footnote{These SEDs were obtained from http://www.iasf-milano.inaf.it/$\sim$polletta/templates/swire\underline{ }templates.html}. The unobscured Seyferts are assigned SEDs from the online library of models computed by Nenkova et al.~(2008) using the {\it CLUMPY} code for dust radiative transfer in a clumpy AGN torus. To narrow down the parameter space of
available models, we use the results of Thompson et al.~(2009) which are based on analysis of Seyfert 1 AGN mid-infrared spectra. They prefer a range of {\it CLUMPY} parameter space defined
by the parameters $Y=30$, $q=1$, $\sigma = 45$~deg, $m=2$, $i=30$~deg, $\tau_{\rm{V}}=30-60$ and $N_{\rm{0}} \leq 6$ (see Nenkova et al.~2008 for the meaning of these parameters). 18 CLUMPY
models satisfy these constraints and were assigned at random to our type I Seyferts, with the SED consisting of the reprocessed torus emission and direct AGN broken power-law emission
(defined as a power-law $F_{\rm{\lambda}} \propto \lambda^{-1.5}$ over 0.1--1{\rm\thinspace $\mu$m}~with a Rayleigh-Jeans slope at longer wavelengths).
For the type II Compton-thin AGN, we assigned one of the type 2 QSO templates given by Polletta et al.~(2006) and Polletta et al.~(2007) (termed QSO2 and Torus, respectively) for the quasars,
and one of the aforementioned CLUMPY models (but without the direct AGN emission component) to the Seyferts.
For the Compton-thick AGN, we assigned half the population SEDs using the same prescription as for the type II Compton-thin sources; for the other half of the population, we reflect the growing
body of evidence which suggests that a large proportion of Compton-thick AGN are obscured by an extended distribution of cold dust in the host galaxy and not by a nuclear torus (see e.g.
Polletta et al., 2006, 2008; Martinez-Sansigre et al. 2006, Lacy et al. 2007). Accordingly we assign these sources an intrinsic type I or Compton-thin type II SED, extinguished by a
dust screen with $A_{\rm{V}} = 4- 27$~mag~(as found by Polletta et al.~2008 for obscured quasars). An extinction curve extending into the infrared is taken from McClure~(2009).
These AGN infrared SED components are normalised using the correlation between 12 micron and hard X-ray emission found by Gandhi et al.~(2009), which appears to hold for obscured and unobscured
AGN (including Compton-thick sources) and to extend into the quasar regime:
\begin{equation}
\rm{ log L_{\rm{MIR}}} = (-4.37 \pm 3.08) + (1.106 \pm 0.071)\rm{log} L_{\rm{X}},
\end{equation}
where L$_{\rm{MIR}} = \lambda L_{\rm{\lambda}}$ at 12.3{\rm\thinspace $\mu$m}~in \hbox{$\erg\s^{-1}\,$}. Gaussian scatter of $\sigma = 0.36$ is assumed on the relation.
\subsubsection{Star-formation}
There is abundant evidence for ongoing star formation in the host galaxies of radio-quiet AGN and this is likely to dominate the far-infrared SED. The consensus
emerging from targeted observations of local Seyferts (Buchanan et al.~2006; Thompson et al.~2009) and studies of more distant samples and cosmological surveys (e.g. Silverman et al.~2009; Hern\'{a}n-Caballero et al.~2009; Serjeant \& Hatziminaoglou~2009) is that the star formation is generally higher in obscured AGN compared with their unobscured counterparts, and increases sub-linearly with AGN luminosity. We embody these findings with a star formation rate (SFR) $\propto \sqrt{L_{\rm{X}}} (1+z)^{1.6}$, but assume different
normalizations for the type I and II AGN. This luminosity and redshift dependence is similar to that measured for quasars by Serjeant \& Hatziminaoglou~(2009) and the luminosity
dependence matches that of local Seyfert 1s (Thompson et al.~2009).
For type I AGN, we normalise the relation using the measurements of type I SDSS AGN at z=0.15 by Kim et al.~(2006), which have a median SFR of 0.5\hbox{$\Msun\yr^{-1}\,$}. The median log L([OIII]$\lambda5007$) of this sample, as inferred from Fig.~4(b) of Kim et al., is 41.4 (\hbox{$\erg\s^{-1}\,$}) which equates to log L$_{\rm{X}}=42.6$ using the log L$_{\rm{X}}$--log L([OIII]$\lambda5007$) relation of eqn.~(1) of Silverman et al.~(2009). Thus for type I AGN:
\begin{equation}
{\rm SFR} = 0.63 \sqrt{L_{\rm{X}} /10^{43}} (1+z)^{1.6} \hbox{$\Msun\yr^{-1}\,$}.
\end{equation}
For the type II AGN (Compton-thin and Compton-thick), we base our results on the $z=0.1$ SDSS type II AGN sample of Kauffmann et al.~(2003), focussing on the sub-sample of strong-lined sources with log L([OIII])$>$ 40.5 indicated by the red dots in Fig.~10 of Silverman et al.~(2009). Referring once more to the log L([OIII]) distributions in Fig.~4(b) of Kim et al.~(2006), we infer for this particular luminosity-limited sample a median log L([OIII])=41.1, which translates into log L$_{\rm{X}}=42.25$ and hence boosts the pre-factor of our assumed SFR(L$_{\rm{X}}$,z) relation from 0.63 to 2\hbox{$\Msun\yr^{-1}\,$}. A scatter of 1 dex is assumed on the SFR(L$_{\rm{X}}$,z) relation and the star-formation rates are assumed to decline as $(1+z)^{-7.9}$ at $z > 4.8$ following the assumed evolution of the global star formation
rate density. SEDs for the star formation component are assigned from the Rieke et al.~(2009) library, using the relation given by Yun, Reddy \& Condon~(2001) (their eqn.~13) to relate the star formation rate and 1.4~GHz radio luminosity; the far-infrared--radio relation (eqn.~1) was then used to assign L(FIR) and hence an appropriate template from the Rieke library using Fig.~\ref{fig:LTIRtoFIR_Rieke}.
\subsection{Radio-loud AGN}
Radio-loud AGN were modelled by W08 using the Willott et al.~(2001) 151~MHz luminosity function comprising a weakly-evolving low luminosity component and a rapidly-evolving high luminosity component, identified with FRI and FRII radio sources, respectively. We consider them separately for the purposes of infrared SED assignment.
\subsubsection{FRIs}
There is scant evidence for the presence of intrinsically luminous optical-UV AGN in the radio population we associate with FRI sources (e.g. Vardoulaki et al.~2008). We therefore model their infrared SEDs with a template representing host galaxy starlight only, which we produced using the GALSYNTH online interface to the GRASIL code developed by Silva et al.~(1998). The models are based on their `Elliptical model', which gives the evolving SED (taking account of dust reprocessing) of a stellar population formed in a burst of 1~Gyr duration (see description of this particular model at http://adlibitum.oat.ts.astro.it/silva/grasil/modlib/modlib.html). We generated models with ages separated by 0.1~Gyr for 0.1--1.5~Gyr, and by 1~Gyr for ages 2--13~Gyr. We assume formation redshifts distributed uniformly in the range z=10--20, and select the template nearest in age at the redshift of the galaxy under consideration. To normalise the SEDs, we draw upon the bivariate 1400~MHz--K-band luminosity
function derived by Mauch \& Sadler~(2007) for a local sample: given a radio power, the absolute K-band magnitude is drawn from a Gaussian distribution of specified mean and variance and this is used
to normalise the SED. For simplicity, we assume the same (local) relation between K-band luminosity and radio power at all redshifts. An artifact of this choice of model is that the mid-far infrared regions
of the SED template increase sharply for ages below the assumed duration of the starburst (1~Gyr). This causes discontinuities in the mid- and far-infrared fluxes at $z \simeq 4$ where the galaxy age first drops below this figure.
\subsubsection{FRIIs}
The FRIIs are first split into populations of quasars and radio galaxies for which the nuclear emission is seen directly and through obscuration, respectively. Willott et al.~(2000) found that for log L(151~MHz)[W/Hz/sr]$>26.5$ the quasar fraction is 0.4, independent of redshift and luminosity, and around 0.1 at lower radio luminosity. Taking into account the contamination by FRIs at lower luminosity, we simply assume that
40 per cent of our FRIIs are seen as quasars, i.e. those seen at viewing angle of $\leq 53$~degrees to the jet axis. To assign AGN and starburst infrared SEDs, we recall that quasars are, on average, more intrinsically luminous AGN than the radio galaxies (see e.g. Simpson 2003). To work with a parameter that is closely tied to the intrinsic strength of the nuclear emission, we convert from L(151~MHz) [W/Hz/sr] to L([OIII]$\lambda5007$) [W] using a relation
from Grimes, Rawlings \& Willott~(2004) (GRW04):
\begin{equation}
\rm{log L([OIII]\lambda5007}) = 7.53 + 1.045\rm{log L(151~MHz).}
\end{equation}
Eqn.~(6) applies to the population as a whole, but following GRW04 we assume that the quasars and radio galaxies lie 0.3 dex systemtically above and below it, respectively, with an additional 0.5 dex of scatter on each distribution. To assign an AGN component to the infrared SED, we draw upon the findings of Haas et al.~(2008) who presented {\em Spitzer} 3--24{\rm\thinspace $\mu$m}~photometry of 3CR steep-spectrum quasars and radio galaxies at redshift $1<z<2.5$. The quasar mid-infrared SEDs are very uniform, roughly constant in $\nu F_{\rm{\nu}}$. There is a near proportionality between $\nu F_{\rm{\nu}}$(8~{\rm\thinspace $\mu$m}, rest) and $\nu F_{\rm{\nu}}$(178~MHz, rest) with 0.5 dex scatter. The radio galaxies are
dominated by host galaxy starlight below 3{\rm\thinspace $\mu$m}~(rest) and at longer wavelengths by a quasar nucleus reddened by up to A$_{\rm{V}}=50$~mag. The radio galaxies are a factor 3--10 less luminous in this spectral range than the
quasars, which they attribute entirely to extinction. For a sample of 3CR sources at $0.4 \leq z <1.2$, Cleary et al.~(2007) find that the quasars are 4 times as luminous as the radio galaxies at 15{\rm\thinspace $\mu$m}, half of which they
attribute to extinction, and half to synchrotron contamination in the quasars. However, their sample includes a significant number of steep-spectrum compact and super-luminal quasars. They also assumed a spherically symmetric
dust model, which is likely to be inadequate.
We assume that the quasars and radio galaxies follow the same intrinsic relation between $\nu L_{\rm{\nu}}$(8~{\rm\thinspace $\mu$m}~rest) [W] and L([OIII]$\lambda5007$) [W], derived from the quasar points in Fig.~1 of Haas et al.~(2008) using eqn.~(6) to convert from L(151~MHz) to L([OIII]$\lambda5007$). We assume:
\begin{equation}
\rm{log \nu L_{\rm{\nu}} (8~{\rm\thinspace $\mu$m}~rest)} = 3.3 +
0.9569\rm{log L([OIII]\lambda5007}).
\end{equation}
No intrinsic scatter is assumed on this, as the scatter in eqn.~(6) can reproduce the scatter in Fig.~1 of Haas et al.~(2008). Using this normalisation, we assign the quasars randomly to one of the three type I quasar templates from
Polletta et al.~(2007). The radio galaxies are also assigned one of these template SEDs in an analogous manner, and then extinction of $A_{\rm{V}}$=0--40~mag is applied. An elliptical host galaxy model from the GRASIL library is assigned as for the FRIs but in this case normalised using the K--z relation of Willott et al.~(2003) with an additional offset of $\Delta K = -0.36(\rm{log L(151 MHz)[W/Hz/sr]} - 26.55)$~mag (following Willott et al.~2003, Fig.~3).
In principle, the evolving elliptical template from the GRASIL library should self-consistently reproduce the far-infrared dust emission. To compare with this, we included in addition a far-infrared
dust component to explicitly match the starburst properties to the sub-millimetre observations of radio galaxies and quasars by Archibald et al.~(2001), Willott et al.~(2002) and Reuland et al.~(2004).
From Willott et al.~(2002), the radio quasars appear to be a factor of several brighter than the radio galaxies at 850{\rm\thinspace $\mu$m}. In Grimes, Willott \& Rawlings~(2005), this was enshrined as a universal
relationship between L([OIII]$\lambda5007$) [W] and L(850{\rm\thinspace $\mu$m}) [W/Hz/sr]:
\begin{equation}
\rm{log L(850{\rm\thinspace $\mu$m})} = -2.5753 + 0.682\rm{log L([OIII]\lambda5007}),
\end{equation}
with additional scatter drawn from a gaussian of $\sigma = 0.44$. With this normalisation, we adopt a modified blackbody with $\beta = 1.95$, $T=41${\rm\thinspace K}, as found by Priddey \& McMahon~(2001) for high-redshift quasars. Archibald et al.~(2001) found L(850{\rm\thinspace $\mu$m}~(rest)) to be a strongly increasing function of redshift, but for our adopted SED parameters, the redshift dependence is a more modest $(1+z)^{1.5}$ (see Reuland et al., Fig.~4); this dependence is added to eqn.~(8) and normalised at $z=2.5$. There is evidence for a decline in L(850{\rm\thinspace $\mu$m}~(rest)) at $z>4$, so in common with other parts of the simulation we assume a decline of the form $(1+z)^{-7.9}$ at $z>4.8$. We note, however, that Rawlings et al.~(2004) found any explicit redshift dependence to be small or absent. Although Willott et al.~(2002) found tentative evidence for an anti-correlation between radio-source size and sub-mm luminosity for quasars, no such trend was found for the radio galaxy sample of Reuland et al.~(2004) and we have chosen not to incorporate one.
Finally, we remark that our treatment of the infrared SED is confined to reprocessed dust emission due to star formation and AGN heating. As a result, we have not endeavoured to incorporate an infrared synchrotron emission component in the beamed radio-loud AGN. If required, such a component could be added as an extrapolation of the radio SED given in the $S^{3}$ database, following the beaming treatment described by W08. It would only have a significant effect on the predicted source counts at the brightest flux densities.
\section{COMPARISON WITH EXISTING OBSERVATIONS}
The prescriptions outlined in section 2 constitute our `baseline model'. Here we compare its source counts and other derived quantities against observations. In Fig.~\ref{fig:BASELINEdsc} we show the normalised differential source counts at 24, 70 and 160{\rm\thinspace $\mu$m}~for comparison with the most recent {\em Spitzer} MIPS survey results. To generate these plots we used a 4~deg$^{2}$ area of the input radio simulation, coupled with a shallow subset of the full 400~deg$^{2}$ radio simulation area to generate the bright flux ends (above 1.6{\rm\thinspace mJy}~at 24{\rm\thinspace $\mu$m}, 30{\rm\thinspace mJy}~at 70{\rm\thinspace $\mu$m}~and 100{\rm\thinspace mJy}~at 160{\rm\thinspace $\mu$m}). The simulated counts fall decisively short of the observations below 1{\rm\thinspace mJy}~and 10{\rm\thinspace mJy}~at 24 and 70{\rm\thinspace $\mu$m}, respectively. There is a modest excess of simulated sources above $\sim 100${\rm\thinspace mJy}~at 70 and 160{\rm\thinspace $\mu$m}, and above $\sim 10${\rm\thinspace mJy}~at 850{\rm\thinspace $\mu$m}~(Fig.~\ref{fig:BASELINE850csc}) where we compare with results from the {\em SCUBA} Half Degree Extragalactic Survey (SHADES; Coppin et al.~2006).
Since the source counts are dominated by star-forming galaxies, we show in Fig.~\ref{fig:BASELINEqparams} the quantity $q_{\rm{IR}} = log (F_{\rm{IR}}/F_{\rm{20 cm}})$ in the 24 and 70{\rm\thinspace $\mu$m}~bands for the simulated normal galaxy population. $F_{\rm{IR}}$ and $F_{\rm{20 cm}}$ are the simulated infrared and radio fluxes which would be observed in these bands, i.e. without the application of any K-corrections. The simulated sample is dominated by galaxies substantially fainter than those in current observational samples, but our $q_{\rm{24}}$ and $q_{\rm{70}}$ values compare reasonably well with those measured locally (Appleton et al.~2004, Beswick et al.~2008 and Garn et al.~2009a,b). For a better comparison with the observations, we applied cuts in flux density to yield the sub-samples also shown in Fig.~\ref{fig:BASELINEqparams}: at 24{\rm\thinspace $\mu$m}, we selected galaxies with $F_{\rm{24}} > 0.3${\rm\thinspace mJy}~and $F_{\rm{20 cm}} > 90${\rm\thinspace $\mu$Jy}~to mimic the selection of Appleton et al.~(2004), but we note that these authors only included sources with spectroscopic redshifts and appear not to discriminate between star-forming galaxies and AGN, biases which we cannot easily accommodate. At 70{\rm\thinspace $\mu$m}, we apply cuts of $F_{\rm{70}} > 6${\rm\thinspace mJy}~and $F_{\rm{20 cm}} > 30${\rm\thinspace $\mu$Jy}~to match the selection of Seymour et al.~(2009), who use photometric and spectroscopic redshifts and apply star-forming galaxy/AGN classification. The resulting sample shows only a very modest decline in $q_{\rm{70}}$ out to $z=2$, consistent with the findings of Seymour et al.
\begin{figure*}
\includegraphics[width=0.47\textwidth,angle=0]{24micDSC_RIEKEtemps_AUG1_2010FINAL.ps}
\includegraphics[width=0.47\textwidth,angle=0]{70micDSC_RIEKEtemp_AUG1_2010FINAL.ps}
\includegraphics[width=0.47\textwidth,angle=0]{160micDSC_RIEKEtemp_AUG1_2010FINAL.ps}
\caption{\normalsize Source counts for the baseline model simulation (section 2) showing normal (asterisks) and starburst galaxies (cross), radio-quiet AGN (diamonds), and the total (plus signs) [FRIs and FRIIs are also included in the totals]. The large squares on the 24{\rm\thinspace $\mu$m}~panel show the {\em Spitzer} counts of Papovich et al.~(2004); the thick lines in the 70 and 160{\rm\thinspace $\mu$m}~panels bound the observed counts in the S-COSMOS and SWIRE fields (Frayer et al.~2009). The plots were generated using a $4$~deg$^{2}$ region of the W08 input radio simulation and the full $400$~deg$^{2}$ to generate the bright flux end (see section 2 for details).}
\label{fig:BASELINEdsc}
\end{figure*}
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{850micronCUMSC_AUG1_2010FINAL.ps}
\caption{\normalsize The thin continuous line shows the simulated integral source counts at 850{\rm\thinspace $\mu$m}~for the baseline model using a 4~deg$^{2}$ region of the input radio simulation; the contributions of normal and starburst galaxies are shown by the dashed and dot-dashed lines, respectively. The asterisks and thick lines show the measured counts and uncertainty region for the SCUBA Half-Degree Extragalactic Survey (Coppin et al. 2006).}
\label{fig:BASELINE850csc}
\end{figure}
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{q24_Rieketemplates_NG_AUG1_plusOBS_tempex.ps}
\includegraphics[width=0.47\textwidth,angle=0]{q70_Rieketemplates_NG_AUG1_plusOBS.ps}
\caption{\normalsize The quantity $q_{\rm{IR}} = log (F_{\rm{IR}}/F_{\rm{20 cm}})$ is used to characterise the far-infrared:radio correlation for star-forming galaxies and is plotted here for the entire population of simulated normal galaxies (black dots) at 24 and 70{\rm\thinspace $\mu$m}, using the baseline model of section 2. The red crosses and diamonds are the sources remaining after the application of infrared and radio flux cuts to mimic the selection of Appleton et al.~(2004) and Seymour et al.~(2009), respectively. }
\label{fig:BASELINEqparams}
\end{figure}
\section{REVISIONS TO THE BASELINE MODEL}
We now adapt the baseline model to provide a better match to the existing infrared source count data. We focus entirely on the star-forming galaxies and perform the modifications in two steps, motivated by recent
observational findings. Firstly, we use post-processing to modify the cosmological luminosity evolution of the star-forming galaxies to remove the shortfall in the 70{\rm\thinspace $\mu$m}~source counts. Secondly, we allow the SED template
assignment procedure to evolve towards cooler templates at higher redshift in an attempt to rectify the deficit in the 24{\rm\thinspace $\mu$m}~counts.
\subsection{Modifications to the luminosity evolution of the star-forming galaxies}
In W08 we assumed that the radio luminosity function of the star-forming galaxy population undergoes pure luminosity evolution (PLE) of the form $(1+z)^{p}$ with p=3.1 out to z=1.5,
applied in an {\em Einstein -- de Sitter cosmology}, but adapted to the flat $\Lambda$ cosmology of the simulation. Since then, new observational constraints on the evolutionary form have been published. At
70{\rm\thinspace $\mu$m}, Huynh et al.~(2007) constrained the evolution out to $z \sim 1$ and found a degeneracy between luminosity and density evolution with $p=2.78$ in the PLE case; using several {\em Spitzer} surveys,
Magnelli et al.~(2009) found redshift evolution consistent with PLE with $p=3.6 \pm 0.4$ out to $z \sim 1.3$. Both these studies derived the evolution for a flat $\Lambda$ cosmology, which for a given functional form of PLE, leads to a higher
abundance of galaxies than present in the W08 simulation. The discrepancy is especially pronounced at and below the break in the luminosity function. We can, however, correct for it by boosting the luminosity
of the star-forming galaxies in the W08 catalogue by an amount $\Delta log L(L,z)$. The boost required to bring the W08 luminosity function into agreement with PLE of $p=3.1$ to $z=1.5$ in a $\Lambda$ universe is shown
in Fig.~\ref{fig:lumboost}. [We note, however, that W08 assumed that the local luminosity function is flat below 20.6 W/Hz, in which regime no luminosity boost can compensate for the deficit in space density; for continuity,
the boost is accordingly held constant below this evolving break luminosity]. After the application of this luminosity boost, the 24 and 70{\rm\thinspace $\mu$m}~source counts are as shown in Fig.~\ref{fig:LUMBOOSTdsc}. The 70 {\rm\thinspace $\mu$m}~counts below 10{\rm\thinspace mJy}~are now in satisfactory agreement with the observational data but a deficit remains at 24{\rm\thinspace $\mu$m}, which we address in the next subsection.
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{LFboost_evol4_corrected.ps}
\caption{\normalsize The boost in luminosity, $\Delta log L(L,z)$ (dex), which must be applied to the star-forming galaxies in the W08 catalogue to transform to pure luminosity evolution of the form $(1+z)^{3.1}$ to $z=1.5$ in a
$\Lambda$-Universe. The lines show redshifts z=0.08 (solid), 0.45 (short dashed), 0.83 (dotted), 1.2 (dot-dashed) and 1.5 (long dashed).}
\label{fig:lumboost}
\end{figure}
\begin{figure*}
\includegraphics[width=0.47\textwidth,angle=0]{24micDSC_RIEKEtemps_OCT100_2010FINAL.ps}
\includegraphics[width=0.47\textwidth,angle=0]{70micDSC_RIEKEtemp_OCT100_2010FINAL.ps}
\includegraphics[width=0.47\textwidth,angle=0]{160micDSC_RIEKEtemp_OCT100_2010FINAL.ps}
\caption{\normalsize 24, 70 and 160{\rm\thinspace $\mu$m}~source counts for the modified model of section 4.1, in which the luminosity of the star-forming galaxies is boosted to mimic pure luminosity evolution in a $\Lambda$-universe.
Symbols and observed data are as in Fig.~\ref{fig:BASELINEdsc}.}
\label{fig:LUMBOOSTdsc}
\end{figure*}
\subsection{Evolution towards cooler dust templates and the final model}
There is growing evidence that the one-to-one correspondence between infrared luminosity and template SED established in the local Universe breaks down at earlier cosmic epochs. To cite several
examples, Symeonidis et al.~(2009) showed that luminous and ultraluminous galaxies out to redshift $z \sim 1$ have much cooler dust distributions (i.e. SEDs peaking at longer wavelengths)
than their local counterparts in the IRAS Bright Galaxy Sample (see also Symeonidis et al.~2008); and in the mid-infrared, Farrah et al.~(2008) showed that ULIRGs at $z \sim 1.7$ have
spectral features characteristic of starbursts with luminosities $10^{10-11}$\hbox{$\rm\thinspace L_{\odot}$}, rather than local ULIRGs. Such trends may reflect more spatially extended star-forming regions at
higher redshifts with lower dust optical depths. Based on a study of sub-mm galaxies at $z \sim 2.5$, Chapin et al.~(2009) argued for luminosity evolution in the colour-luminosity relation
of the form $(1+z)^{3}$, in the sense that the SED becomes cooler at higher redshift (see also Pope et al.~2006).
Building on these findings, we now introduce phenomenological modifications to the SED template assignment procedure with the aim of reducing the remaining deficit in the 24{\rm\thinspace $\mu$m}~source counts
(Fig.~\ref{fig:LUMBOOSTdsc}). With reference to the Rieke templates in Fig.~\ref{fig:RiekesameFIR}, we see that for a given L(FIR), evolution towards cooler dust templates with increasing redshift
can significantly boost the observed 24{\rm\thinspace $\mu$m}~flux relative to that at 70{\rm\thinspace $\mu$m}. Accordingly, the L(FIR) is scaled down by a factor $f(z)$ for the purposes of selecting the far-infrared colour, $C$,
from the $C$--L(FIR) distribution (see section 2.1). The chosen template is, however, still normalised to the original value of L(FIR). After some experimentation, we chose a function which evolves as
$f(z) = (1 + z)^{2.5}$ out to $z=1$. Beyond $z \sim 1$, the evolution towards cooler templates must stop or go into reverse in order not to overproduce the 850{\rm\thinspace $\mu$m}~counts. We chose a decline of
the form $(1+z)^{-1.5}$ from $z=1-2$, flat thereafter. Even with this in place, the starbursts still overproduce the bright end of the 850{\rm\thinspace $\mu$m}~source counts above 10{\rm\thinspace mJy}~if we retain the `default
post-processing option', i.e. a $(1+z)^{-7.9}$ decline in space density at $z>4.8$. As shown in Fig.~\ref{fig:TEMPEVdsc850}, one possible way to avoid this is to remove the entire starburst population (i.e. the high-luminosity component) at $z>1.5$ and we adopt this solution for our final model.
The removal of the high-redshift starburst component of the luminosity function implies that the evolving normal galaxy population in the simulation is sufficient to account for the star formation at these epochs. Although we used the terms `normal galaxy' and `starburst' merely as labels to refer to the two Schechter function components of the local luminosity function, this trend may find some physical justification in the work of Obreschkow \& Rawlings~(2009). The latter authors showed that, with increasing redshift, galaxy disks are smaller, leading to higher gas densities and pressures and hence higher molecular gas fractions and star formation rates. In the local Universe, star formation rates in excess of $\sim 10$\hbox{$\Msun\yr^{-1}\,$}~can only occur in dense gas disks created by zero angular momentum gas in major mergers characteristic of starburst galaxies, e.g. the merger-triggered local ULIRGs.
Even with the starburst cut-off in place, the modest (factor $\sim 2$) excesses of simulated sources at the bright flux ends of all three {\em Spitzer} bands persist. In Fig.~\ref{fig:fluxredshift}, we show the flux--redshift planes for the normal galaxy and starburst populations for 70 and 160{\rm\thinspace $\mu$m}~fluxes exceeding 100{\rm\thinspace mJy}; much of the excess normal galaxy population occurs at redshifts $z<0.5$, particularly at 70{\rm\thinspace $\mu$m}. It may arise from a combination of effects, such as a departure from strictly power-law pure luminosity evolution or from an inaccurate treatment of the scatter on the far-infrared relation in eqn.~(1) (i.e. from galaxies scattered into the high-L(FIR) gaussian wing of the ``q'' parameter distribution). In support of the first possibility, Huynh et al.~(2007) found no evidence for significant evolution at infrared luminosities below $10^{11}$\hbox{$\rm\thinspace L_{\odot}$}~for redshifts $z<0.4$.
A modest (factor $\sim 2$) deficit also persists in the 24{\rm\thinspace $\mu$m}~counts around 0.3{\rm\thinspace mJy}, which suggests that the required template evolution may be more extreme than assumed, possibly requiring a dependence on luminosity in addition to redshift. It is also possible that the Rieke templates underestimate the PAH strength in the higher redshift galaxies.
Finally, we note that gravitational lensing biases (e.g. Negrello et al.~2007) are not accounted for in these simulations, but can be critical when the source counts are very steep.
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{850micronCUMSC_OCT107_2x2_test1pt5plus.ps}
\includegraphics[width=0.47\textwidth,angle=0]{850micron_zhist_OCT107_2pt4mJy_and5mJy_zed1pt5CUTOFF_2x2deg_newbin.ps}
\caption{\normalsize Upper panel: 850{\rm\thinspace $\mu$m}~source counts for the final model of section 4.2 as determined from a 4~deg$^{2}$~simulation area; the observed SHADES counts are the same as in Fig.~\ref{fig:BASELINE850csc}; the dashed line shows the contribution of the normal galaxies and the dot-dashed line represents the starbursts, assuming a sharp cut-off in the latter population at $z>1.5$; the dotted line shows the contribution of the starbursts under the assumption of the `default post-processing' cut-off in the space density, i.e. $(1+z)^{-7.9}$ at $z>4.8$. Lower panel: 850{\rm\thinspace $\mu$m}~redshift distributions for flux limits of 2.4 and 5{\rm\thinspace mJy}, assuming the $z>1.5$ cut-off in the starburst population.}
\label{fig:TEMPEVdsc850}
\end{figure}
\begin{figure*}
\includegraphics[width=0.47\textwidth,angle=0]{24micDSC_RIEKEtemps_OCT107_2010FINAL.ps}
\includegraphics[width=0.47\textwidth,angle=0]{70micDSC_RIEKEtemps_OCT107_2010FINAL.ps}
\includegraphics[width=0.47\textwidth,angle=0]{160micDSC_RIEKEtemps_OCT107_2010FINAL.ps}
\caption{\normalsize Spitzer MIPS source counts for the final model, after application of a luminosity boost (section 4.1) and template evolution (section 4.2) to
the star-forming galaxies and with the starburst population removed at $z>1.5$; symbols and observed data are as in Fig.~\ref{fig:BASELINEdsc}.}
\label{fig:TEMPEVdsc}
\end{figure*}
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{flux_redshift20x20.ps}
\caption{\normalsize Flux-redshift distributions for simulated normal galaxies and starbursts contributing to the excess differential source counts above 100{\rm\thinspace mJy}~in Fig.~\ref{fig:TEMPEVdsc} at 70 and 160{\rm\thinspace $\mu$m}, demonstrating that much of this excess arises from a local population at $z<0.5$; determined using a shallow subset of the full $20 \times 20$~deg$^{2}$ input radio simulation.}
\label{fig:fluxredshift}
\end{figure}
\subsection{Comparison with observed redshift distributions}
Having achieved reasonable agreement with the {\em Spitzer} MIPS source counts, particularly at the faint end of the observed flux range, we now compare the redshift distributions. In Fig.~\ref{fig:24micZDIST}, we show them for 24{\rm\thinspace $\mu$m}~sources above flux limits of 0.08, 0.15 and 0.3~mJy. The observational data were taken from Le Floc'h et al.~(2009) and are based on photometric redshift analysis of 30,000 sources in a $\sim 1.68$~deg$^{2}$ region of the COSMOS field. The simulated distributions were derived from a 0.25~deg$^{2}$ region and are in good agreement with the observations. We note, however, that Le Floc'h et al.~excluded galaxies brighter than $I_{\rm{AB}}=20$~mag to minimise cosmic variance fluctuations at low redshift. As they also remarked, the small cosmic volume covered by the observations below $z \sim 0.4$ prevents any useful comparison in this regime.
At 70{\rm\thinspace $\mu$m}~we compare with results from the COSMOS and GOODS-N fields (Fig.~\ref{fig:70micZDIST}) using a 4~deg$^{2}$ simulation area. The COSMOS results are based on photometric redshift analysis by Hickey et al.~(in prep.) for a shallow field covering 1.43~deg$^{2}$ and a deep sub-region of 0.165~deg$^{2}$. The deep field is complete to 10{\rm\thinspace mJy}, but the shallow field is
only 60~per cent complete at this flux level; in calculating dn/dz for the shallow field we compensated for this by weighting each source by $1/c(f)$, where $c(f)$ is the completeness level at its flux, as read off from Fig.~3 of Frayer et al.~(2009). The shallow field dn/dz agrees better with the simulation at $z<0.4$, whilst at $z>0.75$ the match to the deeper field is more favourable. Redshift distributions for 70{\rm\thinspace $\mu$m}~COSMOS sources have also been compiled by Kartaltepe et al.~(2009).
At a deeper flux level of 2{\rm\thinspace mJy}, we compare with the raw redshift distribution derived by Huynh et al.~(2007) for the 0.05~deg$^{2}$ GOODS-N field, using a 0.25~deg$^{2}$ simulation area; there is again a deficit of observed sources but the GOODS-N source catalogue is only $\sim 50$~per cent complete at the flux limit (see Frayer et al.~2006a, Table 1), and the raw redshift distribution of Huynh et al.~(2007; their Fig.~1) appears not to correct for this incompleteness. The effects of cosmic variance for the small field observed may also be important.
At 850{\rm\thinspace $\mu$m}, the redshift distributions for flux limits of a few {\rm\thinspace mJy}~(Fig.~\ref{fig:TEMPEVdsc850}) have median redshifts at $z = 2-3$ and high-redshift tails extending beyond $z=4$. Thus, even though the final model has eliminated the starburst (high-L) population at $z>1.5$, it is nevertheless able to reproduce the population of sub-mm (SCUBA) galaxies which reside on average at $z \simeq 2.3$, as deduced from spectroscopic (Chapman et al.~2005) and radio-mm-FIR photometric measurements (Aretxaga et al.~2007). There is also observational evidence that the distribution extends to redshift $z=4$ and beyond (Coppin et al.~2009). This reinforces our earlier statement that the terms `normal galaxy' and `starburst', which we use to label the two Schechter function components of the $z=0$ star-forming galaxy luminosity function, should not be invested with much physical significance beyond the local Universe.
\begin{figure*}
\includegraphics[width=0.47\textwidth,angle=0]{24micron_zhist_OCT106_flim0.ps}
\includegraphics[width=0.47\textwidth,angle=0]{24micron_zhist_OCT106_flim1.ps}
\includegraphics[width=0.47\textwidth,angle=0]{24micron_zhist_OCT106_flim2.ps}
\caption{\normalsize Comparison of the predicted 24{\rm\thinspace $\mu$m}~redshift distributions for the final model of section 4.2 (solid lines) in the central 0.25~deg$^{2}$ of the simulation with those measured in a 1.68 deg$^{2}$ region of the COSMOS field by Le Floc'h et al.~(2009). The latter authors excluded galaxies brighter than $I_{\rm{AB}}=20$~mag to minimize cosmic variance
fluctuations at $z<0.4$. }
\label{fig:24micZDIST}
\end{figure*}
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{70micron_zhist_OCT107_COSMOS_both.ps}
\includegraphics[width=0.47\textwidth,angle=0]{70micron_zhist_OCT107_Huynh.ps}
\caption{\normalsize Comparison of the simulated redshift distribution for the final model of section 4.2 (solid lines) with observations at 70{\rm\thinspace $\mu$m}.
Upper panel: results from the COSMOS field from Hickey et al.~(in prep) for S(70{\rm\thinspace $\mu$m})$>10${\rm\thinspace mJy}~(dashed and dotted histograms are for the deep and shallow fields, respectively -- see
section 4.3 for details); lower panel: the observed data (histogram) are the raw redshift distribution of Huynh et al.~(2007, Fig.~1) for the GOODS-N field above 2{\rm\thinspace mJy}, for which the 70{\rm\thinspace $\mu$m}~source catalogue is,
however, only about 50 per cent complete.}
\label{fig:70micZDIST}
\end{figure}
\subsection{Possible evolution in the far-infrared--radio correlation?}
{\em Our default assumption is that the luminosity boost introduced in section 4.1 applies equally to the radio and infrared luminosities and that it thus reflects a
modification to the cosmological evolution of the star-forming galaxy population as a whole.} Indeed, as we show in Fig.~\ref{fig:RadioSCcheck}, the 1.4~GHz source counts for the
star-forming galaxies come into better agreement with the observations after the application of the luminosity boost. The relevant observations are taken from Seymour et al.~(2008) and provide a break-down of the 1.4~GHz sources counts into contributions from AGN and star-forming galaxies, with the latter dominating below 0.1~mJy. With reference to the radio source counts in Fig.~4 of W08, the increase in the star-forming galaxy contribution brings the total radio source count at 1.4~GHz into better agreement with the observations at this flux level. In light of this, the luminosity boost should be applied when using the simulated radio catalogues, as described in Appendix A.
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{RadioSCcheck2x2deg.ps}
\caption{\normalsize Differential 1.4~GHz source counts for the star-forming galaxies in the simulation with (dashed line) and without (continuous line) the
application of the luminosity boost introduced in section 4.1. The observational data points are the star-forming galaxy source counts measured by Seymour et al.~(2008).}
\label{fig:RadioSCcheck}
\end{figure}
Nevertheless, we cannot exclude the possibility that some contribution to $\Delta log L(L,z)$ may arise from redshift evolution and non-linearity in the far-infrared:radio correlation.
Observational evidence for redshift evolution is at present limited, as discussed in section 3, but strong upward evolution in the relation is expected at $z>3$ due to increased inverse Compton scattering off the cosmic microwave background (e.g. Murphy~2009). In the local Universe, there is evidence that the relation may become non-linear at low luminosities (e.g. Condon~1992; Yun, Reddy \& Condon~1992; Best et al.~2005). Best et al.~(2005) found that below log L(1.4~GHz) = 22.5~W~Hz$^{-1}$ the measured local radio luminosity function falls below that derived in the far-infrared. Such non-linearity could partially account for the luminosity boosting, but it would need to evolve strongly with redshift in order to match that required in Fig.~\ref{fig:lumboost}.
At higher redshift, major contributions in this area are expected from {\em Herschel} surveys, as suggested by initial results from {\em BLAST} (Ivison et al.~2009). To compare with the latter
we show in Fig.~\ref{fig:qIvison} the simulated quantity $q_{\rm{IR}}$, defined analogously to eqn.~(1) but with the far-infrared luminosity replaced by the rest-frame 8--1000{\rm\thinspace $\mu$m}~luminosity, L(TIR),
(eqn.~4 of Ivison et al.). The simulated values are in good accord with the value of $q_{\rm{IR}}=2.41 \pm 0.2$ for the 250{\rm\thinspace $\mu$m}-selected sample of Ivison et al.~(2009), with no evidence of
redshift dependence. Ivison et al.~also determined $q_{\rm{IR}}$ for a sample derived by stacking at the positions of 24{\rm\thinspace $\mu$m}~sources and found tentative evidence for a decline of the
form $(1+z)^{-0.15 \pm 0.03}$, but we do not attempt to replicate the selection effects associated with the definition of this particular sample. The lack of redshift dependence in the
simulated $q_{\rm{IR}}$ primarily reflects the redshift-invariance we assumed for the far-infrared--radio relation in eqn.~(1), in conjunction with the only mildly non-linear dependence of L(TIR) on
L(FIR) in Fig.~\ref{fig:LTIRtoFIR_Rieke}.
For completeness, we also show in Fig.~\ref{fig:qIvison} the quantities $q_{\rm{24}}$ and $q_{\rm{70}}$ as functions of redshift, for comparison with those of the baseline model shown in
Fig.~\ref{fig:BASELINEqparams}. Reflecting the imposed template evolution, $q_{\rm{24}}$ increases more strongly with redshift out to $z=1$ than in the baseline model, whilst $q_{\rm{70}}$ remains
effectively constant out to $z=1$.
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{qIR_Ivison.ps}
\includegraphics[width=0.47\textwidth,angle=0]{q24_Rieketemplates_NG_OCT107_plusOBS_tempex.ps}
\includegraphics[width=0.47\textwidth,angle=0]{q70_Rieketemplates_NG_OCT107_plusOBS_tempex.ps}
\caption{\normalsize Upper panel: the bolometric (8--1000{\rm\thinspace $\mu$m}) $q_{\rm{IR}}$ for the simulated star-forming galaxies. The dashed lines represent the $\pm 3 \sigma$ range measured by Ivison
et al.~(2009) for a 250{\rm\thinspace $\mu$m}~sample. The continuous line shows the tentative evolution measured by Ivison et al.~from stacking measurements. Middle and lower panels: $q_{\rm{24}}$ and $q_{\rm{70}}$
for the final model, with symbols defined as in Fig.~\ref{fig:BASELINEqparams}. }
\label{fig:qIvison}
\end{figure}
\section{PREDICTIONS FOR HERSCHEL SURVEYS}
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{Herschel_CUMSC_OCT107_2x2_colour_zed1pt5SBcutoff.ps}
\includegraphics[width=0.47\textwidth,angle=0]{HerschelBLAST_DSC_OCT107_BLAST_DEVLIN_2x2_colour_zed1pt5SBcutoff.ps}
\caption{\normalsize Upper panel: predicted {\em Herschel} integral source counts at 100~(black solid line), 250~(red dotted line), 350~(green dashed) and 500{\rm\thinspace $\mu$m}~(blue dot-dashed line); lower panel: in differential format for comparison with the {\em BLAST} counts from Table S2 of Devlin et al.~(2009), which are shown by the points with error bars (asterisks, 250{\rm\thinspace $\mu$m}; squares, 350{\rm\thinspace $\mu$m}; triangles, 500{\rm\thinspace $\mu$m}).}
\label{fig:Herschel_SC}
\end{figure}
We now present some predictions for forthcoming {\em Herschel} surveys and a comparison with results from the {\em BLAST} mission which flew
a prototype of the SPIRE instrument on a telescope half {\em Herschel}'s size. In Fig.~\ref{fig:Herschel_SC}, we show predicted integral source counts for surveys at 100, 250, 350
and 500{\rm\thinspace $\mu$m}~which correspond closely with those predicted by the backward-evolution model of Valiante et al.~(2009). Also shown in this figure are normalised differential source
counts at 250, 350 and 500{\rm\thinspace $\mu$m}, which compare reasonably well with the {\em BLAST} counts as determined from the `P(D)' analysis of Devlin et al.~(2009) (see also Patanchon et al.~2009). At 250 and 350{\rm\thinspace $\mu$m}, the model marginally exceeds the {\em BLAST} measurements in the highest flux bins, as also hinted at in the 160{\rm\thinspace $\mu$m}~comparison above $\sim 100${\rm\thinspace mJy}. In Fig.~\ref{fig:BLAST250zdist} we compare with the `complete' redshift distribution derived by Dunlop et al.~(2009) for the 150~arcmin$^{2}$~GOODS-S field at 250{\rm\thinspace $\mu$m}; the observed
sample consists of 20 sources with redshifts down to a nominal flux limit of 35{\rm\thinspace mJy}, although small number statistics and cosmic variance limit the scope of the comparison.
\begin{figure}
\includegraphics[width=0.47\textwidth,angle=0]{250micron_zhist_OCT107_BLAST_N_zed1pt5SBcutoff.ps}
\caption{\normalsize A comparison of the redshift distribution of {\em BLAST} 250{\rm\thinspace $\mu$m}~sources brighter than 35{\rm\thinspace mJy}~in GOODS-S (Dunlop et al.~2009) (histogram) with the simulated distributions for flux limits of
25, 30, 35, 40 and 50{\rm\thinspace mJy}~(solid lines, top to bottom), determined from a 0.25~deg$^{2}$ simulation area.}
\label{fig:BLAST250zdist}
\end{figure}
Although the $S^{3}$ online interface allows the user to query the database to generate redshift distributions for arbitrary radio and
infrared flux cuts and survey areas, we follow Lacey et al.~(2009) and show in Fig.~16 dn/dz plots for a selection of
{\em Herschel} Key Programme\footnote{http://herschel.esac.esa.int/Key\underline{ }Programmes.shtml} surveys: (i) the ultra-deep 0.012~deg$^{2}$~`pencil beam' GOODS-{\em Herschel} field\footnote{http://herschel.esac.esa.int/Docs/KPOT}; (ii) Levels 1 and 5 of the {\em Herschel} Multi-tiered Extragalactic Survey (HERMES \footnote{http://astronomy.sussex.ac.uk/$\sim$sjo/Hermes}); (iii) the wide ($\sim 600$~deg$^{2}$) but shallow ATLAS survey (Eales et al.~2009). Table~1 shows their area coverages, multi-band sensitivities, predicted numbers of galaxies and the fractions at $z>1$ and $z>2$. Our dn/dz predictions were derived from a $0.25$~deg$^{2}$ simulation area and scaled to the required survey areas, except for the ATLAS and HERMES Level 5 predictions for which we used a 4~deg$^{2}$ area . The predicted redshift distributions correspond quite well with those of Lacey et al.~(2009) (L09), despite the differences in our simulation methodologies. For GOODS-{\em Herschel}, our predicted
redshift distributions exhibit broader peaks than those of L09, extending beyond $z=1$; similarly, for HERMES Level 1 and 5 at 250{\rm\thinspace $\mu$m}~our distribution peaks just above $z=1$ and that of L09 just below. At 100 and 160{\rm\thinspace $\mu$m}~in HERMES Level 5, we predict broad redshift distribution with a tail extending beyond $z=1$ whereas L09 predict distributions strongly peaked at $z=0.2$. For the ATLAS survey, the distributions of L09 exhibit marginally higher peaks than ours at $z \leq 0.2$; the L09 distributions cut-off sharply above $z \sim 1.5$, as do ours, due to the step function cut-off in the starburst population above $z=1.5$.
\begin{table*}
\caption{A selection of {\em Herschel} Key Programme surveys for which we provide dn/dz plots in section 5.}
\begin{tabular}{|lllllll|}\hline
Survey & Area & Band & Depth & Total$^{\star}$ & f(z$>$1)$\dagger$ & f(z$>$2)$\dagger\dagger$ \\
& (deg$^{2}$) & ({\rm\thinspace $\mu$m}) & ({\rm\thinspace mJy}) & & & \\ \hline
GOODS-{\em Herschel} & 0.012 & 100 & 0.6 & 320 & 0.61 & 0.14 \\
Ultra-deep & & 160 & 0.9 & 460 & 0.69 & 0.25 \\
Hermes Level 1 & 0.11 & 250 & 4.2 & 1850 & 0.69 & 0.24 \\
& & 350 & 5.7 & 1080 & 0.75 & 0.29 \\
& & 500 & 4.9 & 680 & 0.82 & 0.37 \\
Hermes Level 5 & 27 & 100 & 27 & 11440 & 0.24 & 0.0018 \\
& & 160 & 39 & 13420 & 0.41 & 0.033 \\
& & 250 & 14 & 85700 & 0.63 & 0.18 \\
& & 350 & 19 & 27000 & 0.69 & 0.24 \\
& & 500 & 16 & 10470 & 0.79 & 0.34 \\
ATLAS & 600 & 100 & 67 & 59700 & 0.15 & 1.6E-04 \\
& & 160 & 94 & 57170 & 0.30 & 8.3E-03 \\
& & 250 & 46 & 165100 & 0.51 & 0.10 \\
& & 350 & 62 & 28530 & 0.51 & 0.12 \\
& & 500 & 53 & 5720 & 0.63 & 0.19 \\ \hline
\end{tabular} \\
$\star$ Total number of simulated galaxies in survey area at $z=1-4$ \\
$\dagger$ Fraction at $z>1$ \\
$\dagger\dagger$ Fraction at $z>2$ \\
\end{table*}
\begin{figure}
\begin{centering}
\includegraphics[width=7.8cm,angle=0]{ZDIST_GOODSULTRADEEP_colour_zed1pt5SBcutoff.ps}
\includegraphics[width=7.8cm,angle=0]{ZDIST_HERMES_LEVEL1_colour_zed1pt5SBcutoff.ps}
\includegraphics[width=7.8cm,angle=0]{ZDIST_HERMES_LEVEL5_colour_2x2_10nJycat_zed1pt5SBcutoff.ps}
\includegraphics[width=7.8cm,angle=0]{ZDIST_ATLAS_2x2_colour_10nJycatNG_zed1pt5SBcutoff.ps}
\caption{\normalsize Predicted redshift distributions for a selection of {\em Herschel} Key Programme surveys (listed in Table 1) at 100{\rm\thinspace $\mu$m}~(black solid lines), 160{\rm\thinspace $\mu$m}~(purple dot-dot-dashed lines), 250{\rm\thinspace $\mu$m}~(red dotted lines), 350{\rm\thinspace $\mu$m}~(green dashed lines) and 500{\rm\thinspace $\mu$m}~(blue dot-dashed lines)}
\end{centering}
\label{fig:Herschelzdist}
\end{figure}
\section{CONCLUSIONS}
We have post-processed the S$^{3}$-SEX semi-empirical simulation of the extragalactic radio continuum sky (W08) to make
predictions for {\em Herschel} surveys at far-infrared wavelengths. Existing observations in the mid-infrared with {\em Spitzer} at 24, 70 and 160{\rm\thinspace $\mu$m}~and at 850{\rm\thinspace $\mu$m}~with {\em SCUBA}, together impose strong constraints on the assignment of infrared SEDs to
the star-forming galaxies as a function of redshift. Our principal findings, incorporated into the final model, are as follows:
(i) In order to match the 70{\rm\thinspace $\mu$m}~counts, the star-forming galaxies are required to undergo stronger luminosity evolution than assumed by W08 for the radio simulation.
It may be that W08 simply used an inaccurate evolutionary prescription based on the information available at that time; alternatively, it could be
that there is genuine differential evolution between the far-infrared and radio populations, as a result of an evolving non-linearity in the far-infrared:radio correlation which is already apparent locally at low
luminosity . Results from {\em Herschel} and {\em SKA} precursors will resolve this issue.
(ii) From the local Universe to redshift $z=1$, star-forming galaxies are required to develop progressively cooler SED templates at fixed L(FIR) once the latter
has been set by the far-infrared--radio correlation; the rest-frame 60--100{\rm\thinspace $\mu$m}~colour is assumed to evolve as $\rm{log} L_{\rm{60}}/L_{\rm{100}} \sim (1+z)^{-2.5}$;
beyond $z=1$, this evolution must go into reverse in order not to overproduce the 850{\rm\thinspace $\mu$m}~source counts.
Our chosen model compares favourably with recent far-infrared survey results from {\em BLAST}. This inspires confidence in using it to make predictions for {\em Herschel} Key Programme surveys. Our predicted source counts and
redshift distributions correspond closely with those of Valiante et al.~(2009) and Lacey et al.~(2009), respectively, despite the clear differences in the methodologies of our models. This suggests that the
combination of {\em Spitzer} mid-infrared and 850{\rm\thinspace $\mu$m}~data impose strong constraints on the allowable model parameter space. Data products from our simulation are available on the
S$^{3}$ website\footnote{http://s-cubed.physics.ox.ac.uk} and we expect them to serve as a valuable resource for the interpretation of {\em Herschel} surveys.
\section*{ACKNOWLEDGMENTS}
RJW is supported by the Square Kilometre Array Design Study. MJJ acknowledges a Research Councils UK Fellowship.
|
\section{Introduction}
The fast and precise localizations of Gamma-ray Bursts (GRBs) provided by {\em Swift}/BAT \citep{gcgmn04} in hard X-rays have enabled follow-up observations in longer wavelengths during the burst or shortly after its cessation for an increasing number of events. These observations, both from ground and from the narrow-field instruments onboard {\em Swift}, have revealed features that were not observed in the pre-{\em Swift} era. In X-rays, a canonical lightcurve \citep{nkgpg06,zfdkm06,owogp06} starts with a rapidly decaying phase that is too steep to be explained by the standard external blast wave model. This phase is often followed by a slowly decaying period, shallower than expected from the same external shock model. Only after this, the lightcurves show the normal decay and sometimes a late break, often interpreted as a jet break \citep[see e.g.][]{woogp07,lrzzb08,rlbfs09,ebpoo09}, as observed for the pre-{\em Swift} bursts. Another common but unexpected phenomena are the X-ray flares \citep{cmrfm07,fmrcm07} superimposed on the smoothly decaying segments observed for about half of the bursts. Their short time-scale variability is hard to produce externally at a large distance from the progenitor. All these features, perhaps explained by the delayed large angle internal emission \citep[e.g.][]{kpa00,lzowa06,zlz07}, continuous injection of energy \citep[e.g.][]{rm98,sm00,zfdkm06,lzz07}, delayed internal shocks or central engine activity \citep[][and references therein]{cmrfm07,fmrcm07,kgmkr09}, suggest associations with central engine properties. It is likely that the central engine is active for much longer than previously believed.
At optical wavelengths, an initial rising optical counterpart is detected for a significant fraction of bursts \citep{mvmcd07,pv08,opsdk09,kbag09,eaaab09,kgmkr09,gkmag09,gckgm09}. The temporal behaviors generally agree with predictions of the fireball forward shock model \citep{sp99, mesza06} before the onset of the shock deceleration. Alternatively, a rising afterglow can be attributed to emission in a collimated afterglow, viewed from outside the initial jet opening angle, when the shock decelerates and the relativistic beaming angle widens \citep{pv08}. A variety of lightcurve shapes, including ones with a wide peak or plateau phase, can be accommodated in this model by combination of the jet geometry and the viewing angle. Such signature of early external shock emission is not observed in X-ray, because the accompanying X-ray emission is likely obscured by the dominating internal emission. Optical observations thus provide important clues about the onset of the afterglow because emission from the external shock appears to dominate the optical emission from early times.
During the burst, optical flares apparently correlated with the hard X-ray emission are observed for several events \citep{vwwfs05,vwwag06,pwozg07,yrsrg08}. High time-resolution data plays a key role in studying the real correlation between the low and high energy emission. Such observations in optical, however, are only feasible for either a once-in-a-decade event, like the "naked eye" burst GRB 080319B \citep{rksgw08}, which was bright enough to be detected by a surveying very-wide-field optical instrument, or a bright burst lasting long enough in a usual follow-up scenario (including when the {\em Swift}/BAT is triggered on a pre-cursor). For a number of other events, an optical flare unrelated to the high energy emission is detected \citep[e.g.][]{abbbb99} or a monotonic decaying transient is observed at a flux level consistent with back-extrapolation from later observations \citep[e.g.][]{rykaa05}. These provide evidence that, at least in some cases, emission from the external shock dominates the prompt optical observations. Such diverse behaviors suggest that both internal and external shock emission may contribute to the prompt optical detections but their relative strengths vary from event to event.
GRB 081008, at a redshift of 1.967, is a typical long-lasting burst detected by {\em Swift}/BAT that provides a rare opportunity to study the optical characteristics during the prompt phase. ROTSE-IIIc started imaging only 42 s after the burst trigger and before the second epoch of major $\gamma$-ray emission. An initially rising optical transient was observed, followed by two peaks, the latter coincident with a gamma-ray peak. {\em Swift} slewed immediately, and the XRT and the UVOT started observing during the second optical peak, providing well-sampled data at early times and followed the event until it dropped below the detection threshold (see Figure~\ref{fig1}). At about 13~ks after the burst, GROND started observing the afterglow in its 7 optical/IR channels. The excellent energy coverage from IR to X-ray allows us to model the spectral energy distribution with an intrinsic broken power-law, although the evolution of the spectral shape is not restricted. The overall behavior of the observed afterglow is consistent with being achromatic, while there is some hint of a slightly steeper temporal decay in the X-rays after the break.
\begin{figure}[h]
\epsscale{1.0}
\plotone{081008_bat_xrt_opt.eps}
\caption{Multi-wavelength lightcurve of GRB 081008. The UVOT flux densities are scaled to the u-band using the normalizations determined from the simultaneous temporal fit. The ROTSE-III unfiltered data are scaled to the GROND $r^\prime$-band using the normalization from the same temporal fit. The exponential to power-law decay models \citep{woogp07} for the afterglow components are over-plotted in dashed lines. Selected epochs or time intervals for spectral analysis are labeled with corresponding indices in Figure~\ref{fig2}. \label{fig1}}
\end{figure}
In the following sections, we first summarize the {\em Swift}, ROTSE-III and GROND observations in $\S$ 2. We then present the multi-wavelength spectral and temporal analysis at selected epochs during the prompt and afterglow phases in $\S$ 3. We next discuss the interpretations of the observations in $\S$ 4 and finally conclude in $\S$ 5. Throughout the paper, we adopt the convention that the flux density can be described as $f_{\nu} \propto t^{-\alpha}\nu^{-\beta}$, where $\alpha$ and $\beta$ are the corresponding temporal and spectral power-law indices. Note that a negative $\alpha$ represents a rising lightcurve while a positive $\beta$, equal to the photon index minus one, corresponds to a spectrum that falls with increasing energy. We assume a standard cosmology model with the Hubble parameter H$_{\rm 0}$=70~km~s$^{-1}$ Mpc$^{-1}$ and the density parameters $\Omega_{\rm m}$=0.3, $\Omega_{\Lambda}$=0.7. All quoted errors are 1-sigma (68\% confidence), unless otherwise stated.
\section{Observations and Data Reduction}
\subsection{BAT}
{\em Swift} BAT was triggered by GRB 081008 on 19:58:09.4 UT \citep[Trigger 331093,][]{rbbbe08}. We refer to this time as the trigger time, T$_{\rm trigger}$. The BAT data were analyzed using the standard analysis software distributed within FTOOLS, version 6.5.1. The command $\it{batgrbproduct}$ was first run to provide the basic set of products. This analysis found the emission to be detected from T$_{\rm trigger}$-65.2~s and we therefore adopt this time as the onset of the burst T$_{\rm 0}$. Such a shift of the reference time only affects our analysis of the prompt and the very early afterglow, but we use it throughout the paper to be consistent.
Figure~\ref{fig1} shows the BAT flux density (at 71~keV) interpolated using the mean spectral index. The counts are binned with a minimum signal-to-noise ratio of 6 and maximum bin size of 20~s. The lightcurve displays a weak component decaying from T$_{\rm 0}$, two bright structured peaks at T$_{\rm 0}$+$\sim$65~s and T$_{\rm 0}$+$\sim$175~s and a fainter peak at T$_{\rm 0}$+$\sim$125~s. The emission drops below the BAT detection threshold at T$_{\rm 0}$+$\sim$270~s. The T90 (15-350~keV) of the burst is measured to be $185\pm39$~s. Spectral hardening and subsequent softening are observed during the first broad peak at T$_{\rm 0}$+$\sim$70~s. The average spectrum of the burst (from T$_{\rm 0}$ to T$_{\rm 0}$+263.3~s) is well fit by a power-law with exponential cutoff model. The best fit photon index and cutoff energy are 1.26(-0.32,+0.24) and 117(-50,+83)~keV (with 90\% confidence). At a redshift of 1.967 \citep{cfcb08g1,ddc08g2}, the isotropic equivalent energy release (0.1~keV-10~MeV) is $\sim$6.3$\times10^{52}$~erg, given a total fluence estimate of $\sim$6.7$\times10^{-6}$~erg~cm$^{-2}$.
\subsection{XRT}
{\em Swift} slewed to this burst immediately and the XRT began observing in window timing (WT) mode at T$_{\rm 0}$+152.3~s. A bright uncatalogued X-ray counterpart was identified and its position distributed via a GCN alert notice. As the count rate dropped, XRT switched to photon counting (PC) mode at T$_{\rm 0}$+476~s.
The XRT lightcurve was obtained from the {\em Swift}/XRT GRB lightcurve repository \citep{ebpto07}. The spectra were extracted from the XRT team repository, from the new products outlined in \citet{ebpoo09}. The X-ray flux densities at 2.77~keV were calculated using mean photon indices for WT and PC mode accordingly. This chosen frequency is the weighted mean of the X-ray emission assuming $\beta$=1, a typical spectral index observed, which is also similar to the afterglow spectral index of GRB 081008. The interpolations are thus not very sensitive to small changes in the photon index and the data from the two modes join smoothly. Three peaks are seen in the XRT lightcurve in Figure~\ref{fig1}. The first two have complex structures and the first peak is at a similar time as the last BAT detected pulse. The third peak is a flare superimposed on the steep decay after the second peak. After that, the lightcurve smoothly declined and steepened at T$_{\rm 0}$+$\sim$15~ks.
As observed for the first bright BAT peak, the XRT hardness ratio (1.5-10~keV/0.3-1.5~keV) curve shows a similar structure as the early lightcurve, displaying visible hardening during the flares \citep[see ][for other examples of such behavior]{brfkz05,fblck06,rmbbc06}. In general, there is also a trend of softening over time. This behavior, together with the rapid variation, indicates that the early X-ray detections are dominated by continuing burst activity, e.g. emission produced in the internal shocks. Detailed spectral analysis will be presented in $\S$ 3.
\subsection{UVOT}
UVOT began observations of the field at T$_{\rm 0}$+142.7~s. The observing sequence started with a 10~s $v$ band settling exposure and a 150~s $white$ (160-650~nm) finding chart exposure. These were followed by a 250~s $u$ band exposure, after which the filter wheel rotated through all filters, taking short 20~s exposures. One more 150~s $white$ band finding chart observation was taken at T$_{\rm 0}$+792~s. After 5~ks, longer exposures were obtained in all filters.
The optical counterpart was detected in $white$, $v$, $b$, $u$ and $uvw1$ bands, but not in uvm2 and uvw2 bands. This is consistent with the Lyman-$\alpha$ break at the measured redshift. In Figure~\ref{fig1}, all UVOT observed fluxes are scaled to $u$ band using the normalizations determined from the simultaneous temporal fit described in $\S$3.4. Initial analysis of the late time $white$ filter data shows the decay of the afterglow becomes very shallow after 100~ks. \citet{cfcb08g2} pointed out the presence of an extended source 2 arcsec away from the afterglow, which is within the 3 arcsec extraction radii used for UVOT. To measure the effect of this probable host galaxy, target-of-opportunity observations in the $white$ filter were requested on Nov.~13th. After removing the host contribution from the $white$ band photometry, no detections after 90~ks remain above 3 sigma. Such a correction is not attempted in other bands, given the negligible contribution from the underlying host galaxy during these earlier observations.
\subsection{ROTSE-III}
ROTSE-IIIc \citep{akmrs03}, a 45~cm robotic optical telescope located at the H.E.S.S. site in Namibia, responded to GRB 081008 automatically upon receiving the GCN notice. The first 5-seconds exposure started 107.1~s after T$_{\rm 0}$ (8.3~s after receiving the alert). An initially brightening optical counterpart was detected and reported in \citet{r08}.
ROTSE-IIIc took a total of 10$\times$5~s, 10$\times$20~s and 144$\times$60~s exposures before the target elevation dropped below the observing limit. The first ten 5-seconds exposures were taken in sub-frame mode with 3 seconds readout gap inbetween, instead of the 6 seconds full-frame readout time for later images. We took the first 70 single images and co-added the later images into sums of 5 to 10 depending on the image quality. The ROTSE-IIIc images were bias-subtracted and flat-fielded by the automated pipeline. Initial object detections were performed by SExtractor \citep{ba96}. The images were then processed with our custom RPHOT photometry program based on the DAOPHOT \citep{stetson87} PSF-fitting photometry package \citep{qryaa06}. A reference image was constructed from images taken on October 18th when the OT had faded below our detection limit.
Due to ROTSE-III's large pixel scale (3.4 arcsec), the optical counterpart is slightly blended with four surrounding objects in the ROTSE-IIIc images. To remove the possible contamination, we tried several different methods. First, we used the RPHOT functionality to subtract four PSF-scaled point sources around the OT. The lightcurve did appear to be steeper than obtained without subtraction. The two objects north of the OT are blended sources themselves but not resolved by ROTSE-IIIc. They may not be completely removed using the point source approximation, but the residual should be negligible compared to the bright OT. To confirm this, we attempted an alternative subtraction method described in \citet{ya08}. Each burst image was cross-convolved with the reference to a common PSF and the difference between the two was taken. The brightness of the OT in the difference images were then measured by matching their PSF to the stars in the convoluted reference images. The results were consistent with the RPHOT subtraction method within the uncertainties. We report here the RPHOT PSF-fitting results with the 4 contaminating sources removed. Given the above comparison, the systematic error from subtraction is smaller than the statistical errors and is not included in the reported uncertainties.
The unfiltered thinned ROTSE-III CCDs have a peak sensitivity in the $R$ band wavelength range. We thus calibrate the zero point magnitudes using the median offset from the USNO B1.0 $R$ band measurements of selected field stars. In Figure~\ref{fig1}, the ROTSE-III flux densities are scaled to match the GROND $r^\prime$-band measurements using the offsets from the simultaneous temporal fit described in $\S$3.4.
\subsection{GROND}
The Gamma-Ray Burst Optical/NearInfrared Detector (GROND, \citealp{gbcdh08}) mounted at the 2.2~m ESO/MPI telescope at LaSilla observatory (Chile) imaged the field of GRB 081008 simultaneously in $g^\prime r^\prime i^\prime z^\prime JHK$ starting 13.6 ks after T$_{\rm 0}$, after local sunset under clear sky conditions. A total of 28 $g^\prime r^\prime i^\prime z^\prime$ images with integration times of 66~s, 115~s and 375~s were obtained during the first night post burst. At the same time, the NIR channels were operated with a constant exposure of 10~s. In addition, the field was observed with GROND at day 2, 5 and 10 after the trigger. Data reduction and relative photometry was carried out using standard IRAF tasks \citep{tody93}, similar to the procedure outlined in \citet{kgmkr09}. To exclude a possible contamination of the measured afterglow brightness with light from the host 2.2$\arcsec$ away, the results from the PSF photometry were checked against measurements taken using small apertures (0.3 and 0.5$\times$FWHM) where an aperture correction was applied. Image subtraction, using the frame taken 10 days after the burst as reference, further confirmed that the contamination is negligible. Absolute photometry was measured with respect to the primary Sloan standards SA107-1006 and SA112-805 in $g^\prime r^\prime i^\prime z^\prime$ \citep{stkrf02}, observed under photometric conditions and 2MASS field stars in $JHK$ \citep{scsws06}.
The position of the afterglow is determined as R.A. = $18^{h}39^{m}49^{s}.88$ Decl. = $-57^{\circ}25^{'}52^{"}.8$ (J2000), with a 90\% uncertainty of 0.3 arcsec. We adopt this as the best burst localization. Along the line of sight to this position, the Galactic extinction is E(B-V)=0.095 \citep{sfd98}, and the Galactic column density is $7.1\times10^{20}~{\rm cm}^{-2}$ \citep{kbhab05}.
\section{Spectral and Temporal analysis}
\subsection{Prompt Spectral Evolution}
We first examine the spectral evolution of the burst. Available BAT and/or XRT data are extracted in three time intervals, T$_{\rm 0}$+55~s to T$_{\rm 0}$+85~s, T$_{\rm 0}$+161~s to T$_{\rm 0}$+195~s and T$_{\rm 0}$+245~s to T$_{\rm 0}$+270~s (as marked in Figure~\ref{fig1}), centered on the two broad peaks detected by BAT and the one X-ray peak shortly afterward. Because of the strong hardness ratio evolution, we limited the width of the chosen time intervals. The spectra were modeled with either a power-law or a power-law with exponential cutoff. When X-ray data were available, we included in the model a fixed Galactic absorption component and an additional variable system at the host redshift. The fit parameters for the best models are listed in Table~\ref{burstspec}. We tried modeling the spectra including a high energy power-law tail using a Band function \citep{bmfsp93}, but the power-law index above the cut-off energy is not well constrained by our data. The best fit spectral models at each epoch are plotted (with labeled offset for clarity) in Figure~\ref{fig2} as solid lines.
\begin{figure}[h]
\epsscale{1.0}
\plotone{081008_sed_model_offset.eps}
\caption{Spectral models in the selected time intervals. Interval 1, 2 and 3 (offset for clarity) correspond to the first 3 prompt peaks. Interval 4 corresponds to the bright X-ray flare seen by the XRT and 5 is the X-ray afterglow. Optical fluxes are corrected for Galactic extinction, but are not included in the model fit in intervals 1 to 4, because the optical fluxes are likely dominated by the external forward shock and thus not related to the high energy emission. We include these points as upper limits to constrain the internal emission component in the optical. The dashed lines, connecting the data points in the X-ray and the optical regions in the same interval, do not represent the actual spectral shape. During interval 5, the emission is entirely from the afterglow and the spectrum is plotted in dashed-dot line.\label{fig2}}
\end{figure}
\begin{deluxetable}{cccccccc}
\tablecaption{Best-fit spectral parameters to the burst and the flares \label{burstspec}}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.18in}
\tablehead{\colhead{Interval} & \colhead{T$_{\rm range}$-T$_{\rm 0}$} & \colhead{Host nH} & \colhead{$\beta$} & \colhead{$E_{cut}$} & \colhead{$\chi^2/dof$} & \colhead{Instrument} \\
\colhead{} & \colhead{(s)} & \colhead{($10^{22}~{\rm cm}^{-2}$)} & \colhead{} & \colhead{(keV)} & \colhead{} & \colhead{} }
\startdata
1 & 55 - 85 & \nodata & $0.33\pm0.07$ & \nodata & 58/56 & BAT \\
2 & 161 - 195 & $1.06_{-0.27}^{+0.32}$ & $0.26_{-0.05}^{+0.04}$ & $63.2_{-9.8}^{+13.3}$ & 201/184 & BAT \& XRT \\
3 & 240 - 275 & $1.01_{-0.26}^{+0.29}$ & $0.58_{-0.14}^{+0.12}$ & $14.9_{-5.3}^{+8.5}$ & 200/194 & BAT \& XRT \\
4 & 345 - 385 & $1.77_{-0.32}^{+0.38}$ & $1.51_{-0.12}^{+0.13}$ & \nodata & 97/75 & XRT \\
\enddata
\tablecomments{All errors quoted correspond to 90\% confidence intervals.}
\end{deluxetable}
In both interval 2 and 3, the high energy spectra are well described by a power-law with exponential cutoff model where the cutoff energy drops from above 60~keV in interval 2, to 15~keV in interval 3, and the photon index softens. The measured local column density (1.06(-0.27,+0.32)$\times10^{22}~{\rm cm}^{-2}$ and 1.01(-0.26,+0.29)$\times10^{22}~{\rm cm}^{-2}$) shows good agreement between intervals. During interval 1, the BAT data are fit by a power-law model with a relatively hard photon index ($\Gamma\sim$1.33). Given the general trend of softening, it is likely that the cutoff energy at this early epoch is close to or above 150~keV and thus not well constrained by the BAT energy coverage. \citet{ssbbc09} has demonstrated that a reasonable estimate of the peak energy (in ${\nu}f_{\nu}$ space) can be achieved using the power-law photon index in the BAT energy range if the latter is between 1.3 and 2.3. Our value of 1.33 for the 1st interval is just inside this range and corresponds to a peak energy of 196(-47,+587)~keV, assuming a cutoff power-law model as the actual spectral form. For the later two intervals, the peak energies can be explicitly calculated as $(1-\beta)E_{cut}$, which are $\sim$47~keV and $\sim$6~keV respectively. These numbers agree with the estimates following the method in \citet{ssbbc09}, although the BAT-alone power-law photon index in the 3rd interval is not well constrained and exceeds 2.3.
During the 2nd and 3rd intervals, optical data are also available, but they are not included in the spectral fit for two reasons. First, emission from a different origin (e.g. external forward shock) is likely dominating the optical data (see $\S$3.3 and $\S$3.4). Second, only data in the $white$ filter (or unfiltered) is available, thus any optical deficit below the extrapolation from high energy observations can be compensated by an arbitrary value of local extinction, while any optical excess is hard to accommodate within the chosen models. Instead, we plot the $white$ band fluxes (corrected for Galactic extinction and plotted with the same offsets as the corresponding high energy spectra) in Figure~\ref{fig2} and connect them directly to their high energy counterparts with dashed lines, just to show how they compare. The optical detections are considered as upper limits to constrain the internal emission component. If internal emission has contributed to at least part of the optical observations, they do not exceed the extrapolations from XRT and BAT detections at any time.
\subsection{X-ray Flare}
A bright single X-ray flare is superimposed on the rapid decay phase after the second broad X-ray peak. It has a similar hardness ratio evolution as the earlier peaks. It also shows a fast-rise-exponential-decay (FRED) profile that is typical of a GRB pulse. Its rapid rise and decay supports an identification with extended central engine activity. We thus analyze the flare within the same frame as the prompt emission.
We extract the X-ray spectrum from T$_{\rm 0}$+345~s to T$_{\rm 0}$+385~s (noted as interval 4), centering on the flare. UVOT $v$ band and ROTSE-III detections are available during this period, but there is no sign of a flare in the optical indicating that the emission is dominated by the afterglow. So we fit the XRT data alone. The spectrum is reasonably fit by a single power-law model with a photon index of 2.5 (see Table~\ref{burstspec}), which is much softer than the prompt emission and agrees with the softening trend. The inferred column density is somewhat higher than those from the earlier times, probably caused by curvature in the real spectral shape. This would suggest that $E_{peak}$ is near or within the XRT energy range at this time.
\subsection{Early Optical Afterglow}
The ROTSE-III observations show an initially rising optical transient with two peaks at T$_{\rm 0}+\sim$135~s and T$_{\rm 0}+\sim$170~s (see Figure~\ref{fig1} and \ref{fig3}). The later one is almost simultaneous with the BAT and XRT peaks and may be attributed to a related prompt optical flare. The first one, however, does not appear to be correlated with the high energy emission. As shown in $\S$3.4, the overall optical lightcurve is well-fit by one dominating component. We therefore interpret this first peak in terms of external shock emission. A detailed discussion of this interpretation will be presented in $\S$4.1.
\begin{figure}[h]
\epsscale{1.0}
\plotone{081008_prompt_2.eps}
\caption{Contemporaneous BAT, XRT and optical (ROTSE-III and UVOT $white$) observations. Time bins for the BAT and XRT data are determined by Bayesian block analysis (see text in $\S$3.3). The high energy data points are connected by dotted lines as guidance to the eye. The UVOT event data is binned in 10s intervals. The smooth rise and decay model is plotted over the optical to show that no significant deviation is detected. \label{fig3}}
\end{figure}
\begin{deluxetable}{cccccc}
\tablecaption{Afterglow temporal best-fit parameters \label{tempfit}}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.18in}
\tablehead{\colhead{Energy Range} & \colhead{Rise time} & \colhead{Start of Decay} & \colhead{$\alpha_{1}$} & \colhead{T$_{\rm break}$} & \colhead{$\alpha_{2}$} \\
\colhead{} & \colhead{$t_a$(s)} & \colhead{$T_a$(s)} & \colhead{}& \colhead{(s)} & \colhead{} }
\startdata
X-ray & 130\tablenotemark{*} & 1800\tablenotemark{**} & $0.96\pm0.10$ & $15870\pm8623$ & $1.78\pm0.25$ \\
Optical/IR & 130\tablenotemark{*} & $136\pm18$ & $0.96\pm0.03$ & $7834\pm2160$ & $1.313\pm0.048$ \\
X-ray \& Optical/IR \tablenotemark{***} & \nodata & \nodata & $0.81\pm0.07$ & $15139\pm8150$ & $1.71\pm0.13$ \\
\enddata
\tablenotetext{*}{Fixed parameters, to a value estimated from the optical fit.}
\tablenotetext{**}{The plateau is hidden by the prompt emission, thus no error is estimated.}
\tablenotetext{***}{Only data after T$_{\rm 0}$+750s are fit by a smoothly joined broken power-law model, with a smooth factor, s=1.}
\tablecomments{All errors quoted correspond to 90\% confidence intervals.}
\end{deluxetable}
Figure~\ref{fig3} shows a close look at the BAT, XRT and optical lightcurves before T$_{\rm 0}$+345~s. The BAT and XRT data are binned in intervals determined by Bayesian block analysis \citep{s98}. For BAT, this analysis is carried out by $\it{battblock}$ on the 1~s-binned lightcurve data. For XRT, only WT data are used and the background ($\sim$1~ct/s) is negligible compared to the bright signal. We thus adapted the algorithm used in $\it{battblock}$ for event files and consider only the change points resulting in time bins longer than half a second to ensure comparable signal to noise ratios. The high energy lightcurves show significant variations on a similar time-scale; while no convincing short-time-scale variability is detected in optical. This is not surprising though, as the optical curve connects smoothly into a monotonic decline afterward.
\subsection{Multi-wavelength Afterglow}
We adopt the two component exponential to power-law decay model in \citet[][Eq~(2) and descriptions within]{woogp07} to characterize the temporal flux evolution in X-ray and optical. In this model, the first component corresponds to the prompt emission, and the second relates to the afterglow emission. All optical observations are sufficiently fit by one afterglow component. In the X-rays, the second component is not revealed until the end of the last flare. So the rise time ($t_a$) of this second component is fixed to 130~s according to the optical fit. The best fit models for the afterglow components are over-plotted in Figure~\ref{fig1} and the parameters are tabulated in Table~\ref{tempfit}.
For both optical and X-ray, a later temporal break is required in the model. The break times in the two energy ranges (15870$\pm$8623~s and 7834$\pm$2160~s) are consistent at 90\% confidence, but the final decay ($\alpha_2$ in Table~\ref{tempfit}) in the X-ray is significantly faster than in the optical. We further attempted fitting all energy bands after T$_{\rm 0}$+750~s simultaneously with a smoothly joined broken power-law model (f(t)=$a[(t/t_b)^{-s\alpha_1}+(t/t_b)^{-s\alpha_2}]^{1/s}$), with a small smoothness factor, s$\sim$1. Although the overall fit is acceptable, giving a reduced Chi-squared of 1.17 for 239 degrees of freedom (see Table~\ref{tempfit} for the best fit parameters), the fit parameters are dominantly determined by the well-confined optical data. Some systematic deviations from the model, not reflected in the Chi-squared statistics, are noticed in the X-rays. Between T$_{\rm 0}$+7~ks and T$_{\rm 0}$+28~ks, the X-ray points are mostly above the model, and after T$_{\rm 0}$+30~ks, the X-ray data are all below the model. Our data are thus consistent with displaying a smooth achromatic temporal break, but a chromatic behavior is not confidently ruled out in case of a subtle evolution.
We then examine the spectral energy distributions (SEDs) in selected afterglow epochs. In the X-rays, no significant spectral or hardness ratio changes can be detected. In the optical, the UVOT detections before the break and the GROND observations after the break are poorly fit by one power-law spectral index plus the same amount of local extinction. This is caused either by spectral evolution or a spectral break at the red (blue) end of the UVOT (GROND) coverage, at around 2.5~eV. This latter conjecture is confirmed by further analysis.
\begin{deluxetable}{ccccccccc}
\tablecaption{Best-fit afterglow spectral parameters \label{afterglowspec}}
\tabletypesize{\scriptsize}
\setlength{\tabcolsep}{0.03in}
\tablehead{\colhead{T-T$_{\rm 0}$} & \colhead{Energy Range} & \colhead{Host Type} & \colhead{Host Extinction} & \colhead{nH} & \colhead{$\beta_{opt}$} & \colhead{$E_{break}$} & \colhead{$\beta_{x}$}& \colhead{$\chi^2/dof$} \\
\colhead{(ks)} & \colhead{} & \colhead{} & \colhead{E(B-V)} & \colhead{($10^{22}~{\rm cm}^{-2}$)} & \colhead{} & \colhead{(eV)}& \colhead{} & \colhead{}}
\startdata
1.5 & XRT only & \nodata & \nodata & $0.47_{-0.33}^{+0.36}$ & \nodata & \nodata & $1.06_{-0.13}^{+0.14}$ & 46/43 \\
1.5 & XRT \& UVOT & MW & $0.15_{-0.03}^{+0.04}$ & $0.62_{-0.23}^{+0.25}$ & $1.14\pm0.05$ & \nodata & $1.14\pm0.05$ & 52/46 \\
1.5 & XRT \& UVOT & LMC & $0.11_{-0.03}^{+0.02}$ & $0.58_{-0.22}^{+0.25}$ & $1.12\pm0.04$ & \nodata & $1.12\pm0.04$ & 53/46 \\
1.5 & XRT \& UVOT & SMC & $0.08\pm0.02$ & $0.54_{-0.22}^{+0.24}$ & $1.10\pm0.04$ & \nodata & $1.10\pm0.04$ & 55/46 \\
29.5 & XRT, UVOT \& GROND & MW & $0.18_{-0.07}^{+0.09}$ & $0.97_{-0.60}^{+0.85}$ & $-0.13_{-0.41}^{+0.30}$ & $2.4\pm0.2$ & $1.12_{-0.08}^{+0.11}$ & 11/14 \\
1.5 \& 29.5 & XRT, UVOT \& GROND & MW & $0.16_{-0.03}^{+0.04}$ & $0.65_{-0.21}^{+0.24}$ & $-0.06_{-0.18}^{+0.17}$ & $2.4_{-0.1}^{+0.2}$ & $1.14\pm0.05$ & 72/63 \\
\enddata
\tablecomments{All errors quoted correspond to 90\% confidence intervals.}
\end{deluxetable}
The overall shapes of the SEDs are modeled from optical/IR to X-ray with all available data before and after the break. Assuming they belong to the same synchrotron emission component, we allow either one or no spectral break within the energy range. Before the break, we extract an X-ray spectrum in the time interval from T$_{\rm 0}$+765~s to T$_{\rm 0}$+1778~s (first orbit) and from T$_{\rm 0}$+5286~s to T$_{\rm 0}$+7562~s (second orbit) to maximize the signal to noise ratio, given that no significant spectral or hardness ratio changes is detected in the X-rays. This spectrum is scaled to the X-ray count rate at T$_{\rm 0}$+1.5~ks to minimize the interpolation of the UVOT data. The UVOT mean count rates at T$_{\rm 0}$+1.5~ks are estimated by fitting the nearby data points (within 500~s on each side) in each band with power-laws. The errors are estimated using a bootstrap approach, where each data point is randomly re-sampled from a Gaussian distribution centered on the observed count rate with a sigma equal to the measured error. The UVOT count rates are converted to XSPEC compatible spectral files with the corresponding response matrices downloaded from the HEASARC Calibration Database (version 20041120v104).
The SED can be fit by a single power-law model with fixed Galactic dust and gas absorption and additional dust and gas absorption at the host redshift (see Table~\ref{afterglowspec}). This model is supported by the well-matched initial temporal power-law decay indices ($\alpha_1\sim$0.96) in the optical and the X-ray. The relatively small amount of host extinction required and the lack of coverage at the rest-frame 2175$\mbox{\AA}$ Galactic absorption feature do not allow us to distinguish between the Milky Way (MW), Large Magellanic Cloud (LMC) or Small Magellanic Cloud (SMC) environment.
After the break, another X-ray spectrum is constructed between T$_{\rm 0}$+16~ks and the last detection at T$_{\rm 0}$+400~ks. This spectrum alone has a relatively low signal to noise. We scale this X-ray spectrum to the latest epoch where the afterglow was detected in the $J$, $H$ and $K$ bands (at T$_{\rm 0}$+$\sim$29.5~ks) and convert the contemporaneous GROND detections to XSPEC compatible files. The UVOT $u$ and $b$ band rates are also interpolated to this epoch by fitting the neighboring two detections (about 30~ks apart) with power-laws as described above.
A single power law model does not produce an acceptable fit to the overall SED, regardless of the choice of local dust and gas absorption. At least one spectral break is required between X-ray and optical. Our best fit model (see Table~\ref{afterglowspec}) constrains a break at $\sim$2.4~eV, in the UVOT v-band. Such a low break frequency is not well covered by the earlier UVOT data, and the two SEDs thus have consistent shapes. A MW extinction curve is used because it provides a best fit and is marginally preferred during the earlier epoch. Nevertheless, we do not consider this as a conclusive property of the environment because of the possible uncertainty in our selected intrinsic spectral shape, e.g. the smoothness of the break is not considered in our model.
Figure~\ref{fig4} shows a simultaneous fit of the two SEDs with a broken power-law model with Galactic dust and gas absorption and dust and gas absorption at the host redshift. We tie the dust (following a MW extinction law) and gas properties for a consistent environment throughout the afterglow phase. The fit is fairly good except in the $uvw1$ band. A plausible explanation for the under-estimate in $uvw1$ is that our model applies a sharp cutoff above the Lyman-break which might not be true. The best fit model parameters are consistent within errors with the estimates obtained in narrower energy bands and smaller time coverages as described above. Also, the column density derived is consistent with the estimate during prompt interval 2 and 3 (see Table~\ref{burstspec}).
\begin{figure}[h]
\epsscale{1.0}
\plotone{081008_afterglow_sed.eps}
\caption{Combined XRT, UVOT and GROND spectral energy analysis during the afterglow (epoch 5 and 6). The local extinction and absorption properties are tied. The fit parameters are listed in Table~\ref{afterglowspec}. The absorbed models are over-plotted in the upper panel. Dotted lines show the intrinsic power-law and broken-power-law components. The location of a spectral break is indicated by an upward arrow. The residual ratios are plotted in the bottom panels. \label{fig4}}
\end{figure}
In summary, the observed optical and X-ray afterglow emission started off following similar temporal decay rates ($\alpha_{1}\sim$0.96) and displayed smooth temporal breaks at similar times (at $\sim$15~ks). While the two SEDs before and after the break were characterized intrinsically by a power-law model and a broken power-law model respectively, they had consistent shapes because the low break frequency was not well covered during the first epoch. Besides, the optical/IR observations by GROND between T$_{\rm 0}$+13~ks and T$_{\rm 0}$+30~ks do not show evidence of a moving spectral break. The overall afterglow behavior was consistent with being achromatic. Nonetheless, there is some hint of a slightly steeper decay in the X-rays after the break. Because the afterglow was not observed in IR and near-IR at early times, probable evolution of the spectral break is not constrained, e.g. it could have increased over time from below optical to the observed 2.4~eV.
\section{Discussion}
\subsection{Early Optical Afterglow}
\subsubsection{Reverse or Forward Shock}
For the initial optical observations, fixing the beginning of the external shock to the burst time T$_{\rm 0}$ yields a temporal rise index of $\alpha = -5.0\pm0.6$. This is too rapid to be considered as the pre-deceleration phase of the forward shock \citep[$t^2$ or $t^3$ in ISM, slower than $t^{1/2}$ in Wind medium, see][]{pv08}, but agrees well with the prediction for a reverse shock rising in a thin shell case when the burst duration is less than the deceleration time \citep{k00,zkm03}. Back-extrapolation from late time afterglow observation shows that the reverse shock and forward shock must have comparable peak luminosities. The analysis of the later afterglow emission suggests the forward shock characteristic synchrotron frequency ($\nu_m$) may have passed below the optical by T$_{\rm 0}$+200~s. If the micro-physical parameters are comparable in the reverse and forward shock regions, the reverse shock should have a much lower characteristic synchrotron frequency \citep{sp99} and thus have a very weak optical emission. The detection of reverse shock emission would then point to a more magnetized reverse shock region \citep{zkm03,kp03,jf07}.
However, there are several questions to be asked for this interpretation. First, is the thin shell assumption correct? Although the overall prompt BAT and XRT detections extend to after the beginning of the optical detections, the second epoch of strong burst activity emerged once the initial rapid rise in the optical had stopped. The temporal and spectral similarity between the flares indicate that the second epoch of high energy emission was from the same emitting region as the earlier bursts of emission, i.e. closer to the central engine than external medium. This later part of the ejecta would then run into the external medium after the initial shock has been established. We thus can ignore the effect of this later injected kinetic energy when considering the early rising lightcurve. Suppose the bulk of kinetic energy was injected into the external medium by the materials producing the first flare ending at T$_{\rm 0}$+100~s, the shock can be reasonably considered impulsive.
This leads to the second question: whether the onset of the afterglow is chosen correctly as T$_{\rm 0}$? \citet{qryaa06} have pointed out that different choices of time origin may lead to different interpretations of the early data \citep[also see ][for a discussion in the X-rays]{tgcmc05}. \citet{lb06} explored how varying central engine activity would affect the self-similar evolution in the early afterglow and its lightcurve. Their simulation is limited to a power-law energy release, but gives a good schematic picture. They found that if the burst energy release is concentrated toward the end of the burst duration, the self-similar evolution would deviate from one simple power-law behavior. For GRB 081008, the emission at T$_{\rm 0}$+65~s is about one order of magnitude higher than at T$_{\rm 0}$, it is therefore proper to choose a delayed time origin in a power-law fit. Such a small time shift has a negligible effect on the temporal index inferred for the later smooth decay. But if we choose the reference time to be T$_{\rm 0}$+45~s, the initial rise phase would have a temporal index of $-3.1\pm0.4$, fully consistent with the pre-deceleration of the forward shock. For any choice of reference between T$_{\rm 0}$+25~s and T$_{\rm 0}$+62~s (including the beginning of the first big pulse detected by BAT), the estimated temporal power-law index is consistent with -3 within 2~sigma. Given the earlier discussion about the reverse shock and the fact that the lightcurve is well fit by one smooth rise and fall component, we prefer the external forward shock interpretation for this rising lightcurve observed by ROTSE-III.
Assuming the first emission peak in the optical marks the onset of the forward shock deceleration, we can apply the adjusted rise time (T$_{\rm 0}$+135~s~-~T$_{\rm 0}$+45~s) of 90~s and the energy released before T$_{\rm 0}$+135~s (see $\S$4.1.2) into Eq (1) in \citet{mvmcd07} to obtain a Lorentz factor of $\Gamma_{\rm dec}\sim$225 at the deceleration radius and an initial Lorentz factor $\Gamma_{\rm 0}$ of $\sim$450 \citep[$\sim2\Gamma_{\rm dec}$;][]{mesza06}. We have assumed a particle density n=1~${\rm cm}^{-3}$ and a radiative efficiency $\mu$=0.2. The deceleration radius is then estimated as ${\sim}2{\Gamma}^{2}ct{\sim}3\times10^{17}$~cm.
\subsubsection{Refreshed Shock?}
After the forward shock start decelerating, the passing of the synchrotron frequency and/or cooling frequency across the optical band may produce temporal breaks. But a second peak, as observed for GRB 081008, is hard to explain in a model involving only one external forward shock component. The temporal coincidence with a high energy flare suggests possible contribution from internal emission. This scenario is hard to rule out given the slight optical excess ($\sim$4~sigma) above the afterglow model seen in Figure~\ref{fig3}. However, the lightcurve also shows a probable dip below the smooth, one component model, and the decay after the second peak joins smoothly into the later phase. We thus explore other interpretations for the double peaked lightcurve.
In a reasonable contemplation in the frame of the fireball model, the shells of material producing the late burst activity would reach the external medium and catch up with the early decelerated shells at some moment and re-energize the forward shock. This later scenario results in an irregular ``step'' structure, as additional energy is injected over a time period comparable or longer than the second epoch of burst activity. It has been noted that refreshed shocks from slow or late shells may alter the shape of the afterglow lightcurves \citep{rm98,kp00}. It is, however, hard to directly relate such structure with the internal emission as they are produced by different mechanisms at different regions. Here, we provide a simplest possible test on this speculation.
If a second batch of kinetic energy is injected into the forward shock, a new self-similar solution would be established sometime after the cessation of all burst activity. Indeed, we didn't observe any apparent deviation from a power-law decay after T$_{\rm 0}$+200~s. We try to estimate the synchrotron emission before and after the refreshed shocks. Without the additional energy injection, we can assume the afterglow would decay monotonically after the first peak at the same rate as observed in the later afterglow. At a time after T$_{\rm 0}$+200~s, the observed flux is 50-60\% higher than this extrapolation, including uncertainties from the flux measurement at the first peak and the normalization of the later decaying lightcurve.
We then estimate the energy release in the prompt emission before and after T$_{\rm 0}$+135~s (the first optical peak). We estimate the fluence between 0.1~keV and 10~MeV in the rest frame. In the observer's frame, the 0.1~keV lower bound fits nicely within the XRT sensitivity range. Before T$_{\rm 0}$+135~s, only BAT observations are available and the average spectrum can be fit by a power-law with exponential cutoff model. The cutoff energy at 157~keV, although above the BAT energy threshold, agrees well with calculation using the power-law photon index following \citet{ssbbc09}. We estimate the fluence during this period to be $\sim4.7\times10^{-6}$~erg~cm$^{-2}$. The hard X-ray emission dropped below the BAT detection threshold at T$_{\rm 0}$+$\sim$265~s, but it is likely that the central engine activity continued with softer emission detected mainly in soft X-rays. Between T$_{\rm 0}$+155~s and T$_{\rm 0}$+210~s, simultaneous XRT and BAT detections constrain the spectral shape well, and using a power-law model with an exponential cutoff, we estimate a total fluence of $\sim2.5\times10^{-6}$~erg~cm$^{-2}$. After this, the peak energy shifts further toward lower energies and the fluence in the next 100s is only one fourth of the previous interval (with cutoff energy at $\sim$10~keV). The additional energy release between T$_{\rm 0}$+135~s and T$_{\rm 0}$+155~s is even smaller. From T$_{\rm 0}$+135~s to T$_{\rm 0}$+$\sim$310~s, we estimate a total prompt fluence of $\sim3.4\times10^{-6}$~erg~cm$^{-2}$. This is about 70\% of the total energy released before T$_{\rm 0}$+135~s, and consistent with the flux increase calculated within the large uncertainties. Note that the overall energy output is about 20\% higher than the estimate in $\S$2.1 using an average gamma-ray spectrum and the BAT detected burst duration.
The X-ray flare at T$_{\rm 0}$+$\sim$400~s is likely generated by the same internal mechanism. We can estimate its emission in the same energy range as above. The spectrum is so soft (with photon index greater than 2) that we can directly integrate it in the high energy range. No extrapolation below the XRT energy is necessary with our choice of energy range, so we do not have to worry about a spectral break below 0.3~keV. The ``bolometric'' fluence calculated is two orders of magnitude lower than the previous prompt emission. Therefore, if the contribution to the forward shock is proportional to the prompt luminosity, the change in the forward shock emission is comparable to the error bars in the optical measurements and too small to be noticed.
A caveat for this calculation is the assumption that the relative radiative efficiency in the internal and external processes is constant although the relationship between the two is not straightforward. During the burst, the peak emission frequency down-shifted from $\gamma$-ray to X-ray. The X-ray emission could be from slower or delayed ejecta, or alternatively from collisions with a smaller Lorentz factor dispersion and thus lower efficiency in converting the kinetic energy. At the forward shock front, the conversion efficiency may depend on the condition of the previously shocked medium and therefore requires a detailed simulation to determine. Another overlooked factor is the evolution of the synchrotron spectral shape, as the flux is estimated in a narrow optical band.
\citet{vwwag06} showed that the early optical lightcurve of GRB 050820A can be modeled with a prompt component tracing the $\gamma$-ray emission and two afterglow components. The afterglow components, correspondingly produced by the precursor and the main event ejecta, have consistent ratios between the prompt fluence and afterglow amplitude. This agrees with our finding although their first afterglow component is faint and not well-sampled for a detailed study. A similar double peaked prompt lightcurve is observed for GRB 061007 \citep{eaaab09}. The overlapping high energy pulses indicate continuous energy input into the external shock and thus modeling of the early lightcurve in a thick shell case may be more appropriate. For that event, the initial very rapid rise of the optical emission is also problematic in the current forward shock model.
\subsection{Late Afterglow Evolution}
Before the temporal break and above $\sim$2.4~eV, optical and X-ray appear on the same segment of a synchrotron spectrum, we thus use the method described in \citet{rlbfs09} to find all scenarios consistent with the observed temporal and spectral indices. The temporal decay ($\alpha$=0.96$\pm$0.03) marginally agrees with the value expected ((3$\beta$-1)/2$\sim$1.2, for $\beta\sim$1.1) for above the cooling and characteristic synchrotron frequencies, regardless of slow or fast cooling regime and the external medium density profile. In such a scenario, both the cooling frequency and the characteristic frequency are below or close to the optical by T$_{\rm 0}$+200~s.
For all the other closure relations considered, the temporal decay is shallower than predicted if no energy injection is present. In terms of the canonical X-ray lightcurve \citep{nkgpg06,zfdkm06,owogp06}, this decay index is within the range observed for the shallow decay phase (phase II). This phase, detected in a significant fraction of the bursts, is generally explained as due to prolonged energy injection. However, the source of energy is unclear as well as the mechanism for fine tuning the input energy into the forward shock smoothly over such long period of time.
We apply the evolution constraints in \citet{rlbfs09} to the X-ray data and find the cause of the break is consistent with either a jet geometry or the cessation of energy injection. For both types of break, achromatic behavior extending to the optical range is expected, consistent with our observations.
After the temporal break, our data tentatively suggest a spectral break around 2.4~eV, although it is hard to constrain the exact spectral shape and the environment properties at the same time. Such a break is hard to be accommodated in the standard fireball model. A synchrotron spectrum is characterized by three possible break frequencies, a cooling frequency ($\nu_c$), a characteristic frequency ($\nu_m$) and a self-absorption frequency ($\nu_a$) \citep[e.g.][]{sp99,gs02}. All three frequencies should evolve over time. The observed break at 2.4~eV is too high to be a self-absorption frequency. Given a spectral index below the break (-0.13\{-0.41,+0.30\}) consistent with -1/3, the shape of the spectrum can be expected in a slowing cooling regime ($\nu_m<\nu_c$) around $\nu_m$ \citep{gs02}. However, $\nu_m$ should always decease over time as the forward shock decelerates, unless the shock is reenergized by significant late-time injection, which does not seem to be the case here. We do not observe the expected rise or flat evolution in IR (below the break) and do not detect a higher break frequency in our earlier SED. Finally, our data do not rule out an increasing break frequency, as expected for a cooling break in a wind medium \citep{cl00}, but the spectral index difference across the break ($\sim$1) is much larger than the predicted amount of 0.5.
If some chromatic effect exists around the temporal break, the passage of a cooling frequency in a wind medium produces the correct trend of a break moving from optical to X-ray. The expected temporal index change of 0.25 is similar to the observed value in the optical ($\sim$0.3) but not enough to account for the amount of change detected in the X-rays ($\sim$0.8). In any case, the temporal break at $T_{\rm 0}$+15~ks is not likely to be caused by the movement of the observed spectral break, because there is no evidence that the break frequency has yet reached the X-ray regime by $T_{\rm 0}$+29.5ks, as would be expected in such a case.
In short, the observed temporal break is consistent with being a jet break or the cessation of energy injection. Before the break, the multi-wavelength observations are best interpreted by the cooling and characteristic synchrotron frequency being below the optical. No satisfactory explanation is found for the later spectral shape and its probable evolution. It is viable to assume that the spectral break has increased over time, as the cooling frequency in a wind medium, but such evolution is not likely to be responsible for the observed temporal break.
\subsection{Burst Energetics}
If the temporal break at $T_{\rm 0}$+$\sim$15~ks is interpreted as a jet break, the jet opening angle can be estimated as $2.1^\circ$ or $2.4^\circ$ for ISM \citep{sph99} or a wind medium \citep{bfk03} respectively. In the calculations, we have assumed a particle density n=1${\rm cm}^{-3}$, a characteristic wind density A$_{*}$=1, a radiative efficiency $\mu$=0.2 and a kinetic energy $E_{k}=E_{\rm iso}/\mu$. The collimation corrected total gamma-ray energy release (E$_\gamma$) is hence $4.2\times10^{49}$~erg and $5.5\times10^{49}$~erg for ISM and wind medium respectively. This number would be slightly larger if early X-ray detections are also considered (see $\S$4.1.2).
For a jet break, the change of the temporal index ($\sim$0.8) favors an ISM-like medium \citep{pana05}. Both the inferred opening angle and the collimated energy output are on the small side but well within the distribution for {\em Swift} detected GRBs \citep[e.g.][]{lrzzb08,rlbfs09}. The nature of the break is, however, uncertain given the previous discussions on a probable secondary chromatic effect. A later jet break, occurring close to or shortly after the afterglow dropped below the detection threshold of all our instruments ($\sim$400~ks), would also give a sensible estimate of the opening angle and E$_\gamma$.
\section{Conclusions}
With energetics typical among {\em Swift} detected events, GRB 081008 happens to last long enough to have overlapping phases with the early follow-up observations in X-ray and optical. It is well behaved in the sense that all early canonical phases in X-ray and optical were detected. A key feature of the high energy emission is the hardness ratio evolution in accordance with the flux fluctuation throughout the burst and the X-ray flares. The temporal continuity and spectral softening suggest that early X-ray observations originate from extended internal activities, but either with a gradually decreasing efficiency or a declining energy output from the central engine.
In the optical, the detections seem to be dominated by afterglow emission at all times. An intriguing question to ask is what happens to the material producing the late burst emission. In the frame of the fireball model, they will run into the ambient medium and refresh the external shock. Whether this signature is observed for GRB 081008 is uncertain. If such early refreshed shocks can be observed, their relative strength to the prompt emission provide important clues to the burst mechanism and production of the afterglow. Refreshed shocks have been used to explain late-time lightcurve bumps \citep{gnp03,jbg06}. In those cases, slow shells catch up with the forward shock front hours to days after the initial burst. The existence and properties of the slow shells are only inferred from the lightcurve shapes. Observations directly linking the ejecta material and refreshed shocks are achievable for long bursts with discrete episodes of activities or for events with very bright X-ray flares if they are produced in a similar way as the prompt burst emission. The latter cases may produce cleaner chasers of the refreshed shock with less contamination from internal optical emission \citep[however, see][]{kgmkr09}. \citet{fblck06} have attempted to relate the subsequent bumps in the X-ray lightcurve with the bright X-ray flare in GRB 050502B. No solid evidence of refreshed shocks associated with flares has been observed in the optical so far, probably due to limited time coverage of the data or incomplete understanding of the modification to the afterglow lightcurve.
The importance of well-sampled multi-wavelength data has been demonstrated. In general, the observed afterglow emission is consistent with displaying an achromatic temporal break. Such a break can be interpreted as due to a collimated outflow (jet) or the end of energy injection. However, there is a hint of a steeper decay in the X-rays after the break. While for GRB 081008, the deviation is rather subtle, similar steeper decay in the X-rays than in the optical has been observed for some other events \citep[e.g.][]{rmysa06,pwbdh09}. Using data from GROND, UVOT and XRT, we are able to model the SED of the afterglow from IR to X-ray and constrain a probable spectral break within the optical regime. As many other well-studied bursts, the overall temporal and spectral behavior of the afterglow is hard to explain in the frame of the standard fireball model.
\acknowledgments
FY gratefully acknowledges useful discussions with S.~B.~Pandey. This work made use of data supplied by the UK {\em Swift} Science Data Centre at the University of Leicester. FY is supported by the NASA {\em Swift} Guest Investigator grants NNX-08AN25G. JLR acknowledges support for this work from NASA contract NAS5-00136. RW and PTO gratefully acknowledge funding from the STFC for {\em Swift} at the University of Leicester. TK acknowledges support by the DFG cluster of excellence `Origin and Structure of the Universe'. ROTSE-III has been supported by NASA grant NNX-08AV63G, NSF grant PHY-0801007, the Australian Research Council, the University of New South Wales, the University of Texas, and the University of Michigan. Special thanks to the H.E.S.S. staff, especially Toni Hanke. Part of the funding for GROND (both hardware as well as personnel) was generously granted from the Leibniz-Prize (DFG grant HA 1850/28-1) to Prof. G. Hasinger (IPP).
|
\section{Introduction}
In these notes we provide the foundation for the deformation theoretic
parts of \cite{VdB28,VdB26}. In \cite{VdB26} we construct
non-commutative analogues of quadrics and in \cite{VdB28} we define
non-commutative ${\blb P}^1$-bundles over commutative varieties. A notable
special case of the latter are non-commutative analogues of Hirzebruch
surfaces.
Indeed \cite{VdB28} contains a proof that any formal deformation of a Hirzebruch
surface (in a suitable sense) is given by a non-commutative
Hirzebruch surface. Similarly the original (privately circulated) version of
\cite{VdB26} contains a proof that any formal deformation of a quadric is
a non-commutative quadric (see \cite[\S
11.2]{VdBSt} for a sketch). I deleted this proof when I
first put the paper on the arXiv (8 years after it was written) since I was
unhappy with the deformation theoretic setup that was used.
Meanwhile a satisfactory infinitesimal deformation theory for abelian
categories has been developed in \cite{lowenvdb2,lowenvdb1}.
In
the noetherian setting (which is sufficient for the applications we
have in mind) the passage from the infinitesimal context to the
formal context is an application of Jouanolou's results
in \cite{Joua}. Nonetheless Jounalou's expos\'e is written for a different
purpose so some translation
is necessary. After several (not very satisfactory) attempts to
rewrite the deformation theoretic parts of \cite{VdB28,VdB26} using
Jouanolou's language of ``AR-J-adic systems'' I decided that it was better to
write a self contained paper on formal deformations of noetherian abelian
categories, which resulted
in the current paper.
On the purely mathematical level there is very little originality in
what we will do. Besides Jouanolou's expos\'e we have also borrowed
from \cite{AZ2} (which basically discusses trivial deformations) and
\cite[\S5]{EGA31} (which discusses formal schemes). On the expository
level we deviate from the aforementioned references by systematically
using Pro-objects instead of adic objects. Pro-objects form a
co-Grothen\-dieck category so in particular they have very well
behaved inverse limits.
\medskip
We now give a more detailed exposition of our setup. Let $R$ be a
commutative noetherian ring and let $J$ be an ideal in $R$.
\begin{definition}
Let ${\cal C}$ be a noetherian $R$-linear abelian category. The \emph{completion}
$\widehat{{\cal C}}$ of ${\cal C}$ is the full subcategory of
$\operatorname{Pro}({\cal C})$ consisting of the pro-objects $M$ over ${\cal C}$ such that $M/MJ^n\in
{\cal C}$ for all $n$ and such that the canonical map $M\rightarrow \invlim_n
M/MJ^n$ is an isomorphism.
\end{definition}
We reproduce Jouanolou's proof
(recast in our language) that $\widehat{{\cal C}}$ is a noetherian
abelian category (see Proposition \ref{ref-2.2.4-6} below).
There is an obvious exact functor $\Phi:{\cal C}\rightarrow\widehat{{\cal C}}:M\mapsto \invlim_n M/MJ^n$ and we say that
${\cal C}$ is \emph{complete} if this functor is an equivalence.
Roughly speaking an \emph{$(R,J)$-deformation} of an $R/J$-linear abelian
category ${\cal E}$ will be a complete $R$-linear category ${\cal F}$
together with an equivalence ${\cal E}\cong{\cal F}_{R/J}$ where
${\cal F}_{R/J}$ is the full subcategory of ${\cal F}$ consisting of objects
annihilated by $J$. To make this definition work one has to impose
certain flatness conditions. See \S\ref{ref-2.1-1} and \S\ref{ref-3-35} for
more details.
\medskip
So to understand formal deformations of abelian categories we
have to understand completion. We first note that
completion extends to functors. We show that if $(T^i)_i:{\cal C}\rightarrow
{\cal D}$ is $\partial$-functor between noetherian abelian $R$-linear categories
then this functor extends to a $\partial$-functor
$(\hat{T}^i)_i:\widehat{{\cal C}}\rightarrow \widehat{{\cal D}}$ (see Theorem
\ref{ref-2.3.1-15}). This is a slight improvement over \cite{Joua} as
Jouanolou imposes some extra conditions on $T$ which seem to be
superfluous.
\medskip
We use the good behaviour of $\partial$-functors to study
$\operatorname {Ext}$-groups. Let ${\cal C}_t$ be the full subcategory of ${\cal C}$
consisting of the objects which are annihilated by some power
of~$J$. Then we define the \emph{completed $\operatorname {Ext}$-groups} between objects
$M,N\in \widehat{{\cal C}}$ as follows
\begin{equation}
\label{new-1}
`\operatorname {Ext}^i_{\widehat{{\cal C}}}(M,N)=\operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_t)}(M,N)
\end{equation}
An alternative point of view to this definition is that we consider
the full subcategory $D_c({\cal C})$ of $D(\operatorname{Pro}({\cal C}_t))$ of complexes whose
cohomology lies in $\widehat{{\cal C}}$. Thus $D_c({\cal C})$ has a $t$-structure whose heart is
$\widehat{{\cal C}}$.
Then the completed $\operatorname {Ext}$-groups for $M,N\in \widehat{{\cal C}}$ may be reinterpreted as
\[
{}`\operatorname {Ext}_{\widehat{{\cal C}}}(M,N)=\operatorname {Hom}_{D_c({\cal C})}(M,N[n])
\]
In the case that ${\cal C}$ is the category of torsion $l$-adic constructible
sheaves it would be interesting to compare this derived category
to the standard derived category of $l$-adic sheaves \cite{Behrend1,BBD,Ek}.
\medskip
Obviously $`\operatorname {Ext}_{\widehat{{\cal C}}}(-,-)$ is a $\partial$-functor in both arguments
but
apart from this we don't have anything to say about it. However in the
event that ${\cal C}_{R/J}$ (the objects in ${\cal C}$ annihilated by $J$)
has finitely generated $\operatorname {Ext}$-groups over $R/J$ and ${\cal C}$ is ``formally flat'' (see
\S\ref{ref-2.4-18}) then we have the expected formula
\[
`\operatorname {Ext}^i_{\widehat{{\cal C}}}(M,N)=\invlim_k \dirlim_l \operatorname {Ext}_{{\cal C}_{R/J^l}}^i(M/MJ^l,N/NJ^k)
\]
An important theorem in algebraic geometry is Grothendieck's existence
theorem~\cite{EGA31}. This theorem extends to the current setting (see also
\cite{AZ2}). In \S\ref{ref-4-36} we introduce the notion of a \emph{strongly
ample sequence}.\footnote{We could have used ample
sequences \cite{Polishchuk1} but to have good behaviour of higher
$\operatorname {Ext}$-groups it is convenient to use a slightly stronger notion.}
By definition a sequence $(O(n))_{n\in{\blb Z}}$ of objects in a noetherian abelian
category ${\cal E}$ is strongly ample if the following conditions hold
\begin{itemize}
\item[(A1)] For all $M\in {\cal E}$ and for all $n$
there is an epimorphism $\oplus_{i=1}^t O(-n_i)\rightarrow M$ with $n_i\ge n$.
\item[(A2)] For all $M\in {\cal E}$ and for all $i>0$ one has
$\operatorname {Ext}^i_{\cal E}(O(-n),M)=0$ for $n\gg 0$.
\end{itemize}
A strongly ample sequence $(O(n))_{n\in {\blb Z}}$ in ${\cal E}$ is ample in
the sense of \cite{Polishchuk1}. Hence using the methods of \cite{AZ}
or \cite{Polishchuk1} one obtains $ {\cal E}\cong \operatorname {gr}(A)/{\mathrm{f.l.}} $ if ${\cal E}$ is
$\operatorname {Hom}$-finite, where $A$ is the noetherian ${\blb Z}$-algebra $\oplus_{ij}
\operatorname {Hom}_{\cal E}(O(-j),O(-i))$.
The following is our version of Grothendieck's existence theorem.
\begin{proposition} (see Proposition \ref{ref-4.1-37}) Assume that $R$ is
$J$-adically complete. Let ${\cal E}$ be an $\operatorname {Ext}$-finite $R$-linear
noetherian category with a strongly ample sequence $(O(n))_n$. Then
${\cal E}$ is complete and furthermore if ${\cal E}$ is flat (see \S\ref{ref-2.1-1}) we have for
$M,N\in{\cal E}$:
\begin{equation}
\label{ref-1.1-0}
\operatorname {Ext}^i_{\cal E}(M,N)=`\operatorname {Ext}^i_{\cal E}(M,N)
\end{equation}
\end{proposition}
The property for a sequence to be strongly ample lifts well under deformations.
\begin{theorem} (an extract of Theorem \ref{ref-4.2-44}) Let ${\cal D}$
be an $R$-deformation of an $\operatorname {Ext}$-finite flat $R/J$-linear
noetherian abelian category ${\cal C}$ and $(O(n))_n$ be a sequence of
$R$-flat objects in ${\cal D}$. Then $(O(n)/O(n)J)_n$ is strongly ample
in ${\cal C}$ if and only if $(O(n))_n$ is strongly ample in ${\cal D}$.
\end{theorem}
Many algebraic varieties (e.g.\ Del Pezzo surfaces) have a strongly
ample sequence consisting of exceptional objects. Such a
sequence can then be lifted to any deformation (see \S\ref{seclb}). This idea is basically
due to Bondal and Polishchuk and is described explicitly in \cite[\S
11.2]{VdBSt}. It was used to define non-commutative quadrics in
\cite{VdB26} and indirectly in the classification of non-commutative
Hirzebruch surfaces in \cite{VdB28}. See also the recent paper
\cite{LDeD}.
\medskip
Let us also mention that a very complete treatment of deformations of algebraic
varieties as ringed spaces (including their derived categories) over $k[[t]]$ has
been given in \cite{HMS}.
\section{Completion of abelian categories}
\subsection{Base extension}
\label{ref-2.1-1}
We recall briefly some notions from \cite{lowenvdb1}. Throughout $R$ will be a commutative
noetherian ring and $\operatorname{mod}(R)$ is its category of finitely generated modules.
Let ${\cal C}$ be an $R$-linear abelian category. Then we have bifunctors $-\otimes_R-:{\cal C}\times \operatorname{mod}(R)\rightarrow
{\cal C}$, $\operatorname {Hom}_R(-,-):\operatorname{mod}(R)^\circ\times {\cal C}\rightarrow {\cal C}$ defined in the usual
way. These functors may be derived in their $\operatorname{mod}(R)$-argument to
yield bi-delta-functors $\operatorname {Tor}^R_i(-,-)$, $\operatorname {Ext}_R^i(-,-)$. An object $M\in {\cal C}$ is
\emph{$R$-flat} if $M\otimes_R -$ is an exact functor, or equivalently if $\operatorname {Tor}_i^R(M,-)=0$ for
$i>0$.
By
definition (see \cite[\S3]{lowenvdb1}) ${\cal C}$ is $R$-\emph{flat} if
$\operatorname {Tor}^R_i$ or equivalently $\operatorname {Ext}^i_R$ is effaceable in its
${\cal C}$-argument for $i>0$. This implies that $\operatorname {Tor}^R_i$ and
$\operatorname {Ext}_R^i$ are universal $\partial$-functors in both arguments.
If $f:R\rightarrow S$ is a morphism of commutative rings and ${\cal C}$ is an
$R$-linear abelian category then ${\cal C}_S$ denotes the (abelian)
category of objects in ${\cal C}$ equipped with an $S$-action. We
usually refer to objects in ${\cal C}_S$ as $(S,{\cal C})$-objects and if
$S$ is graded then we also talk about graded $(S,{\cal C})$-objects. If
$f$ is surjective then ${\cal C}_S$ identifies with the full subcategory
of ${\cal C}$ given by the objects annihilated by $\operatorname {ker} f$. If $R$ is
noetherian and $S$ is module finite over $R$ then the inclusion
functor ${\cal C}_S\rightarrow {\cal C}$ has right and left adjoints given
respectively by $\operatorname {Hom}_R(S,-)$ and $-\otimes_R S$.
\subsection{Completion of noetherian abelian categories}
Below we refer to a pair $(R, J)$ where
$R$ is a commutative noetherian ring and $J\subset R$ is an ideal
as a $J$-\emph{adic noetherian ring}.
Below $(R,J)$ is a $J$-adic noetherian ring. We put $R_n=R/J^n$ and we denote
the $J$-adic completion of $R$ by $\hat{R}$. This is also a noetherian ring.
Using a slight abuse of notation we denote the extended ideal $J\hat{R}$ by $J$.
Recall that an abelian category ${\cal C}$ is said to be \emph{noetherian} if it
is essentially small and all objects are noetherian. Below ${\cal C}$ is an $R$-linear
noetherian category.
If ${\cal D}$ is an essentially small abelian
category then the category $\operatorname{Pro}({\cal D})$ of \emph{pro-objects} over ${\cal D}$ is the category
whose objects are filtered inverse systems $(M_\alpha)_\alpha$ and
whose $\operatorname {Hom}$-sets are given by
\begin{equation}
\label{ref-2.1-2}
\operatorname {Hom}_{\operatorname{Pro}({\cal D})}((M_\alpha)_\alpha,(N_\beta)_\beta)=
\invlim_\beta\dirlim_\alpha\operatorname {Hom}_{\cal D}(M_\alpha,N_\beta)
\end{equation}
In other words if we identify ${\cal D}$ with the one-object inverse systems in $\operatorname{Pro}({\cal D})$ then
$(M_\alpha)_\alpha=\invlim_\alpha M_\alpha$ in $\operatorname{Pro}({\cal D})$.
\begin{lemmas} \cite[\S I.8]{SGA41} Assume that ${\cal D}$ is an essentially small abelian
category. Then $\operatorname{Pro}({\cal D})^{\text{opp}}$ is a Grothendieck category
and in particular $\operatorname{Pro}({\cal D})$ has exact filtered inverse limits
and enough projectives. The natural functor ${\cal D}\rightarrow \operatorname{Pro}({\cal D})$ is
fully faithful exact and its essential image is closed under
extensions. If $D\in {\cal D}$ then $\operatorname {Hom}_{\cal D}(-,D)$ sends inverse limits
to direct limits (in other words $D$ is co-finitely presented).
\end{lemmas}
\begin{definitions}
The \emph{completion} $\widehat{{\cal C}}$ of ${\cal C}$ is the full subcategory of $\operatorname{Pro}({\cal C})$
consisting of the objects $M$ such that $M/MJ^n\in {\cal C}$ for all $n$ and
such that the canonical map $M\rightarrow \invlim_n M/MJ^n$ is an
isomorphism.
\end{definitions}
It is easy to see that $\widehat{{\cal C}}$ is a $\widehat{R}$-linear category.
To study objects in $\widehat{{\cal C}}$ we need to consider filtrations.
By
definition a filtration on an object $M$ of an $R$-linear category
${\cal D}$ is a descending chain of subobjects $M=\cdots \supset F_0M\supset
F_1M\supset\cdots$. The associated graded objects
$\operatorname {gr}_F M$ is the ${\blb Z}$-graded object over~${\cal D}$ defined by the
formal direct sum $\bigoplus_n F_{n}M/F_{n+1}M$. By $F_J$ we denote the $J$-adic
filtration. I.e.\ $F_{J,i}M=MJ^i$ for $i\ge 0$ and $F_{J,i}M=M$ for $i\le 0$.
We say that the filtration $F$ is \emph{adapted} to $J$ if
$(F_iM)J\subset F_{i+1}M$ (see \cite[\S 4.2]{Joua}). In that case $\operatorname {gr}_F M$ is
a graded $(\operatorname {gr}_J R,{\cal D}_{R/J})$-object.
\begin{lemmas} \label{ref-2.2.2-3} If $M$ is a noetherian object in an
$R$-linear abelian category ${\cal D}$ and~$S$ is a positively graded
noetherian $R$-algebra such that $S_0=R$. Then $M\otimes_R S$ is a noetherian graded
$(S,{\cal D})$-object
\end{lemmas}
\begin{proof} This follows from a variant of Hilbert's basis theorem. See e.g.\
\cite[Thm 5.1.4]{Joua} and \cite[Lemma 4.2.4]{Joua}.
\end{proof}
\begin{lemmas} \label{ref-2.2.3-4} (compare with \cite[Thm 4.2.6]{Joua}) Assume that $K\in \operatorname{Pro}({\cal C})$ is
equipped with a $J$-adapted filtration $F$ such that
\begin{enumerate}
\item $\invlim_n K/F_nK=K$.
\item $\operatorname {gr}_F K$ is a noetherian graded $(\operatorname {gr}_J R,{\cal C}_{R/J})$-object.
\end{enumerate}
Then $K\in \widehat{{\cal C}}$.
\end{lemmas}
\begin{proof}
We follow somewhat the idea of \cite[Lemma
4.2.7]{Joua}. For any $r\ge 0$ define $F^{(r)}_i K=KJ^{i-r}\cap
F_{i-1} K+F_i K$ (with $J^{i-r}=R$ for $r\ge i$).
Then we have $F_i K\subset F^{(r)}_i K\subset F_{i-1}K$ and $(F^{(r)}_i K)J\subset
F^{(r)}_{i+1} K\subset F_i K$. In other words $\operatorname {gr}^{(r)}_FK\overset{\text{def}}{=}\bigoplus
F^{(r)}_{i+1} K/F_{i+1} K$ is an ascending chain of graded $(\operatorname {gr}_J R,{\cal C}_{R/J})$-subobjects of
$\operatorname {gr}_F K$ which must be stationary. Thus there is an $r$ such that for all $i$
\[
KJ^{i-r}\cap
F_{i-1} K+F_i K=KJ^{i-r-1}\cap
F_{i-1} K+F_i K=\cdots=F_{i-1}K
\]
and in particular $F_{i-1}K\subset KJ^{i-r}+F_i K$. Iterating this inclusion and renumbering
we get that there exists an $r$ such that
\[
F_iK\subset KJ^{i-r}+F_jK
\]
for all $j\ge i$. Fix $i$ and choose generators $f_1,\ldots,f_p$ for $J^{i-r}$. Then
we get diagrams for $j\ge i$
\begin{equation}
\label{ref-2.2-5}
\xymatrix{
&&0\\
(K/F_jK)^p\ar^{(f_i)_i}[r] &K/F_j K\ar[r] &K/(F_j K+J^{i-r} K)\ar[r]\ar[u]&0\\
&K\ar[r]\ar[u]&K/F_i K\ar[u]
}
\end{equation}
Using exactness
of filtered inverse limits we get from \eqref{ref-2.2-5}
\[
\xymatrix{
&&0\\
K^p\ar^{(f_i)_i}[r] &K\ar[r]&\invlim_j K/(F_j K+J^{i-r} K)\ar[r]\ar[u]&0\\
&K\ar[r]\ar@{=}[u]&K/F_i K\ar[u]
}
\]
and hence from the upper exact sequence we obtain
\[
K/KJ^{i-r}=\invlim_j K/(F_j K+J^{i-r} K)
\]
In other words the identity map $K\rightarrow K$ induces a map $K/F_iK\rightarrow
K/KJ^{i-r}$ which yields $F_i K\subset KJ^{i-r}$.
\medskip
The fact that $\operatorname {gr}_F K$ is noetherian implies easily that it has left bounded grading.
Since ${\cal C}$ is closed under extensions inside $\operatorname{Pro}({\cal C})$ it follows that $K/F_iK\in{\cal C}$
for all~$i$. Furthermore since ${\cal C}$ is an abelian subcategory of $\operatorname{Pro}({\cal C})$ it
is also closed under $-\otimes_RM$ for $M\in \operatorname{mod}(R)$.
Hence $K/KJ^i=(K/F_{i+r}K)
\otimes_R R/J^i\in {\cal C}$. Furthermore since the $J$-adic filtration and the
$F$-filtration are cofinal we also get $\invlim_n K/KJ^n=\invlim_n K/F_nK=K$.
This shows that indeed $K\in \widehat{{\cal C}}$.
\end{proof}
\begin{propositions} \label{ref-2.2.4-6} (compare with \cite[Thm 5.2.3]{Joua}) $\widehat{{\cal C}}$ is a noetherian abelian subcategory of $\operatorname{Pro}({\cal C})$.
\end{propositions}
\begin{proof}
We first prove that $\widehat{{\cal C}}$ is an abelian
subcategory of $\operatorname{Pro}({\cal C})$. It is obviously closed under cokernels
(using the exactness of $\invlim$ and right exactness of $-\otimes_R R/J^n$) so we must prove it is closed
under kernels.
Let
\[
0\rightarrow K\rightarrow M\rightarrow N
\]
be an exact sequence in $\operatorname{Pro}({\cal C})$ with $M$, $N\in
\widehat{{\cal C}}$. We must prove $K\in \widehat{{\cal C}}$. Put
$F_iK\overset{\text{def}}{=}MJ^{i}\cap K\supset KJ^i$. This is a filtration on $K$
which is adapted to $J$. Furthermore we have exact sequences
\begin{equation}
\label{ref-2.3-7}
0\rightarrow K/F_nK\rightarrow M/MJ^n\rightarrow N/NJ^n
\end{equation}
By exactness of filtered inverse limits we deduce $K=\invlim_n K/F_nK$.
Furthermore we obtain exact sequences
\[
0\rightarrow \operatorname {gr}_F K\rightarrow \operatorname {gr}_{F_J} M\rightarrow \operatorname {gr}_{F_J} N
\]
Since $M/MJ\in {\cal C}$ it follows from Lemma \ref{ref-2.2.2-3} that $\operatorname {gr}_{F_J}M$ is
a noetherian graded $(\operatorname {gr}_J R,{\cal C}_{R/J})$ object.
Hence $\operatorname {gr}_F K$ is also a noetherian graded $(\operatorname {gr}_J
R,{\cal C}_{R/J})$-object. By Lemma \ref{ref-2.2.3-4} we conclude $K\in\widehat{{\cal C}}$.
It remains to show that $\widehat{{\cal C}}$ is noetherian. Since any
object $M$ in $\widehat{{\cal C}}$ satisfies $M=\invlim_n M/J^nM$ and the
category of ${\blb Z}$-indexed inverse systems over ${\cal C}$ is essentially small it
follows that $\widehat{{\cal C}}$ is essentially small as well. Thus it remains to
show that any $M\in \widehat{C}$ is noetherian.
Let $N\hookrightarrow M$ be a subobject of $M$ in $\widehat{{\cal C}}$. Put
$F_n N=N\cap MJ^n$. Then $N/F_n N$ is the image of $N/NJ^n\rightarrow M/MJ^n$ and so
it lies in ${\cal C}$. Furthermore taking the inverse limits of the maps
\[
\xymatrix{
N/NJ^n\ar@{->>}[r]&N/F_n N\ar@{^(->}[r] & M/MJ^n
}
\]
and using exactness of filtered inverse limits we get $N=\invlim_n N/F_n N$. Now
assume that we have inclusions $N_1\subset N_2\subset M$ in $\widehat{{\cal C}}$ such that if
we equip $N_1$, $N_2$ with the filtrations induced from the $J$-adic
filtration on $M$ then the map $\operatorname {gr}_F N_1\rightarrow \operatorname {gr}_F N_2$ is an
isomorphism. We claim that then necessarily $N_1=N_2$. Indeed from the
five lemma we obtain $N_1/F_i N_1=N_2/F_i N_2$. It then suffices to
take inverse limits.
Now let $M^{(r)}\subset M$ be an ascending chain of subobjects and equip them
with the filtrations induced from the $J$-adic filtration on $M$. As indicated above
$\operatorname {gr}_{F_J} M$ is a noetherian graded object over $(\operatorname {gr}_J R,{\cal C}_{R/J})$ and hence
the chain $(\operatorname {gr}_F M^{(r)})_r$ is stationary. By the discussion in the previous paragraph
the chain $(M^{(r)})_r$ is stationary as well.
\end{proof}
We may compare our definition of $\widehat{{\cal C}}$ with the notion of \emph{$J$-adic inverse systems.}
\begin{definitions} (see \cite[\S3.1]{Joua}) Let ${\cal C}$ be an $R$-linear noetherian abelian
category. The category of \emph{$J$-adic inverse systems} $\check{{\cal C}}$ over ${\cal C}$ is
defined as the full subcategory of inverse systems $(M_n,\phi_n)$
over ${\cal C}$ such that $M_nJ^n=0$ and such that the transition maps
$\phi_n:M_n\rightarrow M_{n-1}$ induce isomorphisms $M_{n}/M_{n}J^{n-1}\rightarrow M_{n-1}$.
\end{definitions}
\begin{propositions} \label{ref-2.2.6-8} The functor
\[
\Sigma:\widehat{{\cal C}}\rightarrow \check{{\cal C}}:M\mapsto (M/MJ^n)_n
\]
is an equivalence of categories. Its inverse is given by
\[
\Psi:\check{{\cal C}}\rightarrow \widehat{{\cal C}}: (N_n)_n\mapsto \invlim_n N_n
\]
\end{propositions}
\begin{proof} We first show that $\Psi$ is well defined. Let
$(N_n)_n\in \check{{\cal C}}$ and let $N$ be its inverse limit in $\operatorname{Pro}({\cal C})$. Using
exactness of filtered inverse limits in $\operatorname{Pro}({\cal C})$ we get
\[
N/NJ^i=(\invlim_n N_n)\otimes_R R/J^i=\invlim_n (N_n/N_n J^i)=N_i
\]
Thus we have indeed $N=\invlim_i N_i=\invlim_i N/NJ^i$. From this reasoning we
also get $\Sigma\Psi(N_i)_i=(N_i)_i$.
The fact that $\Psi\Sigma$ is the identity is by definition.
\end{proof}
The following easy result motivates the definition of $\widehat{{\cal C}}$.
\begin{propositions} \label{ref-2.2.7-9} One has $\operatorname{mod}(R)\,\hat{}=\operatorname{mod}(\hat{R})$.
\end{propositions}
\begin{proof} In the proof we must distinguish between inverse limits
in $\operatorname{Mod}(R)$ and $\operatorname{Pro}(\operatorname{mod}(R))$. Therefore we will temporarily
denote the latter by ${}^p\!\invlim$.
Let ${}^p\!\hat{R}$ be the object of $\operatorname{mod}(R)\,\hat{}$ given by ${}^p\!\invlim_n
R/J^n$. Its endomorphism ring is equal to $\invlim_n R/J^n=\hat{R}$. It
suffices to prove that ${}^p\!\hat{R}$ is a projective generator of
$\operatorname{mod}(R)\,\hat{}$.
We first show that ${}^p\!\hat{R}$ is projective. Let $M\in \operatorname{mod}(R)\,\hat{}$. Then
\begin{align*}
\operatorname {Hom}_{\operatorname{Pro}(\operatorname{mod}(R))}({}^p\!\hat{R},M)&=\invlim_n\operatorname {Hom}_{R/J^n}(R/J^n,M/MJ^n)\\
&=
\invlim_n M/MJ^n
\end{align*}
Hence we must prove that $M\mapsto \invlim_n M/MJ^n$ is exact. Now let
\[
0\rightarrow K\rightarrow M\rightarrow N\rightarrow 0
\]
be an exact sequence in $\operatorname{mod}(R)\,\hat{}$. By lemma \ref{ref-2.2.8-10} below we have that
\[
0\rightarrow K/KJ^n \rightarrow M/MJ^n\rightarrow N/NJ^n\rightarrow 0
\]
is exact up to essentially zero systems. From this one easily deduces that
its inverse limit is exact.
Now we prove that ${}^p\!\hat{R}$ is a generator. Let $M$ be a object of
$\operatorname{mod}(R)\,\hat{}$. Choose $R/J$-generators for $M/JM\in \operatorname{mod}(R/J)$
and lift those to $R/J^n$-generators for $M/MJ^n\in\operatorname{mod}(R/J^n)$. By Nakayama we
get compatible epimorphisms $(R/J^n)^t\rightarrow M/MJ^n$ for some fixed $t$.
Taking inverse limits we obtain an epimorphism
${}^p\!\hat{R}^t\rightarrow M$
and we are done.
\end{proof}
The following lemma was used.
\begin{lemmas} \label{ref-2.2.8-10} Assume that ${\cal D}$ is an $R$-linear abelian
category and $M\subset N$ is an inclusion of noetherian objects in
${\cal D}$. Then these objects satisfy the Artin-Rees condition in the sense
that there exists an $r$ such that for all $i$ we have
$NJ^{n+r}\cap M\subset MJ^n$.
\end{lemmas}
\begin{proof}
This is proved in the standard way.
Let $\tilde{R}=\bigoplus_{n\ge
0} J^n$ be the Rees ring of~$R$
The graded ring $\tilde{R}$ is finitely generated over $R$ and it follows from
Lemma \ref{ref-2.2.2-3} that $\bigoplus_i NJ^i$ is a noetherian graded object
over $(\tilde{R},{\cal D})$.
Hence so is $C=\bigoplus_i NJ^i\cap M$.
Inside $C$ we have an ascending chain of subobjects
$C^{(r)}=\bigoplus_i NJ^i \cap MJ^{i-r}$ (with $J^p=R$ for $p\le
0$) which must be stationary. Hence for a certain $r$ we have for any $i$:
$NJ^i \cap MJ^{i-r}=NJ^i \cap MJ^{i-r-1}=\cdots=NJ^i\cap M$. Putting $i=n+r$ yields
$MJ^{n}\supset NJ^{n+r}\cap MJ^n=NJ^{n+r}\cap M$.
\end{proof}
There is a canonical functor
\begin{equation}
\label{ref-2.4-11}
\Phi:{\cal C}\rightarrow \widehat{{\cal C}}:M\mapsto \invlim_n(M/MJ^n)
\end{equation}
\begin{propositions}
\label{ref-2.2.9-12}
The functor $\Phi$ introduced above is exact. It induces an equivalence
\[
{\cal C}_{R/J^n}\cong (\widehat{{\cal C}})_{R/J^n}
\]
\end{propositions}
\begin{proof}
Exactness is a consequence of Lemma \ref{ref-2.2.8-10}. It is similar to the
proof of exactness of $M\mapsto \invlim_n M/MJ^n$ in the proof of
Proposition \ref{ref-2.2.7-9}.
The second statement is a tautology when written out formally.
\end{proof}
\begin{lemmas}\label{ref-2.2.10-13} Let $\Phi:{\cal C}\rightarrow {\cal D}$ be a functor between $R$-linear noetherian abelian categories
which induces equivalences $\Phi_n:{\cal C}_{R/J^n}\rightarrow {\cal D}_{R/J^n}$. Then
\[
\widehat{\Phi}:\widehat{{\cal C}}\rightarrow \widehat{{\cal D}}:M\mapsto \invlim_n \Phi_n(M/MJ^n)
\]
is an equivalence.
\end{lemmas}
\begin{proof} One easily checks that $\widehat{\Phi}$ is well defined and that its
inverse is given by $\widehat{\Phi}^{-1}(N)=\invlim_n \Phi_n^{-1}(N/NJ^n)$.
\end{proof}
\begin{definitions} A noetherian $R$-linear abelian category ${\cal C}$ is \emph{complete}
if the functor $\Phi:{\cal C}\rightarrow \widehat{{\cal C}}$ is an equivalence.
\end{definitions}
\begin{propositions} $\widehat{{\cal C}}$ is complete.
\end{propositions}
\begin{proof} This follows from the second statement of Proposition \ref{ref-2.2.9-12} combined
with Lemma \ref{ref-2.2.10-13}.
\end{proof}
For completeness let us recall the following result
\begin{lemmas} \label{ref-2.2.13-14} (Nakayama) Let $M\in \widehat{{\cal C}}$ be
such that $MJ=M$. Then $M=0$.
\end{lemmas}
\begin{proof} We have $M=\invlim_n M/MJ^n=0$ since $M=MJ=MJ^2=\cdots$.
\end{proof}
\subsection{Functors}
Now we consider functors. Let $T=(T^i)_i$ be a $\partial$-functor
between~$R$-linear noetherian abelian categories ${\cal C}$ and
${\cal D}$. We extend $T$ to a $\partial $-functor~$\hat{T}$ commuting
with filtered inverse limits between $\operatorname{Pro}({\cal C})$ and $\operatorname{Pro}({\cal D})$.
In this section we prove the following strengthening of \cite[Prop.\ 5.3.1]{Joua}.
\begin{theorems} \label{ref-2.3.1-15} The functor $\hat{T}$ sends $\widehat{{\cal C}}$ to $\widehat{{\cal D}}$.
\end{theorems}
\begin{proof} This is a variant of \cite[Prop.\ 5.3.1]{Joua}. For the convenience
of the reader we adapt the proof in loc.\ cit.\ to our setting.
We need some rudiments from the foundation of the theory of spectral sequences.
In its abstract form a spectral sequence over an abelian category ${\cal E}$ is a
sequence of complexes $\mathbb{E}=(E^\bullet_r,d_r)_{r\ge 1}$ together with isomorphisms
$H^\bullet(E_r^\bullet,d)\cong (E_{r+1}^\bullet,0)$. If the terms of the complexes
$E^\bullet_r$ carry a grading then we assume that $d_r$ is homogeneous. Note that this
setup is shifted with respect to the usual indexing of
spectral sequences. This is more convenient for filtered objects.
Starting from spectral sequence $\mathbb{E}$ we may construct subobjects
\[
0=B_1^n\subset\cdots \subset B^n_r\subset\cdots \subset Z_r^n\subset \cdots \subset Z_1^n=E^n_1
\]
with $E^n_{r}=Z_r^n/B_r^n$. The subobjects $B^n_r,Z^n_r$ are constructed
recursively using the following exact sequences.
\begin{equation}
\label{ref-2.5-16}
0\rightarrow Z_{r+1}^n/B_r^n\rightarrow Z_r^n/B_r^n\xrightarrow{d_r^n} Z_r^{n+1}/B_r^{n+1}\rightarrow
Z_{r}^{n+1}/B_{r+1}^{n+1}\rightarrow 0
\end{equation}
If $Z_\infty^n=\invlim_r Z^n_r$, $B_\infty^n=\dirlim_r B^n_r$ exist then
we say that $\mathbb{E}$ \emph{converges} to $E_\infty^n=Z^n_\infty/B^n_\infty$. The graded object
$E^\bullet_\infty$ is called the limit of the spectral sequence.
The spectral sequence is said to \emph{degenerate} at $E^n_{r_0}$ if $d^n_r=0$, $d^{n-1}_{r}=0$
for $r\ge r_0$.
In that case it follows from \eqref{ref-2.5-16} that $B^n_{r+1}=B^n_{r}$, $Z^n_{r+1}=Z^n_{r}$ for $r\ge r_0$ and thus
$E_\infty^n$ exists and is equal to $E_{r_0}^n$.
If $(T^n)_n:{\cal E}\rightarrow {\cal F}$ is a $\partial$-functor between abelian
categories and $X\in{\cal E}$ is an object equipped with a descending
filtration $F_{k+1}X\subset F_kX$ indexed by ${\blb Z}$ then the method of
exact couples yields a spectral sequence starting with
$E^n_1=T^n(\operatorname {gr}_F X)$. Here $ T^n(\operatorname {gr}_F
X)\overset{\text{def}}{=}\bigoplus_k T^n(F_kX/F_{k+1}X)$ is viewed as
a ${\blb Z}$-graded object over ${\cal F}$ (through the $k$-index). Hence
this spectral sequence lives in the abelian category of ${\blb Z}$-graded objects
over ${\cal F}$. The expressions for $Z^n_r$ and $B^n_r$ are
\begin{align*}
Z^n_r&=\bigoplus_k\operatorname {ker}(T^n(F_kX/F_{k+1}X)\rightarrow T^{n+1}(F_{k+1}X/F_{k+r}X))\\
B^n_r&=\bigoplus_k \operatorname {im}(T^{n-1}(F_{k-r+1}X/F_{k}X)\rightarrow T^n(F_kX/F_{k+1}X))
\end{align*}
We now make a number of hypotheses.
\begin{enumerate}
\item ${\cal E}$, ${\cal F}$ are complete with exact filtered limits.
\item $(T^n)_n$ commutes with filtered limits.
\item We have $X=F_kX$ for $k\ll 0$.
\item $X$ is complete. I.e.\ $X=\invlim_k X/F_kX$.
\end{enumerate}
We note that limits and colimits on graded objects can be computed degreewise. Hence
$Z_\infty^n$ exists and is equal to
\[
Z^n_\infty=\bigoplus_k\operatorname {ker}(T^n(F_kX/F_{k+1}X)\rightarrow T^{n+1}(F_{k+1}X))\\
\]
Similarly $B_\infty^n$ exists and is equal to
\[
B^n_\infty=\bigoplus_k \operatorname {im}(T^{n-1}(X/F_kX)\rightarrow T^n(F_kX/F_{k+1}X))
\]
It is now well-known and an easy verification that
\[
Z^n_\infty/B^n_\infty=\bigoplus_k \frac{\operatorname {im}(T^n(F_k X)\rightarrow T^nX)}{\operatorname {im}(T^n(F_{k+1} X)\rightarrow T^n X)}
\]
In other words if we equip $T^nX$ with the filtration $F_k(T^nX)=\operatorname {im} (T^n(F_kX)\rightarrow T^nX)$
then $E_\infty^n\cong \operatorname {gr}_F T^n X$. Note that the conditions also imply
that $T^nX$ is complete for this filtration. Indeed
\begin{align*}
\invlim_k T^nX/F_k T^n X&=\invlim_k\operatorname {coker} (T^n(F_kX)\rightarrow T^nX)\\
&=\operatorname {coker}( T^n(\invlim_k F_kX)\rightarrow T^nX)\\
&=T^nX
\end{align*}
Now revert to the notations in the statement of the proposition. We
apply the previous discussion with ${\cal E}=\operatorname{Pro}({\cal C})$,
${\cal F}=\operatorname{Pro}({\cal D})$ and $X=M$. We equip $M$ with the $J$-adic
filtration. By the above discussion we get a spectral sequence
$\mathbb{E}$ with $E^n_1=T^n(\operatorname {gr}_{F_J} M)$ which converges to $\operatorname {gr}_F
\hat{T}^n(M)$. The terms occuring in this spectral sequence are graded
$(\operatorname {gr}_J R,{\cal D}_{R/J})$ objects. The limit is a priori only a $(\operatorname {gr}_J
R,(\operatorname{Pro}{\cal D})_{R/J})$-object.
By Lemma \ref{ref-2.2.2-3} $\operatorname {gr}_J M$ is a noetherian $(\operatorname {gr}_J R,{\cal C}_{R/J})$ object.
Hence by Lemma \ref{ref-2.3.2-17} below $T^n(\operatorname {gr}_J M)$ is a noetherian graded $(\operatorname {gr}_J R,{\cal D}_{R/J})$-object.
Hence the ascending chain $B_r^n$ must be stationary. By \eqref{ref-2.5-16}
we obtain $d^n_r=0$ for $r\gg 0$. Hence~$\mathbb{E}$ degenerates at $E^n_r$ for $r\gg 0$.
It follows that $E_\infty^n=\operatorname {gr}_F \hat{T}^n(M)$ is a noetherian graded $(\operatorname {gr}_J R,{\cal D}_{R/J})$-object.
Since we had already shown that $\hat{T}^n(M)$
is complete we conclude by Lemma \ref{ref-2.2.3-4}.
\end{proof}
\begin{lemmas} \label{ref-2.3.2-17} Let $S_0$ be a commutative noetherian ring and let $S$ be a
finitely generated positively commutative graded $S_0$-algebra whose part of
degree zero is $S_0$. Let $(T^i)_i$ be a $\partial$ functor between
noetherian abelian categories $S_0$-linear categories ${\cal E}$,
${\cal F}$. Then for any noetherian graded $(S,{\cal E})$-object $N$ and
for any $i$ we have that $T^i(N)\overset{\text{def}}{=}\bigoplus_n
T^i(N_n)$ is a noetherian graded $(S,{\cal F})$-object.
\end{lemmas}
\begin{proof} We perform induction on the minimal number of generators
$d$ of $S$ as $S_0$-algebra. If $d=0$ then $S=S_0$ and hence $N$ is
concentrated in a finite number of degrees. In this case $T^i(N)$ is
obviously noetherian.
Now assume $d>0$ and pick a homogeneous generator $t$ of $S$ over $S_0$ of strictly
positive degree $f$. Then $N$ can be written as an extension
\[
0\rightarrow N'\rightarrow N\rightarrow N''\rightarrow 0
\]
where $N'$ is annihilated by some power of $t$ and $N''$ is $t$-torsion free.
The object $N'$ can itself be written as a repeated extension of objects annihilated by $t$.
Thus it suffices to treat the cases where $N$ is annihilated by $t$ and where $N$ is $t$-torsion free.
If $N$ is annihilated by $t$ then $T^i(N)$ is noetherian by induction (since $N$ is
now an $S/tS$-module and $S/tS$ has one generator less than $S$). Hence we assume
that $N$ is $t$-torsion free. From the exact sequence
\[
0\rightarrow N\xrightarrow{\times t} N\rightarrow N/Nt\rightarrow 0
\]
we obtain an injection
\[
T^i(N)/T^i(N)t\hookrightarrow T^i(N/Nt)
\]
By induction $T^i(N/Nt)$ is noetherian and hence so is $T^i(N)/T^i(N)t$. From
this one easily obtains that $T^i(N)$ itself is noetherian.
\end{proof}
\subsection{Formal flatness}
\label{ref-2.4-18}
The notations $R,J,{\cal C}$ are as above. For use below it would be convenient to assume
that ${\cal C}$ if flat. Unfortunately even if ${\cal C}$ is $R$-flat then
there seems to be no a priori reason for $\widehat{{\cal C}}$ to be flat
(although we do not know an explicit counter example).
To work around this issue we make the following definition
\begin{definitions} The $R$-linear category ${\cal C}$ is \emph{formally flat} if the categories ${\cal C}_{R/J^n}$
are $R/J^n$-flat for all $n$.
\end{definitions}
Since $\widehat{{\cal C}}_{R/J^n}={\cal C}_{R/J^n}$ it immediately follows that if ${\cal C}$ is
formally flat then so is~$\widehat{{\cal C}}$.
The following proposition yields a different characterization of formal flatness.
\begin{propositions} \label{ref-2.4.2-19} Let ${\cal C}_t$ be the full subcategory of objects
in ${\cal C}$ that are annihilated by some power of $J$. This is
naturally an $R$-linear category. Then ${\cal C}$ is formally flat if
and only if ${\cal C}_t$ is flat.
\end{propositions}
\begin{proof}
Assume first that ${\cal C}$ is formally flat. Take $M\in {\cal C}_t$, $N\in \operatorname{mod}(R)$. We must
prove that $\operatorname {Tor}_i^R(M,N)$ is effaceable in its first argument in ${\cal C}_t$ for $i>0$. We
assume $M\in {\cal C}_{R/J^n}$.
By dimension shifting in $N$ we may reduce to the case $i=1$. Take an exact sequence in $\operatorname{mod}(R)$
\[
0\rightarrow N'\rightarrow P\rightarrow N\rightarrow 0
\]
with $P$ projective. Then we get an exact sequence
\begin{equation}
\label{ref-2.6-20}
0\rightarrow \operatorname {Tor}_1^R(M,N)\rightarrow M\otimes_R N'\rightarrow M\otimes_R P\rightarrow M\otimes_R N\rightarrow 0
\end{equation}
Let $F$ be the filtration on $N'$ induced from the $J$-adic filtration on $P$.
From
\[
0\rightarrow N'/F_lN' \rightarrow P/PJ^l\rightarrow N/NJ^l\rightarrow 0
\]
we obtain an exact sequence (for $l\ge n$)
\[
0\rightarrow \operatorname {Tor}_1^{R/J^l}(M,N/NJ^l)\rightarrow M\otimes_{R/J^l} N'/F_lN'\rightarrow M\otimes_{R/J^l} P/PJ^l\rightarrow M\otimes_{R/J^l} N/NJ^l\rightarrow 0
\]
Combining these two sequence we see that there is a map
\begin{equation}
\label{new-2}
\operatorname {Tor}_1^R(M,N)\rightarrow \operatorname {Tor}_1^{R/J^l}(M,N/NJ^l)
\end{equation}
natural in $M$ (for $l\ge n$).
By the
Artin-Rees condition (Lemma \ref{ref-2.2.8-10}) we may take an $l$ such that $F_lN'\subset N'J^n$.
Then from \eqref{ref-2.6-20} we obtain an exact sequence
\[
0\rightarrow \operatorname {Tor}_1^R(M,N)\rightarrow M\otimes_{R/J^l} N'/F_lN'\rightarrow M\otimes_{R/J^l} P/PJ^l\rightarrow M\otimes_{R/J^l} N/NJ^l\rightarrow 0
\]
and thus we have deduced that for $l$ large \eqref{new-2} is an isomorphism
\begin{equation}
\label{ref-2.7-21}
\operatorname {Tor}_1^R(M,N)= \operatorname {Tor}_1^{R/J^l}(M,N/NJ^l)
\end{equation}
Since ${\cal C}_{R/J^l}$ is flat $ \operatorname {Tor}_1^{R/J^l}(M,N/NJ^l)$ is
effaceable in its first argument in ${\cal C}_{R/J^l}$ by an epimorphism
$M'\rightarrow M$. Thus we get a commutative diagram
\[
\xymatrix{
\operatorname {Tor}_1^{R}(M',N)\ar[r]\ar[d]&\operatorname {Tor}_1^{R/J^l}(M',N/NJ^l)\ar[d]\\
\operatorname {Tor}_1^{R}(M,N)\ar@{=}[r]&\operatorname {Tor}_1^{R/J^l}(M,N/NJ^l)\\
}
\]
with the right most map being zero. It follows that the left most map
is also zero. Thus $\operatorname {Tor}_1^R(M,N)$ is effaceable in $M$ in ${\cal C}_t$ and
hence ${\cal C}_t$ is flat.
Conversely assume ${\cal C}_t$ is flat. Since ${\cal C}_{t,R/J^n}={\cal C}_{R/J^n}$ and flatness is
stable under base change (\cite[Prop.\ 4.8]{lowenvdb1}) we conclude that ${\cal C}_{R/J^n}$ is flat.
\end{proof}
By Lemma \ref{ref-2.2.10-13} the inclusion of abelian categories
$
{\cal C}_t \rightarrow{\cal C}
$
yields an equivalence
\[
\widehat{{\cal C}}_t\cong \widehat{{\cal C}}
\]
Hence when ${\cal C}$ is formally flat we may always reduce to the case that ${\cal C}$ is flat.
\medskip
Flatness on the level of objects does not present any pitfalls as the following proposition
shows.
\begin{propositions} Let $M\in\widehat{{\cal C}}$. Then $M$ is $R$-flat if and only if $M/MJ^n$ is
$R/J^n$-flat for all $n$.
\end{propositions}
\begin{proof} We consider the non-obvious direction. Assume that $M\in\widehat{{\cal C}}$ is
such that all $M/MJ^n$ are flat. We need to prove that $M\otimes_R -$ is exact.
Consider
an exact sequence in $\operatorname{mod}(R)$.
\[
0\rightarrow K\rightarrow L\rightarrow N\rightarrow 0
\]
We have to show that
\[
0\rightarrow M\otimes_R K\rightarrow M\otimes_R L\rightarrow M\otimes_R N\rightarrow 0
\]
is exact.
After tensoring with $R/J^n$ is is sufficient to show that
\[
0\rightarrow M\otimes_R K/KJ^n\rightarrow M\otimes_R L/LJ^n\rightarrow M\otimes_R N/NJ^n\rightarrow 0
\]
is exact up to essentially zero systems. This is the same sequence as
\[
0\rightarrow M/MJ^n\otimes_{R/J^n} K/KJ^n\rightarrow M/MJ^n\otimes_{R/J^n} L/LJ^n\rightarrow M/MJ^n\otimes_{R/J^n} N/NJ^n\rightarrow 0
\]
Hence by flatness of $M/MJ^n$ it is sufficient that
\[
0\rightarrow K/KJ^n\rightarrow L/LJ^n\rightarrow N/NJ^n\rightarrow 0
\]
is exact up to essentially zero systems. This follows from the
Artin-Rees condition (see Lemma \ref{ref-2.2.8-10}).
\end{proof}
\subsection{$\operatorname {Ext}$-groups}
Now we discuss $\operatorname {Ext}$-groups.
by which we always mean Yoneda $\operatorname {Ext}$-groups.
We keep the notations from the previous section. Thus $(R,J)$ is a
noetherian $J$-adic ring and ${\cal C}$ is an $R$-linear noetherian
abelian category.
We make the following definition
\begin{definitions} Let $M,N\in \widehat{{\cal C}}$.
Then the \emph{completed $\operatorname {Ext}$-groups} between $M$,$N$ are defined by
\begin{align*}
`\operatorname {Ext}_{\widehat{{\cal C}}}^i(M,N)&=\operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^i(M,N)
\end{align*}
\end{definitions}
It is clear that $`\operatorname {Ext}_{\widehat{{\cal C}}}^i(M,N)$ is a $\partial$-functor
in both arguments. Apart from this nice property we don't know
if completed $\operatorname {Ext}$-groups are meaningful objects in general. To get better control we will assume
that ${\cal C}$ is formally flat
and we impose an additional finiteness condition
\begin{definitions} An $R$-linear abelian category ${\cal D}$ is \emph{$\operatorname {Ext}$-finite}
if for all objects $M,N\in {\cal D}$ we have that $\operatorname {Ext}^i_{\cal D}(M,N)$ is a finitely
generated $R$-module for all $i$.
\end{definitions}
The following will be the main result of this section
\begin{propositions} \label{ref-2.5.3-22} Assume that ${\cal C}$ is formally flat and that
${\cal C}_{R/J}$ is $\operatorname {Ext}$-finite. Then $`\operatorname {Ext}_{\widehat{{\cal C}}}^i(M,N)\in \operatorname{mod}(\hat{R})$ for $M,N\in
\widehat{{\cal C}}$
and furthermore
\begin{align}
`\operatorname {Ext}_{\widehat{{\cal C}}}^i(M,N)
&=\invlim_l \operatorname {Ext}_{\widehat{{\cal C}}}(M,N/NJ^k)\label{new-4}\\
&=\invlim_k \dirlim_l \operatorname {Ext}_{{\cal C}_{R/J^l}}^i(M/MJ^l,N/NJ^k) \label{new-5}
\end{align}
If $M$ is in addition $R$-flat then
\begin{equation}
\label{new-6}
`\operatorname {Ext}_{\widehat{{\cal C}}}^i(M,N)=\invlim_k \operatorname {Ext}_{{\cal C}_{R/J^k}}^i(M/MJ^k,N/NJ^k)
\end{equation}
\end{propositions}
The proof is a series of lemmas.
\begin{lemmas} \label{ref-2.5.4-23} Assume that ${\cal C}$ is formally flat and that ${\cal C}_{R/J}$ is $R/J$-$\operatorname {Ext}$ finite then ${\cal C}_{R/J^n}$ is
$R/J^n$-$\operatorname {Ext}$-finite for all $n$.
\end{lemmas}
\begin{proof} We need to prove that for $M,N\in {\cal C}_{R/J^n}$ we have
that $\operatorname {Ext}^i_{{\cal C}_{R/J^n}}(M,N)$ is a finitely generated $R/J^n$-module. By
filtering $M$ we may assume $M\in {\cal C}_{R/J}$. Then we conclude by
a change of rings spectral sequence (which depends on flatness, see \cite[Prop.\ 4.7]{lowenvdb1})
\[
E^{pq}_2:\operatorname {Ext}^p_{{\cal C}_{R/J}}(M,\operatorname {Ext}^q_{{\cal C}_{R/J^n}}(R/J,N))\Rightarrow
\operatorname {Ext}^{p+q}_{{\cal C}_{R/J^n}}(M,N)}\end{proof
\]
\def}\end{proof{}\end{proof}
The following lemma is proved in a similar way.
\begin{lemmas} \label{ref-2.5.5-24} Assume that ${\cal C}$ is formally flat and that ${\cal C}_{R/J}$ is $\operatorname {Ext}$-finite. Let $M\in \widehat{{\cal C}}$
and $N\in {\cal C}_{R/J^n}$. Then $\operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^i(M,N)$ is a finitely generated
$R/J^n$-module for all $i$. Furthermore we have
\begin{equation}
\label{ref-2.8-25}
\operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^i(M,N)=\dirlim_l \operatorname {Ext}_{{\cal C}_{R/J^l}}^i(M/MJ^l,N)
\end{equation}
and
\begin{equation}
\label{ref-2.9-26}
\operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^i(M,N)=\operatorname {Ext}_{\widehat{{\cal C}}}^i(M,N)
\end{equation}
Finally if $M$ is flat over $R$ then
\begin{equation} \label{ref-2.10-27}
\operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^i(M,N)=\operatorname {Ext}_{{\cal C}_{R/J^n}}^i(M/MJ^n,N)
\end{equation}
\end{lemmas}
\begin{proof} If ${\cal C}_t$ is flat then so is $\operatorname{Pro}({\cal C}_t)$ (see \cite[Prop.\ 3.6]{lowenvdb1}
for the dual statement).
Now we use the spectral sequence (for $l\ge n$)
\begin{equation}
\label{ref-2.11-28}
E^{pq}_2(l):\operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)_{R/J^l}}^p(\operatorname {Tor}^R_q(M,R/J^l),N)\Rightarrow \operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^{p+q}(M,N)
\end{equation}
which may derived in a similar way as \cite[Prop.\ 4.7]{lowenvdb1} (the existence
depends on flatness of $\operatorname{Pro}({\cal C}_t)$). The formation
of $\operatorname{Pro}$-objects commutes with certain base changes (see \cite[Prop.\ 4.5]{lowenvdb1} for
the dual statement)
and in particular $\operatorname{Pro}({\cal C}_t)_{R/J^l}=\operatorname{Pro}({\cal C}_{R/J^l})$.
Since
$\operatorname {Tor}^R_q(M,R/J^l)$ lies both in $\widehat{{\cal C}}$ and is annihilated by $J^l$ it
lies in ${\cal C}_{R/J^l}$. For an object $K\in {\cal C}_{R/J^l}$ we have $\operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_{R/J^l})}(
K,N)=\operatorname {Ext}^i_{{\cal C}_{R/J^l}}(
K,N)$ (see \cite[Prop.\ 2.14]{lowenvdb1}). Hence the spectral sequence \eqref{ref-2.11-28} becomes
\begin{equation}
\label{ref-2.12-29}
E^{pq}_2(l):\operatorname {Ext}_{{\cal C}_{R/J^l}}^p(\operatorname {Tor}^R_q(M,R/J^l),N)\Rightarrow \operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^{p+q}(M,N)
\end{equation}
To prove finite generation we put $l=n$ and invoke Lemma \ref{ref-2.5.4-23}.
To prove \eqref{ref-2.8-25} we note that for $l\le l'$ there are maps
of spectral sequence $E(l)\rightarrow E(l')$ which are given by the compositions
\[
\operatorname {Ext}_{{\cal C}_{R/J^l}}^p(\operatorname {Tor}^R_q(M,R/J^l),N)\rightarrow \operatorname {Ext}_{{\cal C}_{R/J^{l'}}}^p(\operatorname {Tor}^R_q(M,R/J^l),N)
\rightarrow \operatorname {Ext}_{{\cal C}_{R/J^{l'}}}^p(\operatorname {Tor}^R_q(M,R/J^{l'}),N)
\]
It follows from Lemma \ref{ref-2.5.6-30} below that $E^{pq}_2(l)\rightarrow
E^{pq}_2(l')$ is zero for $q>0$ and $l\ll l'$ (taking into account that $M$ is a noetherian
object in $\widehat{{\cal C}}$). Taking a direct limit over $l$ of \eqref{ref-2.12-29} we find that indeed
\[
\dirlim_l \operatorname {Ext}_{{\cal C}_{R/J^l}}^p(M/MJ^l,N)= \operatorname {Ext}_{\operatorname{Pro}({\cal C}_t)}^{p+q}(M,N)
\]
The claim \eqref{ref-2.9-26} follows from Lemma \ref{ref-2.5.7-31} below together with \eqref{ref-2.8-25}. The claim \eqref{ref-2.10-27} follows from the degeneration of the spectral
sequence \eqref{ref-2.12-29}.
\end{proof}
We have used the next two lemmas.
\begin{lemmas} \label{ref-2.5.6-30}
Let ${\cal D}$ be an $R$-linear abelian category and assume
that $M\in {\cal D}$ is a noetherian object. Then $\operatorname {Tor}_i^R(M,R/J^n)_n$ is an
essentially zero system for $i>0$.
\end{lemmas}
\begin{proof} We first replace ${\cal D}$ with the smallest abelian subcategory of ${\cal D}$
containing~$M$. This is a noetherian abelian category. Then we replace ${\cal D}$ by its
category of $\Ind$-objects. Then ${\cal D}$ becomes a locally noetherian Grothendieck category
and $M$ is still a noetherian object in ${\cal D}$. In particular $\operatorname {Ext}^i_{\cal D}(M,-)$ commutes
with filtered colimits.
As a Grothendieck category ${\cal D}$ has enough injectives. Hence we have to show that
for an arbitrary injective object $E\in {\cal D}$ we have
\[
\dirlim_n \operatorname {Hom}_{\cal D}(\operatorname {Tor}_i^R(M,R/J^n),E)=0
\]
Now we have
\begin{align*}
\dirlim_n \operatorname {Hom}_{\cal D}(\operatorname {Tor}_i^R(M,R/J^n),E)&=\dirlim_n \operatorname {Ext}^i_{\cal D}(M,\operatorname {Hom}_R(R/J^n,E))\\
&=\operatorname {Ext}^i_{\cal D}(M,\dirlim_n \operatorname {Hom}_R(R/J^n,E))
\end{align*}
Thus we have to show that $F=\dirlim_n \operatorname {Hom}_R(R/J^n,E)$ is injective. In a locally
noetherian Grothendieck category we can test this on inclusions of noetherian
objects. Hence let $K\hookrightarrow M$ be such an inclusion. We need to prove that
$\dirlim_n \operatorname {Hom}(M/MJ^n,E)\rightarrow \dirlim_n\operatorname {Hom}(K/KJ^n,E)$ is an epimorphism, or equivalently
that the kernel of $K/KJ^n \rightarrow M/MJ^n$ is an essentially zero system. This follows from
the Artin-Rees condition (Lemma \ref{ref-2.2.8-10}).
\end{proof}
\begin{lemmas} \label{ref-2.5.7-31} Let ${\cal D}$ be a noetherian $R$-linear category and assume that $N\in {\cal D}$
is annihilated by $J^n$. Let $M\in {\cal D}$. Then we have
\[
\operatorname {Ext}^i_{{\cal D}}(M,N)=\dirlim_l \operatorname {Ext}^i_{{\cal D}_{R/J^l}}(M/MJ^l,N)
\]
\end{lemmas}
\begin{proof}
From the Artin-Rees condition (Lemma \ref{ref-2.2.8-10}) we know that $(-\otimes_{R} R/J^l)_l$ is exact
up to essentially zero systems. From this we obtain in the usual way
that $M\mapsto \dirlim_l \operatorname {Ext}^i_{{\cal D}_{R/J^l}}(M/MJ^l,N)$ is a $\partial$-functor with
values in $R$-modules.
To show that this $\partial$-functor coincides with $\operatorname {Ext}^i_{{\cal D}}(M,N)$ it is
sufficient to prove this for $i=0$ and to show that any element of
$\dirlim_l \operatorname {Ext}^i_{{\cal D}_{R/J^l}}(M/MJ^l,N)$ is effaceable for $i>0$. The case $i=0$
is trivial so assume that $a\in \operatorname {Ext}^i_{{\cal D}_{R/J^l}}(M/MJ^l,N)$ represents an element
$\bar{a}$ of $\dirlim_l \operatorname {Ext}^i_{{\cal D}_{R/J^l}}(M/MJ^l,N)$ for $i>0$.
There exists an epimorphism $T\rightarrow M/MJ^l$ which effaces $a$ for some
$T\in {\cal D}_{R/J^l}$. Let $M'$ be the pullback of $T$ for the map
$M\rightarrow M/MJ^l$. Then the epimorphism $M'/M'J^l\rightarrow M/MJ^l$ factors through $T$ and
hence effaces $a$. This means that $M'\rightarrow M$ effaces $\bar{a}$
\def}\end{proof{}\end{proof}
\begin{lemmas} \label{ref-2.5.8-32} Assume that ${\cal C}$ is formally flat and that ${\cal C}_{R/J}$ is $\operatorname {Ext}$-finite and
\label{ref-2.5.8-33}
let $M,N\in \widehat{{\cal C}}$. Then
\[
\invlim_n \operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_t)}(M,N/NJ^n)\in \operatorname{mod}(R)\,\hat{}\quad(\cong \operatorname{mod}(\hat{R}))
\]
\end{lemmas}
\begin{proof} This follows by applying Theorem \ref{ref-2.3.1-15} to the $\partial$-functor
\[
\operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_t)}(M,-)_i:{\cal C}_t\rightarrow \operatorname{mod}(R)_t
\]
To construct this functor we use Lemma \ref{ref-2.5.5-24}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{ref-2.5.3-22}]
Let $P_\bullet$ be a projective resolution of
$M$ in $\operatorname{Pro}({\cal C}_t)$. Since ${\cal C}_t$ is flat by Proposition \ref{ref-2.4.2-19} the $P_m$ are
$R$-flat (see \cite[Prop.\ 3.4]{lowenvdb1} for the dual version). We compute
\begin{align*}
\operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_t)}(M,N)&=H^i(\operatorname {Hom}_{\operatorname{Pro}({\cal C}_t)}(P_\bullet,\invlim_n N/NJ^n))\\
&=H^i(\invlim_n \operatorname {Hom}_{\operatorname{Pro}({\cal C}_t)}(P_\bullet, N/NJ^n))
\end{align*}
We need to exchange $H^i$ and $\invlim_n$. This is possible if the terms of the
inverse system of complexes $\operatorname {Hom}_{\operatorname{Pro}({\cal C}_t)}(P_\bullet, N/NJ^n)$ as well as its cohomology
satisfy the Mittag-Leffler condition. For the terms this follows from the projectivity of
$P_m$. For the cohomology which is equal to $\operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_t)}(M,N/NJ^n)$ we
invoke Lemma \ref{ref-2.5.8-32} together with Lemma \ref{ref-2.5.9-34} below.
Assuming this we now obtain
\begin{align*}
\operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_t)}(M,N)&=\invlim_n H^i( \operatorname {Hom}_{\operatorname{Pro}({\cal C}_t)}(P_\bullet, N/NJ^n))\\
&=\invlim_n\operatorname {Ext}^i_{\operatorname{Pro}({\cal C}_t)}(M,N/NJ^n)
\end{align*}
This implies \eqref{new-4} via \eqref{ref-2.9-26}. Furthermore we obtain
\eqref{new-5} via \eqref{ref-2.10-27}. Finally we obtain \eqref{new-6} from \eqref{ref-2.9-26}.
\end{proof}
The following lemma was used.
\begin{lemmas}
\label{ref-2.5.9-34} Let $(U_n)_n$ be an inverse system in an $R$-linear noetherian
abelian category ${\cal D}$ such
that $U_nJ^n=0$ and such that the pro-object $(U_n)_n$ lies in
$\widehat{{\cal D}}$. Then $(U_n)_n$ satisfies the Mittag-Leffler
condition.
\end{lemmas}
\begin{proof} Let $U$ be the pro-object in $(U_n)_n$. Define $C_n$,
$K_n\in {\cal D}$ as the kernel and cokernel of the natural maps.
\[
U/UJ^n\rightarrow U_n
\]
Taking inverse limits in $\operatorname{Pro}(\operatorname{mod}(R))$ we see that $(U_n)_n$ and
$(C_n)_n$ are zero pro-objects, or equivalently they are essentially
zero systems. From one easily deduces that $(U_n)_n$ satisfies
the Mittag-Leffler condition.
\end{proof}
\subsection{The complete derived category}
We use the same notations as above. In particular $(R,J)$ is a $J$-adic noetherian
ring and ${\cal C}$ is an $R$-linear noetherian abelian category. We define $D_c({\cal C})$ as
the full subcategory of $D(\operatorname{Pro}({\cal C}_t))$ of complexes whose cohomology lies
in $\widehat{{\cal C}}$. Thus $D_c({\cal C})$ has a $t$-structure whose heart is
$\widehat{{\cal C}}$.
Then the completed $\operatorname {Ext}$-groups for $M,N\in \widehat{{\cal C}}$ may be reinterpreted as
\[
{}`\operatorname {Ext}_{\widehat{{\cal C}}}(M,N)=\operatorname {Hom}_{D_c({\cal C})}(M,N[n])
\]
In the case that ${\cal C}$ is the category of torsion $l$-adic constructible
sheaves it would be interesting to compare this derived category
to the standard derived category of $l$-adic sheaves \cite{Behrend1,BBD,Ek}.
\section{Formal deformations of abelian categories}
\label{ref-3-35}
\def\operatorname{Def}{\operatorname{Def}}
Let $R$ be a noetherian $J$-adic ring and ${\cal C}$ a flat $R/J$-linear noetherian abelian category. Then we
define an $R$-deformation of ${\cal C}$ to be a formally flat complete $R$-linear abelian category ${\cal D}$
together with an equivalence ${\cal D}_{R/J}\cong {\cal C}$. It follows from the above discussion
that ${\cal D}$ is specified up to equivalence by specifying the flat
$R/J^n$-linear categories ${\cal D}_n={\cal D}_{R/J^n}$ together with the equivalences
(isomorphisms in this case) ${\cal D}_m={\cal D}_{n,R/J^m}$ for $n\ge m$ and ${\cal D}_1\cong {\cal C}$.
Denote by $\operatorname{Def}_R({\cal C})$ the class of $R$-deformations of ${\cal C}$. This is a 2-groupoid.
The observations in the previous paragraph may be used to construct a 2-equivalence
\[
\operatorname{Def}_R({\cal C})\cong 3{\text{-}}\invlim_n \operatorname{Def}_{R/J^n}({\cal C})
\]
We leave it to the interested reader to formalize this statement. It
will not be used in this form.
\section{Ampleness}
\label{ref-4-36}
We define what we mean by a strongly ample sequence. This is stronger than
strictly necessary but easier to work with.
Let ${\cal E}$ an noetherian abelian category. For us a sequence $(O(n))_{n\in
{\blb Z}}$ of objects in ${\cal E}$ is strongly ample if the following conditions
hold
\begin{itemize}
\item[(A1)] For all $M\in {\cal E}$ and for all $n$
there is an epimorphism $\oplus_{i=1}^t O(-n_i)\rightarrow M$ with $n_i\ge n$.
\item[(A2)] For all $M\in {\cal E}$ and for all $i>0$ one has
$\operatorname {Ext}^i_{\cal E}(O(-n),M)=0$ for $n\gg 0$.
\end{itemize}
A strongly ample sequence $(O(n))_{n\in {\blb Z}}$ in ${\cal E}$ is ample in
the sense of \cite{Polishchuk1}. Hence using the methods of \cite{AZ} or \cite{Polishchuk1} one
obtains
$
{\cal E}\cong \operatorname{qgr}(A)
$
if ${\cal E}$ is $\operatorname {Hom}$-finite, where $A$ is the noetherian ${\blb Z}$-algebra $\oplus_{ij} \operatorname {Hom}_{\cal E}(O(-j),O(-i))$.
\medskip
Below we fix a complete noetherian $J$-adic ring $R$. The following is a version
of Grothendieck's existence theorem \cite{EGA31}.
\begin{proposition}
\label{ref-4.1-37}
Let ${\cal E}$ be an $\operatorname {Ext}$-finite $R$-linear noetherian category with a
strongly ample sequence $(O(n))_n$. Then ${\cal E}$ is complete and
furthermore if ${\cal E}$ is flat we have for $M,N\in{\cal E}$:
\begin{equation}
\label{ref-4.1-38}
\operatorname {Ext}^i_{\cal E}(M,N)=`\operatorname {Ext}^i_{\cal E}(M,N)
\end{equation}
\end{proposition}
\begin{proof}
\begin{step}
\label{ref-1-39}
We first claim that ${\cal E}$ satisfies Nakayama's lemma. This would
follow from Lemma \ref{ref-2.2.13-14} once we knew ${\cal E}$ is complete but we are not there
yet.
Let $M\in{\cal E}$ me such that $MJ=M$. Pick generators for $a_1,\ldots,a_n$ for $J$
and consider the corresponding epimorphisms
\[
M^{\oplus n}\xrightarrow{(a_i)_i}M\rightarrow 0
\]
Applying $\operatorname {Hom}_{\cal E}(O(-m),-)$ for $m$ large we get an epimorphism
\[
\operatorname {Hom}_{{\cal E}}(O(-m),M)^{\oplus n}\xrightarrow{(a_i)_i}\operatorname {Hom}_{{\cal E}}(O(-m),M)\rightarrow 0
\]
and thus by Nakayama's lemma for $R$ and $\operatorname {Ext}$-finiteness
\[
\operatorname {Hom}_{{\cal E}}(O(-m),M)=0
\]
for large $m$. It then follows from (A1) that $M=0$.
\end{step}
\begin{step}
\label{ref-2-40} Let $M$ be an object in ${\cal E}$. We claim that for $i>0$ and for
large $m$ we have
\begin{equation}
\label{ref-4.2-41}
\operatorname {Ext}^i_{\cal E}(O(-m),MJ^n)=0
\end{equation}
for all $n$ and furthermore for large $m$ we also have
\begin{equation}
\label{ref-4.3-42}
\operatorname {Hom}_{\cal E}(O(-m),MJ^n)=\operatorname {Hom}_{\cal E}(O(-m),M)J^n
\end{equation}
for all $n$.
Put $\tilde{R}=R\oplus J\oplus J^2\cdots$. According to Lemma \ref{ref-2.2.2-3}
$\tilde{M}=M\oplus MJ\oplus MJ^2\cdots$ is a noetherian graded $(\tilde{R},{\cal E})$-object.
Now we follow a similar inductive method as in the proof of Lemma
\ref{ref-2.3.2-17}. We only give a sketch. Put $W=\tilde{M}$. We need to prove
$\operatorname {Ext}^i_{{\cal E}}(O(-m),W)=0$ for large $m$. We deduce this from
the noetherian property of $W$. Let $d$ be the minimal number of
generators of $\tilde{R}$ over $R$ (i.e.\ the number of generators
of the $R$-ideal~$J$). If $d=0$ there is nothing to prove. If $d>0$
then we pick a homogeneous generator~$t$ of strictly positive degree in $\tilde{R}$ and we
reduce to the cases $Wt=0$ and $W$ is $t$-torsion free. The first
case is clear by induction and in the second case we obtain (also by induction)
$\operatorname {Ext}^i_{{\cal E}}(O(-m),W)t=\operatorname {Ext}^i_{{\cal E}}(O(-m),W)$ for large $m$. The
conclusion now follows from the fact that $t$ has strictly positive degree.
To prove \eqref{ref-4.3-42} we have to prove that
$\operatorname {Hom}_{\cal E}(O(-m),W)$ is generated in degree zero for $m$ large. We
use induction based on the noetherian property of $W$ and the fact
that $W$ is generated in degree zero. If $d=0$ then $W$ must be
concentrated in degree zero and there is nothing to prove. If $d\neq
0$ then we pick $t$ as above. Using $\operatorname {Ext}$-vanishing (which is already
proved) we obtain for $m$ large an exact sequence
\[
\operatorname {Hom}_{{\cal E}}(O(-m),W)\xrightarrow{\times t} \operatorname {Hom}_{{\cal E}}(O(-m),W)\rightarrow \operatorname {Hom}_{\cal E}(O(-m),W/Wt)\rightarrow 0
\]
We now finish by induction and the fact that $t$ has strictly positive degree.
\end{step}
\begin{step}
Now we show that the functor
\[
\Phi:{\cal E}\rightarrow \widehat{{\cal E}}
\]
is an equivalence. If $M$, $N\in {\cal E}$ then since $\Phi(M)$ is the pro-object
$M/MJ^n$ and similarly for $N$ we deduce from \eqref{ref-2.1-2}
we have
\begin{align*}
\operatorname {Hom}_{\widehat{{\cal E}}}(\Phi(M),\Phi(N))&=\invlim_k \dirlim_l \operatorname {Hom}_{{\cal E}_{R/J^l}}(M/MJ^l,N/NJ^k)\\
&=\invlim_k \operatorname {Hom}_{{\cal E}}(M,N/NJ^k)
\end{align*}
Thus we have to show that the natural map
\begin{equation}
\label{ref-4.4-43}
\operatorname {Hom}_{\cal E}(M,N)\rightarrow \invlim_n \operatorname {Hom}_{\cal E}(M,N/NJ^n)
\end{equation}
is an isomorphism. Using left exactness of $\invlim$ and (A1) we immediately
reduce to $M={\cal O}(-m)$ where $m$ may be chosen arbitrarily large.
We first observe that \eqref{ref-4.4-43} must be a monomorphism. Suppose on the contrary
that there is an $f:M\rightarrow N$ such that $K=\operatorname {im} f\subset NJ^n$ for all $n$. Then
by Lemma \ref{ref-2.2.8-10} we have $K=KJ$. Hence $K=0$ by Step 1 which implies $f=0$.
Now we prove that \eqref{ref-4.4-43} is an epimorphism when $M=O(-m)$ for
$m$ large. Suppose we are given a a compatible system of maps $f_n:M\rightarrow
N/NJ^n$. By \eqref{ref-4.2-41} the $f_n$ may be lifted to maps
$f'_n:M\rightarrow N$ such that $\operatorname {im}(f'_n-f'_{n+1})\subset NJ^n$. Using \eqref{ref-4.2-41}
this implies $f'_n-f'_{n+1}\in \operatorname {Hom}_{\cal E}(M,N)J^n$. Since $\operatorname {Hom}_{\cal E}(M,N)\in \operatorname{mod}(R)$ the
limit $f=\lim_n f'_n$ exists. It has the property that the image of $f$ in $\operatorname {Hom}_{\cal E}(M,N/NJ^n)$
is equal to $f_n$. This proves what we want.
\end{step}
\begin{step} Finally we prove \eqref{ref-4.1-38}. To show
that $`\operatorname {Ext}=\operatorname {Ext}$ in ${\cal E}$ it is sufficient to show that
$`\operatorname {Hom}=\operatorname {Hom}$ and furthermore that $`\operatorname {Ext}$ is effaceable in its first
argument.
The fact that $`\operatorname {Hom}=\operatorname {Hom}$ is the fact that \eqref{ref-4.4-43} is an isomorphism.
So let us show that $`\operatorname {Ext}$ is
effaceable.
Let $N$ be an object in ${\cal E}$. If $m$ is large and $i>0$ then it
follows from \eqref{ref-4.2-41} that $\operatorname {Ext}^i_{{\cal E}}(O(-m),N/NJ^n)=0$ for all $n$.
Hence $`\operatorname {Ext}^i_{{\cal E}}(O(-m),N)=0$ by \eqref{new-4}. Now let $M\in {\cal E}$. To efface $`\operatorname {Ext}^i_{\cal E}(M,N)$
we take an epimorphism $\bigoplus_{i=1}^t O(-n_i)\rightarrow M$ with the $n_i$ sufficiently large. This finishes
the proof.}\end{proof
\end{step}
\def}\end{proof{}\end{proof}
Now we fix an
$\operatorname {Ext}$-finite flat $R/J$-linear noetherian abelian category ${\cal C}$ and
an $R$-deformation ${\cal D}$ of ${\cal C}$.
\begin{theorem}
\label{ref-4.2-44}
Let $O(n)_n$ be a sequence of $R$-flat objects in
${\cal D}$. Then
\begin{enumerate}
\item $(O(n)/O(n)J)_n$ is strongly ample in ${\cal C}$ if and only if $O(n)_n$ is
strongly ample in ${\cal D}$.
\item If $(O(n)/O(n)J)_n$ is strongly ample then ${\cal D}$ is flat (instead of just formally flat).
\item ${\cal D}$ is $\operatorname {Ext}$-finite as $R$-linear category.
\end{enumerate}
\end{theorem}
\begin{proof}
Property (2) follows immediately from (A1) and (1).
To prove (1) we note that since the $O(n)$ are flat we have
\begin{equation}
\label{ref-4.5-45}
\operatorname {Ext}^i_{{\cal D}}(O(n),M)=\operatorname {Ext}^i_{{\cal C}}(O(n)/O(n)J,M)
\end{equation}
for $M\in{\cal C}$. From this
it is easy to see that if $O(n)_n$
is strongly ample in ${\cal C}$ then $(O(n)/O(n)J)_n$ is strongly ample in
${\cal C}$. So our main task is to prove the converse.
So assume that $(O(n)/O(n)J)_n$ is strongly ample. We will first show
that $O(n)_n$ is strongly ample if we replace $\operatorname {Ext}$ by $`\operatorname {Ext}$ in the
definition of strongly ample.
Let $M$ be a noetherian object in ${\cal D}$. Put
$S=R/J\oplus J/J^2\oplus\cdots$. According to Lemma \ref{ref-2.2.2-3}
$N=M/MJ\oplus MJ/MJ^2\oplus\cdots$ is a noetherian graded $(S,{\cal C})$-object.
We claim that for $m$ large we have
\begin{equation}
\label{ref-4.6-46}
\operatorname {Ext}^i_{{\cal D}}(O(-m),MJ^n/MJ^{n+1})=0 \qquad\text{for all $n$}
\end{equation}
This is again proved using a similar inductive method as in the proof of
Lemma \ref{ref-2.3.2-17} and Proposition \ref{ref-4.1-37} (Step 2). We leave the proof to the reader.
In particular we find that for $m$ large one has
\[
\operatorname {Ext}^i_{{\cal D}}(O(-m),M/MJ^n)=0 \qquad\text{for all $n$}
\]
Taking inverse limits and using Proposition \ref{ref-2.5.3-22}
we find $`\operatorname {Ext}^i_{{\cal D}}(O(-m),M)=0$ for all
$i>0$. Hence this proves (A2) for $`\operatorname {Ext}$.
We now prove (A1) (which does not involve any $\operatorname {Ext}$). We first find
$m_0$ such that \eqref{ref-4.6-46} holds for $i=1$ and $m\ge m_0$.
Using (A1) for ${\cal C}$ we find that there is an epimorphism
$F\overset{\text{def}}{=}\bigoplus_{i=1}^tO(-n_i)\rightarrow M/ MJ$ with
$n_i\ge m$. We may lift this map to a compatible series of maps $F\rightarrow
M/MJ^n$. Taking the inverse limit yields a map $F\rightarrow M$. By Nakayama's
lemma (see Lemma \ref{ref-2.2.13-14}) it follows that this must be an
epimorphism.
So we have proved ampleness with $`\operatorname {Ext}$ replacing $\operatorname {Ext}$. Now we
claim that in fact $`\operatorname {Ext}=\operatorname {Ext}$ in ${\cal D}$. This is proved in the same
way as Step 4 of the proof of Proposition \ref{ref-4.1-37} (using condition (A1) which was
already proved).
Property (3) follows from Proposition \ref{ref-2.5.3-22}.
\end{proof}
\section{Lifting and base change}
\label{seclb}
The usual lifting results for infinitesimal deformations generalize
without difficulty to formal deformations. As usual $R$ is a $J$-adic noetherian ring
and we assume that ${\cal D}$ is an $R$-deformation of a noetherian $\operatorname {Ext}$-finite $R/J$-linear
flat abelian category ${\cal C}$. To simplify the notation we assume ${\cal C}={\cal D}_{R/J}$.
\begin{proposition}
\label{ref-5.1-47}
Let $M\in {\cal C}$ be a flat object such that
$\operatorname {Ext}^i_{\cal C}(M,M\otimes_{R/J} J^n/J^{n+1})=0$ for $i=1,2$ and $n\ge 1$.
Then there exists a unique $R$-flat object (up to non-unique
isomorphism) $\overline{M}\in {\cal D}$ such that $\overline{M}/\overline{M}J\cong M$.
\end{proposition}
\begin{proof} This follows in a straightforward way from the infinitesimal lifting
criterion for objects (see \cite[Theorem A]{lowen4} for the dual version).
\end{proof}
\begin{proposition}
\label{ref-5.2-48}
Let $\overline{M},\overline{N}\in {\cal D}$ be flat objects and put
$\overline{M}/\overline{M}J= M$, $\overline{N}/\overline{N}J= N$. Assume that for all $X$
in $\operatorname{mod}(R/J)$ we have $\operatorname {Ext}^i_{\cal C}(M,N\otimes_{R/J} X)=0$ for a certain $i>0$. Then
we have $`\operatorname {Ext}^i_{\cal D}(\overline{M},\overline{N}\otimes_R X)=0$ for all $X\in \operatorname{mod}(R)$.
\end{proposition}
\begin{proof} This follows in a straightforward way from \cite[Prop.\ 6.13]{lowenvdb1}.
\end{proof}
\begin{proposition}
\label{ref-5.3-49}
Let $\overline{M},\overline{N}\in {\cal D}$ be flat objects and put
$\overline{M}/\overline{M}J= M$, $\overline{N}/\overline{N}J=
N$. Assume that for all $X$ in $\operatorname{mod}(R/J)$ we have
$\operatorname {Ext}^1_{\cal C}(M,N\otimes_{R/J} X)=0$. Then $\operatorname {Hom}_{\cal D}(\overline{M},\overline{N})$ is
$R$-flat and furthermore
for all $X$ in $\operatorname{mod}(R)$ we
have $\operatorname {Hom}_{\cal D}(\overline{M},\overline{N}\otimes_R
X)=\operatorname {Hom}_{\cal D}(\overline{M},\overline{N})\otimes_R X$.
\end{proposition}
\begin{proof} This is routine. Choose a short exact sequence
\[
0\rightarrow Y\rightarrow P\rightarrow X\rightarrow 0
\]
with $P$ a finitely generate projective. We then get an exact sequence
\[
0\rightarrow \operatorname {Hom}_{\cal D}(\overline{M},\overline{N}\otimes_R Y)\rightarrow \operatorname {Hom}_{\cal D}(\overline{M},\overline{N}\otimes_R P)\rightarrow \operatorname {Hom}_{\cal D}(\overline{M},\overline{N}\otimes_R X)\rightarrow
\operatorname {Ext}^1_{\cal D}(\overline{M},\overline{N}\otimes_R Y)
\]
By the Proposition \ref{ref-5.2-48} and the hypotheses we get $\operatorname {Ext}^1_{\cal D}(\overline{M},\overline{N}\otimes_R Y)=0$.
We then obtain a commutative diagram
\[
\xymatrix{
0\ar[r]& \operatorname {Hom}_{\cal D}(\overline{M},\overline{N}\otimes_R Y)\ar[r]& \operatorname {Hom}_{\cal D}(\overline{M},\overline{N}\otimes_R P)\ar[r] & \operatorname {Hom}_{\cal D}(\overline{M},\overline{N}\otimes_R X)\ar[r]& 0\\
& \operatorname {Hom}_{\cal D}(\overline{M},\overline{N})\otimes_R Y\ar[u]^{\alpha_Y}\ar[r]& \operatorname {Hom}_{\cal D}(\overline{M},\overline{N})\otimes_R P\ar@{=}[u]\ar[r]
& \operatorname {Hom}_{\cal D}(\overline{M},\overline{N})\otimes_R X\ar[r]\ar[u]_{\alpha_X}& 0 }
\]
We obtain that $\alpha_X$ is an epimorphism for all $X$. But then $\alpha_Y$ is an epimorphism
from which we then deduce that $\alpha_X$ is an isomorphism. But then $\alpha_Y$
is an isomorphism and hence the lower right exact sequence is in fact exact. Thus
$\operatorname {Hom}_R(\overline{M},\overline{N})$ is $R$-flat, finishing the proof.
\end{proof}
\def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\ifx\undefined\leavevmode\hbox to3em{\hrulefill}\,
\newcommand{\leavevmode\hbox to3em{\hrulefill}\,}{\leavevmode\hbox to3em{\hrulefill}\,}
\fi
|
\section{Introduction}
\label{sec_intr}
The foundation of the Ising model became an very important event in the modern
physics. It was the basic tool explaining critical temperatures for which phase
transitions occur in physical systems (see Domb et al~\cite{domb_et_al}). It
is one of the most studied models with wide range of applications in different
sciences. Weidlich used Ising model to explain the polarization phenomena in
sociology. It also has been adapted in economics to explain the diffusion
of technical innovations. New technologies were stated as the result of the
interaction with neighboring firms. Into general equilibrium economics Ising
model was introduced by F\"ollmer~\cite{follmer}. There are several sources
of influence on the price of the stock. One important component is real firm
data; another one, correlation between amount of buyers and sellers. The
action of each trader (he is a buyer or seller) was taken as the value of
the trader. Kaizoji~\cite{kaizoji} introduced an Ising-type model of
speculative activity, which explain bubbles and crashes in stock market. He
introduced the market-maker, who adjusts the price on the market in
dependence of correlation of the buyers and sellers.
After Curie had discovered the critical temperature, Weiss developed a theory
of ferromagnetism based on a spin system. It appears by replacement of the
nearest-neighbor pairs interacting of the Ising model by assumption that
each spin variable interacts with each other spin variable at any site of the
lattice with exactly the same strength.
Some space-time models have been suggested. One model based on the Ising model,
was suggested by J\"arpe~\cite{jarpe1}. The partition function of the model
contains two Hamiltonians, one of which describes previous moment of the time.
This model may be used for describing volatility clustering on the market.
The purpose of this paper is to develop a new space-time model, which is also
describes the volatility clustering without assumptions about the structure
of the lattice. Today information about trading on the markets is available,
e.g.\ in Internet, to everyone. Since there is less space restrictions for
this reason it motivates creating a space-time model baced on the Curie-Weiss
model. Such a model is formally defined in this paper, and some results about
critical temperature of the market is derived according to the distribution of
the Hamiltonian in this model.
In Section~\ref{sec_meth} the model and methods which are used are described.
In Section~\ref{sec_res} all main results are presented. The implications are
discussed in Section~\ref{sec_disc}.
\section{Model and Methods}
\label{sec_meth}
The model which is introduced in this paper
is based on two models. The first is the Curie-Weiss model, which is a simple
modification of the Ising model. It allows all agents in the system to interact
with each other with a constant strength. The second model is the spatio-temporal
model of J\"arpe~\cite{jarpe1} which possesses both spatial and time
dependence. The state of a site in a lattice is depending on the states of
its nearest neighbors and on the global degree of clustering of the previous
pattern.
We took from the Curie-Weiss model the idea of the global interaction and
from the model of J\"arpe the structure of the partition function and obtained
a new space-time model which is appropriate for describing the volatility clustering on
the market.
Let us consider a market which contains $N$ traders symbolically denoted by
$i=1,2,\ldots,N$. In this simple model every trader in a time-period can buy a
fixed amount of stock or sell the same amount. In the first case the "Trader's
decision" is $X_i=1$, in the second case $X_i=-1$. Then $X=(X_1,X_2,\ldots,X_N)$
represents the investment attitude of the market. All traders are neighbors.
That means that each trader knows about the "Trader's decisions" of all others
traders, so his decision is under influence of the others.
A configuration of the model is a specification of "Trader's decisions" of all
traders of the market. With each configuration $x=\{ x_i:i=1,2,\ldots,N\}$
a Hamiltonian or interaction energy,
\[ H(x)\;=\;-\frac 1N\sum_{i=1}^{N-1}\sum_{j=i+1}^N x_ix_j \]
is connected. We will consider this Hamltonian without investment environment.
Let $p(x^k)=P(X=x^k)$ be the probability of observing the state $x^k$ where
$(x^1,x^2,\ldots,x^{2^N})$ is an enumeration of the distinct states of $X$.
Obviously we have that $0\leq p(x^k)\leq 1$ and $\sum_{k=1}^{2^N}p(x^k)=1$.
Further we assume that all states $x$ are possible (i.e $0<p(x^k )<1$).
Now, wanting to minimize the entropy we have an optimization problem of
minimizing
\begin{equation}
\label{opt_problem}
E_P(H(x))\;=\;\sum_{k=1}^{2^N}p(x^k)H(x^k)
\end{equation}
with respect to measure $P$.
Suppose that the energy of each configuration $x_i$ has been determined. The
probability, $P$, that the system has configuration $x$ with energy $H$ if the
configuration at the privious moment of time is given is:
\[ p(x|x')\;=\;P(X_t=x\,|\,X_{t-1}=x')\;=\;Z_{x'}^{-1}\exp\big({\textstyle
\frac{\beta}{N}}H(x)H(x')\big) \]
where $Z_{x'}=\sum_xexp(\frac\beta N H(x)H(x'))$ is a partition function and
$\beta$ is the market temperature describing the strenght of interaction
between the traders.
\subsection{Sufficient statistic}
\begin{prop}
The statistic $H(X_t)$ is minimal sufficient for inference about the
temperature paramter conditional on the previous state, $H(X_{t-1})$.
\end{prop}
\subsection{The Critical Temperature Of The New Time-Space Model}
The behavior of the system in the Curie-Weiss model is described by the equation
\begin{equation}
m\;=\;\tanh m
\end{equation}
where $m(x)=\frac 1N\sum_i x_i$ represents magnetization of the configuration.
This equation allows us to obtain the property of the temperature of the market
and the critical value of the temperature.
We will use the method of Hartmann and Weigt~\cite{hartmann&weigt} to obtain the
equation of the behavior of new model.
\begin{thm}
The equation of the behavior of the system for the new model is
\begin{equation}
m(t)\;=\;\tanh\left(\frac{m(t)m^2(t-1)\beta}{2}\right)
\end{equation}
\end{thm}
\begin{thm}
The critical temperature of the new model is $\beta=2$. For $\beta<2$ the model
$\{X_t:t\in\mathbb{Z}\}$ is stationary in the space-time sense. For $\beta>2$,
the variable $X_t$ is stationary in the spatial sense conditional on $X_{t-1}=
x_{t-1}$ if $|m(t-1)|\leq\sqrt{\frac 2\beta}$.
\end{thm}
When $N\rightarrow\infty$ the partition function has the form
\[ Z\;\approx\;\exp(-N f(m_0(t),m(t-1)))\sqrt{\frac{2\pi}{Nf''(m_0(t),m(t-1))}} \]
We know the function $f$ and the form of the $f''$ is deduced from
\[ f(m(t),m(t-1))\;=\;\textstyle{\frac{1+m(t)}{2}\ln(\frac{1+m(t)}{2})+
\frac{1-m(t)}{2}\ln(\frac{1-m(t)}{2})-\frac\beta 4m^2(t)m^2(t-1)} \]
implying that
\[ f(m(t),m(t-1))\;\big|_{m(t)=0}\;=\;\textstyle{\ln\frac 12} \]
\[ f''(m(t),m(t-1))\;\big|_{m(t)=0}\;=\;\textstyle{1-\frac\beta 2m^2(t-1)} \]
Therefore
\[ Z\;\approx\;2^N\sqrt{\frac{2\pi}{N(1-\frac\beta 2 m^2(t-1))}} \]
Since the Hamiltonian $H$ is sufficient for the temperature parameter, we are
interested in obtaining the distribution of the Hamiltonian. Testing for
dependence will make a null hypothisis assuming independence, and thus we
first consider the distribution of $H$ assuming $\beta=0$. We are interested
in analysis of dependence between $X_i$ and $X_j$ for $i,j=1,\ldots,N$.
\begin{thm}
Let the variables $\{X_i:i=1,\ldots,N\}$ be independent of each other and take
their values in $\{-1, 1\}$ with equal probability $\frac 12$. Then all
non-identical pairwise products are independent, i.e.\ $X_iX_j\bot X_kX_l$
if $i=j,k,l$ or $j=i,k,l$ for any dimension $N$.
\end{thm}
In this paper, we consider a model where $a$ and $b$ are decisions of the
traders. For every trader the probability of their decision to 'buy' or 'sell' is
$\frac 12$ unconditional of the other traders.
Now, let us consider the set of values of the mean field $M=\sum_{i=1}^NX_i$.
For this we have the state space
\[ \left\{{\textstyle -1,-1+\frac 2N,-1+\frac 4N,\ldots,1}\right\} \]
meaning that $M$ can be in $N+1$ states. Consider now values of $M^2$. If $N$
is even, then one value of $M$ is 0 and
\[ M^2\in\left\{{\textstyle 1,(1-\frac 2N)^2,(1-\frac 4N)^2,\ldots,0}\right\} \]
which is to say that $M^2$ could be in $\frac N2+1$ different states. If $N$
is odd, then 0 is not a value of $M$ and
\[ M^2\in\left\{{\textstyle 1,(1-\frac 2N)^2,(1-\frac 4N)^2,\ldots,\frac{1}{N^2}}
\right\} \]
and in that case $M^2$ could be any of $\frac{N+1}2$ states.
Regarding the Hamiltonian $H(X)=-\frac 1N\sum_{i<j}X_iX_j=-\frac N2M^2+\frac 12$
we consequently the statespace of $H$
\[ \left\{{\textstyle\frac{1-N}2,\frac 12-\frac N2(1-\frac 2N)^2,
\frac 12-\frac N2(1-\frac 4N)^2,\ldots,\frac 12}\right\} \]
in the case when $N$ is even. If $N$ is odd we get
\[ \left\{{\textstyle\frac{1-N}2,\frac 12-\frac N2(1-\frac 2N)^2,
\frac 12-\frac N2(1-\frac 4N)^2,\ldots,\frac 12-\frac 2N}\right\} \]
\begin{lem}
If $P(X_i=1)=p$ and $\{X_i:i=1,\ldots,N\}$ are independent, then
$\frac{M(N+1)}{2}\in Bin(N,p)$, where $M=\frac 1N\sum_{i=1}^NX_i$.
\end{lem}
\begin{thm}
The distribution of $\sum_{i<j}X_iX_j$ when $\{X_i\}$ are independent is
\[ P\Bigg(\sum_{i<j}X_iX_j=h\Bigg)\;=\;
\left\{\begin{array}{ll}
{N\choose N/2}\frac{1}{2^N} & \mbox{if }h=-N/2\\
{N\choose(\sqrt{2h+N}+N)/2}\frac{1}{2^{N-1}} & \mbox{if }h\neq -N/2\\
\end{array}\right. \]
\end{thm}
\subsection{Hamiltonian distribution with dependent traders}
Let us now consider the case when the decisions of the traders are not independent.
\begin{thm}
The distribution of $\sum_{i<j}X_iX_j$ when $\{X_i\}$ are not independent is
\[ P\Bigg(\sum_{i<j}X_iX_j=h\Bigg)\;=\;
\left\{\begin{array}{ll}
Z^{-1}e^{-\beta/2}{N\choose N/2}\frac{1}{2^N} & \mbox{if }h=-N/2\\
Z^{-1}e^{h\beta/N}{N\choose(\sqrt{2h+N}+N)/2}\frac{1}{2^{N-1}} & \mbox{if }
h\neq -N/2\\
\end{array}\right. \]
\end{thm}
\subsection{Time dependent process}
From now on we consider the space-time process $X=\{X_t:t\in\mathbb{Z}\}$
where $X_t=\{X_{i,t}:i=1,\ldots,N\}$. Then we have a corresponding sequence
of Mean fields, $\{M_t:t\in\mathbb{Z}\}$ where $M_t=\frac 1N\sum_{i=1}^NX_{i,t}$,
and of Hamiltonians, $\{H_t:t\in\mathbb{Z}\}$ where $H_t=-\frac 1N\sum_{i=1}^{N
-1}\sum_{j=i+1}^NX_{i,t}X_{j,t}$.
\section{Results}
\label{sec_res}
\begin{thm}
The sequence of Mean fields, $\{M_t\}$, is a Markov chain with transition
probabilities
\[ P(M_t=m_t\,|\,M_{t-1}=m_{t-1})\;=\;
{N\choose N(1+m_t)/2}Z^{-1}_{m_{t-1}}\exp\bigg(\frac{\beta(1-Nm_t^2)(1-Nm_{t-1
}^2)}{4N}\bigg) \]
\end{thm}
\begin{thm}
The sequence of Hamiltonians, $\{H_t\}$, is a Markov chain with transition
probabilities
\[ P(H_t=h_t\,|\,H_{t-1}=h_{t-1})\;=\;
\left\{\begin{array}{ll}
{N\choose N/2}Z^{-1}_{h_{t-1}}\exp(-{\textstyle\frac{\beta}{2}}h_{t-1})
& \mbox{if }h_t=\frac 12\\[1mm]
2{N\choose(N+\sqrt{N(1-2h_t)})/2}Z^{-1}_{h_{t-1}}\exp({\textstyle\frac{\beta}{N}}h_t
h_{t-1}) & \mbox{if }h_t\neq\frac 12
\end{array}\right. \]
\end{thm}
\begin{thm}
The conditional expectation of the Hamiltonian is
\[ E(H_t\,|\,H_{t-1}=h_{t-1})\;=\;{\textstyle\frac{N}{h_{t-1}}}\cdot\frac{d\ln
Z_{h_{t-1}}}{d\beta} \]
and the conditional variance is
\[ V(H_t\,|\,H_{t-1}=h_{t-1})\;=\;{\textstyle(\frac{N}{h_{t-1}})^2}\cdot\frac{
d\ln Z_{h_{t-1}}}{d\beta} \]
\end{thm}
\subsection{Asymptotics}
\begin{thm}
In case with independent trander we have for large $N$
\[ P\Bigg(\sum_{i<j}X_iX_j=h\Bigg)\;\approx\;
\frac{1}{\sqrt{\pi(2h+N)}}\exp\bigg(-\frac{2h+N}{4}\bigg) \]
\end{thm}
\begin{thm}
For large $N$ the conditional expectation of the Hamiltonian is
\[ E(H_t\,|\,H_{t-1}=h_{t-1})\;\approx\;\frac{N(1-2h_{t-1})}{4h_{t-1}(
N-\frac\beta 2(1-2h_{t-1}))} \]
and the varinace is
\[ V(H_t\,|\,H_{t-1}=h_{t-1})\;\approx\;\frac{N^2(1-2h_{t-1})^2}{8h_{t
-1}^2(N-\frac\beta 2(1-2h_{t-1}))^2} \]
\end{thm}
\subsection{Stationarity}
The process $\{X_t\}$ is a time-homogeneous Markov chain because
\[ P(X_{t+1}=x\,|\,X_t=x')\;=\;P(X_t=x\,|\,X_{t-1}=x') \]
for all states $x$ and $x'$ and time-points $t$.
\begin{thm}
The process $\{X_t\}$ is time-reversible and has stationary distribution
\[ \pi(x)\;=\;\frac{Z_x}{\sum_xZ_x} \]
where $Z_x=\sum_y\exp(\frac\beta NH(x)H(y))$.
\end{thm}
\subsection{Exact calculations}
\begin{ex}
We obtained the exact distribution of the statistic $\sum_{i<j}X_{i,t}X_{j,t}$.
Now let us calculate this distribution in the case with 10 traders. We wrote a
program in the R language of programming which calculates the probability of all
possible configurations of the system containing $N$ traders, by determining the
value of the Hamiltonian for each configuration and build the matrix {\tt Vec} of
dimensions
\[ 3\,\times\,(\mbox{"the largest possible value of $H$"}-\mbox{"the smallest
possible value of $H$"}) \]
where the first line contains all possible values of $H$, the second line: the
number of the configurations that leads to the corresponding value in the first
line, the third: probability of corresponding value from the first line. In
Figure~\ref{fig_exact_distr} we can see the distribution of $H$ for $N=10$.
\begin{figure}[htbp]
\begin{center}
\leavevmode
\psfig{file=exact_calc.ps,width=65mm}
\caption{\em Exact distitribution of the energy function when there are 10 traders.}
\label{fig_exact_distr}
\end{center}
\end{figure}
In the Table~\ref{table_exact_distr} are the exact numerical values of the
energy distribution.
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|}\hline
$h$ & $-5$ & $-3$ & 3 & 13 & 27 & 45\\ \hline
Number of occurrances & 252 & 420 & 240 & 90 & 20 & 2\\ \hline
Probability & 0.246 & 0.41 & 0.234 & 0.088 & 0.0195 & 0.00195\\ \hline
\end{tabular}
\caption{\em The exact distribution values of the energy function when there are
10 traders.}
\label{table_exact_distr}
\end{center}
\end{table}
\end{ex}
\subsection{Hypothesis test of independence}
If there is only weak interaction between the traders, then the Hamiltonian
is more likely to attain smaller values. If the interactions are stronger,
then larger values of the Hamiltonian is more likely and one talks about
magnetization, which could be an indicator of, or even a cause of, bubbles
and crashes on the market. Therefore methods to state wheter the values of
the observed Hamiltonian is in some dangerous region is of vital importance
to decision makers and inderictly to the whole society.
In order to see if the value of the Hamiltonian deviates from zero to such
an extent that dangerous development is indicated, the correct distribution
is needed. If the Hamiltonian approaches the critical value, the system may
be in a dangerously instable state and a bubble or a crash on the market can
appear.
Assume that we make $n$ observations $x_1,x_2,\ldots,x_n$ of the system.
Then the corresponding Hamiltonian values $h_1,h_2,\ldots,h_n$ are calculated
and categorized into $K$ classes. If the number of traders are 10 (as in the
previous example), then the we may have 6 classes corresponding to the 6
possible states of the Hamiltonian. But of course we may define these classes
in any way we choose.
Strong interactions as opposed to near independence is reflected by the
hypotheses
\[ \left\{\begin{array}{l}H_0:\beta=0\\ H_1:\beta>0\end{array}\right. \]
To have argument for dependence, i.e.\ to prove $H_1$, the statistic
\[ S\;=\;\sum_{k=1}^K\frac{(O_k-E_k)^2}{E_k} \]
may be used, where $O_k$ is the number of observations in klass $k$, and
$E_k$ is the expected number of observations in class $k$ according to the
distribution of $S$ under $H_0$ which is $\chi^2_{k-1}$. Thus the null
hypothesis is rejected at level $\alpha$ of significance for values of
$S$ greater than $C$ which is $1-\alpha$ percentile of the $\chi^2$
distribution with $k-1$ degrees of freedom.
\begin{ex}
Let us consider the data from the Swedish steel market for the October 22,
2008. We have the traders and buyers with certain moments of trading. To
analyze these data we will collect information about trading by dividing the
time on parts of each about 10 minutes. We than have the ten most active
traders (AVA, CSB, DBL, ENS, EVL, MSI, NDS, NON, SHB, SWB) and twenty intervals
of activity. Let us consider the Hamiltonian for these traders. The result is
in Table~\ref{tab_indep_test}.
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|l|c|c|c|c|}\hline
State & $-5$ & $-3$ & $3$ & $[13,45]$\\ \hline
Number of occurances & 2 & 13 & 4 & 1\\ \hline
Expected number of occurance & 4.92 & 8.2 & 4.688 & 2.19\\ \hline
\end{tabular}
\caption{\em The frequencies of observations divided into four classes.}
\label{tab_indep_test}
\end{center}
\end{table}
Thus $N=10$ and $K=4$. The value of the test statistic in this case is
\[ {\textstyle\frac{(2-4.92)^2}{4.92}+\frac{(13-8.2)^2}{8,2}+\frac{(4-4.688)^2}{4.688}+
\frac{(1-2.19)^2}{2.19}}\;=\;5.29\;<\;7,8147\;=\;\chi^2_3 (0.05) \]
This means that we can not reject the hypothesis of independence on level 5\%
of significance (or any lower level). As a matter of fact the $p$-value of
$S$ in this case is $0.15175$ so dependence can not be proved on any
reasonable level of significance.
\end{ex}
\section{Discussion}
\label{sec_disc}
In this thesis we investigated a new space-time model for interacting agents
in the financial market. First we reviewed the history of the Ising model
and some other Ising-type models. Then the Ising model, Curie-Weiss model
and some modifications of these models were formally presented. Also we
considered one way of finding the critical temperature of the market. A
new space-time model was developed and necessary and sufficient conditions
for its stationarity were found. The non-linear sensitivity of market global
properties in terms of temperature parameter changes was investigated. The
critical temperature for this model was analytically derived.
The distribution of the Hamiltonian was analyzed using its dependence with
the magnetization of the market and the exact distribution was calculated.
The conditional expectation and variance of the Hamiltonian were found and
the stationary distribution was obtained. Then the exact distribution of
the Hamiltonian for 10 traders was calculated, and the expected distribution
was confirmed. Hypothesis test for independence between agents was considered
for the Swedish steel market and it showed that there is no evident critical
situation on the market at the time of this dataset.
The parameter reflects how strongly traders are influenced by each other
in the market. It can signalize risk for a crash or a bubble in the market.
Therefore it is very important that its analysis is accurate in such
situations. What remains for future work? We could try a lot of real data
and compare inferential results relying on this model to observable quantities
generally accepted as a measure of health of the situation on the market.
Also we can consider a bigger ammount of traders to have a more exact
$p$-value in the hypothesis testing. Then we can estimate the amount of
interaction for our model using e.g. maximum likelihood estimator. Here
exists also the possibility to develop hypothesis testing based on a time
dependent model. Also interesting to find out how good this model is to
explain volatility clustering.
|
\section{Introduction}
Any finitely generated group can be endowed with a metric, via the Cayley graph of the group with respect to any finite generating set.
The metrics we get on the group from two different finite generating sets may differ greatly, but they will induce the same ``coarse geometry'' on the group, in the sense that there will be a quasi-isometry taking the group endowed with one metric to the group endowed with the other.
The discovery and investigation of algebraic properties of groups that are invariant under quasi-isometries is an active area of geometric group theory.
Though many such properties have been found, there are few results that provide answers to the question that we are interested in here: are there certain types of subgroups whose existence and location are invariant under quasi-isometries?
In this paper, we show that we can answer this in the affirmative in the case of certain commensurizer subgroups of one-ended, finitely presented groups. We recall the definition of a commensurizer (also called a commensurator by some authors):
\begin{defn}
Let $G$ be a group with a subgroup $H$. Then the {\em commensurizer of $H$ in $G$}, denoted Comm$_G(H)$, is the subgroup consisting of all $g \in G$ such that $H \cap gHg^{-1}$ is of finite index in $H$ and in $gHg^{-1}$.
\end{defn}
Our main result is the following. We use $d_{Haus}$ to denote Hausdorff distance.
\\
\noindent {\bf Theorem \ref{Cqiinv}.} {\it
Let $f\colon\thinspace G \to G'$ be a $(\Lambda, K)$-quasi isometry between finitely presented, one-ended groups, and suppose that $H$ is a two-ended subgroup of $G$ with $n$ coends in $G$, for $n\geq 3$.
Then there is a two-ended subgroup $H'$ of $G'$ such that $H'$ has $n$ coends in $G'$ and there exists some constant $y=y(G, H, \Lambda, K)$ such that
$$d_{Haus}(f({\rm Comm}_{G}(H)), {\rm Comm}_{G'}(H'))< y.$$
}
To prove this theorem, we make use of ``quasi-lines'', which were defined by Papasoglu in \cite{papa_qlines}, and were used to prove the main results in that work.
We make use of some of his results, and develop his theory further.
A variety of corollaries follow relatively easily from Theorem \ref{Cqiinv}.
First, we have the following.
\\
\noindent {\bf Corollary \ref{fn_cor}.} {\it
Let $f, G, G', H$ and $H'$ be as in Theorem \ref{Cqiinv}. Then Comm$_G(H)$ is of type $F_n$ if and only if Comm$_{G'}(H')$ is of type $F_n$.
}
\\
This result follows from Theorem \ref{Cqiinv} and the theorem below, which we show by introducing a theory of ``coarse isometries'' (a more general notion than that of quasi-isometries), and applying Kapovich's arguments from \cite{Kap_ggt_book} that show that being of type $F_n$ is a quasi-isometry invariant.
\\
\noindent {\bf Theorem \ref{fn_thm}.} {\it
Let $f \colon\thinspace G \to G'$ be a quasi-isometry between finitely generated groups, and suppose that $C$ is a subgroup of $G$, $C'$ is a subgroup of $G'$, and that $d_{Haus}(f(C), C')<\infty$.
Then $C$ is of type $F_n$ if and only if $C'$ is of type $F_n$, for $n \geq 1$.
}
\\
For the next consequence of Theorem \ref{Cqiinv}, we recall that a JSJ decomposition of a group is a graph of groups decomposition that encapsulates the structure of the different splittings the group admits over two-ended subgroups, in the same way that a JSJ decomposition of a 3-manifold encapsulates the structure of the essential embeddings of annuli and tori into the manifold.
Many different versions of these decompositions of groups have been defined (see \cite{KropJSJ}, \cite{SelaJSJ}, \cite{RipsSela}, \cite{BowCutPts}, \cite{BowJSJ}, \cite{DunSa}, \cite{DunSwen}, \cite{FujiPapa}).
Scott and Swarup defined a version in \cite{SS}, which exists for all one-ended, finitely presented groups.
Papasoglu's work in \cite{papa_qlines} proves the invariance under quasi-isometries of many of the vertex groups of the Scott-Swarup JSJ decomposition.
Theorem \ref{Cqiinv} implies the invariance of some of the remaining vertex groups:
\\
\noindent {\bf Corollary \ref{SSJSJ_cor}.} {\it
Let $f \colon\thinspace G \to G'$ be a quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$, and suppose that $C$ is a vertex group of commensurizer type of the Scott-Swarup JSJ decomposition of $G$.
Then there is a vertex group, $C'$, of commensurizer type of the Scott-Swarup JSJ decomposition of $G'$, which is such that
$$d_{Haus}(f(C), C')<\infty.$$
}
In the case when $G$ and $G'$ are 3-manifold groups, the Scott-Swarup JSJ decompositions are closely related to the JSJ decompositions of the associated 3-manifolds.
In particular, Corollary \ref{SSJSJ_cor} implies the following.
\\
\noindent {\bf Corollary \ref{3man_app}.} {\it
Let $M$ and $M'$ be connected orientable Haken 3-manifolds with incompressible boundary, and with $f\colon\thinspace \pi_1(M) \to \pi_1(M')$ a quasi-isometry. Suppose that $N$ is a nonexceptional Seifert fibered component of the characteristic submanifold of $M$ that meets the boundary of $M$.
Then there is a nonexceptional Seifert fibered component, $N'$, of the characteristic submanifold of $M'$ that meets the boundary of $M'$.
Moreover, if $C$ denotes the subgroup of $\pi_1(M)$ induced by the inclusion of $N$ into $M$ and $C'$ denotes the subgroup of $\pi_1(M')$ induced by the inclusion of $N'$ into $M'$, then
$$d_{Haus}(f(C), C')<\infty.$$
}
In \cite{KapLeeb_Haken}, M. Kapovich and Leeb prove a stronger result that implies Corollary \ref{3man_app}, using different methods. See also \cite{KapLeeb_SFS} for a related result.
Two more corollaries to Theorem \ref{Cqiinv} are results about groups of quasi-isometries of groups.
We denote the set of quasi-isometries from a metric space $X$ to a metric space $Y$, after identifying any maps of finite sup distance from one another, by $QI(X,Y)$, and we denote $QI(X,X)$ by $QI(X)$.
Of particular interest are the sets $QI(G)$, for $G$ a finitely generated group.
These sets form groups themselves, and are often quite complicated and difficult to study.
The following two results are implied immediately from the proof of Theorem \ref{Cqiinv}.
\\
\noindent {\bf Corollary \ref{qi1}.} {\it
Suppose that $G$ is a one-ended, finitely presented group such that $G = $ Comm$_G(H)$ for a two-ended subgroup $H$ of $G$ that has at least three coends in $G$.
Consider $G/H$, with the metric described in section \ref{qi(g)_section}.
Then there is a canonical map $QI(G) \to QI(G/H)$ that is surjective.
}
\\
\noindent {\bf Corollary \ref{qi2}.} {\it
Let $G$ be a one-ended finitely presented group and let $N= \{ 3,4,5,\ldots\} \cup \{ \infty\}$.
For any $n \in N$, let $K_n$ be a maximal collection of two-ended subgroups of $G$ with $n$ coends that have mutually infinite Hausdorff distance, and let $\{ K_n^j\}_{j \in J_n}$ be the partition of $K_n$ into sets of subgroups of quasi-isometric commensurizers.
Then there is a canonical map
$$QI(G) \to \left\{ \prod_{n \in N} \prod_{H \in K_n} QI(M_H, M_{\sigma(H)}) \ : \ \sigma \in \prod_{n \in N} \prod_{j \in J_n} Sym(K_n^j)\right\} ,$$
where, for any two-ended subgroup $H$ of $G$, $M_H$ denotes the set Comm$_G(H)/H$, with a metric given in section \ref{qi(g)_section}, and $Sym(K_n^j)$ denotes the symmetric group on the set $K_n^j$.
}
\\
We note that Corollary \ref{qi1} is very similar in spirit to the main result in \cite{SoucheWiest}.
In section \ref{semi_section}, we prove one more corollary to Theorem \ref{Cqiinv}, which characterizes when a semidirect product of a free group with $\mathbb{Z}$ is quasi-isometric to the direct product of the free group with $\mathbb{Z}$:
\\
\noindent {\bf Corollary \ref{semi}.} {\it
Let $F_n$ be the free group on $n$ generators for any $n \geq 1$.
Then $F_n \rtimes \mathbb{Z}$ is quasi-isometric to $F_n \times \mathbb{Z}$ if and only if it is virtually $F_n \times \mathbb{Z}$.
}
\\
Other results on the topic of quasi-isometric semi-direct products and extensions have been obtained by Alonso and Bridson \cite{AlonsoBridson}, Bridson \cite{Bridson_optimal_isoper_ab_by_free}, and Bridson and Gersten \cite{BridsonGersten_optimal_isoper_torus_bundles}.
Alonso and Bridson give a sufficient condition for a group extension of two arbitrary groups to be quasi-isometric to the direct product of the groups, while the latter two works give necessary conditions for two semi-direct products (of the form $A \rtimes F$ for $A$ abelian and $F$ free in \cite{Bridson_optimal_isoper_ab_by_free} and of the form $\mathbb{Z}^n \rtimes \mathbb{Z}$ in \cite{BridsonGersten_optimal_isoper_torus_bundles}) to be quasi-isometric.
\\
\noindent {\bf Acknowledgements} Much of the following work was done as part of the author's dissertation.
The author gratefully thanks her advisor, Peter Scott, for his help and guidance.
The author also thanks Panos Papasoglu for several helpful discussions, and Mario Bonk, for his help with the proof of Theorem \ref{fn_thm} in the case that $n=1$.
In addition, thanks to Matt Brin and Ross Geoghegan for their encouragement and advice.
\section{Basic definitions and facts about quasi-lines in finitely presented groups}
In this section, we will introduce ``quasi-lines'', which were defined by Papasoglu in \cite{papa_qlines}, and were main objects of study in that work.
We will prove some basic properties of quasi-lines, including justifying why we may think of two-ended subgroups of finitely generated groups as quasi-lines inside the ambient groups.
We will then prove several results about complementary components of certain quasi-lines - in particular, that quasi-lines satisfy conditions $ess(m_0)$ and $iness(m_1)$ for some number $m_0$ and function $m_1$ in the settings that we are interested in.
Some of these results were used implicitly in \cite{papa_qlines}.
We will first set some basic notation and conventions.
Let $X$ be any metric space and let $x \in X$, $Y \subset X$ and $0 \leq r < \infty$.
Then $\overline{Y}$ shall denote the closure of $Y$ in $X$.
We set balls to be open and neighborhoods closed, i.e. let $B_r(x) = \{ z \in X : d(z,x)<r\}$ and $N_r(Y) = \{ z \in X : d(z,Y) \leq r\}$.
Let $Y'$ be another subset of $X$, and we shall denote by $d_{Haus} (Y,Y')$ the Hausdorff distance between $Y$ and $Y'$. I.e.,
$$d_{Haus}(Y,Y') = \inf \{ r \geq 0 : Y \subset N_r(Y') \ {\rm and } \ Y' \subset N_r(Y)\}.$$
If $X'$ is another metric space, then a map $f \colon\thinspace X \to X'$ is a {\em $(\Lambda , K)$ quasi-isometry} if $\Lambda \geq 1 $ and $K \geq 0$ are such that, for any $x_1, x_2 \in X$,
$$\frac{1}{\Lambda} d_X(x_1, x_2) -K \leq d_{X'} (f(x_1), f(x_2)) \leq \Lambda d_X(x_1, x_2) + K,$$
and $X' = N_K(f(X)).$
We say that $f$ is a quasi-isometry if $f$ is a $(\Lambda, K)$ quasi-isometry for some $\Lambda$ and $K$, and in this case we say that $X$ and $X'$ are quasi-isometric.
A map $g \colon\thinspace X' \to X$ is a {\em quasi-inverse} to $f$ if both $\sup_{x \in X}d(x, g \circ f (x))$ and $\sup_{x' \in X'} d(x', f \circ g (x'))$ are finite.
Suppose further that $X$ is a locally finite CW complex.
Then the number of {\em ends} of $X$ is
$$e(X) = \sup |\{ {\rm infinite\ components\ of \ }X-K\} |,$$
where the supremum is taken over all finite subcomplexes $K$ of $X$.
Thus $e(X)$ can take on any value in $\mathbb{Z}_{\geq 0} \cup \{ \infty\}$.
The number of ends of a metric space is a quasi-isometric invariant: if $X$ and $X'$ are quasi-isometric, then $e(X) = e(X')$.
\begin{convention}
We shall assume throughout this paper that all finitely generated groups that we deal with come equipped with a chosen finite generating set, and that all finitely presented groups come equipped with a chosen finite collection of defining relations.
\end{convention}
If $G$ is a finitely generated group, then we shall denote by $\mathscr{C}^1(G)$ the Cayley graph for $G$ with respect to the associated finite generating set.
If $G$ is a finitely presented group, then we shall denote by $\mathscr{C}^2(G)$ the Cayley complex for $G$ with respect to the chosen finite generating set and relations.
We recall that $\mathscr{C}^2(G)$ is a simply connected, 2-dimensional CW complex with 1-skeleton equal to $\mathscr{C}^1(G)$, and that $G$ is identified with the vertex sets of $\mathscr{C}^1(G)$ and $\mathscr{C}^2(G)$.
The group $G$ acts cocompactly and by cellular isometries on $\mathscr{C}^1(G)$, and on $\mathscr{C}^2(G)$ if it exists.
Furthermore, this action is faithful, and transitive on vertices.
We will take these actions of $G$ to be on the left.
Recall that the Cayley graphs (and Cayley complexes, if they exist) of a group $G$ with respect to different finite generating sets (finite presentations, respectively) are all quasi-isometric.
Thus there is a well-defined notion of the number of ends of a finitely generated group: if $G$ is finitely generated, then the number of ends of $G$, $e(G)$, is defined to equal $e(\mathscr{C}^1(G))$.
It is a fact that the number of ends of any finitely generated group can be one of only $0,1,2$ or $\infty$.
We have that $e(G) = 0$ if and only if $G$ is finite, $e(G) = 2$ if and only if $G$ contains a finite index infinite cyclic subgroup, and, by the work of Stallings, $e(G) = \infty$ if and only if $G$ splits over a finite subgroup and does not have a finite index infinite cyclic subgroup.
See \cite{ScottWall} for more details.
We will always take CW complexes to be metrized to have edges of length one, and to have the interiors of 2-cells be isometric to regular polygons.
Let $X$ be a CW complex and consider $\mathbb{R}$ as a graph by taking the integer points to be the vertices.
Then let $l\colon\thinspace \mathbb{R} \to X^{(1)}$ be continuous, injective and cellular, hence parametrized by arc length, i.e. with $length (l([s,t])) = d_\mathbb{R}(s,t)$, for all $s,t \in \mathbb{R}$. Suppose further that $l$ is a uniformly proper map, so for every $M>0$, there exists an $N>0$ such that if $A \subset X$ with diam$(A)<M$, then diam$(l^{-1}(A))<N$. Then we shall say that $l$ is a {\it line}, and we shall sometimes use $l$ to denote the image of $l$.
To a line $l$, we associate the distortion function $D_l(t)\colon\thinspace \mathbb{R}_{\geq 0} \to \mathbb{R}_{\geq 0}$, where
$$D_l(t) = \sup \{ {\rm diam}(l^{-1}(A)) : {\rm diam}(A) \leq t \}.$$
As $l$ is parameterized by arc length, we note that $D_l(t) \geq t$ for all $t \in \mathbb{R}$.
Let $L$ be a closed, path connected subspace of $X$ containing a line $l$, with $N>0$ such that any point in $L$ can be joined to $l$ by a path in $L$ of length less than or equal to
$N$. If $\phi\colon\thinspace \mathbb{R}_{\geq 0} \to \mathbb{R}_{\geq 0}$ is a proper, increasing function and, for all $t >0$, $D_l(t) \leq \phi(t)$, then we will say that $L$ is a {\it $(\phi, N)$ quasi-line}, or simply, a {\it quasi-line}. We shall refer to $\phi$ and $N$ as {\em parameters for $L$}, and $l$ as the line {\em associated to $L$}.
We note that the assumption that $\phi$ be proper and increasing is not a strong one, for any line $l$ has $D_l$ bounded by some increasing function, and, as $D_l(t) \geq t$ for all $t$, any such function must be proper.
Observe that if $L$ is a $(\phi, N)$ quasi-line and $R>0$, then $N_R(L)$ is a $(\phi, N+R)$ quasi-line, and we may take the line associated to $L$ to also be associated to $N_R(L)$.
The following lemma shows that the restriction that a line be embedded is not an important one.
\begin{lem}\label{get_line}
Let $l'$ be a uniformly proper cellular map from $\mathbb{R}$ into a CW complex $X$ (taking $\mathbb{R}$ to be a simplicial complex with vertex set $\mathbb{Z}$).
Then there is a line $l \colon\thinspace \mathbb{R} \to X$ with Im$(l) \subset $ Im$(l')$, and such that $d_{Haus}($Im$(l),$Im$(l'))$ is finite, and bounded above by a function of $D_l$.
\end{lem}
\begin{proof}
If $l'$ is an embedding, then it suffices to take $l = l'$. So suppose that $l'$ is not an embedding.
As $l'$ is uniformly proper, there is some maximal $n = n(l') \in \mathbb{R}$ such that the preimage of some point in Im$(l')$ has diameter $n$.
Because $l'$ is cellular, note that $n \in \mathbb{Z}$.
We shall induct on $n$.
Note that $n>0$, as $l'$ is not an embedding.
Let $S$ denote a maximal disjoint set of closed intervals of size $n$ in $\mathbb{R}$ such that the endpoints of each interval are sent to the same point of $X$ by $l'$.
Note that we can take $S$ to be such that the endpoints of each interval in $S$ are mapped by $l'$ to vertices of $X$.
Let $\iota \colon\thinspace \mathbb{R} \to \mathbb{R}$ denote the quotient map attained by identifying each component of $S \subset \mathbb{R}$ to a point, and define a right inverse to $\iota$, $\iota'$, to take each such point to an endpoint of its full preimage.
Since the endpoints of each component of $S$ are identified by $l'$, there is a well-defined, continuous map $l_1: \mathbb{R} \to X$ defined by $l_1(t) = l'\circ \iota(t)$.
Intuitively, $l_1$ is the map we get by removing a disjoint collection of maximal loops from $l'$.
Clearly $l_1$ is cellular, thus is parameterized by arc length, and we note that
$l_1$ is uniformly proper, for if $A$ is any subset of $X$, then diam$(l_1^{-1}(A)) \leq $ diam$((l')^{-1}(A))$.
Furthermore, we have that $d_{Haus}($Im$(l_1), $ Im$(l'))\leq \frac{1}{2}n(l')$.
It remains to show that $n(l_1)<n(l')$.
To see this, let us suppose that there are $t_0, t_1 \in \mathbb{R}$ such that $|t_0-t_1| \geq n(l')$, and $l_1(t_0) = l_1(t_1)$.
Suppose that $t_0$ is the image of a collapsed segment under $\iota$.
Then there exist two points $s, s' \in \mathbb{R}$ that are the endpoints of this segment, with $\iota (s) = \iota(s') = t_0$, $l'(s) = l'(s')$, and $|s-s'| = n(l')$.
If $t_1$ is the image under $\iota$ of only a point, then let $s_1$ denote that point.
If $t_1$ is the image of a segment under the map $\iota$, then let $s_1$ denote an endpoint of that segment.
Then $l'(s_1) = l'(s) = l'(s')$, and either $|s_1-s|>|s-s'|$ or $|s_1-s'|>|s-s'|$. But $|s-s'|=n(l')$, so this contradicts the definition of the function $n$.
Thus we may suppose that $t_0$ and $t_1$ are images under $\iota$ of single points, say $s_0$ and $s_1$ respectively. If $\iota$ collapses no segments in the interval $[s_0, s_1]$ then we reach another contradiction, for $[s_0, s_1]$ must be an interval of size $n(l')$, whose endpoints are mapped to the same vertex of $X$ by $l'$, and that is disjoint from $S$. This contradicts the maximality of $S$.
Finally, suppose that $\iota$ collapses a segment $[s,s']$ in $[s_0, s_1]$. Then $|s-s'| = n(l')$, so $|s_0- s_1|>n(l')$. The endpoints $s_0$ and $s_1$ must share the same image under $l'$, and again this contradicts the definition of the function $n$.
Thus $n(l_1)<n(l')$. If $l_1$ is not an embedding, then we can repeat this process on $l_1$, getting a map $l_2 \colon\thinspace \mathbb{R} \to X$ such that $n(l_2)<n(l_1)$, and so on.
Eventually we must get a map $l_k$ such that $n(l_k) = 0$, and hence $l_k=l$ is the desired line.
\end{proof}
Next, we will establish some basic facts about quasi-lines.
First, we give the connection between quasi-lines and infinite cyclic subgroups.
Recall that $G$ is identified with the vertex set of $\mathscr{C}^1(G)$.
\begin{lem}\label{2ended_subgps_are_qlines}
Let $G$ be a finitely generated group, and $H \subset G$ a two-ended subgroup. If $R>0$ is large enough so that $N_R(H)\subset \mathscr{C}^1(G)$ is connected, then $N_R(H)$ is a quasi-line.
\end{lem}
\begin{proof}
Let $H$ be a two-ended subgroup of $G$, let $\langle h \rangle \cong \mathbb{Z}$ be a finite index subgroup of $H$, and let $R>0$ be such that $N_R(H)\subset \mathscr{C}^1(G)$ is connected. Let $p$ be a simplicial path in $N_R(H)$ from the identity to $h$. Let $l' \colon\thinspace \mathbb{R} \to N_R(H) \subset \mathscr{C}^1(G)$ be the natural map onto $\cup_{n \in \mathbb{Z}} h^n\cdot p$ that is parameterized by arc length.
Note that there is some $N>0$ depending on $\langle h \rangle$ and $R$ such that each point of $N_R(H)$ can be connected to Im$(l')$ by a path in $N_R(H)$ of length less than or equal to $N$. If we can show that $l'$ is uniformly proper, then it will follow from Lemma \ref{get_line} that $N_R(H)$ is a quasi-line.
Fix any $M>0$, and recall that $p$ is a finite path and that $H$ acts freely on $\mathscr{C}^1(G)$.
It follows that, for any point $p_0 \in p$, the image under $(l')^{-1}$ of the vertices contained in $B_M(p_0)$ is finite, and hence $diam((l')^{-1}(B_M(p_0)))<\infty$.
Let $D_M = \max \{ diam((l')^{-1}(B_M(p_0))) \}$, where the maximum is taken over all vertices $p_0$ of $p$.
Now suppose that $A \subset \mathscr{C}^1(G)$ has diameter less than $(M-N)$ and meets $N_R(H)$.
Then for any $a \in (A \cap N_R(H))$, certainly $A \subset B_{M-N}(a)$, and hence there is some $p_1 \in l'$ of distance no more than $N$ from $a$, and $A \subset B_M(p_1)$.
Thus there exists some $n \in \mathbb{Z}$ such that $h^nA \subset B_M(h^np_1)$ and $h^np_1$ is a vertex of $p$.
Therefore we have that
$$diam((l')^{-1}(A)) = diam((l')^{-1}(h^nA)) \leq diam ((l')^{-1}(B_M(h^np_1))) \leq D_M,$$
and it follows that $l'$ is uniformly proper, with $D_{l'}(M-N)$ bounded by $D_M$ for each $M>0$.
\end{proof}
The next lemma shows that quasi-lines are suitably invariant under quasi-isometries.
\begin{lem}\label{imgs_of_qlines_are_qlines}
Let $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a $(\Lambda, K)$ quasi-isometry and let $L \subset \mathscr{C}^1(G)$ be a $(\phi, N)$ quasi-line. Then there is some $R = R(\Lambda, K, \phi, N)> 0$ such that $N_R(f(L))$ is a quasi-line.
\end{lem}
\begin{proof}
Let $l$ be the line associated to $L$ and let $\pi$ denote nearest point projection of $\mathscr{C}^1(G')$ onto its vertex set.
Let $l'$ denote $\pi \circ f \circ l |_\mathbb{Z}$, and note that $l'$ is uniformly proper.
As $l$ is a line, for each $i \in \mathbb{Z}$, $d(l(i), l(i+1)) = 1$.
Hence $d (f \circ l(i), f\circ l(i+1)) \leq (\Lambda + K)$ so $d(l'(i), l'(i+1)) \leq (\Lambda + K + 1)$.
Let $\psi \colon\thinspace \mathbb{Z} \to \mathbb{R}$ be such that, for all $i\in \mathbb{Z}$, $\psi(i+1)-\psi(i) = d(l'(i), l'(i+1))$, so $\psi$ increases distances by no more than a factor of $(\Lambda + K + 1)$.
Let $l'' \colon\thinspace Im(\psi) \to \mathscr{C}^1(G')$ be such that $l' = l'' \circ \psi$, and it follows that $l''$ is uniformly proper.
Next, we extend the definition of $l''$ to all of $\mathbb{R}$ by mapping each $[\psi(i), \psi(i+1)]$ isometrically to a geodesic segment from $l'(i)$ to $l'(i+1)$.
Thus $l''$ is a unit speed, cellular map of $\mathbb{R}$ into $\mathscr{C}^1(G')$.
As any point in $l''(\mathbb{R})$ is of distance no more than $\frac{1}{2}(\Lambda + K + 1)$ from a point of $l''(Im(\psi))$ and $l''|_{Im(\psi)}$ is uniformly proper, it follows that $l'' \colon\thinspace \mathbb{R} \to \mathscr{C}^1(G')$ is uniformly proper.
Thus, by Lemma \ref{get_line}, there is a line $\hat{l} \colon\thinspace \mathbb{R} \to \mathscr{C}^1(G')$ such that $Im(\hat{l}) \subset Im(l'')$ and $d_{Haus}(Im(\hat{l}), Im(l''))$ is bounded by a constant depending only on $\Lambda, K, \phi$ and $N$.
There is some $R>0$ depending on the same constants such that $N_{R}(f(L))$ is connected and contains $Im(\hat{l})$, and thus is a quasi-line.
\end{proof}
\begin{rmk}\label{qlines_have_same_consts}
In the argument to prove our main result, Theorem \ref{Cqiinv}, we shall work with a quasi-isometry $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$, a two-ended subgroup $H$ of $G$, and its translates $\{ gH\}$.
We will also be concerned with $f(H)$ and its translates $\{ (g')f(H)\}$.
The previous results show that $H$ and $f(H)$ have neighborhoods that are quasi-lines.
Since left multiplication is an isometry in any Cayley graph, it follows that the translates of $H$ and $f(H)$ have neighborhoods that are quasi-lines, with the same parameters as those containing $H$ and $f(H)$ respectively.
\end{rmk}
We note next that quasi-lines are two-ended:
\begin{lem}\label{e(L)=2}
Let $L$ be a quasi-line contained in a locally finite CW complex $X$. Then $e(L)=2$.
\end{lem}
\begin{proof}
Let $\phi$ and $N$ be parameters for $L$, and let $l \subset L$ be the line associated to $L$.
Then every point in $L$ can be connected by a path of length less than or equal to $N$ to $l$, and $e(l) = 2$, so $e(L) \leq 2$.
As the line $l$ is an injective and cellular map into the 1-skeleton of $X$, and $X$ is locally finite, we must have that $e(L) \geq 1$.
To see that $e(L)=2$, first note that if $a, b \in \mathbb{R}$ are such that $|a-b|>\phi(2N)$, then $d(l(a),l(b))>2N$. Thus if we fix such $a$ and $b$, say with $a<b$, then, for any $q \in l((-\infty, a])$ and $q' \in l([b, \infty ))$, $d(q,q')>2N$. Let $K$ be the set of all points $p \in L$ such that there is a path of length less than or equal to $N$ contained in $L$ that connects $p$ to $l((a,b))$. Since $X$ is locally finite, $K$ is compact. Thus $L-K$ contains two infinite components - one intersecting $l((-\infty, a])$ and one intersecting $l([b,\infty))$.
\end{proof}
The following definitions will be important, particularly to the remainder of this section.
\begin{defn} \label{ess_iness_ness_def}
If $L$ is a quasi-line in a metric space $X$, then a connected component $C$ of $X-L$ is said to be {\em essential} if $C \cup L$ has one end. Otherwise, $C$ is said to be {\em inessential}.
If $C$ is not contained in $N_R(L)$, for any $R \geq 0$, then we shall say that $C$ is {\em nearly essential}.
\end{defn}
\begin{defn}
If there is some $m_0>0$ such that, for each $p \in L$, each essential component of the complement of $L$ intersects a vertex of $B_{m_0}(p)$, then we say that $L$ satisfies {\em ess}$(m_0)$.
\end{defn}
\begin{defn}
If $m_1 \colon\thinspace \mathbb{R}_{\geq 0} \to \mathbb{R}_{\geq 0}$ is such that, for each $R \geq 0$, each inessential component of the complement of $N_R(L)$ is contained in the $m_1(R)$-neighborhood of $N_R(L)$, then we say that $L$ satisfies {\em iness}$(m_1)$.
\end{defn}
We will see shortly that all the quasi-lines in which we are interested (see Remark \ref{qlines_have_same_consts}) will satisfy $iness(m_1)$ for some $m_1$. It follows that the components of the complements of these quasi-lines are essential if and only if they are nearly essential.
\begin{defn} $L$ is said to be {\it $n$-parting} if the complement of $L$ has at least $n$ essential components.
\end{defn}
Note that, if $L$ is $n$-parting and $R>0$, then $N_R(L)$ is also $n$-parting.
Next, we shall show that any quasi-line in a one-ended finitely presented group has only finitely many essential components in its complement, and moreover satisfies $ess(m_0)$ for some $m_0$.
\begin{defn} \label{defn_chain}
Let $L$ and $C$ denote subsets of a metric space $X$, let $n>0$, and let $x,y \in C \cap L$. Then we shall say that $x$ and $y$ are connected by an {\em $(L,n)$-chain in $C \cap L$} if there are points $x=z_0, z_1, \ldots, z_k =y$ in $C\cap L$ such that, for each $i$, there is a path in $L$ connecting $z_i$ to $z_{i+1}$ of length less than or equal to $n$.
\end{defn}
We will begin by working in $\mathscr{C}^2(G)$.
\begin{lem}\label{CC_lem}
Let $G$ be a finitely presented group, let $L$ be a $(\phi, N)$ quasi-line in $\mathscr{C}^1(G)$, and let $0<\epsilon \ll 1$.
Then let $L'$ be an open set in $\mathscr{C}^2(G)$ such that $L \subset L'$, each point in $L'$ can be connected to $L$ by a path in $\mathscr{C}^2(G)$ of length less than $\epsilon$, and each point in $L' \cap \mathscr{C}^1(G)$ can be connected to $L$ by a path in $\mathscr{C}^1(G)$ of length less than $\epsilon$.
Suppose in addition that both $\overline{L'}$ and $\overline{L'} \cap \mathscr{C}^1(G)$ deformation retract onto $L$.
Then there is some $n_0 = n_0 (G, \phi, N, \epsilon)$ such that, if $C'$ denotes any component of $\mathscr{C}^2(G)-L'$, and $x, y \in C'\cap \overline{L'}$, then $x$ and $y$ are connected by an $(\overline{L'}, n_0)$-chain $\{ z_i\}$ in $C' \cap \overline{L'}$. Moreover, the path in $\overline{L'}$ connecting any $z_i$ and $z_{i+1}$ can be taken to be in $L$, outside of an initial segment containing $z_i$ and a final segment containing $z_{i+1}$, each of length less than or equal to $\epsilon$.
\end{lem}
To prove this lemma, we first need the following.
\begin{lem}\label{L''_simp_conn}
Let $G, \mathscr{C}^2(G), L, L'$, and $\epsilon$ be as in Lemma \ref{CC_lem}.
Let $(\mathscr{C}^2)''$ denote the union of $\mathscr{C}^2(G)$ together with a disk added at each closed edge path of $\overline{L'}$ of length less than or equal to $\phi (2(N+\epsilon)+1) + 2(N + \epsilon)+1$, and let $\overline{L''}$ denote the union of $\overline{L'}$ with these disks.
Then $\overline{L''}$ is simply connected.
\end{lem}
\begin{proof}
Let $l$ be the line associated to $L$, so that $l \subset L \subset L'$, and note that $\overline{L'}$ is a $(\phi , N+\epsilon)$ quasi-line.
Recall that the edges of $\mathscr{C}^2(G)$ are of length one, and metrically, the 2-cells of $\mathscr{C}^2(G)$ are regular polygons.
Let the disks added to create $(\mathscr{C}^2)''$ and $\overline{L''}$ be regular polygons as well.
Let $\gamma \colon\thinspace S^1 \to \overline{L''}$ be any closed curve in $\overline{L''}$.
In the following we shall replace $\gamma$ with homotopic curves (which we shall also call $\gamma$), and we shall assume at each stage that $\gamma$ is parameterized by arc length.
Thus we shall consider copies of $S^1$ with different metrics as appropriate below.
Note first that $\gamma$ can be homotoped in $\overline{L''}$ to a cellular path that is contained in $\overline{L'} \cap [\mathscr{C}^2(G)]^{(1)} =\overline{L'} \cap ((\mathscr{C}^2)'')^{(1)}$. Let $\gamma$ now denote the resulting curve under this homotopy.
Fix a finite collection $P$ of points in $Im(\gamma) - l$ such that the $(1/2)$-neighborhood of $P$ contains $Im(\gamma) - l$.
At each point $p_0$ in $P$, consider a spike from $p_0$ to $l$ in $\overline{L'}$ of length no more than $(N+\epsilon)$.
We shall homotope $\gamma$ to traverse the appropriate spike each time it meets a point of $P$.
We have thereby constructed a homotopy of $\gamma$ to a curve that is contained in $l$ except for finitely many segments in $\overline{L'} \cap [\mathscr{C}^2(G)]^{(1)}$ of length bounded by $2(N+\epsilon)+1$
Let $\Sigma$ denote the components of the complement of $\gamma^{-1}(l)$ in $S^1$, so each component of $\Sigma$ has length no more than $2(N+\epsilon) + 1$.
We will prove, by induction on $|\Sigma|$, that $\gamma$ must be null-homotopic in $\overline{L''}$.
Note that, after adding the spikes, $\gamma$ must meet $l$, so $\Sigma$ must consist of at least one segment.
If $|\Sigma| = 1$, then we shall let $\sigma$ denote the element in $\Sigma$, and let $p$ and $q$ denote the endpoints of $\sigma$.
Then the distance between $\gamma(p)$ and $\gamma(q)$ in $\overline{L'} \cap ((\mathscr{C}^2)'')^{(1)}$ must be bounded by $2(N+\epsilon) + 1$, so $\gamma(p)$ and $\gamma(q)$ are connected in $l$ by a path of length no more than $\phi(2(N+\epsilon)+1)$.
It follows that this path in $l$ together with $\gamma(\sigma)$ make up a circuit in $\overline{L'}$ of length no more than $\phi(2(N+\epsilon) + 1) + 2(N+\epsilon) + 1$, and thus must bound a disk in $\overline{L''}$.
Thus $\gamma$ may be homotoped in $\overline{L''}$ so that $\gamma(\sigma)$ is taken to this path in $l$, i.e. $\gamma$ may be homotoped within $\overline{L''}$ to lie entirely inside of $l$.
But $l$ is an embedding of $\mathbb{R}$, so $\gamma$ is null-homotopic in $\overline{L''}$.
Next, assume that $|\Sigma|=i>1$. By choosing any element $\sigma$ of $\Sigma$, the same argument as was given above shows that we can homotope $\gamma$ within $\overline{L''}$ so that $\sigma$ is replaced by a path in $l$.
Thus we have homotoped $\gamma$ so that now $|\Sigma| = i-1$.
By repeating this process, we reduce to the case of $|\Sigma|=1$ above, and hence $\gamma$ must have originally been null-homotopic in $\overline{L''}$.
It follows that $\overline{L''}$ is simply connected, as desired.
\end{proof}
\begin{proof}[Proof of Lemma \ref{CC_lem}]
Again let $l$ be the line associated to $L$, and let $(\mathscr{C}^2)''$ and $\overline{L''}$ be as in Lemma \ref{L''_simp_conn}.
Recall that $\overline{L''}$ differs from $\overline{L'}$ only by disks with boundary in $\overline{L'}$, and that these disks are not contained in $\mathscr{C}^2(G)$.
It follows that the intersection of $\overline{L'}$ with the closures of the components of $\mathscr{C}^2(G)-L'$ is the same as the intersection of $\overline{L''}$ with the closures of the components of $(\mathscr{C}^2)''-\overline{L''}$.
Thus, if we can prove that the conclusion of Lemma \ref{CC_lem} holds for $\overline{L''}$ in $(\mathscr{C}^2)''$, then the lemma will follow for $\overline{L'}$ in $\mathscr{C}^2(G)$.
Let $\{ C_\alpha \}$ denote the components of the complement of $\overline{L''}$.
As $\overline{L''}$ is simply connected, we can apply Van Kampen's theorem to $\{ \overline{C_\alpha} \cup \overline{L''}\}$ to see that each $\overline{C_\alpha} \cup \overline{L''}$ is simply connected. Since $\overline{L''}$ and each $\overline{C_\alpha}$ are connected, we have that $\overline{C_\alpha} \cap \overline{L''}$ is connected.
Thus for any fixed $x,y \in \overline{C_\alpha} \cap \overline{L''}$, there exists a path $p$ from $x$ to $y$ contained in $\overline{C_\alpha} \cap \overline{L''}$.
Without loss of generality, we can assume that the frontier of $\overline{L'}$ (which equals the frontier of $\overline{L''}$) meets any edge of $\mathscr{C}^1(G) = (\mathscr{C}^2)^{(1)} = ((\mathscr{C}^2)'')^{(1)}$ in only finitely many points.
The group $G$ has one end and $\mathscr{C}^2(G)$ is simply connected, thus each edge of $\mathscr{C}^2(G)$ is contained in a 2-cell. The same is true for $(\mathscr{C}^2)''$, thus we can take $p$ to be transverse to $((\mathscr{C}^2)'')^{(1)}$, with $|\{ p \cap ((\mathscr{C}^2)'')^{(1)}\} |$ finite and $p$ still contained in $\overline{C_\alpha} \cap \overline{L''}$.
Let $z_0, z_1, \ldots , z_{k}$ denote the elements of $p \cap ((\mathscr{C}^2)'')^{(1)}$, numbered in the order in which they are traversed by $p$, traveling from $x$ to $y$, with $z_0=x, z_k=y$.
Recall that $G$ is finitely presented, so the 2-cells of $\mathscr{C}^2(G)$ are of bounded perimeter.
Also the additional 2-cells added to create $(\mathscr{C}^2)''$ have diameter bounded by a function of $\phi, N$ and $\epsilon$, so there is a bound $n_0 = n_0(G, \phi, N, \epsilon)$ on the perimeters of the 2-cells of $(\mathscr{C}^2)''$.
Any component of $p-(\mathscr{C}^2)^{(1)}$ must be contained in such a cell, thus its interior can be replaced by a segment in $(\mathscr{C}^2)^{(1)}$
with length less than $n_0$.
Recall that $p$ was originally contained in the frontier of $\overline{L'}$, $\overline{L'} \subset N_\epsilon (L)$ with $\epsilon \ll 1$, and both $\overline{L'}$ and $\overline{L'} \cap \mathscr{C}^1(G)$ deformation retract onto $L$.
Hence we can replace each segment of $p-(\mathscr{C}^2)^{(1)}$ with a segment in $(\mathscr{C}^2)^{(1)}$ of length less than $n_0$ that is contained in $L$ except for initial and terminal segments of length bounded by $\epsilon$.
Thus the $z_i$'s form an $(\overline{L''}, n_0)$-chain from $x$ to $y$ as desired.
\end{proof}
We can now prove that quasi-lines in finitely presented groups satisfy $ess(m_0)$.
\begin{lem}\label{compl_has_fin_many_comps}
Let $G$ be a one-ended finitely presented group, with $L$ a $(\phi,N)$ quasi-line in $\mathscr{C}^1(G)$. Then $\mathscr{C}^1(G)-L$ contains only finitely many essential components. Moreover, there is some $m_0=m_0(G, \phi , N)$ such that $L$ satisfies $ess(m_0)$.
\end{lem}
\begin{proof}
We shall prove that $L$ satisfies $ess(m_0)$, for some $m_0>0$. Since $\mathscr{C}^1(G)$ is locally finite, it will follow that the complement of $L$ contains finitely many essential components.
Let $C$ be a component of the complement of $L$ in $\mathscr{C}^1(G)$. We shall use Lemma \ref{CC_lem} to show that there is some $n>0$ (not depending on our choice of $C$) such that any $x,y \in (\overline{C} \cap L)$ are connected by an $(L,n)$-chain in $(\overline{C} \cap L) \subset \mathscr{C}^1(G) \subset \mathscr{C}^2(G)$.
Let $L'$ be as defined in Lemma \ref{CC_lem}, so, for some $0<\epsilon \ll 1$, $L \subset L' \subset N_\epsilon (L)$, $L'$ is open in $\mathscr{C}^2(G)$, and both $\overline{L'}$ and $\overline{L'} \cap \mathscr{C}^1(G)$ deformation retract onto $L$.
Recall that in Lemma \ref{CC_lem}, we proved a result similar to that desired now, but for $\overline{L'}\subset \mathscr{C}^2(G)$.
Fix any such $x$ and $y$, and, as $\overline{C}$ is connected, there is a simple oriented edge path $p$ in $\overline{C}$ connecting them.
Note that our assumptions on $L'$ imply that each edge in $p$ must meet one component of $\mathscr{C}^2(G)-L'$.
Thus $p$ is a union of edge paths $p_1, p_2, \ldots , p_k$ such that, for each $i$, the terminal vertex of $p_i$ is equal to the initial vertex of $p_{i+1}$, and each $p_i$ intersects $L'$ in components of length no more than $\epsilon$ containing its initial and terminal vertices, with the rest of $p_i$ contained in some component $C'$ of $\mathscr{C}^2(G)-L'$.
Let $x_i$ and $y_i$ denote the two points of $(C' \cap \overline{L'})$, so each is within $\epsilon$ of a different endpoint of $p_i$.
By Lemma \ref{CC_lem}, each pair $x_i$ and $y_i$ can be connected by an $(\overline{L'}, n_0)$-chain $\{ z_j'\}$ in $C' \cap \overline{L'}$. Recall that, moreover, a path of length no more than $n_0$ between any two consecutive points in the chain is in $L$, outside of initial and final segments of length no more than $\epsilon$, and that $\overline{L'} \cap \mathscr{C}^1(G)$ deformation retracts onto $L$.
Thus, each $z_j'$ can be connected by a path of length no more than $\epsilon$ in $\overline{L'} \cap \mathscr{C}^1(G)$ to a point $z_j \in \overline{C} \cap L$ such that $\{ z_j\}$ forms an $(L, n_0 + 2\epsilon)$-chain in $\overline{C}\cap L$, connecting the endpoints of $p_i$. Concatenating these chains, we see that, for $n=n_0+2\epsilon$, $x$ and $y$ can be connected by an $(L,n)$-chain in $\overline{C}\cap L$ as desired.
From now on, we shall work only in $\mathscr{C}^1(G)$, not $\mathscr{C}^2(G)$.
We shall now find an $m_0>0$ such that there is an $(L,n)$-chain in the frontier of each essential component $C$ of the complement of $L$ that must intersect the $m_0$-ball about any given point of $L$.
Fix any $a \in L$ and $R\gg N, n$. As $C$ is essential, $e(C \cup L)=1$, and, from Lemma \ref{e(L)=2}, recall that $L$ must have two ends. It follows then that $C$ must intersect both unbounded components of $L-B_R(a)$;
let $x$ be in the intersection of $\overline{C}\cap L$ with one, and $y$ in the intersection of $\overline{C}\cap L$ with the other.
By the work above, there exists an $(L,n)$-chain, $\{ z_i\}$, from $x$ to $y$ in $\overline{C}\cap L$.
Recall that $L$ is a $(\phi,N)$ quasi-line, and let $l$ be the line associated to $L$. Then, for each $i$, there is a path in $L$ of length less than or equal to $N$ connecting $z_i$ to some $w_i \in l$. For each $i$, $d(z_i, z_{i+1})\leq n$, thus $d(w_i,w_{i+1})\leq n+2N$, and thus the path in $l$ between any two adjacent $w_i$'s has length less than or equal to $\phi (n+2N)$.
Let $a_0 \in l$ be of distance less than or equal to $N$ from $a \in L$. As $R\gg 0$, $x$ and $y$ are such that there is some $i$ with $l^{-1}(w_i) \leq l^{-1}(a_0) \leq l^{-1}(w_{i\pm 1})$, and hence, for some $j$, $d(a_0, w_j)\leq \frac{1}{2}\phi (n+2N)$. Thus
$$d(a,z_j) \leq d(a, a_0) + d(a_0, w_j) + d(w_j, z_j) \leq \frac{1}{2}\phi (n+2N)+2N.$$
Since $z_j \in C$, and $z_j$ is of distance less than 1 from a vertex of $C$, it follows that, for any $m_0\geq [\frac{1}{2}\phi (n+2N)+2N+1]$, $C$ intersects $B_{m_0}(a)$ in a vertex. Thus $L$ satisfies $ess(m_0)$.
\end{proof}
We note that, in particular, the argument above proves the following:
\begin{cor}\label{from_ess_proof}
Let $G$ be a one-ended finitely presented group, with $L$ a $(\phi, N)$ quasi-line in $\mathscr{C}^1(G)$ and $C$ a component of $\mathscr{C}^1(G)-L$, which need not be essential. Let $m_0 = m_0(G, \phi, N)$ be as in Lemma \ref{compl_has_fin_many_comps}.
If $K\subset L$ is such that $K$ separates $L$ into two infinite components and $C$ meets both of those components, then $B_{m_0}(x)$ meets $C$ in a vertex, for each $x \in K$.
\end{cor}
\begin{rmk}\label{all_qlines_sat_ess_w_same_m_0}
By Lemma \ref{compl_has_fin_many_comps}, and since we saw in Remark \ref{qlines_have_same_consts} that all quasi-lines with which we are concerned in any one Cayley graph will have the same parameters, they will all satisfy $ess(m_0)$ for some fixed $m_0$.
\end{rmk}
In the remainder of this section, we will show that any quasi-line $L$ that we are concerned with satisfies $iness(m_1)$ for some $m_1$.
\begin{lem}\label{NRH_sats_iness}
Let $G$ be a one-ended finitely generated group, and let $H$ be a two-ended subgroup of $G$. Then any neighborhood of $H$ that is a quasi-line satisfies $iness(m_1)$, for some $m_1$ depending only on $H$ and the size of the neighborhood.
\end{lem}
\begin{proof}
Fix $R>0$ such that $N_R(H)$ is connected, hence is a quasi-line.
It will suffice to find some number $m_1(R)$ such that each inessential component of the complement of $N_R(H)$ is contained in the $m_1(R)$-neighborhood of $N_R(H)$.
We will prove this by first showing that a component $C$ of $\mathscr{C}^1(G)-N_R(H)$ is essential if and only if $C$ is nearly essential.
As $N_R(H)$ has two ends, it follows that $C$ is nearly essential if $C$ is essential. We now prove the converse.
Suppose that $C$ is not essential.
Then $C \cup N_{R}(H)$ has more than one end, so there is a compact $K \subset (C \cup N_{R}(H))$ such that $(C \cup N_{R}(H))-K$ has more than one infinite component.
Let $m$ denote the number of infinite components of $(C \cup N_{R}(H))-K$, and suppose that $m>2$.
Since $e(G)=1$, each of these components must meet $N_R(H)$, and
as $e(N_R(H))=2$, the intersection of $N_R(H)$ with at least $(m-2)$ of these components must be finite.
Let $M$ be the union of $K$ with these finite regions of $N_R(H)$, and thus at least $(m-2)$ components of the complement of $M$ in $(C\cup N_R(H))$ do not intersect $N_R(H)$.
Moreover, note that each such component, of which there is at least one, must be a component of $\mathscr{C}^1(G)-M$.
At least one other component of $\mathscr{C}^1(G)-M$ has infinite intersection with $N_R(H)$, and hence $e(G) > 1$, a contradiction.
Thus we must have that $e(C \cup N_R(H))=2$. We shall show that we can find a finite index subgroup of $H$ that fixes $C$.
Note that since $C$ is a component of the complement of $N_R(H)$, then, for any $g \in H$, $gC$ is also a component of the complement of $N_R(H)$.
Let $\langle h \rangle$ be a finite index subgroup of $H$, and suppose that the $\langle h \rangle$-orbit of $C$ contains infinitely many components of the complement of $N_R(H)$.
Suppose, in addition, that $C$ does not meet $N_R(H)$ along its entire length, i.e. that there is some compact region $K' \subset N_R(H)$ and an infinite component $L_+$ of $N_R(H)-K'$ such that $C$ does not meet $L_+$.
As $e(\mathscr{C}^1(G)) = 1$, the intersection of $C$ with $N_R(H)$ must be infinite, so $N_R(H)-K'$ must have another infinite component, call it $L_-$, and $C$ must meet $L_-$.
Moreover, for any point $q \in N_R(H)$ and any $r>0$, $C$ must meet $L_-$ outside of $B_r(q)$.
Let $\phi$ and $N$ be parameters for $N_R(H)$ and let $m_0=m_0(G, \phi, N)$ be as in Lemma \ref{compl_has_fin_many_comps}.
Then, by Corollary \ref{from_ess_proof}, for any point $p \in L_-$ that is sufficiently far from $K'$, $C$ must meet $B_{m_0}(p)$ in a vertex. Fix such a point $p$.
As we have assumed that $\langle h \rangle \cdot C$ consists of infinitely many components, choose $\{ n_i\}$ such that $\{ h^{n_i} C\}$ are distinct. We can moreover choose the $\{ n_i\}$ such that $L_- \subset h^{n_i} L_-$, for all $i$.
But then each $h^{n_i} C$ must meet $B_{m_0}(p)$ in a vertex. $\mathscr{C}^1(G)$ is finitely generated, hence there are only finitely many vertices in $B_{m_0}(p)$, but the translates $h^{n_i}\cdot C$ are disjoint, thus we have reached a contradiction.
If instead, for any compact subset $K'$ of $N_R(H)$, $C$ meets both infinite components of $N_R(H)-K'$, then a similar argument, with $L$ taking the role of $L_-$, also gives a contradiction.
Thus the $\langle h \rangle$-orbit of $C$ must be a finite collection of components.
By passing to a finite index subgroup of $\langle h \rangle$ if necessary, we can assume that $\langle h \rangle$ fixes $C$.
Recall that we showed above that $(C\cup N_R(H))$ has two ends. The subgroup $\langle h \rangle$ acts on this union by isometries, so the quotient of $(C \cup N_R(H))$ by this action must be compact.
It follows that $C$ is contained in a finite neighborhood of $N_R(H)$, hence is not nearly essential.
Thus $C$ is essential if and only $C$ is nearly essential.
It remains to argue that each inessential component of the complement of $N_R(H)$ is contained in the $m_1(R)$-neighborhood of $N_R(H)$ for some $m_1(R)$.
Note that any inessential component of $\mathscr{C}^1(G)-N_R(H)$ is not nearly essential, hence projects onto a bounded component of $H\backslash \mathscr{C}^1(G)- H\backslash N_R(H)$.
As $H \backslash N_R(H)$ is compact and $H \backslash \mathscr{C}^1(G)$ is locally finite, there are only finitely many components of $H \backslash \mathscr{C}^1(G) - H \backslash N_R(H)$, thus there is some $m_1(R)>0$ such that each bounded component is contained in the $m_1(R)$-neighborhood of $H \backslash N_R(H)$.
It follows that any inessential component of the complement of $N_R(H)$ is contained in the $m_1(R)$-neighborhood of $N_R(H)$, as desired.
\end{proof}
We show next that the property of satisfying $iness(m_1)$ for some function $m_1$ is invariant under quasi-isometries.
\begin{lem}\label{iness_preserved_under_qi}
Let $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a $(\Lambda, K)$ quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$, and let $L\subset \mathscr{C}^1(G)$ be a quasi-line satisfying $iness(m_1)$.
If $R'\geq 0$ is such that $L'=N_{R'}(f(L))$ is a quasi-line in $\mathscr{C}^1(G')$, then $L'$ must satisfy $iness (m_1')$, for some $m_1'$ depending on $\Lambda, K,$ $m_1$, and $R'$.
\end{lem}
\begin{proof}
Recalling Lemma \ref{imgs_of_qlines_are_qlines}, we fix $R'$ so that $L' = N_{R'}(f(L))$ is a quasi-line.
As was the case previously, it suffices to prove that there is some number $m_1'(R')>0$ (dependent on $R'$) such that the inessential components of the complement of $L'$ are contained in the $m_1'(R')$-neighborhood of $L'$.
Again we will begin by showing that any component of $\mathscr{C}^1(G')-L'$ is essential if and only if it is nearly essential.
Recall that we always have that essential implies nearly essential.
Let $f^{-1}$ be a quasi-inverse to $f$, and note that, for any $R>0$, each component of $\mathscr{C}^1(G')-L'$ gets mapped by $f^{-1}$ either into $N_R(L)$ or into the union of $N_R(L)$ with components of its complement.
We claim that we may choose $R$ large enough that, if $C'$ is a component of $\mathscr{C}^1(G')-L'$ such that $f^{-1}(C')$ meets a component $C$ of $\mathscr{C}^1(G)-N_R(L)$, then the image under $f^{-1}$ of no other component of $\mathscr{C}^1(G')-L'$ will meet $C$.
To see this, let $\Lambda', K', \delta$ be such that $f^{-1}$ is a $(\Lambda', K')$ quasi-isometry, with $f^{-1}(L') \subset N_\delta (L)$.
Let $\{ C_\alpha'\}$ be the components of $\mathscr{C}^1(G')-L'$, and let $R_1> \Lambda' K'$. Note that, if $\alpha \neq \beta$, and $C_\alpha' -N_{R_1}(L')$ and $C_\beta' -N_{R_1}(L')$ are nonempty, then any points $p_\alpha \in C_\alpha' -N_{R_1}(L')$, $p_\beta \in C_\beta' -N_{R_1}(L')$ are at least a distance of $2\Lambda' K'$ apart.
Let $R> (\delta + \Lambda' R_1 + K')$, and
note that $f^{-1}(N_{R_1}(L')) \subset N_R(L)$.
Recall that $f^{-1}$ is coarsely surjective, with $N_{K'}(f^{-1}(\mathscr{C}^1(G'))) = \mathscr{C}^1(G)$.
Suppose that there is some component of $\mathscr{C}^1(G)-N_R(L)$ that is met by more than one image $f^{-1}(C_\alpha')$.
Then there are two such, call them $f^{-1}(C_\alpha')$ and $f^{-1}(C_\beta')$, with some $p_\alpha \in C_\alpha' - N_{R_1}(L'), p_\beta \in C_\beta' - N_{R_1}(L')$, such that $d(f^{-1}(p_\alpha),f^{-1}( p_\beta))<K'$.
But this means that
$\frac{1}{\Lambda'}d(p_\alpha, p_\beta)-K' < K'$, i.e. that
$d(p_\alpha, p_\beta)<2\Lambda' K'$, which is a contradiction.
Thus, with $R$ chosen as above, we have that the images under $f^{-1}$ of different components of the complement of $L'$ shall not meet the same component of the complement of $N_R(L)$.
Suppose now that $C'$ is a component of $\mathscr{C}^1(G')-L'$ that is nearly essential. Let $C_0$ be the union of the components of $\mathscr{C}^1(G)-N_R(L)$ that are met by $f^{-1}(C')$.
Since $L$ satisfies $iness (m_1)$, $C_0$ must contain an essential component of the complement of $N_R(L)$.
Let $C_0^e$ denote the essential components in the complement of $N_R(L)$ that are met by $f^{-1}(C')$, and now we have that $C_0^e$ is nonempty.
Observe that $(C' \cup L')$ is quasi-isometric to $C_0 \cup f^{-1}(L')$, which is quasi-isometric to $(C_0 \cup L)$. Certainly this is quasi-isometric to $(C_0 \cup N_R(L))$, which in turn must be quasi-isometric to $(C_0^e \cup N_{R}(L))$, since the components of $C_0$ that are inessential must be contained in the $m_1(R)$-neighborhood of $N_R(L)$.
We claim that $e(C_0^e \cup N_{R}(L))=1$.
This is immediate if $C_0^e$ contains only one component, so assume that $C_0^e = \{ C_i\}$ contains more than one, and suppose for a contradiction that $e(C_0^e \cup N_{R}(L)) >1$.
Then there is some finite subgraph $K$ of $(C_0^e \cup N_{R}(L))$ whose complement has more than one infinite component.
For each $i$, $C_i$ is essential so $(C_i \cup N_R(L))-K$ has exactly one infinite component which we shall call $D_i$, and as $\mathscr{C}^1(G)$ is locally finite, we note that $N_R(L) - (D_i \cap N_R(L))$ must therefore be finite.
On the other hand, as $G$ is finitely presented, so $N_R(L)$ satisfies $ess(m_0)$ for some $m_0$.
Thus there are only finitely many $C_i$'s, so there must be indices $i$ and $j$ such that $D_i$ and $D_j$ are disconnected in $(C_0^e \cup N_{R}(L))$ by $K$.
As $N_R(L)$ is finite outside of $D_i$ and is finite outside of $D_j$, we have reached a contradiction.
Thus $e(C_0^e \cup N_{R}(L))=1$, so $e(C' \cup L') =1$ and $C'$ is essential.
Hence components of the complement of $L'$ are essential if and only if they are nearly essential.
It remains to conclude that $L'$ satisfies $iness(m_1)$.
As $f^{-1}$ is coarsely surjective, we have that a finite neighborhood of the image under $f^{-1}$ of any component of the complement of $L'$ is equal to a subset of $N_R(L)$, together with a collection of components of the complement of $N_R(L)$.
As the inessential components of the complement of $N_R(L)$ are contained in the $m_1(R)$-neighborhood of $N_R(L)$, it follows that there is some $m_1'(R')>0$ such that any component $C'$ of the complement of $L'$ is either contained in the $m_1'(R')$-neighborhood of $L'$, or is contained in no finite neighborhood of $L'$.
Thus, $C'$ must be contained in $N_{m_1'(R')}(L')$ or else is nearly essential hence essential, as desired.
\end{proof}
\begin{rmk}\label{iness_preserved}
Recall that we are concerned with the following quasi-lines.
If $G$ is a one-ended finitely presented group with a two-ended subgroup $H$, then we will consider a quasi-line in $\mathscr{C}^1(G)$ of the form $N_R(H)$ and its translates under the action of $G$.
We will also consider a quasi-line that is the $R'$-neighborhood of the image of $H$ under a quasi-isometry $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$, and the translates of that quasi-line in $\mathscr{C}^1(G')$.
$G$ acts on $\mathscr{C}^1(G)$ by isometries, and hence it follows from Lemma \ref{NRH_sats_iness} that any collection of quasi-lines that are translates of $N_R(H)$ by elements of $G$ will all satisfy $iness(m_1)$ for the same function $m_1$.
Simiarly, and using Lemma \ref{iness_preserved_under_qi}, we have that any translates of the $R'$-neighborhood of $f(H)$ will all satisfy $iness(m_1')$ for the same function $m_1'$.
\end{rmk}
\section{The quasi-isometry invariance of two-ended subgroups with at least three coends} \label{section3}
In this section, we will prove that, up to finite Hausdorff distances, quasi-isometries take two-ended subgroups with at least three coends to other two-ended subgroups with at least three coends (Theorem \ref{existence_qi_inv}).
Coends will be defined below, and we will see that two-ended subgroups having at least three coends will be exactly those whose corresponding quasi-lines are 3-parting (Lemma \ref{coends_and_ess_comps}).
The main ingredient in the proof of the quasi-isometry invariance of two-ended subgroups with at least three coends is Proposition \ref{LnearZ}, which shows that any $3$-parting quasi-line satisfying $iness(m_1)$ is a finite Hausdorff distance from an infinite cyclic subgroup.
For this, we use the proof of a similar result from \cite{papa_qlines}.
There, Papasoglu shows that, given a 3-parting quasi-line in the Cayley graph of a one-ended, finitely presented group, either the quasi-line is a finite distance from an infinite cyclic subgroup, or a related limit of translates of quasi-lines is.
We will see below that this latter possibility can be eliminated.
\\
We first note that quasi-lines satisfying $iness(m_1)$ but that are not 3-parting need not be a finite Hausdorff distance from a copy of $\mathbb{Z}$.
For example, consider the nearest-point projection of a line $l_0$ in $\mathbb{R}^2$ with irrational slope into the Cayley graph of $\mathbb{Z}^2$, where the vertices are taken to be the integer lattice points in $\mathbb{R}^2$ and we take the standard generating set.
Let $L$ denote a connected neighborhood in $\mathscr{C}_{\mathbb{Z}^2}$ of the projection of $l_0$. Then $L$ is a 2-parting quasi-line in $\mathscr{C}_{\mathbb{Z}^2}$ that satisfies $iness(m_1)$ for some $m_1$.
However, the infinite cyclic subgroups of $\mathbb{Z}^2$ correspond to lines in $\mathbb{R}^2$ with rational slope, hence $L$ is an infinite Hausdorff distance from any subgroup of $\mathbb{Z}^2$.
\\
In order to prove Proposition \ref{LnearZ}, we will need to know that 3-parting quasi-lines do not cross one another in an essential way.
We shall say that $a,b \in \mathscr{C}^1(G)$ are {\it $K$-separated by a quasi-line $L$} if $B_K(a)$ and $B_K(b)$ are in different components of the complement of $L$.
The following is Proposition 2.1 from \cite{papa_qlines}:
\begin{thm} {\rm \cite{papa_qlines}} \
\label{2.1}
Let $G$ be a one-ended, finitely presented group, and let $L, L_1$ be $(\phi',N')$ quasi-lines in $\mathscr{C}^1(G)$ that satisfy $iness(m_1)$.
Suppose that $L$ is 3-parting.
Then there is some $K = K(G, \phi', N', m_1)$ such that no two points $a,b \in L$ are $K$-separated by $L_1$.
\end{thm}
We include a proof of this theorem in the appendix, in order to clarify some points from the proof given in \cite{papa_qlines}.
To restate Theorem \ref{2.1}, if there is some $a \in L$ that is in an essential component $C$ of the complement of $L_1$ and is more than a distance of $K$ from $L_1$, and $b \in L$ is in a different essential component of the complement of $L_1$, then $b$ is no more than a distance of $K$ from $L_1$.
It follows that $L$ is contained in the $K$-neighborhood of $L_1 \cup C$.
Let $m_0$ be such that $L$ and $L_1$ satisfy $ess(m_0)$ (see Lemma \ref{compl_has_fin_many_comps}), and let $K' = K + 2m_0$.
Then it follows that $L$ is contained in the $K'$-neighborhood of $C$.
Thus we have the following corollary to Theorem \ref{2.1}:
\begin{cor} \label{q_lines_are_K_disjoint}
Let $G$, $L$ and $L_1$ be as in Theorem \ref{2.1}.
Then there is some $K'=K'(G, \phi , N, m_1)$ such that $L$ is contained in the $K'$-neighborhood of an essential component of the complement of $L_1$.
\end{cor}
The next observation will be needed in Lemma \ref{seq_of_qlines_is_fin}.
\begin{lem}\label{get_haus}
Let $L$ and $L'$ be $(\phi, N)$ quasi-lines in a metric space $X$.
Then for any $x_2$, there is some $x_1 = x_1(\phi, N, x_2)>x_2$ such that if $L \nsubseteq N_{x_1}(L')$, then $L' \nsubseteq N_{x_2}(L)$.
\end{lem}
\begin{proof}
Given $x_2$, let $x_1 > \frac{1}{2}\phi(2(x_2+N)) + 2N + x_2$.
Suppose for a contradiction that $L \nsubseteq N_{x_1}(L')$ and that $L' \subset N_{x_2}(L)$.
Let $l$ and $l'$ be the lines associated to $L$ and $L'$ respectively, and it follows that there is some $t \in \mathbb{R}$ such that the $(x_1-N)$-ball about $l(t)$ does not meet $l'$.
As $L' \subset N_{x_2}(L)$, hence $l' \subset N_{x_2+N}(l)$, it follows that there are $t_1 < t < t_2$ such that $|t-t_i| \geq ((x_1 - N)-(x_2+N))$ for each $i$ and $d(l(t_1), l(t_2)) \leq 2(x_2+N)$.
But our assumption on $x_1$ implies that $\phi (2(x_2+N)) < 2(x_1-2N-x_2)$, a contradiction.
Thus $L' \nsubseteq N_{x_2}(L)$.
\end{proof}
We shall need the following lemma, both to prove Proposition \ref{LnearZ} and also to prove another later result.
\begin{lem}\label{seq_of_qlines_is_fin}
Let $G$ be a one-ended, finitely presented group, and let $\{ L^i\}$ be a collection of 3-parting $(\phi, N)$ quasi-lines in $\mathscr{C}^1(G)$ satisfying $iness(m_1)$. Suppose that $\cap_i L^i$ contains a vertex.
Then there is some constant $x_1 = x_1(G, \phi, N, m_1)$ such that if, for all $i,j$, $d_{Haus}(L^i, L^j) > x_1$,
then $\{ L^i \}$ is finite.
\end{lem}
\begin{proof}
Let $m_0 = m_0(G, \phi, N)$ be as in Lemma \ref{compl_has_fin_many_comps}, so that each $L^i$ satisfies $ess(m_0)$.
Let $K' = K'(G, \phi , N, m_1)$ be as in Corollary \ref{q_lines_are_K_disjoint}, so that, for each $i, j$, $L^i$ is contained in the $K'$-neighborhood of an essential component of the complement of $L^j$.
Furthermore, let $m_0' = m_0(G, \phi, N+K')$, so that, for any $i$, $N_{K'}(L^i)$ (which is a $(\phi, N+K')$ quasi-line) satisfies $ess(m_0')$.
Let $x_2 > \max \{ K', m_0'\}$ and let $x_1$ be from Lemma \ref{get_haus}.
Thus $d_{Haus}(L^i, L^j) > x_1$ implies that $L^i \nsubseteq N_{x_2}(L^j)$ and $L^j \nsubseteq N_{x_2}(L^i)$.
Let $\mathscr{L}_0$ denote $\{ L^i\}$, and suppose that $\mathscr{L}_0$ is infinite.
Then choose any element $L_0$ from $\mathscr{L}_0$.
As $L_0$ satisfies $ess(m_0)$, the complement of $L_0$ has only finitely many essential components, so there is some essential component $B_0$ whose $K'$-neighborhood contains infinitely many elements of $\mathscr{L}_0$.
Let $\mathscr{L}_1 = \{L \in [\mathscr{L}_0-L_0] : L \subset N_{K'}(B_0)\}$. Choose $L_1$ from $\mathscr{L}_1$, and let $B_1'$ be the essential component of the complement of $L_1$ whose $K'$-neighborhood contains $L_0$. Note that $x_2>K'$ implies that $B_1'$ is unique.
As $\mathscr{L}_1$ is infinite, there is some essential component of the complement of $L_1$ whose $K'$-neighborhood contains infinitely many elements of $\mathscr{L}_1$. Let $B_1$ denote this component, and let $\mathscr{L}_2$ denote $\{ L \in [\mathscr{L}_1-\{L_0, L_1\} ] : L \subset N_{K'}(B_1)\}$. Choose $L_2$ from $\mathscr{L}_2$, and continue on in this manner. This produces an infinite sequence of quasi-lines $\{ L_i\}$ and subsets of $\mathscr{C}^1(G)$, $\{ B_i\}$ and $\{B_i'\}$, such that, for each $i$, $B_i$ is an essential component of the complement of $L_i$ such that $L_j \subset N_{K'}(B_i)$ for all $j>i$, and $B_i'$ is an essential component of the complement of $L_i$ such that $L_j \subset N_{K'}(B_i')$ for all $j<i$ (with perhaps $B_i = B_i'$).
Each $L_i$ is 3-parting, so we may set $D_i$ to be an essential component of the complement of $L_i$ that is not equal to $B_i$ nor $B_i'$, for each $i$.
We shall see next that the $D_i$'s are basically disjoint.
Let $i \neq j$, and note that, since $L_i$ is not contained in the $x_2$-neighborhood of $L_j$,
there must be some point $p \in L_i$ such that $B_{x_2}(p)$ does not intersect $L_j$. Thus $B_{x_2}(p)$ is contained in $B_j$ or $B_j'$.
Note that, for each $i$, $D_i - N_{K'}(L_i)$ is a collection of essential and inessential components of the complement of $N_{K'}(L_i)$.
Since $D_i$ is an essential component of the complement of $L_i$, and $L_i$ satisfies $iness(m_1)$, it follows that $D_i-N_{K'}(L_i)$ must contain an essential component $E_i$ of the complement of $N_{K'}(L_i)$.
As $x_2> m_0'$, $B_{x_2}(p)$ must meet each essential component of the complement of $N_{K'}(L_i)$, so, in particular, $B_{x_2}(p)$ meets $E_i$, hence $B_{x_2}(p) \cup E_i$ is connected.
The quasi-line $L_j$ is disjoint from $D_i-N_{K'}(L_i)$, hence does not meet $E_i$, or the union $B_{x_2}(p) \cup E_i$. It follows that this union is contained in $B_j$ or $B_j'$, so is disjoint from $D_j$, and hence from $E_j \subset D_j$. Thus, the $E_i$'s are disjoint.
Now we recall that $\cap_i L_i$ contains a vertex, say $y \in \mathscr{C}^1(G)$, and hence $B_{m_0'}(y)$ intersects each $E_i$. Since these regions are disjoint, $B_{m_0'}(y)$ must contain a collection of vertices in bijection with $\{ L_i\}$. But $G$ is finitely generated, hence $B_{m_0'}(y)$ has only finitely many vertices, and we have reached a contradiction.
\end{proof}
\begin{prop}\label{LnearZ}
Let $L$ be a $3$-parting $(\phi, N)$ quasi-line in the Cayley graph of a one-ended, finitely presented group $G$, and suppose that $L$ satisfies $iness(m_1)$ for some $m_1$. Then there is some subgroup $H \cong \mathbb{Z}$ of $G$ such that $d_{Haus}(L,H)< \infty$.
\end{prop}
\begin{proof}
Let $L$ be as in the statement of the proposition, and let $x_1$ be as in Lemma \ref{seq_of_qlines_is_fin}, defined with the parameters of $L$.
In case 1 of section 6 of \cite{papa_qlines}, Papasoglu makes the following construction.
Fix some $y \in L$, and choose a sequence $\{y_i \} \subset L$ such that $d(y,y_i) \to \infty$. Let $g_i$ be such that $g_iy_i = y$, and, by passing to a subsequence, we may assume that, for all $i>j$,
$$g_jL \cap B_j(y) = g_iL \cap B_j(y).$$
If there is some $i$ such that $d_{Haus} (g_iL, g_jL) $ is less than or equal to any fixed constant for infinitely many $g_j$, then it is shown in \cite{papa_qlines} that
there is some $g$ contained in the subgroup generated by these $g_j$ such that $\langle g \rangle \cong \mathbb{Z}$, and $g_iL$ is a finite Hausdorff distance from $\langle g \rangle$.
Thus $d_{Haus}(L, g_i^{-1}\langle g \rangle) < \infty$. Since $d_{Haus}(g_i^{-1}\langle g \rangle, g_i^{-1}\langle g \rangle g_i)$ is bounded by the word length of $g_i$, it follows that $L$ is a finite Hausdorff distance from $g_i^{-1}\langle g \rangle g_i \cong \mathbb{Z}$.
So, by passing to a subsequence, we may assume that, for each $i$ and $j$, $d_{Haus}(g_iL, g_jL)>x_1$.
It follows that this infinite subsequence of $\{ g_iL\}$ satisfies the hypotheses of Lemma \ref{seq_of_qlines_is_fin}, which is a contradiction.
\end{proof}
Next, we will give the definition of coends. We will see that Theorem \ref{existence_qi_inv} follows quickly from Proposition \ref{LnearZ} and a few basic facts about coends.
Given a group $G$ with a subgroup $H$ and a subset $Y$, we say that $Y$ is {\em $H$-finite} if $Y$ is contained in finitely many cosets $Hg$ of $H$.
In \cite{KropRol}, Kropholler and Roller defined
$$\tilde{e}(G,H) = {\rm dim}_{\mathbb{F}_2} (\mathcal{P}G / \mathcal{F}_HG)^G,$$
where $\mathcal{P}G$ is the power set of all subsets of $G$, and $\mathcal{F}_HG$ is the set of all $H$-finite subsets of $G$.
The quotient set $\mathcal{P}G/\mathcal{F}_HG$ forms a vector space over $\mathbb{F}_2$, the field with two elements, under the operation of symmetric difference. Thus a subset $X$ of $G$ represents an element of $(\mathcal{P}G / \mathcal{F}_HG)^G$ if and only if the symmetric difference $X+Xg$ is $H$-finite for all $g \in G$.
Following Bowditch \cite{BowJSJ}, we shall call $\tilde{e}(G,H)$ the number of {\em coends} of $H$ in $G$.
(Kropholler and Roller called $\tilde{e}(G,H)$ the number of relative ends of $H$ in $G$, and we note that this is also sometimes referred to as the number of filtered ends of $H$ in $G$.)
If $X$ is a subset of $G$, then we can think of $X$ as a subset of the vertex set of $\mathscr{C}^1(G)$, and thus $\delta X$, the coboundary of $X$, is the set of edges in $\mathscr{C}^1(G)$ that have exactly one vertex contained in $X$.
It is a fact that $X$ represents an element of $(\mathcal{P}G / \mathcal{F}_HG)^G$ exactly when $\delta X$ is $H$-finite.
(See Cohen \cite{Cohen_cohom_dim_one} for a proof of this in the case when $H$ is trivial.)
The following lemma shows that we can characterize the number of coends of a two-ended subgroup in terms of essential components:
\begin{lem}\label{coends_and_ess_comps}
Let $G$ be a one-ended, finitely generated group with two-ended subgroup $H$, and let $n<\infty$. Then $\tilde{e}(G,H)\geq n$ if and only if there is some $R>0$ such that $N_R(H)$ is a quasi-line in $\mathscr{C}^1(G)$ that is $n$-parting.
Moreover, $\tilde{e}(G,H)=\infty$ if and only if, for each $n< \infty$ there is some $R = R(n)$ such that $N_R(H)$ is an $n$-separating quasi-line.
\end{lem}
\begin{proof}
A subset $X$ of $G$ represents an element in the $\mathbb{F}_2$-vector space $(\mathcal{P}G/\mathcal{F}_HG)^G$ if and only if $\delta X$ is an $H$-finite set of edges in $\mathscr{C}^1(G)$.
Note that this happens precisely when $\delta X$ is contained in a finite neighborhood of $H$ in $\mathscr{C}^1(G)$.
Essential components of the complement of any quasi-line of the form $N_R(H)$ naturally correspond to elements of $(\mathcal{P}G/\mathcal{F}_HG)^G$: let $\hat{Y}$ be an essential component of the complement of $N_R(H)$, and let $Y$ denote the vertex set of $\hat{Y}$. Then for any $\epsilon >0$, the boundary of $\hat{Y}$ is contained in $N_{R+\epsilon}(H)$, hence $\delta Y \subset N_{R+1}(H)$, thus $Y$ represents an element of $(\mathcal{P}G/\mathcal{F}_HG)^G$.
Note that $Y$ is not $H$-finite, so the element it represents must be nontrivial in $(\mathcal{P}G/\mathcal{F}_HG)^G$.
By Lemma \ref{2ended_subgps_are_qlines}, we can fix $R>0$ such that $N_R(H)$ is a quasi-line.
Suppose that $N_R(H)$ be $n$-parting, and let $Y_1, \ldots, Y_{n}$ be essential components of the complement of $N_R(H)$. They are disjoint, hence represent independent elements of $(\mathcal{P}G/\mathcal{F}_HG)^G$, and thus $\tilde{e}(G,H) \geq n$.
If $\tilde{e}(G,H)\geq n$ for some $n<\infty$, then we can find representatives $X_1, \ldots , X_n$ of elements of a basis for $(\mathcal{P}G/\mathcal{F}_HG)^G$. Thus there is some $R>0$ such that, in $\mathscr{C}^1(G)$, $\delta X_i \subset N_R(H)$, for all $i$. Then note that each $X_i$ is equivalent in $(\mathcal{P}G/\mathcal{F}_HG)^G$ to a union of components of $\mathscr{C}^1(G)-N_R(H)$. Recall from Lemma \ref{NRH_sats_iness} that, for some $m_1>0$, $N_R(H)$ satisfies $iness (m_1)$, hence each $X_i$ is equivalent to a union of essential components of $\mathscr{C}^1(G)-N_R(H)$.
Since the $X_i$'s are independent, $n$ of these essential components must be disjoint, so the complement of $N_R(H)$ has at least $n$ distinct essential components, i.e. $N_R(H)$ is $n$-parting.
Now suppose that $\tilde{e}(G,H)= \infty$, and fix any $n < \infty$. Then in particular $\tilde{e}(G,H)\geq n$ so, by the previous paragraph, there is some $R = R(n)$ such that $N_R(H)$ is an $n$-parting quasi-line.
Lastly, suppose now that $H$ is such that, for any $n<\infty$ there exists some $R(n)$ such that $N_R(H)$ is an $n$-parting quasi-line, and let $Y_1^n, Y_2^n, \ldots , Y_n^n$ denote the essential components of the complement of $N_R(H)$.
Fix any sequence $n_1, n_2, n_3, \ldots $ such that $R(n_i) < R(n_{i+1})$ for all $i$.
Then we note that there are indices $j_i \neq k_i$ such that $1 \leq j_i, k_i \leq n_i$ and such that $Y_{j_i}^{n_i} \subset Y_{k_l}^{n_l}$ for all $i > l$, and hence $\{ Y_{j_i}^{n_i}\}_{i=1}^\infty$ are a disjoint collection of representatives of elements of $(\mathcal{P}G / \mathcal{F}_H(G))^G$.
It follows that $\tilde{e}(G,H)=\infty$.
\end{proof}
\begin{lem}\label{ess_comps_go_to_ess_comps}
Let $f\colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a $(\Lambda, C)$ quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$, and let $L$ be a $(\phi, N)$ quasi-line in $\mathscr{C}^1(G)$ satisfying $iness(m_1)$.
Then there is some $R'=R'(\Lambda, C, \phi, N, m_1) >0$ such that, if $L$ is $n$-parting, then $N_{R'}(f(L))$ is also $n$-parting.
\end{lem}
\begin{proof}
Lemma \ref{imgs_of_qlines_are_qlines} shows that we can find some $R''>0$ such that $N_{R''}(f(L))$ is a quasi-line. Thus so is $N_{R_0'}(f(L))$ for any $R_0' \geq R''$, and, by Lemma \ref{iness_preserved_under_qi}, we also have that $N_{R_0'}(f(L))$ satisfies $iness(m_1')$ for some $m_1'$ (depending on $R_0'$).
For any such $R_0'$, the image under $f$ of any component of $\mathscr{C}^1(G)-L$ will be contained in the union of $N_{R_0'}(f(L))$ and components of its complement.
As in the proof of Lemma \ref{iness_preserved_under_qi}, there is some $R' \geq R''$ such that the images under $f$ of distinct components of $\mathscr{C}^1(G)-L$ do not meet the same components of $\mathscr{C}^1(G') - N_{R'}(f(L))$.
Let $L' = N_{R'}(f(L))$.
As $f$ is coarsely surjective and $L'$ satisfies $iness(m_1')$, the image of any essential component in the complement of $L$ meets an essential component in the complement of $L'$. As no two components of the complement of $L$ meet the same components of the complement of $L'$, it follows that the complement of $L'$ contains at least as many essential components as the complement of $L$.
\end{proof}
We now can prove the following:
\begin{thm}\label{existence_qi_inv}
Let $f\colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$, and assume that $G$ contains a 2-ended subgroup $H$ that has $n$ coends in $G$, for $n \in \{ 3, 4, \ldots \} \cup \{ \infty\}$.
Then there is a two-ended subgroup $H'$ of $G'$ that has $n$ coends in $G'$, and furthermore
$$d_{Haus}(f(H), H')<\infty.$$
\end{thm}
\begin{proof}
Suppose first that $n<\infty$.
By Lemmas \ref{2ended_subgps_are_qlines} and \ref{NRH_sats_iness}, for any $R$ such that $N_R(H)$ is connected, we ahve that $N_R(H)$ is a $(\phi, N)$ quasi-line that satisfies $iness(m_1)$, where $\phi, N$, and $m_1$ all depend on $R$.
As $\tilde{e}(G,H)= n$, it follows from Lemma \ref{coends_and_ess_comps} that we can further choose $R$ so that $N_R(H)$ is $n$-parting.
Let $L = N_R(H)$ for some such $R$.
Then, by Lemmas \ref{imgs_of_qlines_are_qlines} and \ref{iness_preserved_under_qi}, and \ref{ess_comps_go_to_ess_comps}, there is some $R'$ such that $N_{R'}(f(L))$ is a quasi-line satisfying $iness(m_1')$, and $N_{R'}(f(L))$ is $n$-parting. Let $L'$ denote $N_{R'}(f(L))$ for some such $R'$.
Proposition \ref{LnearZ} implies that there is some $H' \cong \mathbb{Z}$ that is a finite Hausdorff distance from $L'$. Let $L'' = N_{R''}(H')$, with $R''>0$ such that $L''$ contains $L'$. Then $L''$ is $n$-parting, so, by Lemma \ref{coends_and_ess_comps}, $\tilde{e}(G', H') \geq n$.
If $\tilde{e}(G',H')>n$, then Lemma \ref{coends_and_ess_comps} and Lemma \ref{ess_comps_go_to_ess_comps}, applied to a quasi-inverse of $f$, implies that there is some quasi-line that is a finite Hausdorff distance from $H$ and is $m$-parting for some $m>n$. There is a neighborhood of $H$ contains this quasi-line, thus is $m$-parting, so by Lemma \ref{coends_and_ess_comps}, $\tilde{e}(G,H) \geq m$, a contradiction.
Thus $\tilde{e}(G', H')=n$, so $H'$ is the desired subgroup.
Suppose then that $\tilde{e}(G,H)=\infty$.
Then $\tilde{e}(G,H) \geq m$ for any $m < \infty$, and the above argument shows that there is some $R''>0$ such that $N_{R''}(H')$ is an $m$-parting quasi-line.
Thus Lemma \ref{coends_and_ess_comps} implies that $\tilde{e}(G', H')=\infty$.
\end{proof}
\section{The quasi-isometry invariance of commensurizer subgroups}
In this section, we will see that commensurizers of two-ended subgroups with at least three coends are invariant under quasi-isometries.
More specifically, we saw in the last section that, if $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ is a quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$, and $H$ is a two-ended subgroup of $G$ with at least three coends, then there is a two-ended subgroup $H'$ with at least three coends, that is a finite Hausdorff distance from $f(H)$ in $\mathscr{C}^1(G')$.
We will now see that in fact Comm$_{G'}(H')$ is a finite Hausdorff distance from the image under $f$ of Comm$_G(H)$ in $\mathscr{C}^1(G')$ (Theorem \ref{Cqiinv}).
We first observe the geometric structure of commensurizers:
\begin{lem}\label{commcharacterization}
If $G$ is a finitely generated group with subgroup $H$, then
$${\rm Comm}_G(H) = \{ g \in G : d_{Haus}(H,gH)<\infty \}.$$
\end{lem}
\begin{proof}
Let $l(g)$ be the minimal word length of representatives for $g \in G$, with respect to the given finite generating set for $G$. Then note that, for all $x, g \in G$, $d(x,xg) =d(e,g) = l(g)$.
Thus $d_{Haus}(gH, gHg^{-1}) \leq l(g^{-1})$, so it suffices to show that $g \in Comm_G(H)$ if and only if $d_{Haus}(H, gHg^{-1})<\infty$.
Let $H^g$ denote $gHg^{-1}$. If $d_{Haus}(H, H^g)=M<\infty$, then, for any $x \in H$, there is some $y \in H^g$ such that $d(x,y) \leq M$, i.e. $d(y^{-1}x, e) = l(y^{-1}x)\leq M$.
Let $L(M) = \{ k \in G : l(k) \leq M\}$. It follows that
\begin{equation}\label{comm1}
H \subset \cup_{k\in L(M)}H^gk
\end{equation}
and similarly that
\begin{equation}\label{comm2}
H^g \subset \cup_{k\in L(M)}Hk.
\end{equation}
Observe that in fact (\ref{comm1}) and (\ref{comm2}) are equivalent to having $d_{Haus}(H, H^g)\leq M$.
If $H$ meets $H^gk$, then there is some $h_1 \in H$ with $H^gk=H^gh_1$.
$G$ is finitely generated, so $L(M)$ is finite, and it follows that (\ref{comm1}) implies that there are finitely many elements $h_1, \ldots , h_n$ in $H$ such that
$$H \subset \cup_{i=1}^n H^gh_i.$$
Thus $H = \cup_{i=1}^n (H \cap H^g)h_i$, i.e. $(H \cap H^g)$ is of finite index in $H$.
Similarly $(H \cap H^g)$ is of finite index in $H^g$, so $H$ and $H^g$ are commensurable, hence $g \in $ Comm$_G(H)$.
Conversely, if $g \in $ Comm$_G(H)$, then there are elements $h_1, \ldots , h_n$ in $H$ such that $H = \cup_{i=1}^n (H \cap H^g)h_i$, and elements $h_1', \ldots , h_n'$ in $H^g$ such that $H^g = \cup_{i=1}^{n'} (H \cap H^g)h_i'$.
In particular, (\ref{comm1}) and (\ref{comm2}) hold if we take $M$ to be the maximal word length of the $h_i$'s and $(h_i')$'s.
Thus $d_{Haus}(H, H^g)\leq M$, so we have shown the lemma.
\end{proof}
\begin{rmk}\label{char_comm}
As we saw in the proof of Theorem \ref{existence_qi_inv}, if $H$ is a two-ended subgroup of $G$ with at least $n$ coends, $n < \infty$, then there is some $R$ such that $N_R(H)$ is an $n$-parting $(\phi, N)$ quasi-line satisfying $iness(m_1)$, for some $\phi, N$, and $m_1$.
Thus, by Lemma \ref{commcharacterization} and since $G$ acts on its Cayley graph by isometries on the left,
$$N_R({\rm Comm}_G(H)) = \cup_{g \in {\rm Comm}_G(H)}N_R(gH) = \cup_{g \in {\rm Comm}_G(H)}g\cdot N_R(H) $$
is a union of isometric copies of $N_R(H)$ that are pairwise of finite Hausdorff distance from one another. Hence we may think of Comm$_G(H)$ as a collection of ``parallel'' $n$-parting $(\phi, N)$ quasi-lines that satisfy $iness(m_1)$.
\end{rmk}
Consider the following.
\begin{prop}\label{fin_H_dist}
Let $G$ be a one-ended, finitely presented group with a two-ended subgroup $H$ that has at least three coends, and let $C= $ Comm$_G(H) = \coprod_i g_iH$.
Given quasi-line parameters $\phi$ and $N$, and a function $m_1$, there exists a constant $x = x(\phi, N, m_1, H)$ such that, if $L$ is a $3$-parting $(\phi, N)$ quasi-line in $\mathscr{C}^1(G)$ satisfying $iness(m_1)$ and such that $d_{Haus}(L,H)<\infty$, then, for some $i$, $d_{Haus}(L, g_iH)<x$.
\end{prop}
Assuming this proposition for the moment, we shall see how it implies the invariance of these commensurizer subgroups under quasi-isometries.
Suppose that $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ is a quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$, that $H$ is a two-ended subgroup of $G$ with at least $n$ coends, for some $3 \leq n < \infty$, and that $C = $ Comm$_{G}(H)$.
Then, by Remark \ref{char_comm}, we have that some neighborhood $N_R(C)$ of $C$ is a union of pairwise finite Hausdorff distance, $n$-parting $(\phi ,N)$ quasi-lines $\{ L_i\}$, all of which satisfy $iness(m_1)$ for some $m_1$.
By Theorem \ref{existence_qi_inv}, there is a two-ended subgroup $H'$ of $G'$ such that $\tilde{e}(G,H) = \tilde{e}(G', H')$ and $d_{Haus}(f(H), H')$ is finite.
Let $C' = $ Comm$_{G'}(H')$.
By Lemma \ref{imgs_of_qlines_are_qlines}, there exists $R'$ such that the $R'$-neighborhood of each $f(L_i)$ is a $(\phi', N')$ quasi-line, for some $\phi'$ and $N'$ depending on $R'$.
By Lemma \ref{ess_comps_go_to_ess_comps}, we can choose $R'$ so that each $N_{R'}(f(L_i))$ is $3$-parting.
By Lemma \ref{iness_preserved_under_qi} and Remark \ref{iness_preserved}, we can further suppose that each $N_{R'}(f(L_i))$ satisfies $iness(m_1')$, for some fixed $m_1'$.
Thus we may apply Proposition \ref{fin_H_dist} to get some $x$ such that each $N_{R'}(f(L_i))$ is contained in $N_x(C')$.
It follows that $N_{R'}(f(N_R(C))) \subset N_x(C')$, i.e. that $f(C)$ is contained in a finite neighborhood of $C'$.
As was the case for $C$, recall that a neighborhood of $C'$ is a union of quasi-lines as above.
Thus, by running the same argument on a quasi-inverse to $f$, it follows that $d_{Haus}(f(C), C')<\infty$. Hence we have the following.
\begin{thm} \label{Cqiinv}
Let $f\colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a $(\Lambda, K)$-quasi isometry between finitely presented, one-ended groups, and suppose that $H$ is a two-ended subgroup of $G$ with $n$ coends in $G$, for $n \geq 3$.
Then there is a two-ended subgroup $H'$ of $G'$ such that $H'$ has $n$ coends in $G'$ and there exists some constant $y=y(G, H, \Lambda, K)$ such that
$$d_{Haus}(f({\rm Comm}_{G}(H)), {\rm Comm}_{G'}(H'))< y.$$
\end{thm}
\begin{proof}[Proof of Proposition \ref{fin_H_dist}]
Let $\mathscr{L}$ be the set of $3$-parting $(\phi, N)$ quasi-lines in $\mathscr{C}^1(G)$ that satisfy $iness(m_1)$, and are a finite Hausdorff distance from $H$. If $\mathscr{L}$ is finite, then we are done, so assume that $\mathscr{L}$ is infinite, and that no such $x$ exists. Then we can find a sequence $\{L_i\}$ of elements of $\mathscr{L}$ such that
$$ \min_{g \in C} d_{Haus}(L_i, gH) \to \infty,$$
as $i \to \infty$.
Let $c_{i}=g \in C$ realize the minimum above for $L_i$, and fix $x_1=x_1(G,\phi, N, m_1)$ from Lemma \ref{seq_of_qlines_is_fin}.
Then we can pass to a subsequence so that, for all $j>i$,
\begin{equation}
d_{Haus}(L_j, c_{j}H)>d_{Haus}(L_i, c_{i}H)+x_1.
\label{assumption}
\end{equation}
Then, by the following argument we will have that, for all $g,g' \in G$ and $i \neq j$, we have
\begin{equation}
d_{Haus} (gL_i, g'L_j)>x_1.
\label{eq1}
\end{equation}
Firstly, note that it suffices to show that $d_{Haus}(L_i, gL_j)>x_1$, for any $g \in G$ and $i<j$. If $g \notin C$, then $d_{Haus}(H, gH) = \infty$. But $d_{Haus}(L_i, H)$ and $d_{Haus}(gL_j, gH)$ are finite, so $d_{Haus}(L_i, gL_j) = \infty$.
Assume then that $g \in C$, and $d_{Haus} (L_i, gL_j)\leq x_1$. Then
$$d_{Haus}(gL_j, c_{i}H) \leq d_{Haus}(gL_j, L_i) + d_{Haus}(L_i, c_iH) \leq x_1 + d_{Haus}(L_i, c_{i}H).$$
Thus $$d_{Haus}(L_j, g^{-1}c_{i}H) = d_{Haus}(gL_j, c_{i}H) \leq x_1 + d_{Haus}(L_i, c_{i}H).$$ But note that $d_{Haus}(L_j, c_{j}H) \leq d_{Haus}(L_j, g^{-1}c_{i}H)$ by the definition of $c_j$, so we have that
$$d_{Haus}(L_j, c_{j}H) \leq x_1+d_{Haus}(L_i, c_{i}H),$$
contradicting (\ref{assumption}).
Thus (\ref{eq1}) holds for all $g \in G$.
By translating the $L_i$'s, we can obtain a new set of quasi-lines that each contain $e \in G$, and for which (\ref{eq1}) holds for all $g \in G$, though the quasi-lines may no longer be a finite Hausdorff distance from $H$.
This new sequence of quasi-lines satisfies the hypotheses of Lemma \ref{seq_of_qlines_is_fin}. This leads to a contradiction, since we had assumed $\mathscr{L}$ to be infinite.
\end{proof}
\section{Commensurizers of type $F_n$}
Recall that a group $C$ is of type $F_n$ if there is a $K(C,1)$ with finite $n$-skeleton.
Being of type $F_1$ is equivalent to being finitely generated, and being of type $F_2$ is equivalent to being finitely presented.
We shall prove the following result in this section:
\begin{thm}\label{fn_thm}
Let $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a quasi-isometry between finitely generated groups, and suppose that $C$ is a subgroup of $G$, $C'$ is a subgroup of $G'$, and that $d_{Haus}(f(C), C')<\infty$.
Then $C$ is of type $F_n$ if and only if $C'$ is of type $F_n$, for $n \geq 1$.
\end{thm}
In light of Theorem \ref{Cqiinv}, we have, as an immediate corollary:
\begin{cor}\label{fn_cor}
Let $f, G, G', H$ and $H'$ be as in Theorem \ref{Cqiinv}. Then Comm$_G(H)$ is of type $F_n$ if and only if Comm$_{G'}(H')$ is of type $F_n$.
\end{cor}
We can prove Theorem \ref{fn_thm} in the case that $n=1$ with a short and simple argument, using a coarse geometric characterization for a subgroup to be finitely generated (Lemma \ref{char_fg_subgps}).
To prove the theorem for $n>1$, we will introduce some new terminology and basic facts about ``uniformly distorting'' maps and ``coarse isometries''.
In \cite{Kap_ggt_book}, Kapovich gives a proof that being of type $F_n$ is a quasi-isometry invariant, and we will note that his arguments go through in the more general setting of coarse isometries.
First, we shall see that Theorem \ref{fn_thm} holds when $n=1$.
Consider the following.
\begin{lem}\label{char_fg_subgps}
Let $C$ be a subgroup of a finitely generated group $G$. Then $C$ is finitely generated if and only if there exists some $A_0>0$ such that, for any $g,h \in C$, there is some sequence $s_0, s_1, \ldots s_n \subset C$ so that $g=s_0$, $h=s_n$, and for all $i$, $d(s_i, s_{i+1})< A_0$.
\end{lem}
\begin{proof}
Call a sequence $\{ s_i\}$ as in the statement of the lemma an {\it $A_0$-chain from $g$ to $h$}. If $C$ is finitely generated, then fix a generating set $S_C$ for $C$, and note that the generators of $C$ have word length in $\mathscr{C}^1(G)$ less than some constant $A_0$. For any $g,h \in C$, we can represent $g^{-1}h$ by a word $s_1s_2\cdots s_m$ with each $s_i$ in $S_C$, and then the sequence $e$, $s_1$, $s_1s_2$, $\ldots$, $s_1s_2\cdots s_m=g^{-1}h$ is a $A_0$-chain from $e$ to $g^{-1}h$, and hence $g, gs_1, gs_1s_2, \ldots , gs_1s_2\cdots s_m=h$ is a $A_0$-chain in $C$ from $g$ to $h$.
Assume now that $C$ contains a $A_0$-chain between any two of its elements, for some $A_0$, and let $S_C = C \cap B_{A_0}(e)$.
Since $G$ is finitely generated, $S_C$ is finite, and we claim that $S_C$ generates $C$.
Fix any $h \in C$, and let $e=s_0,$ $ s_1,$ $ \ldots$ $ s_{n-1},s_n=h$ be a $A_0$-chain in $C$ from $e$ to $h$.
Then $h = s_0(s_0^{-1}s_1)(s_1^{-1}s_2)\cdots$ $(s_{n-2}^{-1}s_{n-1})(s_{n-1}^{-1}s_n)$, with $s_0 = e$ and $(s_{i}^{-1}s_{i+1})$ in $S_C$ for each $i$.
Thus $S_C$ generates $C$, so we are done.
\end{proof}
\begin{proof}[Proof of Theorem \ref{fn_thm} in the case that $n=1$]
Recall that $f\colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ is a quasi-isometry, and that $d_{Haus}(f(C), C') < \infty$. We shall show that $C$ is finitely generated if and only if $C'$ is.
Let $f$ be a $(\Lambda, \kappa)$-quasi-isometry. Let $n = d_{Haus}(f(C),C')$, and assume that $C$ is finitely generated. Since $f$ has a quasi-inverse, it suffices to prove that $C'$ must also be finitely generated.
Fix $g',h' \in C'$, and let $s_0, \ldots s_m$ be a sequence of vertices in $C$ such that $d(f(s_0), g')$ $<n$, $d(f(s_m), h')<n$, and $d(s_i, s_{i+1}) < A_0$ for all $i$. Let $s_i' = f(s_i)$, and let $s_i'' \in C'$ be such that $d(s_i', s_i'') <n$. As $d(s_i', s_{i+1}') < \Lambda A_0 +\kappa$, we must have that $d(s_i'', s_{i+1}'') < \Lambda A_0 +\kappa+2n$.
Then the consecutive terms of the sequence $g', s_0'',s_1'', \ldots , s_n'', h'$ are less than $(\Lambda A_0 + \kappa + 2n)$ apart, thus by the lemma above, $C'$ is finitely generated.
\end{proof}
Next, we will introduce a few new notions, and then see that the theorem holds for $n>1$.
If $C$ is a finitely generated subgroup of a finitely generated group $G$, then we shall use $d_G$ to denote the metric on $G$, and hence $C$, induced from the finite generating set that is fixed for $G$, and we shall denote by $d_C$ the metric on $C$ that is induced from the finite generating set fixed for $C$.
For any $w \in C$, we shall write $|w|_C$ for $d_C(e,w)$, and we shall write $|w|_G$ for $d_G(e,w)$, for any $w \in G$.
For simplicity, in the remainder of this section we shall work with metric spaces such as $(G, d_G)$, instead of with Cayley graphs.
\begin{defn}
We shall say that a map between metric spaces, $f \colon\thinspace (X,d_X) \to (Y,d_Y)$, is {\em $(\phi, \Phi)$-uniformly distorting}, or {\em $(\phi, \Phi)$-u.d.} if $\phi$ and $\Phi$ are weakly increasing proper maps from $\mathbb{R}_{\geq 0}$ to $\mathbb{R}_{\geq 0}$, or from $Im(d_X)$ to $Im(d_Y)$, such that, for any $x, x' \in X$ and any $r$,
\begin{enumerate}
\item{if $d_X(x,x') \geq r$ then $d_Y(f(x), f(x')) \geq \phi (r)$, and}
\item{if $d_X(x,x') \leq r$ then $d_Y(f(x), f(x')) \leq \Phi(r)$.}
\end{enumerate}
We will say that $f$ is {\em u.d.} if $f$ is $(\phi, \Phi)$-u.d. for some $\phi$ and $\Phi$.
\end{defn}
Note that the composition of u.d. maps is u.d.
\begin{convention}
The metric spaces that we are interested in are groups with word metrics, hence all distance functions below will take on only integer values.
Therefore we will only consider $(\phi, \Phi)$-u.d. maps where we shall take the domain and range of $\phi$ and $\Phi$ to be $\mathbb{Z}_{\geq 0}$.
\end{convention}
We note the following fact:
\begin{lem}\label{inclusion_is_wup}
Let $G$ be a finitely generated group with finitely generated subgroup $C$. Then the indentity map $i_C \colon\thinspace (C, d_C) \to (C, d_G)$ is u.d.
\end{lem}
\begin{proof}
We shall see that the geometric action of $C$ on itself ensures that all metric distortion is uniform.
Let
$$\phi (r) = \min \{ |c|_G : c \in C, |c|_C \geq r\},$$
and let
$$\Phi(r) = \max \{ |c|_G : c \in C, |c|_C \leq r\}.$$
As $C$ is finitely generated, $\Phi$ is finite valued, and note that both functions are weakly increasing.
In addition, for any $c_1, c_2 \in C$, let $c = c_1^{-1}c_2$ and note that $d_C(c_1, c_2) = |c|_C \geq r$ implies $d_G(c_1, c_2) = |c|_G \geq \phi(r)$ and $d_C(c_1, c_2) = |c|_C \leq r$ implies $d_G(c_1, c_2) = |c|_G \leq \Phi(r)$.
As $G$ is finitely generated, hence locally finite, it follows that $\lim_{r \to \infty} \phi(r) = \lim_{r \to \infty} \Phi(r) = \infty$, and hence that $\phi$ and $\Phi$ are proper.
\end{proof}
We note, though we shall not make use of this fact, that in the above proof, the function $\Phi$ is bounded above by a linear function.
For let $S_C$ denote the finite generating set for $C$ and let
$$L = \max_{s \in S_C} |s|_G.$$
It follows that, for any $c \in C$, $|c|_G \leq L|c|_C$, hence $\Phi(r) \leq Lr$.
\begin{defn}
If $f \colon\thinspace (X, d_X) \to (Y,d_Y)$ is a map between metric spaces and $t \geq 0$, then we will say that $f$ is {\em $t$-onto} if the $t$-neighborhood of $Im(f)$ in $Y$ is equal to $Y$.
If $f$ is $t$-onto for some $t$, then we will say that $f$ is {\em coarsely onto}.
If $f$ is both u.d. and coarsely onto, then we shall say that $f$ is a {\em coarse isometry}.
\end{defn}
We note that any quasi-isometry is a coarse isometry.
\begin{defn}
We say that a function $f_1 \colon\thinspace (X, d_X) \to (Y,d_Y)$ has {\em finite distance } from a function $f_2 \colon\thinspace (X, d_X) \to (Y,d_Y)$ if
$$\sup_{x \in X} d_Y(f_1(x), f_2(x)) < \infty.$$
\end{defn}
Justification for the terminology ``coarse isometry'' is in the following fact:
\begin{lem}\label{coarse_inverse}
If $f \colon\thinspace (X, d_X) \to (Y,d_Y)$ is a coarse isometry between metric spaces, then there is a coarse isometry $f' \colon\thinspace (Y,d_Y) \to (X,d_X)$ such that $f'f$ and $ff'$ have finite distances from $id_X$ and $id_Y$ respectively.
\end{lem}
\begin{defn}
We shall call any function $f'$ satisfying the conclusion of the above lemma a {\em coarse inverse} to $f$.
\end{defn}
\begin{proof}[Proof of Lemma \ref{coarse_inverse}]
Let $\phi, \Phi$ and $t$ be such that $f$ is $t$-onto and $(\phi, \Phi)$-u.d., and note that we can assume that $\Phi(0)= 0$.
Let $\pi$ denote nearest point projection from $Y$ to $Im(f)$, and define $f' \colon\thinspace Y \to X$ to take any $y \in Y$ to a point $x \in X$ such that $f(x) = \pi(y)$.
Let
$$\phi'(r) = \min \{ s \in \mathbb{Z}_{\geq 0} : \Phi(s) \geq r\}$$
and let
$$\Phi'(r) = \max \{ s \in \mathbb{Z}_{\geq 0} : \phi (s) \leq r\}.$$
Note that both $\phi'$ and $\Phi'$ are weakly increasing, and that $\phi'$ is proper.
As $\phi$ is a proper map, it follows that $\Phi'$ is as well.
Suppose that $y_1, y_2 \in Y$ and $r \geq 0$ are such that $d(y_1,y_2) \geq r$. Then we have that $d(\pi(y_1), \pi(y_2)) \geq r-2t$, and it follows that $d(f'(y_1), f'(y_2)) \geq \phi' (r-2t)$.
Similarly if $d(y_1, y_2) \leq r$, then $d(\pi(y_1), \pi(y_2)) \leq r+2t$ and hence $d(f'(y_1), f'(y_2)) \leq \Phi'(r+2t)$.
Hence if we let $\phi''(r) = \phi'(r-2t)$ (taking $\phi'$ to be zero on the negative integers) and $\Phi''(r) = \Phi'(r+2t)$, then we have that $f'$ is $(\phi'', \Phi'')$-u.d.
To see that $f'$ is coarsely onto, note that if $f(x) = f(x')$ then $d(x,x') \leq \Phi'(0)$, and hence for any $y \in Y$, $diam(f^{-1}(y)) \leq \Phi'(0)$.
Note that $Im(f')$ meets $f^{-1}(y)$ for each $y \in Y$, and it follows that $f'$ is $\Phi'(0)$-onto, and hence a coarse isometry.
The above argument also implies that $\sup_{x \in X} d(f'f(x), x) \leq \Phi'(0)$, so the composite map $f'f$ is a finite distance from $Id_X$.
On the other hand, for any $y \in Y$, $ff'(y) = \pi(y)$, hence $\sup_{y \in Y} d(ff'(y), y) \leq t$, so $ff'$ is a finite distance from $Id_Y$ as desired.
\end{proof}
We note also the following, the proof of which is left to the reader:
\begin{lem}
Suppose that $f \colon\thinspace X \to Y, g \colon\thinspace Y \to Z$ are coarse isometries. Then $g f \colon\thinspace X \to Z$ is also a coarse isometry.
\end{lem}
We can now show the following, which explains our interest in coarse isometries:
\begin{prop}\label{cc'_ci}
Let $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a quasi-isometry between finitely generated groups, and suppose that $C$ is a subgroup of $G$, $C'$ is a subgroup of $G'$, and that $d_{Haus}(f(C), C')<\infty$.
Then there is a coarse isometry between $(C, d_C)$ and $(C', d_{C'})$.
\end{prop}
\begin{proof}
Let $i_C \colon\thinspace (C, d_C) \to (C, d_G)$ and $i_{C'} \colon\thinspace (C', d_{C'}) \to (C', d_{G'})$ denote the identity maps on $C$ and $C'$.
As we saw in Lemma \ref{inclusion_is_wup}, both $i_C$ and $i_{C'}$ are u.d. and hence coarse isometries.
Let $i_{C'}'$ be a coarse inverse to $i_{C'}$.
Let $t$ be such that $d_{Haus}(f(C), C')\leq t$. Then we can extend $i_{C'}'$ to a coarse isometry $j_{C'}$ from the $t$-neighborhood of $C'$ in $G'$ to $(C', d_{C'})$ by defining the projection map $\pi$ to take each point $n$ in the neighborhood to a point $c' \in C'$ such that $d_{G'}(n,c') \leq t$, and then setting $j_{C'}(n) = i_{C'}'(c')$.
We have that $j_{C'}$ is a coarse isometry, hence so is $j_{C'} \circ f \circ i_C \colon\thinspace (C, d_C) \to (C', d_{C'})$.
\end{proof}
For any discrete metric space $(X,d_X)$ and $d \geq 0$, we let $Rips_d(X)$ denote the $d$-Rips complex of $X$, i.e. the simplicial complex whose vertex set is equal to $X$, and such that any finite collection $X_0$ of vertices spans a simplex if and only if $d_X(x, x') \leq d$ for all $x,x' \in X_0$.
It is immediate that the proof of Lemma 2.9 of \cite{Kap_ggt_book} extends to the following:
\begin{lem}\label{simpl_maps}
Let $f \colon\thinspace (X, d_X) \to (Y,d_Y)$ be $(\phi, \Phi)$-u.d. Then $f$ induces a simplicial map $Rips_d(X) \to Rips_{\Phi(d)}(Y)$ for each $d \geq 0$.
\end{lem}
We recall Definition 2.10 of \cite{Kap_ggt_book}:
\begin{defn}
A metric space $X$ is said to be {\em coarsely $n$-connected} if, for each $r \geq 0$ there exists some $R \geq r$ such that the map $Rips_r(X) \to Rips_R(X)$ induces the trivial maps on $i^{th}$ homotopy groups, for every $0 \leq i \leq n$.
\end{defn}
Corollary 2.15 of \cite{Kap_ggt_book} shows that coarse $n$-connectedness is a quasi-isometry invariant.
In light of Lemma \ref{simpl_maps}, a minor alteration of that proof shows the following.
\begin{thm}\label{cnc_is_ci_inv}
Coarse $n$-connectedness is a coarse isometry invariant.
\end{thm}
Thus we have
\begin{thm}\label{fn_is_ci_inv}
The property of a finitely generated group being of type $F_n$, $n > 1$, is a coarse isometry invariant.
\end{thm}
\begin{proof}
Suppose that $C$ and $C'$ are finitely generated groups, with $C$ of type $F_n$, and that there is a coarse isometry between $(C,d_C)$ and $(C', d_{C'})$.
Then Theorem 2.17 of \cite{Kap_ggt_book} implies that $(C,d_C)$ is coarsely $(n-1)$-connected, hence, by Theorem \ref{cnc_is_ci_inv}, so is $(C', d_{C'})$. It is shown in the proof of Theorem 2.21 of \cite{Kap_ggt_book} that each coarsely $(n-1)$-connected group has type $F_n$, so the theorem follows.
\end{proof}
Thus, Theorem \ref{fn_thm} in the case that $n>1$ is immediate from Proposition \ref{cc'_ci} and Theorem \ref{fn_is_ci_inv}.
\section{The quasi-isometry invariance of the vertices of commensurizer type of the Scott-Swarup JSJ decomposition}
The proof of Theorem \ref{Cqiinv} was motivated by the goal of showing that certain vertex groups of the Scott-Swarup JSJ decomposition for finitely presented, one-ended groups are invariant under quasi-isometries.
We will see that our result follows immediately from Theorems \ref{existence_qi_inv} and \ref{Cqiinv}, given the theory of Scott and Swarup.
We shall begin with some basic facts about the Scott-Swarup decomposition, and then discuss Papasoglu's results from \cite{papa_qlines} that show the invariance under quasi-isometry of certain parts of the decomposition.
Next we will discuss Scott and Swarup's theory in more detail, and see that our invariance results follow from Theorems \ref{existence_qi_inv} and \ref{Cqiinv}.
For an introduction to splittings and graphs of groups decompositions of groups, and group actions on trees, the reader is referred to \cite{ScottWall}.
In \cite{SS}, Scott and Swarup construct a canonical JSJ decomposition $\Gamma_1(G) $ of any one-ended, finitely presented group $G$, in which the vertex groups ``enclose'' all splittings of $G$ over two-ended subgroups, and moreover, enclose all nontrivial almost invariant subsets of $G$ over two-ended subgroups (see Definition \ref{def_enclosing2}).
If $v$ is a vertex of $\Gamma_1(G)$, then we shall denote its vertex group by $G(v)$, which is defined up to conjugacy. Similarly if $e$ is an edge of $\Gamma_1(G)$, then we shall let $G(e)$ denote its edge group.
$\Gamma_1(G)$ is a regular neighborhood, as defined in \cite{SS}, of all the nontrivial almost invariant subsets of $G$ over two-ended subgroups.
Thus $\Gamma_1(G)$ is a bipartite graph of groups with fundamental group $G$, and with the vertices in the complementary subsets called {\it $V_0$-vertices} and {\it $V_1$-vertices}.
If $v$ is a vertex of $\Gamma_1(G)$, then $G(v)$ is said to be either a {\it $V_0$-} or {\it $V_1$-vertex group}, depending on whether $v$ is a $V_0$- or $V_1$-vertex.
Furthermore, each nontrivial almost invariant subset of $G$ over a two-ended subgroup is ``enclosed'' by some $V_0$-vertex.
We shall define nontrivial almost invariant sets and the notion of enclosure below, but in the case that such an almost invariant set is associated to a splitting of $G$, this means that the enclosing $V_0$-vertex group admits a splitting that is compatible with $\Gamma_1(G)$.
Moreover, when $\Gamma_1(G)$ is refined by this splitting, the added edge is associated to the given splitting of $G$.
Each $V_0$-vertex of $\Gamma_1(G)$ encloses at least one such splitting of $G$ over a two-ended subgroup.
Each $V_0$-vertex $v$ is one of three types:
\begin{enumerate}
\item{$v$ is isolated}
\item{$v$ is of Fuchsian type, or}
\item{$v$ is of commensurizer type.}
\end{enumerate}
If $v$ is isolated, then $v$ is of valence two. Moreover, if we let $e_1$ and $e_2$ denote the edges incident to $v$, then the inclusions of $G(e_1)$ and $G(e_2)$ into $G(v)$ are isomorphisms, and all three subgroups are two-ended.
If $v$ is of Fuchsian type, then $G(v)$ is finite-by-Fuchsian, where the Fuchsian group is a discrete group of isometries of the hyperbolic plane or of the Euclidean plane, but is not finite nor two-ended. Associated to each peripheral subgroup of $G(v)$ there is exactly one corresponding edge $e$ incident to $v$, and $G(e)$ is conjugate to that subgroup.
Lastly, if $v$ is of commensurizer type, then $v$ is not isolated nor of Fuchsian type, and there is a two-ended subgroup $H$ of $G$ with $\tilde{e}(G,H) >3$ such that $G(v) = $ Comm$_G(H)$. Only in this case is it possible that the subgroups carried by the edges incident to $v$ are not two-ended, and in fact they may not even be finitely generated. It follows that the $V_1$-vertex groups of $\Gamma_1(G)$ may not be finitely generated either.
We say that a subgroup $C$ of $G$ is a vertex group of isolated, Fuchsian or commensurizer type respectively if $C$ is the vertex group of a vertex of $\Gamma_1(G)$ of isolated, Fuchsian or commensurizer type respectively.
\\
It is natural to ask if $\Gamma_1(G)$ is somehow invariant under quasi-isometries.
While the underlying graph of $\Gamma_1(G)$ need not be invariant, one could ask whether or not the existence of vertex groups of certain types is invariant under quasi-isometries.
If the answer to this is `yes', then one could ask if the locations of these vertex groups is also invariant under quasi-isometries, in the sense of a quasi-isometry being forced to take a vertex group to within a finite Hausdorff distance of a vertex group of the same type.
Papasoglu has addressed these questions for the Dunwoody-Sageev JSJ decomposition of one-ended, finitely presented groups.
The JSJ decomposition of a group $G$ as given by Dunwoody and Sageev in \cite{DunSa} is a graph of groups decomposition of $G$, say $\Gamma_{DS}(G)$, which is bipartite.
Call the two types of vertex groups white and black, and then all the black vertex groups are either of Fuchsian type or of isolated type (see above).
$\Gamma_{DS}(G)$ describes all the splittings of $G$ over two-ended subgroups, in the sense that if $G$ splits over a two-ended subgroup $C$, either as $A*_CB$ or $A*_C$, then $C$ is conjugate into a vertex group of $\Gamma_{DS}$, has a finite index subgroup which is contained in a black vertex group, and each white vertex group is conjugate into $A$ or $B$.
All of the edge groups of $\Gamma_{DS}(G)$ are two-ended, and it is this that Papasoglu exploits in \cite{papa_qlines} to prove the quasi-isometry invariance of this JSJ decomposition. Specifically, the author proves the following.
\begin{thm} {\rm \cite{papa_qlines}} \
\label{papa_jsj}
Let $G$ and $G'$ be one-ended, finitely presented groups.
Suppose that
$f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ is a quasi-isometry.
Then there is a constant $C>0$ such that if $A$ is a subgroup of $G$ conjugate to a vertex group, a vertex group of Fuchsian type, or an edge group of the graph of groups $\Gamma_{DS}(G)$, then $f(A)$ has Hausdorff distance less than or equal to $C$ from a subgroup of $G'$ conjugate to, respectively, a vertex group, a vertex group of Fuchsian type, or an edge group of $\Gamma_{DS}(G')$.
\end{thm}
Given any one-ended, finitely presented group $G$, $\Gamma_{DS}(G)$ differs from $\Gamma_1(G)$ as follows. The Fuchsian type vertex groups of $\Gamma_{DS}(G)$ and the Fuchsian type vertex groups of $\Gamma_1(G)$ are the same (up to conjugacy), and have the same edge groups.
Also, the isolated vertex groups of $\Gamma_1(G)$ are vertex groups of $\Gamma_{DS}(G)$, and have the same edge groups.
Thus $V_1$-vertices adjacent only to Fuchsian and isolated vertices of $\Gamma_1(G)$ are the same as the corresponding white vertex groups of $\Gamma_{DS}(G)$.
So $\Gamma_1(G)$ differs from $\Gamma_{DS}(G)$ only at the vertices of commensurizer type, and the adjacent edges and $V_1$-vertices.
Thus Theorem \ref{papa_jsj} shows the invariance under quasi-isometries of the $V_0$-vertex groups of $\Gamma_1(G)$ of isolated and Fuchsian types, as well as those $V_1$-vertices that are adjacent only to the isolated and Fuchsian $V_0$-vertices.
It does not, however, answer the question of the invariance of the vertex groups of commensurizer type.
In fact the vertices of commensurizer type are invariant under quasi-isometries, and this fact is an immediate corollary to Theorems \ref{existence_qi_inv} and \ref{Cqiinv}, in light of the following fact about $\Gamma_1(G)$:
\begin{thm}\label{SSJSJ_comm_char}
Let $G$ be a one-ended finitely presented group, and let $H$ be a two-ended subgroup of $G$.
Then Comm$_G(H)$ is a vertex group of $\Gamma_1(G)$ of commensurizer type if and only if $\tilde{e}(G,H) \geq 4$.
\end{thm}
This theorem is not stated explicitly in \cite{SS}, so we digress in order to explain how it follows from that work.
\\
We will need a more detailed description of $\Gamma_1(G)$ in order to do this.
We begin with some definitions from \cite{SS}.
Let $H$ be a subgroup of $G$, let $X$ be a subset of $G$ and let $X^*$ denote the complement of $X$ in $G$.
Recall that $X$ is said to be $H$-finite if it is contained in the union of finitely many cosets $Hg$ of $H$.
\begin{defn}\label{def_ai}
$X$ is an {\em $H$-almost invariant subset of $G$}, or an {\em almost invariant subset of $G$ over $H$}, if $HX=X$ and the symmetric difference of $X$ and $Xg$ is $H$-finite for all $g \in G$.
We say that an $H$-almost invariant set $X$ is {\em nontrivial} if neither $X$ nor $X^*$ is $H$-finite.
\end{defn}
(We note that any $H$-almost invariant subset of $G$ represents an element of $(\mathcal{P}G/\mathcal{F}_HG)^G$ from Section \ref{section3}, though a representative of an element of $(\mathcal{P}G/\mathcal{F}_HG)^G$ need not be fixed by the left action of $H$.)
\begin{rmk} \label{splitting_gives_ai_set}
$H$-almost invariant subsets of $G$ generalize splittings of $G$ over $H$, for there is a natural way to associate to any splitting of $G$ over $H$ an $H$-almost invariant set as follows.
Suppose that $G$ admits a splitting over $H$, and let $T$ be the associated tree.
Let $V(T)$ denote the vertex set of $T$, fix a basepoint $w \in V(T)$ and let $e$ be an oriented edge of $T$ with stabilizer $H$.
Then $e$ determines a partition of $V(T)$: consider the two subtrees of $T$ resulting from the removal of the interior of $e$.
Let $Y_e$ denote the vertex set of the subtree containing the terminal vertex of $e$ and let $Y_e^*$ denote the vertex set of the other subtree.
Let $\varphi \colon\thinspace G \to V(T)$ be defined by setting $\varphi (g) = g\cdot w$, let $Z_e = \varphi ^{-1}(Y_e)$ and let $Z_e^* = \varphi^{-1}(Y_e^*)$.
Lemma 2.10 of \cite{SS} shows that $Z_e$ and $Z_e^*$ are $H$-almost invariant subsets of $G$.
Moreover, these sets are canonically associated to the given splitting, up to complementation and $H$-finite symmetric difference.
(A different choice of basepoint, for instance, would result in an $H$-almost invariant set that has $H$-finite symmetric difference with $Z_e$.) See section 2 of \cite{SS} for more details.
By looking at the translates $gZ_e$ of $Z_e$ by elements of $G$, one can recover the action of $G$ on $T$, but in general one does not get a tree action by looking at translates of almost invariant sets.
\end{rmk}
There is a notion of almost invariant sets crossing:
\begin{defn}\label{def_crossing}
Let $X$ be an $H$-almost invariant subset of $G$ and let $Y$ be a $K$-almost invariant subset of $G$.
We say that {\em $Y$ crosses $X$} if none of the four sets $X \cap Y$, $X^* \cap Y$, $X \cap Y^*$ and $X^* \cap Y^*$ is $H$-finite.
\end{defn}
Scott shows in \cite{Scott_intersection_numbers} that the crossing of nontrivial almost invariant sets is symmetric:
\begin{thm}{\rm \cite{Scott_intersection_numbers}} \
Let $G$ be a finitely generated group with subgroups $H$ and $K$, let $X$ be a nontrivial $H$-almost invariant subset of $G$ and let $Y$ be a nontrivial $K$-almost invariant subset of $G$.
Then $Y$ crosses $X$ if and only if $X$ crosses $Y$.
\end{thm}
Recall that, if $X$ is a subset of $G$ then we may think of $X$ as a subset of the vertex set of $\mathscr{C}^1(G)$, hence we may think of the coboundary of $X$, $\delta X$, as a collection of edges in $\mathscr{C}^1(G)$.
In \cite{SS}, notions of strong and weak crossings play an important role.
\begin{defn}
Let $X$, $Y$ be as in Definition \ref{def_crossing}.
Then we say that {\em $Y$ crosses $X$ strongly} if $\delta Y \cap X$ and $\delta Y \cap X^*$ project to infinite sets in $H \backslash \mathscr{C}^1(G)$.
If $Y$ crosses $X$, but not strongly, then we say that {\em $Y$ crosses $X$ weakly}.
\end{defn}
\begin{rmk}
If $Y$ crosses $X$ strongly, then $Y$ crosses $X$.
Whether or not $Y$ crosses $X$ strongly does not depend on our choice of finite generating set for $G$.
In general, strong crossing is not symmetric: it is possible to have $Y$ cross $X$ strongly and $X$ cross $Y$ weakly (see Example 2.26 of \cite{SS}). However, if $H$ and $K$ are both two-ended, then the notion is symmetric (\cite{SS}, Proposition 7.2).
\end{rmk}
Finally, we must discuss the notion of enclosing. Recall that, given a based $G$-tree $T$ with an edge $e$, we defined almost invariant sets $Z_e$ and $Z_e*$ in Remark \ref{splitting_gives_ai_set}.
\begin{defn}\label{def_enclosing2}
Let $X$ be a nontrivial $H$-almost invariant subset of a group $G$, let $T$ be a $G$-tree and let $\Gamma$ denote the graph of groups decomposition of $G$ associated to $T$.
Pick a basepoint in $V(T)$, so that, for any oriented edge $e$ of $T$, we can define the sets $Z_e$, $Z_e*$ as in Remark \ref{splitting_gives_ai_set}.
Suppose that $u \in V(T)$ is such that, for all edges $e$ of $T$ that are incident to $u$ and directed towards $u$, $X \cap Z_e*$ and $X^* \cap Z_e*$ are $H$-finite.
Then we say that the vertex $v = G\backslash u$ in $\Gamma$ {\em encloses} $X$.
\end{defn}
Suppose for a moment that $X$ is associated to a splitting $\sigma$ of $G$ (see Remark \ref{splitting_gives_ai_set}).
In this special case, a vertex $v$ in $\Gamma$ enclosing $X$ is equivalent to $\Gamma$ and $\sigma$ having a common refinement $\Gamma'$, that differs from $\Gamma$ only in that $v$ is replaced by an edge, and the splitting of $G$ associated to that edge is $\sigma$.
The reader is referred to Section 4 of \cite{SS} for more detail and motivation for the concept of enclosing.
We have defined everything we need in order to discuss $\Gamma_1(G)$ in enough detail to explain how Theorem \ref{SSJSJ_comm_char} follows from \cite{SS}.
All almost invariant subsets of $G$ discussed in the remainder of this section are over two-ended subgroups.
Recall that $\Gamma_1(G)$ is such that its $V_0$-vertices enclose all nontrivial almost invariant subsets of $G$, and that each $V_0$-vertex encloses at least one such subset.
If $v$ is isolated, then the only almost invariant sets enclosed by $v$ are those from the splitting of $G$ associated to the edges incident to $v$, hence $v$ does not enclose any crossing almost invariant sets. Conversely, if a $V_0$-vertex $v$ does not enclose any crossing almost invariant sets over two-ended subgroups, then $v$ is isolated.
Moreover, if $X$ is an $H$-almost invariant set of $G$ which is enclosed by $v$, then $\tilde{e}(G,H)$ must be $2$ or $3$ (see part 1 of Theorem 1.9 from \cite{SS_PDn_pairs}).
If $v$ is of Fuchsian type, then $v$ is not isolated, and any almost invariant sets enclosed by $v$ that cross do so strongly. (See Propositions 7.2, 7.4 and 7.5 of \cite{SS}.)
Also, Theorem 7.8 of \cite{SS}, tells us that, if $X$ is an $H$-almost invariant set that is enclosed by a vertex $v$ of Fuchsian type, then we have that either $\tilde{e}(G,H) = 2$, or $X$ is associated to the splitting given by an edge incident to $v$.
If $v$ is of commensurizer type, then any two almost invariant sets enclosed by $v$ that cross do so weakly. Moreover, if $X$ and $Y$ are almost invariant sets over subgroups $H$ and $K$ that are enclosed by $v$, then $H$ and $K$ are commensurable (see Propositions 7.3 and 7.5 of \cite{SS}), $G(v)=$ Comm$_G(H)=$ Comm$_G(K)$, and $\tilde{e}(G,H) = \tilde{e}(G,K)$ is at least 4.
Conversely, if $H$ is a two-ended subgroup of $G$ such that $\tilde{e}(G,H) \geq 4$ and there exists a nontrivial $H$-almost invariant subset of $G$, then there is a commensurizer vertex group of $\Gamma_1(G)$ that is equal to Comm$_G(H)$.
(See part 1 of Theorem 1.9 from \cite{SS_PDn_pairs}.)
We note that if $\tilde{e}(G,H) \geq 4$ and there are no nontrivial $H$-almost invariant subsets of $G$, then Lemma 2.40 of \cite{SS} implies that $H$ contains a (finite index) subgroup $H'$ such that $\tilde{e}(G,H') = \tilde{e}(G,H)$ and there exists a nontrivial $H'$-almost invariant subset of $G$.
Hence in this case $\Gamma_1(G)$ contains a commensurizer vertex group equal to Comm$_G(H') = $ Comm$_G(H)$.
Thus if $H$ is a two-ended subgroup of $G$, then Comm$_G(H)$ is a vertex group of $\Gamma_1(G)$ of commensurizer type if and only if $\tilde{e}(G,H) \geq 4$.
This proves Theorem \ref{SSJSJ_comm_char}, and thus we have the following immediate corollary to Theorem \ref{existence_qi_inv}:
\begin{cor}
Let $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$.
Then $\Gamma_1(G)$ has a vertex group of commensurizer type if and only if $\Gamma_1(G')$ does.
\end{cor}
Moreover, the next fact follows immediately from Theorem \ref{Cqiinv}.
\begin{cor}\label{SSJSJ_cor}
Let $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G')$ be a quasi-isometry between the Cayley graphs of one-ended, finitely presented groups $G$ and $G'$, and suppose that $C$ is a vertex group of $\Gamma_1(G)$ of commensurizer type.
Then there is a vertex group, $C'$, of $\Gamma_1(G')$ of commensurizer type such that
$$d_{Haus}(f(C), C')<\infty.$$
\end{cor}
In addition, ``small'' and ``large'' vertex groups of commensurizer type are invariant under quasi-isometries.
We recall from \cite{SS} that a vertex group $C=$ Comm$_G(H)$ of $\Gamma_1(G)$ of commensurizer type is said to be {\it small} if $H$ is of finite index in $C$, and otherwise, $C$ is said to be {\it large}.
Thus $C$ is small if and only if $e(C)=2$.
Hence Theorem \ref{Cqiinv} also implies the following.
\begin{cor}
If $f \colon\thinspace \mathscr{C}^1(G_1) \to \mathscr{C}^1(G_2)$ is a quasi-isometry between the Cayley graphs of finitely presented, one-ended groups $G_1$ and $G_2$, then $\Gamma_1(G_1)$ has a vertex group of small commensurizer type if and only if $\Gamma_1(G_2)$ does, and $\Gamma_1(G_1)$ has a vertex group of large commensurizer type if and only if $\Gamma_1(G_2)$ does.
\end{cor}
\section{On the quasi-isometry invariance of the topological JSJ decomposition}
We recall that any orientable Haken 3-manifold $M$ with incompressible boundary has a JSJ decomposition, and the ``characteristic'' pieces of this decomposition essentially make up the characteristic submanifold of $M$, $V(M)$.
(See below for a description of $V(M)$.)
In this section, we shall discuss how Scott and Swarup's theory of JSJ decompositions of groups, together with Corollary \ref{SSJSJ_cor}, imply the invariance under quasi-isometries of the Seifert fibered components of $V(M)$ that meet $\partial M$.
We first remark that M. Kapovich and Leeb have used different methods to prove a stronger result that implies this one.
In Theorem 1.1 of \cite{KapLeeb_Haken}, the authors prove the quasi-isometry invariance of all components of characteristic submanifolds of Haken manifolds with zero Euler characteristic.
We also note that, in the earlier work \cite{KapLeeb_SFS}, the authors proved the quasi-isometry invariance of the existence of (not necessarily peripheral) Seifert fibered components of characteristic submanifolds for Haken manifolds with zero Euler characteristic that are not Nil nor Sol.
Our main result in this section is the following.
\begin{cor}\label{3man_app}
Let $M$ and $M'$ be connected orientable Haken 3-manifolds with incompressible boundary, and with $f\colon\thinspace \pi_1(M) \to \pi_1(M')$ a quasi-isometry. Suppose that $N$ is a nonexceptional Seifert fibered component of the characteristic submanifold of $M$ that meets the boundary of $M$.
Then there is a nonexceptional Seifert fibered component, $N'$, of the characteristic submanifold of $M'$ that meets the boundary of $M'$.
Moreover, if $C$ denotes the subgroup of $\pi_1(M)$ induced by the inclusion of $N$ into $M$ and $C'$ denotes the subgroup of $\pi_1(M')$ induced by the inclusion of $N'$ into $M'$, then
$$d_{Haus}(f(C), C')<\infty.$$
\end{cor}
This corollary will follow immediately from Corollary \ref{SSJSJ_cor} and the following proposition.
\begin{prop} \label{3man_prop}
Let $M$ be a connected orientable Haken 3-manifold with incompressible boundary and let $G=\pi_1(M)$.
Then there is a one-to-one correspondence between the nonexceptional Seifert fibered components of $V(M)$ that meet $\partial M$ and the commensurizer vertex groups of $\Gamma_1(G)$, given by taking a Seifert fibered component $N$ to the subgroup of $G$ given by the inclusion of $\pi_1(N)$ into $\pi_1(M)$.
\end{prop}
Before proving Proposition \ref{3man_prop}, we recall some basic definitions.
We say that a 3-manifold $M$ is {\em irreducible} if every 2-sphere in $M$ bounds a 3-ball.
We call a map of a surface $S$ into $M$ {\em proper} if the map takes $\partial S$ into $\partial M$.
A proper embedding of an orientable surface $S$ that is not the disk or 2-sphere into $M$ is said to be {\em incompressible} if it induces an injection on fundamental groups.
An embedding of the 2-sphere into $M$ is {\em incompressible} if the image does not bound a 3-ball.
We say that $M$ has {\em incompressible boundary} if the inclusion of $\partial M$ into $M$ induces an injection on fundamental groups.
We say that $M$ is an {\em orientable Haken} 3-manifold if $M$ is compact, orientable, irreducible, and contains an incompressible surface.
A map of the torus into $M$ is said to be {\em essential} if it is incompressible and not homotopic into $\partial M$, and a proper map of the annulus into $M$ is said to be {\em essential} if it is incompressible and is not properly homotopic into $\partial M$.
Following \cite{SS}, we shall say that an embedded essential annulus or torus $S$ in $M$ is {\em canonical} if any essential map of the annulus or torus into $M$ can be properly homotoped until it is disjoint from $S$.
We shall say that a submanifold $N$ of $M$ is {\em simple} if any essential map of an annulus or torus into $M$ with image in $N$ can be properly homotoped into the frontier of $N$.
Jaco and Shalen \cite{JacoShalen} and Johannson \cite{Johannson} proved that there is a unique finite collection $\mathcal{T}$ of disjoint canonical annuli and tori in $M$ such that $\mathcal{T}$ contains one representative from each isotopy class of canonical annuli and tori in $M$.
These authors also showed that the pieces obtained by cutting $M$ along $\mathcal{T}$ are $I$-bundles over surfaces, Seifert fibered, or simple; we shall consider these pieces to be submanifolds of $M$.
We define the characteristic submanifold of $M$, $V(M)$, to be the collection of $I$-bundle and Seifert fibered submanifolds as above, except that if two such submanifolds meet one another at some surface $S \in \mathcal{T}$, then we shall remove a regular neighborhood of $S$ from $V(M)$.
Also if two simple submanifolds meet at some $S \in \mathcal{T}$, then we shall add a regular neighborhood of $S$ to $V(M)$.
Thus $V(M)$ is a submanifold of $M$, each component of which is a regular neighborhood of an annulus or a torus, an $I$-bundle over a surface, or is Seifert fibered.
We shall say that a component of $V(M)$ is {\em exceptional} if it is a solid torus with frontier 3 annuli of degree 1, or 1 annulus of degree 2 or 3, or is a twisted I-bundle over the Klein bottle as described in \cite{SS}.
A component of $V(M)$ is called {\em peripheral} if it meets $\partial M$.
We can now show the correspondence between the nonexceptional peripheral Seifert fiber components of $V(M)$ and the commensurizer vertex groups of $\Gamma_1(\pi_1(M))$.
\begin{proof}[Proof of Proposition \ref{3man_prop}]
Let $M$ be a connected orientable Haken 3-manifold with incompressible boundary and let $G=\pi_1(M)$.
For any subgroup $H$ of a group $G$, we shall denote by $N_G(H)$ the normalizer of $H$ in $G$.
The proposition is vacuously true if $G$ is finite, so it suffices to take $M$ such that $G$ is infinite, hence torsion-free.
Thus the two-ended subgroups of $G$ are infinite cyclic.
If $N$ is a Seifert fiber space with infinite fundamental group, then we have the following short exact sequence:
$$1 \to \mathbb{Z} \to \pi_1(N) \to \pi_1(B) \to 1,$$
where the $\mathbb{Z}$ is generated by a regular fiber of $N$ and $B$ denotes the base 2-orbifold of $N$ (see, for example, \cite{Scott_geometries}).
In fact, \cite{Waldhausen_gruppen} and \cite{Tollefson_involutions_of_SFSs} imply that if $N$ is an orientable Haken 3-manifold with infinite fundamental group, then the converse holds: if $\pi_1(N)$ has a normal infinite cyclic subgroup then $N$ is Seifert fibered.
We will be interested in manifolds $N$ that are orientable, irreducible and with nonempty boundary.
Any such $N$ is Haken, so this result will apply.
Consider again $M$ and $G$, and suppose that $H = \langle h \rangle \subset G$ is infinite cyclic.
Recall that $N_G(H) \subset $ Comm$_G(H)$; it follows from work of Jaco and Kropholler that Comm$_G(H) = N_G(\langle h^m \rangle)$ for some $m \geq 1$.
To see this, we have that Jaco showed in \cite{Jaco_roots} that any $g \in $ Comm$_G(H)$ is contained in $N_G(\langle h^n \rangle)$ for some $n$, thus any finitely generated subgroup of Comm$_G(H)$ is contained in $N_G(\langle h^n \rangle)$, for some $n$ depending on the subgroup.
In \cite{Kropholler_note}, Kropholler showed that ascending chains of centralizers in $G$ must terminate, and thus ascending chains of normalizers of infinite cyclic subgroups must also terminate.
Since Comm$_G(H)$ can be exhausted by finitely generated subgroups, it follows that Comm$_G(H) = N_G(\langle h^m \rangle)$ for some $m\geq 1$, and that Comm$_G(H)$ is finitely generated.
We note that if $M$ is Seifert fibered and $H \cong \mathbb{Z}$ denotes the subgroup of $\pi_1(M)$ that is carried by a regular fiber of $M$, then $H$ has infinite index in $\pi_1(M)$, and has more than three coends in $G$ if and only if $M$ is a nonexceptional Seifert fiber space with nonempty boundary.
Thus the proposition follows if $M$ is Seifert fibered or $G$ is itself of commensurizer type.
We shall now prove the proposition, assuming neither of these are the case.
Suppose that $N$ is a Seifert fibered component of $V(M)$ and let $C$ denote the subgroup of $G$ that is carried by $N$.
Thus $N$ is orientable, Haken and with boundary, and $C$ must be finitely generated.
Let $H = \langle h \rangle$ denote the subgroup of $G$ generated by a regular fiber of $N$, so that $H$ is normal in $C$.
As we noted above, $H$ is of infinite index in $C$, and $N$ is peripheral and nonexceptional (hence a peripheral component of $V(M)$) exactly when $\tilde{e}(G,H)>3$.
In this case, $C$ is contained in a commensurizer vertex group $C' = $ Comm$_G(H)$ of $\Gamma_1(G)$, with $C' = N_G(\langle h^m \rangle)$ finitely generated and $m \geq 1$.
Suppose that this containment is proper.
Consider the cover of $M$ with fundamental group $C'$, which we shall call $M_{C'}$, and denote by $\rho$ the projection from $M_{C'}$ to $M$.
By the theory of Scott \cite{Scott_compact_core}, $M_{C'}$ contains a compact submanifold, let's say $N'$, with fundamental group $C'$.
Thus $N'$ is a Seifert fiber space, and it follows from \cite{JacoShalen} or \cite{Johannson} that $\rho$ can be homotoped so that $\rho(N')$ is contained in a Seifert fibered component of $V(M)$.
As we have assumed that $C$ is properly contained in $C'$, this component must be different from $N$. Let us call it $S$, and note that $S$ is nonexceptional and peripheral, for $\tilde{e}(G, \langle h^m \rangle ) > 3$ and hence the subgroup carried by a regular fiber of $S$ will also have more than three coends in $G$.
It follows that there is a collection of canonical annuli and tori that separates $N$ from $S$ in $M$, which we shall call $\Sigma_1, \ldots , \Sigma_k$.
Hence $G$ has a graph of groups decomposition over the surface groups $\pi_1(\Sigma_i)$, with $C$ and $C'$ contained in distinct vertex groups.
But $C \subset C'$, so $C$ must be contained, up to conjugacy, in an edge group $\pi_1(\Sigma_i)$.
This is not possible, hence we must have that $C = C'$.
Now suppose that $C' = $ Comm$_G(H)$ is a commensurizer vertex group of $\Gamma_1(G)$, so $C' = N_G(\langle h^m \rangle)$ for some $m \geq 1$ and $C'$ is finitely generated.
Consider the cover of $M$ with fundamental group $C'$, which we shall call $M_{C'}$, and denote by $\rho$ the projection from $M_{C'}$ to $M$.
As we saw above, $M_{C'}$ contains a Seifert fiber space $N'$, and $\rho$ can be homotoped so that $\rho(N')$ is contained in a Seifert fibered component $S$ of $V(M)$.
Let $D$ denote the inclusion of $\pi_1(S)$ into $G$, and we have that $C' \subset D$ and $D = N_G(H')$, where $H'$ is a finite index subgroup of $\langle h^m \rangle$.
We note that $\tilde{e}(G,H') > 3$, so $S$ is nonexceptional peripheral.
But $N_G(H') \subset $ Comm$_G(H') = $ Comm$_G(H)$, so $D=C'$, and the proposition follows.
\end{proof}
\section{Application to the groups $QI(G)$}\label{qi(g)_section}
Proposition \ref{fin_H_dist} gives us some insight into the structure of groups of quasi-isometries of one-ended, finitely presented groups.
In this section, we provide an analogue (Corollary \ref{qi1}) to a result of Souche and Wiest, who investigate $QI(T\times \mathbb{R}^n)$ for infinite trees $T$ in \cite{SoucheWiest}.
In addition, we note Corollary \ref{qi2}, which is a weaker, but far more general result.
We first introduce the notion of the group of quasi-isometries of a group.
Given metric spaces $X$ and $Y$, one may consider all quasi-isometries from $X$ to $Y$, modulo the relation that $f \sim f'$ when
\begin{equation}\label{fin_sup_dist}
\sup_{x \in X}d(f(x), f'(x))<\infty.
\end{equation}
We shall denote this set by $QI(X,Y)$.
It is standard to denote $QI(X , X)$ by $QI(X)$, and $QI(\mathscr{C}^1(G))$ by $QI(G)$, for any finitely generated group $G$.
These latter sets form groups, and are generally very complicated--for instance, Sankaran showed in \cite{Sankaran} that $QI(\mathbb{Z})$ contains Thompson's group F and the free group of continuous rank.
Our first corollary is about the quasi-isometries of commensurizer groups.
Suppose that $G$ is finitely presented, one-ended, and equal to the commensurizer of a two-ended subgroup $H$ such that $\tilde{e}(G,H) \geq 3$.
Then recall that we can think of $H$ as a subset of the vertex set of $\mathscr{C}^1(G)$.
The vertex set of $\mathscr{C}^1(G)$ is equal to the disjoint union of the translates (that is, cosets) of $H$, and any two translates $gH, g'H$ are of finite Hausdorff distance from one another (see Lemma \ref{commcharacterization}).
Thus we can define a metric on $G/H$ such that the distance between $gH$ and $g'H$ is equal to the Hausdorff distance between $gH$ and $g'H$ in $\mathscr{C}^1(G)$.
Note that if $H$ happens to be a normal subgroup of $G$, then this recovers the metric on the vertex set of the Cayley graph for $G/H$, with respect to the given generating set for $G$.
If $f$ is a quasi-isometry from $G$ to itself, then, by Theorem \ref{existence_qi_inv} and Proposition \ref{fin_H_dist}, there is some infinite cyclic $H' \subset G$ with $\tilde{e}(G,H')\geq 3$ and a constant $y$ such that, for each $g \in G$, $d_{Haus}(f(gH), g'H')<y$ for some $g' \in G$.
In addition, it follows from Theorem \ref{2.1} that $H$ and $H'$ are a finite Hausdorff distance from one another in $\mathscr{C}^1(G)$. Hence there is some constant $z \geq y$ such that, for each $g \in G$, $d_{Haus}(f(gH), g'H)<z$ for some $g' \in G$.
Thus the quasi-isometry $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G)$ induces a map from $G/H$ to itself, that takes any $gH$ to some point $g'H$ such that $d_{Haus} (f(gH), g'H)$ $< z$.
As $f$ is a quasi-isometry, it follows that this new map is as well.
We note moreover that any quasi-isometry of $G/H$ to itself induces a natural quasi-isometry of $G$ to itself.
Hence we have the following.
\begin{cor}\label{qi1}
Suppose that $G$ is a one-ended, finitely presented group such that $G = $ Comm$_G(H)$ for a two-ended subgroup $H$ of $G$ that has at least three coends in $G$.
Consider $G/H$, with a metric defined by setting the distance between $gH$ and $g'H$ to equal the Hausdorff distance between the vertex sets $gH$ and $g'H$ in $\mathscr{C}^1(G)$.
Then there is a canonical map $QI(G) \to QI(G/H)$ that is surjective.
\end{cor}
\begin{rmk}
We note that the kernel of this map $QI(G) \to QI(G/H)$ is exactly the set of equivalence classes of quasi-isometries $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G)$ for which the distances $d_{Haus}(f(gH), gH)$ are uniformly bounded.
\end{rmk}
We can generalize this idea to get a weaker result for general one-ended finitely presented groups $G$.
Suppose that $G$ contains at least one two-ended subgroup $H$ with $\tilde{e}(G,H) \geq 3$, and for any such $H$, let $M_H$ denote the metric space with underlying set equal to Comm$_G(H)/H$ and the distance between any two points $gH$ and $g'H$ defined to equal the Hausdorff distance between those sets in $\mathscr{C}^1(G)$.
Fix any such $H$, and any quasi-isometry $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G)$.
Then Theorem \ref{existence_qi_inv} shows that there is some two-ended subgroup of $G$, $H'$, (possibly equal to $H$) such that $d_{Haus}(f(H), H') < \infty$ and $\tilde{e}(G,H) = \tilde{e}(G,H')$.
Proposition \ref{fin_H_dist} implies that $f$ induces a quasi-isometry from $M_H$ to $M_{H'}$ as in the above argument.
Thus, not only do we get a natural map taking $f$ into $QI(M_H, M_{H'})$, but we also get a natural map of $f$ into the symmetric group of a suitable collection of two-ended subgroups of $G$ with a fixed number of coends.
We shall fix some more notation so that we can say this more carefully.
For each $n \in \{ 3,4,5,\ldots \} \cup \{\infty\}$, let $K_n$ denote a maximal collection of two-ended subgroups of $G$ that have $n$ coends in $G$ and are of pairwise infinite Hausdorff distance in $\mathscr{C}^1(G)$.
Thus, for any $n$, $f$ induces an element of $\prod_{H \in K_n} QI(M_H, M_{\sigma(H)})$, for some $\sigma$ in the symmetric group on $K_n$.
In fact, we can say a bit more.
Note that by Theorem \ref{Cqiinv}, any groups $H$ and $\sigma(H)$ as above must have quasi-isometric commensurizers.
Thus if we let $\{K_n^j\}_{j \in J_n}$ denote the partition of $K_n$ into collections of subgroups with quasi-isometric commensurizers, i.e. $H, H' \in K_n^j$ for some $j \in J_n$ if and only if Comm$_G(H)$ and Comm$_G(H')$ are quasi-isometric, then the permutation $\sigma$ above must be contained in $\prod_{j \in J_n}Sym(K_n^j)$, where $Sym (K_n^j)$ denotes the symmetric group on $K_n^j$.
Moreover, $f$ induces such a map for all $n$.
Thus we have the following corollary:
\begin{cor}\label{qi2}
Let $G$ be a one-ended finitely presented group and let $N= \{ 3,4,5,\ldots\} \cup \{ \infty\}$.
For any $n \in N$, let $K_n$ be a maximal collection of two-ended subgroups of $G$ with $n$ coends that have mutually infinite Hausdorff distance, and let $\{ K_n^j\}_{j \in J_n}$ be the partition of $K_n$ into sets of subgroups of quasi-isometric commensurizers.
Then there is a canonical map
$$QI(G) \to \left\{ \prod_{n \in N} \prod_{H \in K_n} QI(M_H, M_{\sigma(H)}) \ : \ \sigma \in \prod_{n \in N} \prod_{j \in J_n} Sym(K_n^j)\right\} .$$
\end{cor}
\begin{rmk}
In contrast to Corollary \ref{qi1}, we expect that the map given in Corollary \ref{qi2} will typically not be surjective.
\end{rmk}
\begin{rmk}
As was the case in Corollary \ref{qi1}, the kernel of the map given in Corollary \ref{qi2} is not hard to describe at the following level: a quasi-isometry $f \colon\thinspace \mathscr{C}^1(G) \to \mathscr{C}^1(G)$ is in this kernel if and only if, for each two-ended subgroup $H$ with at least three coends in $G$, there is a constant $y \geq 0$ such that, for all $g \in$ Comm$_G(H)$, $d_{Haus}(f(gH), gH)<y$.
\end{rmk}
\section{When $F_n \rtimes \mathbb{Z}$ is quasi-isometric to $F_n \times \mathbb{Z}$} \label{semi_section}
Our final application of Theorem \ref{Cqiinv} characterizes when a semi-direct product of a free group with $\mathbb{Z}$ is quasi-isometric to the direct product of the two groups.
Specifically, we shall prove the following.
\begin{cor}\label{semi}
Let $F_n$ be the free group on $n$ generators for any $n \geq 1$.
Then $F_n \rtimes \mathbb{Z}$ is quasi-isometric to $F_n \times \mathbb{Z}$ if and only if it is virtually $F_n \times \mathbb{Z}$.
\end{cor}
Related results have been proven by Bridson and his coauthors.
In Proposition 3.7 of \cite{AlonsoBridson}, Alonso and Bridson show that a group $G$, that is an extension of the form $1 \to H \to G \to Q \to 1$ defined by a function $\varphi \colon\thinspace Q \to Aut(H)$ and a cocycle $f \colon\thinspace Q \times Q \to H$, is quasi-isometric to $H \times Q$ if $\varphi$ and $f$ have finite images.
In \cite{Bridson_optimal_isoper_ab_by_free}, Bridson considers groups of the form $A \rtimes F$ for $A$ finitely generated and abelian, and $F$ finitely generated and free.
By characterizing when the Dehn functions of these groups are polynomial and determining their degrees in those cases, Bridson is able to give necessary conditions for these groups to be quasi-isometric to one another.
In \cite{BridsonGersten_optimal_isoper_torus_bundles}, Bridson and Gersten further analyze the groups of the form $\mathbb{Z}^n \rtimes \mathbb{Z}$, and find further necessary (and in special cases, sufficient) conditions for two such groups to be quasi-isometric (see section 5 of \cite{BridsonGersten_optimal_isoper_torus_bundles}).
We shall begin the proof of Corollary \ref{semi} with a few standard lemmas.
\begin{lem}\label{free_gp_lemma}
Let $F_n$ be the free group on $n$ generators, $n \geq 1$, and let $H$ be an infinite cyclic subgroup of $F_n$. Then Comm$_{F_n}(H)$ is the unique maximal infinite cyclic subgroup of $F_n$ that contains $H$.
\end{lem}
\begin{proof}
Let $\mathscr{C}^1(F_n)$ be the Cayley graph of $F_n$ with respect to the standard presentation, thus $\mathscr{C}^1(F_n)$ is a regular $2n$-valent tree on which $F_n$ acts simplicially, freely and with quotient a bouquet of $n$ copies of $S^1$.
Let $h$ denote a generator of $H$. As $h$ acts freely on $\mathscr{C}^1(F_n)$, $h$ has an axis, which we shall denote by $\gamma$. Let $H'$ denote the subgroup of $F_n$ that preserves $\gamma$, so $H \subseteq H'$. As $F_n$ is torsion free, it follows that $H'$ is infinite cyclic.
Note that any infinite cyclic subgroup of $G$ containing $H$ must preserve $\gamma$ (see Proposition II.6.2(2) of \cite{BridsonHaefliger}), hence $H'$ must be the unique maximal infinite cyclic subgroup of $F_n$ that contains $H$.
Let $h'$ denote a generator of $H'$, and fix $k \in \mathbb{Z}$ such that $h = (h')^k$. It is clear that $h'$ normalizes $H$, and hence that $H' \subseteq $ Comm$_{F_n}(H)$.
On the other hand, suppose that $g \in $ Comm$_{F_n}(H)$, hence there exist nonzero integers $n,m$ such that $h^n = g h^m g^{-1}$.
It follows that the axis of $gh^mg^{-1}$ equals $\gamma$.
On the other hand, the axis of $gh^mg^{-1}$ must be $g \gamma$ (see again Proposition II.6.2(2) of \cite{BridsonHaefliger}), thus $g\gamma = \gamma$ so $g \in H'$. Hence $H' = $ Comm$_{F_n}(H)$.
\end{proof}
\begin{lem}\label{roots_in_fn}
Let $F_n$ be the free group on $n$ generators with $n \geq 1$, and suppose that $f_1, f_2 \in F_n$ and $k$ is an integer, $k \neq 0$, such that $f_1^k = f_2^k$. Then $f_1 = f_2$.
\end{lem}
\begin{proof}
We may assume that $k>0$.
For each $i$, note that $f_i$ can be expressed uniquely as a minimal word of the form $u_iw_iu_i^{-1}$, where $w_i$ is cyclically reduced. Thus, for any positive integer $j$, $f_i^j$ is represented by the minimal word $u_iw_i^ju_i^{-1}$, and thus the word length of $f_i^j$ (with respect to the standard generating set for $F_n$) is equal to $l(u_i) + j\cdot l(w_i) + l(u_i^{-1}) = 2l(u_i) + j\cdot l(w_i)$.
If $f_1^k = f_2^k$ then, for any integer $m$, $f_1^{km} = f_2^{km}$. Hence
$$2l(u_1) + mk\cdot l(w_1) = 2l(u_2) + mk\cdot l(w_2)$$
for any integer $m$, and it follows that $l(w_1) = l(w_2)$ and $l(u_1) = l(u_2)$. As each $u_iw_i^ku_i^{-1}$ is a reduced word, it follows that $u_1 = u_2$, and $w_1 = w_2$, thus $f_1 = f_2$.
\end{proof}
We can now prove the main fact needed for Corollary \ref{semi}:
\begin{prop}\label{infinite_prop}
Let $n \geq 1$, let $\alpha \colon\thinspace \mathbb{Z} \to Aut(F_n)$ be a homomorphism, let $\pi \colon\thinspace Aut(F_n) \to Out(F_n)$ be the quotient homomorphism, let $G$ denote the semidirect product $F_n \rtimes_\alpha \mathbb{Z}$ and let $H$ be an infinite cyclic subgroup of $G$.
If $\pi \alpha$ has infinite image, then Comm$_G(H)$ is of infinite index in $G$.
\end{prop}
\begin{proof}
The proposition is vacuously true in the case that $n=1$, so suppose that $n>1$.
Let $S_F = \{ f_1, f_2, \ldots , f_n\}$ be the standard generating set for $F_n$. Let $A$ denote the infinite cyclic subgroup of $G$ such that $G = F_n \rtimes_\alpha A$, and let $t$ denote a generator of $A$.
For any $b \in A$, let $\alpha_b$ denote the image of $b$ under $\alpha$.
Recall that $G$ is generated by $S_F \cup A$, and that, for any $a \in A$ and $f \in F_n$, we have that
$$af = \alpha_a(f)a \ {\rm and } \ fa=a\alpha_{a^{-1}}(f).$$
Thus any $g \in G$ can be written uniquely in the form $fa$, for some $f \in F_n$ and $a \in A$.
Let $H = \langle h \rangle$ be an infinite cyclic subgroup of $G$. If $h \in F_n$, then Lemma \ref{free_gp_lemma} implies that Comm$_{F_n}(H)$ is an infinite cyclic subgroup of $F_n$. Thus $[F_n:$ Comm$_{F_n}(H)] = \infty$.
Note that Comm$_{F_n}(H) = $ Comm$_G(H) \cap F_n$, hence in this case Comm$_G(H)$ is of infinite index in $G$.
Suppose next that $h = a \in A$. If there is some $g $ contained in $F_n \cap $Comm$_G(H)$, then there must be nonzero integers $k, k'$ such that
$$a^k = ga^{k'}g^{-1}.$$
Then $a^k = g\alpha_{a^{k'}}(g^{-1})a^{k'}$. As $g\alpha_{a^{k'}}(g^{-1}) \in F_n$, $a^k, a^{k'} \in A$ and $F_n \cap A = \o$, it follows that $k=k'$, and that $g = \alpha_{a^k}(g)$.
Note that $|$Im$(\pi \alpha )|= \infty$ implies that $|$Im$(\alpha)|=\infty$, and we claim that $|$Im$(\alpha)|=\infty$ implies that there exists some $g_0 \in F_n$ such that $\alpha_b(g_0) \neq g_0$ for all nontrivial $b \in A$. For if no such $g_0$ did exist, then, for each $i=1, 2, \ldots n$, there would be some positive integer $k_i$ such that $\alpha_{t^{k_i}}(f_i) = f_i$. Let $k$ equal the least common multiple of the $k_i$'s, and it follows that $\alpha_{t^k}$ fixes each generator of $F_n$. Thus $\alpha_{t^k} = id$ and $|$Im$(\alpha)| \leq k$, a contradiction.
Hence there is some $g_0 \in F_n$ that is such that $\alpha_b(g_0) \neq g_0$ for all nontrivial $b \in A$, and hence $g_0 \notin $ Comm$_G(H)$.
It follows from Lemma \ref{roots_in_fn} that, for each nonzero integer $r$ and for all nontrivial $b \in A$, $\alpha_b(g_0^r) \neq g_0^r$, and hence $g_0^r \notin $ Comm$_G(H)$.
Thus the elements $g_0^r$, $r \in \mathbb{Z}$, are contained in distinct cosets of Comm$_G(H)$ in $G$, and so $[G:$ Comm$_G(H)]=\infty$ in this case as well.
The general case remains: suppose that $h$ is neither contained in $F_n$ nor contained in $A$. Then there are nontrivial $f \in F_n, a \in A$ such that $h = fa$.
Suppose that there is some element $g \in F_n \cap$Comm$_G(H)$, so there exist nonzero integers $k,k'$ such that
$$(fa)^k = g(fa)^{k'}g^{-1}.$$
Then there exist $f', f'' \in F_n$ such that the above equation becomes
$$f'a^k = f''a^{k'},$$
and hence $k=k'$ as in the previous case.
For any integer $i$, define $m_i = (fa)^ia^{-i}$, and note that $m_ia^i = (fa)^i$, so in particular, $m_i \in F_n$ and, when $i$ is nonzero, $m_i$ is nontrivial.
We claim that, for any $g \in F_n$, $(fa)^k = g(fa)^{k}g^{-1}$ if and only if $\alpha_{a^k}(g) = m_k^{-1}gm_k$.
To see this, note that we have:
$$(fa)^k = g(fa)^{k}g^{-1}$$
$$\Leftrightarrow m_ka^k = gm_ka^kg^{-1} = gm_k\alpha_{a^k}(g^{-1})a^k$$
$$\Leftrightarrow m_k = gm_k\alpha_{a^k}(g^{-1})$$
$$\Leftrightarrow \alpha_{a^k}(g) = m_k^{-1}gm_k$$
as desired.
Now suppose that there is some nontrivial $g \in F_n$ that is not contained in Comm$_G(H)$. Then there is no nonzero $k$ such that $\alpha_{a^k}(g) = m_k^{-1}gm_k$.
Suppose that some nontrivial power of $g$, $g^r$, was contained in Comm$_G(H)$. Then, for some nonzero value of $k$, we would have
$$\alpha_{a^k}(g^{r}) = m_k^{-1}g^rm_k$$
and hence
$$[\alpha_{a^k}(g)]^{r} = [m_k^{-1}gm_k]^r.$$
By Lemma \ref{roots_in_fn}, $\alpha_{a^k}(g) = m_k^{-1}gm_k$, so we have reached a contradiction.
Thus if $g \notin$ Comm$_G(H)$, then for each nonzero integer $r$, $g^r \notin $ Comm$_G(H)$. As in the previous case, it follows that the elements $g^r$ are contained in distinct cosets of Comm$_G(H)$ in $G$ and $[G:$ Comm$_G(H)]=\infty$ as desired.
It remains to consider the case that $F_n \subset $ Comm$_G(H)$.
In this case, we have that in particular $f_i \in $ Comm$_G(H)$ for each $1 \leq i \leq n$; let $k_i$ be a nonzero integer such that
$$(fa)^{k_i} = f_i(fa)^{k_i}f_i^{-1}.$$
Note that we can take each $k_i$ to be positive.
Let $k$ equal the least common multiple of the $k_i$'s. Thus, for each $i$,
$$(fa)^{k} = f_i(fa)^{k}f_i^{-1}$$
and hence $\alpha_{a^k}(f_i) = m_k^{-1}f_im_k$ for each $i$. It follows that $\alpha_{a^k}$ acts on $F_n$ by conjugation by $m_k$ and hence that $|$Im$(\pi \alpha)|$ is finite.
It follows that $|Im(\pi \alpha)|=\infty$ implies that $[G:$ Comm$_G(H)]=\infty$ for any infinite cyclic subgroup $H$ of $G$.
\end{proof}
We can now prove the main result in this section.
\\
\noindent {\bf Corollary \ref{semi}.}{\it
Let $F_n$ be the free group on $n$ generators for any $n \geq 1$.
Then $F_n \rtimes \mathbb{Z}$ is quasi-isometric to $F_n \times \mathbb{Z}$ if and only if it is virtually $F_n \times \mathbb{Z}$.
}
\\
\begin{proof}
Let $\alpha \colon\thinspace \mathbb{Z} \to Aut(F_n)$ be the homomorphism such that $F_n \rtimes \mathbb{Z}=F_n \rtimes_\alpha \mathbb{Z}$.
Let $A = \langle t \rangle$ denote the infinite cyclic subgroup in the statement of the corollary, let $F_n = \langle f_1, \ldots , f_n\rangle$, let $G_\rtimes$ denote $F_n \rtimes_\alpha A$ and let $G_\times$ denote $F_n \times A$.
Certainly if $G_\rtimes$ is virtually $G_\times$, then the groups are quasi-isometric.
We claim that $|$Im$(\pi \alpha)|<\infty$ implies that $G_\rtimes$ is virtually $G_\times$, where $\pi$ denotes the quotient homomorphism from $Aut(F_n)$ to $Out(F_n)$ as in the previous proposition.
For suppose that $|$Im$(\pi \alpha)|<\infty$, so that there is some nonzero $k$ such that $\alpha_{t^k}$ is an inner automorphism of $F_n$, with $m \in F_n$ such that $\alpha_{t^k}$ takes any $f \in F_n$ to $mfm^{-1}$.
Let $G_k$ denote the subgroup of $G_\rtimes$ that is generated by $f_1, \ldots , f_n,$ and $t^k$, and we have that $G_k \cong G_\times$, by an isomorphism from $G_\times$ to $G_k$ that acts as the identity on $F$, and takes $t$ to $m^{-1}t^k$.
To see that $G_k$ is of finite index in $G_\rtimes$, recall that any element of $G_\rtimes$ can be written uniquely in the form $ft^l$ for some $f \in F_n$ and $l \in \mathbb{Z}$. It follows that any $g \in G_\rtimes$ is of the form $(ft^{mk})(t^i)$, for some $m \in \mathbb{Z}$ and $0 \leq i < k$. Thus $G_\rtimes = \cup_{i=0}^{k-1}G_kt^i$, so $G_k$ is of index no more than $k$ in $G_\rtimes$, as desired.
Thus it remains to show that $G_\rtimes$ and $G_\times$ are {\em not} quasi-isometric in the case that $|Im(\pi \alpha)|=\infty$.
Recall from Proposition \ref{infinite_prop} that in this case, for each subgroup $H \cong \mathbb{Z}$ of $G_\rtimes$, $[G_\rtimes:$ Comm$_{G_\rtimes}(H)]=\infty$.
It is clear that the quotient by $A$ of the Cayley graph of $G_\times$ has infinitely many ends, and hence $\tilde{e}(G_\times, A) = \infty$. Note also that $G_\times =$ Comm$_{G_\times}(A)$.
Thus if there were a quasi-isometry from $G_\times$ to $G_\rtimes$, then it would follow from Theorem \ref{Cqiinv} that there is some $z>0$ and some infinite cyclic subgroup $H$ of $G_\rtimes$ such that the $z$-neighborhood of Comm$_{G_\rtimes}(H)$ in the Cayley graph of $G_\rtimes$ is equal to the whole Cayley graph. This, however, would imply that Comm$_{G_\rtimes}(H)$ is of finite index in $G_\rtimes$, a contradiction.
\end{proof}
|
\section{Introduction}
Every simplest cubic field can be generated by adjoining a root of the equation \cite{sd}
\begin{equation}
x^{3}-kx^{2}+(k-3)x+1=0, \tag{1}
\end{equation}
for some suitable \(k\), where $k$ is a parameter belonging to set of rational numbers. Conversely,
the roots of $(1)$, for some rational number $k$, are either rational or generate a simplest cubic field.
It can be observed that two different values of parameters can generate the same simplest cubic field. In this paper we will classify the parameters according to the corresponding simplest cubic fields. We also found the relation between the roots of equation (1) for different values of parameters belonging to the same class.
We assert that these relations could be a step forward to understand about integral points on elliptic curves related to simples cubic fields \cite{ds}
Each element $\omega $ of $Q(\alpha )$ has
the linear fractional representation \(\omega =\frac{a\rho+b}{c\rho +d}\),
where the determinant of $\omega $, denoted by $\det (\omega )$, is defined
as $ad-bc.$ An element $\omega $ of $Q(\alpha )$ is rational if and only if $\det(\omega )=0$.
\section{Relationship between $\protect\alpha $ and $\acute{\protect\alpha}$
\ when $Q(\protect\alpha )=Q(\protect\alpha ^{\prime })$}
First we split the simplest cubic field into classes. Each element$ \omega $
of $Q(\alpha)$ is related to a linear fractional transformation $f \in PSL(2,Q)$ in the
way that all its conjugates can be obtained by applying $f^n$, for some
integer $n$, over $\omega $. We shall call this correspondence as ``linear
fractional transformation related to an element of $Q(\alpha)$''.
This correspondence defines an equivalence relation on $Q(\alpha )$ and each class corresponds to a unique element of \(PSL(2,Q)\).
We call the equivalence class related to a linear fractional transformation $ f$ as $f$-class. The following corollary is straightforward to see.
\begin{corollary}
If $\omega $ $\in $ $Q(\rho )$, then $\omega $ and its conjugates belong to the same $f$-class.
\end{corollary}
The elements belonging to $y$-class, where \(y\) is defined as \(z \longmapsto \frac{z-1}{z}\), are instrumental in establishing a relationship between $\alpha, \alpha'$ \ and $k,k'$ where $\alpha $ is a root of Equation 1 and $\alpha'$ is a root of same equation with $k$ replaced by $k'$ and $Q(\alpha )=Q(\alpha')$. The following example shows an element belonging to $y$-class.
\begin{example}
The roots of the equation $x^{3}-3x^{2}+1=0$ are\ given by
\begin{equation*}
\rho \text{, }\frac{\rho -1}{\rho }\text{, }\frac{-1}{\rho -1}\text{.}
\end{equation*}%
where $\rho =2\cos (\pi /9)+1$. The algebraic number $\rho $ belongs to $y$%
-class
\end{example}
Note that in $y$-class there are more than one elements, for instance, for $%
k=3$ the roots of Equation 1, are $\rho =2\cos (\pi /9)+1$, $\rho
{\acute{}}
=-\cos (\pi /9)+1-\sqrt{3}\sin (\pi /9)$ and $\rho ^{\prime \prime }=-\cos
(\pi /9)+1+\sqrt{3}\sin (\pi /9)$ and the following two numbers belong to $y$-class
\begin{equation*}
\frac{5\rho -2}{2\rho +3}\text{and }\frac{3\rho -1}{\rho +2}
\end{equation*}%
and the following equations
\begin{equation*}
x^{3}-\left( -\frac{51}{73}\right) x^{2}+(-\frac{51}{73}-3)x+1=0\text{, }
\end{equation*}%
\begin{equation*}
x^{3}-\left( \frac{3}{19}\right) x^{2}+(\frac{3}{19}-3)x+1=0
\end{equation*}
are respectively satisfied by them.
\begin{lemma}
An element $\omega $ of $Q( \alpha )$ belong to $y$-class if and only if it is a root of the irreducible equation $x^{3}-kx^{2}+(k-3)x+1=0$, where $k$ is a rational number.
\end{lemma}
\begin{proof}
Let $\omega$ be an element of $y$-class so the other conjugates of $\omega$
are $\frac{\omega -1}{\omega }$, $\frac{-1}{\omega -1}$ and the equation satisfied by them is:
\begin{equation*}
(x-\omega)(x-\frac{\omega -1}{\omega })(x+\frac{1}{\omega -1})=0
\end{equation*}
which can be simplified as
\begin{equation*}
x^{3}-kx^{2}+(k-3)x+1=0
\end{equation*}
where $k=\omega +\frac{\omega -1}{\omega }-\frac{1}{\omega -1}$. Here $k$ must be rational because it is evolved from the equation satisfied by $\omega$,
which is algebraic over $Q$ of degree three.
Conversely, let $\omega$ satisfy an irreducible equation of the form $x^{3}-kx^{2}+(k-3)x+1=0$, where $k$ is a rational number. Then the other two roots of the equation are $\frac{\omega -1}{\omega }$ and $\frac{-1}{\omega -1}$.
Moreover, $(\omega )y=\frac{(\omega -1)}{\omega }$, $(\frac{\omega -1}{\omega})y=\frac{-1}{(\omega -1)}$ and $(\frac{-1}{\omega -1})y=\omega$ imply that $\omega$ and its conjugates lie in the same triangle.
\end{proof}
Thus, corresponding to each ``triplet of conjugates" of $y$-class there is a rational number $k$. We call such correspondence as the element of $y$-class with parameter $k$.
We use the above classification of elements of $y$-class to prove the
following.
\begin{proposition}
Elements of $Q(\alpha)$ belonging to $y$-class are of the form $\frac{(c+d)\alpha -c}{c\alpha+d}$, where $c$ and $d$ are non-zero integers.
\end{proposition}
\begin{proof}
Let $\omega =\frac{a\alpha +b}{c\alpha +d}$ be any primitive element in $Q(\alpha)$, that is, $ad-bc\neq 0$. Then the conjugates of $\omega $ are given by
\begin{equation*}
\omega'=\frac{(a+b)\alpha -a}{(c+d)\alpha -c}\text{,}
\end{equation*}
and
\begin{equation*}
\omega''=\frac{b\alpha -(a+b)}{d\alpha -(c+d)}.
\end{equation*}
Our next aim is to find the values of $a,b,c$ and $d$ such that $\omega$ belongs to \(y\)-class, that is, when
\begin{eqnarray}
(\omega )y &=&\omega'\\
(\omega')y &=&\omega'' \notag \\
(\omega'')y &=&\omega \notag.
\end{eqnarray}
Now from the first of these equations, we get
\begin{equation*}
\frac{(a-c)\alpha +(b-d)}{a\alpha +b}=\frac{\lambda (a+b)\alpha -\lambda a}{\lambda (c+d)\alpha -\lambda c}\text{,}
\end{equation*}
which further yields the following four equations
\begin{equation*}
\begin{array}{cc}
a(1-\lambda)-c-\lambda b=0 \text{, } & \lambda a+b-d=0, \\
a-\lambda (c+d)=0\text{ and } & b+\lambda c=0.
\end{array}
\end{equation*}
But the $det(\frac{\lambda (a+b)\alpha -\lambda a}{\lambda (c+d)\alpha -\lambda c})=det(\omega )$ which is possible if $\lambda = \pm 1$. For $\lambda =-1$,
the above system of equations gives $a=b=c=d=0$, contradicting the
primitivity of $\omega $. The solution of the system for $\lambda =1$ is $a=c+d$, $b=-c$, $c=c$, $d=d$. This solution satisfies the other two
Equations of (2.1). So the elements of \(y\)-class are of the form $\frac{(c+d)\alpha -c}{c\alpha +d}$.
\end{proof}
Now we are in a position to state the theorem which establishes a
relationship between $\alpha $ and $\alpha ^{\prime }$ whenever $Q(\alpha
)=Q(\alpha ^{\prime })$.
\begin{theorem}
If $\alpha $ is a root of the irreducible equation $x^3-kx^2+(k-3)x+1=0$
and $\alpha'$is a root of the same irreducible equation except $k$ is replaced by $k'$ then $Q(\alpha )=Q(\alpha')$ if and only if $\alpha'$ is of the form $\frac{(c+d)\alpha -c}{c\alpha +d}$ or $\alpha$ is of the form $\frac{(c+d)\alpha'-c}{c\alpha'+d}$ for some integer values of $c$ and $d$, not both zero.
\end{theorem}
\subsection{Relationship between $k$ and $k'$ \ when $Q(\alpha )=Q(\alpha')$}
Hence we will answer the next question about simplest cubic field. First we will find the equation satisfied by any element $\omega =%
\frac{a\rho +b}{c\rho +d}$ of the simplest cubic field $Q(\rho )$ where $\rho$
is the root of the equation $x^{3}-kx^{2}+(k-3)x+1=0$. The representation of $\omega $ as a linear combination of $\{1,\rho ,\rho ^{2}\}$ obtained by the method of indeterminate coefficients is given by
\begin{equation*}
\omega =\frac{a_{1}^{2}\rho +b_{1}\rho +c_{1}}{d_{1}}
\end{equation*}
where
\begin{eqnarray*}
a_{1} &=&(-c(bc-da)) \\
b_{1} &=&((ck+d)(bc-da)) \\
c_{1} &=&(ac^{2}-bc^{2}k+3bc^{2}-bdck-d^{2}b) \\
d_{1} &=&(-c^{2}kd+3c^{2}d-d^{2}ck-d^{3}+c^{3})
\end{eqnarray*}
are rational integers. Then equation satisfied by $\omega $ is:
\begin{equation*}
z^{3}+z^{2}\left( \frac{2bdck+3d^{2}b+bc^{2}k-3ac^{2}-3bc^{2}+kd^{2}a+2ckda-6cda}{-c^{2}kd+3c^{2}d-d^{2}ck-d^{3}+c^{3}}\right)
\end{equation*}
\begin{equation*}
-z\left( \frac{-3ca^{2}+cb^{2}k-6cba+2cbka+2bdka+a^{2}kd+3db^{2}-3a^{2}d}{-c^{2}kd+3c^{2}d-d^{2}ck-d^{3}+c^{3}}\right)
\end{equation*}
\begin{equation*}
+\left( \frac{-3ba^{2}+b^{3}+b^{2}ka+bka^{2}-a^{3}}{-c^{2}kd+3c^{2}d-d^{2}ck-d^{3}+c^{3}}\right) =0\text{,}
\end{equation*}
Equation satisfied by elements of $y$-class can be obtained by putting $a=c+d $, $b=-c$ in above equation
\begin{equation*}
z^{3}-z^{2}t+z(t-3)+1=0\text{,}
\end{equation*}
where $t$ is given by
\begin{equation*}
t=\frac{9d^{2}c+c^{3}k+9c^{2}d-3d^{2}ck-kd^{3}}{-c^{2}kd+3c^{2}d-d^{2}ck-d^{3}+c^{3}}\text{.}
\end{equation*}
The following three elements
\begin{equation*}
w=(\frac{(c+d)\alpha -c}{c\alpha +d}),w
{\acute{}}
=(\frac{d\alpha -(c+d)}{(c+d)\alpha -c}),w^{\prime \prime }=(\frac{-c\alpha
-d}{d\alpha -(c+d)})
\end{equation*}
are the roots of the Equation 3.2. Hence if $\alpha $ is a root of $x^{3}-kx^{2}+(k-3)x+1=0$ and $\alpha
{\acute{}}
$ is a root of the same equation, except $k$ is replaced by $k$, such that $Q(\alpha )=Q(\alpha ^{\prime })$ then the relationship between
$k$ and $k
{\acute{}}
$ is given by
\begin{equation*}
k
{\acute{}}
=\frac{9d^{2}c+c^{3}k+9c^{2}d-3d^{2}ck-kd^{3}}{-c^{2}kd+3c^{2}d-d^{2}ck-d^{3}+c^{3}}\text{.}
\end{equation*}
From now onwards, we shall write the above relation between $k$ and $k'$ as $k'=T(c,d,k)$.
Hence we can state:
\begin{theorem}
If $\alpha $ is a root of the equation $x^{3}-kx^{2}+(k-3)x+1=0$ and $\alpha' $
is a root of the same equation, except $k$ is replaced by $k'$, then $Q(\alpha )=Q(\alpha')$ if and only if \(k'=T(c,d,k)\)
for integers $c$ and $d$ not both zero.
\end{theorem}
We define a relation \(R\) on set of rational numbers as $aRb$ if there exist integers \(c\) and \(d\) such that $k'=T(c,d,k)$.
It can be easily seen that the relation \(R\) is equivalence relation. Hence,
\begin{theorem}
There is one-to-one correspondence between the equivalence classes of \(R\) and distinct simplest cubic
fields and vice versa, with the exception of one equivalence class $\{k:$ $k=\frac{(p^{3}-3pq^{2}+q^{3})}{(qp\left( -q+p\right) )}$, for some integers $p$ and $q$ such that $0\neq p\neq q\neq 0\}.$
\end{theorem}
\bigskip
|
\subsection*{Abstract}
Online social networking has quickly become one of the most common activities of Internet users.
As social networks evolve, they encourage users to share more information,
requiring the users, in turn, to place more trust into social networks.
Peer-to-peer (P2P) overlays provide an environment that can return ownership of
information, trust, and control to the users, away from centralized third-party
social networks.
In this paper, we present a novel concept, social profile overlays,
which enable users to share their profile only with trusted peers in a scalable,
reliable, and private manner. Each user's profile consists
of a unique private, secure overlay, where members of that overlay have a
friendship with the overlay owner. Profile data is made available without regard
to the online state of the profile owner through the use of the profile overlay's
distributed data store. Privacy and security are enforced through the use of a
public key infrastructure (PKI), where the role of certificate authority (CA) is
handled by the overlay owner and each member of the overlay has a CA-signed
certificate. All members of the social network join a common public or directory
overlay facilitating friend discovery and bootstrap connections into profile
overlays. We define interfaces and present tools that can be used to implement
this system, as well as explore some of the challenges related to it.
\section{Introduction}
Online social networking has become pervasive in daily life, though as social
networks grow so does the wealth of personal information that they store. Once
information has been released on a social network, known as a user's profile,
the data and the user are at the mercy of the terms dictated by the social network
infrastructure, which today is typically third-party, centrally owned. If the social
network engages in activities disagreeable to the user, due to change of terms or
opt-out programs not well understood by users such as recent issues
with Facebook's Beacon program~\cite{facebook_beacon}, the options presented to the user are limited:
to leave the social network (surrendering their identity and features provided
by the social network), to accept the disagreeable activities, or to petition
and hope that the social network changes its behavior.
As the use of social networking expands to become the primary way in which users
communicate and express their identity amongst their peers, the users become
more dependent on the policies of social network infrastructure owners. Recent
work~\cite{p2p_socialnetwork} explores the coupling between social networks and
P2P systems as a means to return ownership to the users, noting that a social
network made up of social links is inherently a P2P system with the aside that
they are currently developed on top of centralized systems. In this paper, we
extend this idea with focus on the topic of topology; that is, how to self-organize
social profiles that leverage the benefits offered by a structured P2P overlay abstraction.
Structured P2P overlays provide a scalable, resilient, and self-managing
platform for distributed applications. Structured overlays enable users to
easily create their own decentralized systems for the purpose of data sharing,
interactive activities, and other networking-enabled activities. In recent
work~\cite{icdcs10}, we implemented mechanisms that allow users to create
and manage their own private overlays using a common public overlay to assist
in discovery and NAT traversal. This prior work focuses on generic structured
P2P private overlays; in this paper, we
expand upon this approach with in-depth discussion on how to apply this technique to
enable social network overlay profiles.
Social networks consist of users, each has a profile, a set
of friends, and private messaging. The profile contains user's
personal information, status updates, and public conversations, similar to a
message board. Friends are individuals trusted sufficiently by a user to view
the user's profile. Private
messaging enables sending messages discretely between users without leaking the
message to other members. Using this model, we describe how a public
overlay can be used as a directory for finding friends and
joining existing profile overlays. Each user has their own profile overlay,
secured via a public key infrastructure (PKI) to which they are the certificate
authority (CA). The profile overlay stores profile data in its distributed
data store, supporting profile access in scalable mechanisms
regardless of the profile owner's online status. In this paper, we present
the architecture of these overlays, as presented in Figure~\ref{fig:system}, and
how they are used to find and befriend peers, and describe approaches to
handling profile data and establishing initial connections to profile overlays.
\begin{figure}[h]
\centering
\epsfig{file=figs/subrings.eps, width=3.12in}
\caption{An example social overlay network. Alice has a friendship with Bob and
Carol, hence both are members of her profile overlay. Bob has a
friendship with Alice and Dave but not Carol; hence Alice and Dave are members of
his profile overlay, while Carol is not. Each peer has many overlay memberships
but a single root represented by dashed lines in various shades of gray.
For clarity, overlay shortcut connections are not shown.}
\label{fig:system}
\end{figure}
The rest of this paper is organized as follows. Section~\ref{background}
provides background and related work. Section~\ref{social_overlays} describes
our multi-overlay approach, explaining how to map social networks onto structured
P2P overlays. In Section~\ref{outstanding}, we explore some of the remaining
challenges introduced by our approach. We conclude the paper in
Section~\ref{conclusion}.
\section{Background}
\label{background}
In this section, we review structured P2P overlays and distributed,
decentralized online social network approaches.
\subsection{Structured P2P Overlays}
Structured P2P systems provide distributed look-up services with guaranteed
search time with a lower bound of $O(\log N)$, in contrast to unstructured
systems, which rely on global knowledge/broadcasts, or stochastic techniques
such as random walks~\cite{unstructured_v_structured}. Some examples of
structured systems can be found in~\cite{pastry, chord, symphony, kademlia,
can, dynamo}. In general, structured systems are able to make these guarantees by
self-organizing a structured topology, such as a 2D ring or a hypercube.
In the overlay, each node is given a unique node ID drawn from a large address
space. Each node ID must be unique otherwise address collisions will occur,
which can prevent nodes from participating in the overlay. Furthermore, having
the node IDs well distributed assists in providing better scalability as many
shortcut selection algorithms depend on having node IDs uniformly distributed
across the entire address space. Two approaches to ensure this behavior are
to have each node use a cryptographically strong random number generator to
generate the node ID, or to use a trusted third party generate node IDs and
cryptographically sign them~\cite{secure_routing}.
Overlay shortcuts enable efficient routing in ring-structured P2P systems.
Different shortcut selection methods include: maintaining large tables without
using connections and only verifying usability when routing
messages~\cite{pastry, kademlia}, maintaining a connection with a peer every
set distance in the P2P address space~\cite{chord}, or using locations drawn
from a harmonic distribution in the node address space~\cite{symphony}.
Most structured P2P overlays support decentralized storage/look-up of information by
mapping keys to specific node IDs in an overlay. At a minimum, the data is stored
at the node ID either smaller or larger to the data's node ID and for fault
tolerance the data can be stored at other nodes. This sort of mapping
and data storage is called a distributed hash table (DHT). DHTs provide the
building blocks to form more complex distributed data stores as presented in
Past~\cite{past} and Kosha~\cite{kosha}.
In \cite{one_ring, can_multicast}, the authors discuss the concept of a
single overlay supporting services through the use of additional overlays,
which use the underlying overlay to assist in discovery. In \cite{icdcs10}, we
presented a reference implementation of a multiple-overlay system that supports
the use of a public overlay's DHT to store currently active peers in the private
overlays. The system allows users to create their own private overlays without
having to create their own bootstrap network. During evaluation, using both
simulated and real systems, the time for a single peer to join first the public
and thereafter private overlays was small and grew logarithmically with network
sizes. The real system was tested using a public overlay of 600 nodes on
PlanetLab and a random distribution of peers in the private overlay. With
point-to-point security links enabled in the private overlay, the time to
connect was less than 22 seconds for all cases. In simulations, overlays with as
many as 100,000 peers were evaluated. For the 100,000-peer overlay, regardless
of the private overlay size and with security enabled, peers were able to connect
to their private overlays in less than 48 seconds.
In relation to this paper, these results can be interpreted such that the latency
required for a single peer to from being completely disconnected from the social network
to being fully connected to the directory and all profile overlays.
In addition, our system provides both relay-based and hole-punching NAT
traversal~\cite{nsdi10} techniques and supports point-to-point
PKI based security~\cite{icdcs10}, forming a basis for the approach
presented in this paper.
\subsection{Peer-to-Peer Social Networks}
In~\cite{peerson}, a DHT provides the look-up service for storing meta data
pertaining to a peer's profile. Peers query the DHT for updated content from
their friends by hashing their unique identifiers (e.g. friends' email
addresses). The retrieved meta data contains information for obtaining the
profile data such as IP address and file version. Their work relies
on a PKI system that provides identification, encryption, and access control.
In contrast, our approach provides each user their own private overlay secured
by point-to-point encryption and authentication amongst all peers in the profile
overlay. The profile overlay provides a clean abstraction of access control,
whereby once admitted to a private overlay, users can access a distributed data
store which holds the contents of the owners profile.
\cite{vis-a-vis} takes a different approach by depending on virtual individual
servers (VIS) hosted on a cloud infrastructure such as Amazon EC2. Friends
contact each other's VIS directly for updates. A DHT is used as a directory for
groups and interest-based searches. Their approach assumes bidirectional
end-to-end connectivity between each VIS, where a profile is only available
during the up time of the VIS. Because of the demands on network connectivity
and up time, the approach assumes a cloud-hosted VIS and has difficulty being used on user-owned resources.
Our approach enables users to avoid the need for all-to-all connectivity and
constant up time through the use of NAT traversal support and the
ability to store the profile in the overlay's distributed data store.
The approach presented in~\cite{matryoshka} also uses a DHT for looking up a peer's
circle of friends. Once a node in the peer's
outermost circle is found, that node is used to route profile requests to the
innermost circle which contains replicas of a peer's profile. Trust is enabled
through the use of an identification service contacted through the DHT. The
circle of friends concept lacks the simplicity of the abstraction made in our
approach, which can easily be applied to existing structured overlays
unlike the concept of innermost and outermost circles. Our approach
also enables the profile owner to serve as a CA, ensuring that nodes can only
access a profile overlay after having obtained a signed certificate.
Unlike the above approaches, the P2P social network presented in~\cite{tribler-osn}
uses an unstructured overlay without a DHT where peers connect directly to
each other rather than through the overlay establishing unique identifiers to
deal with dynamic IPs. Peers cache each other's data to improve availability.
While helper nodes are used to assist with communication between peers behind
NATs. The approach lacks security and access control considerations and lacks the
guarantees and the simplicity of the abstraction offered by a structured overlay.
\section{Social Overlays}
\label{social_overlays}
In this section, we explain how to map online social networking to our
multi-overlay social network consisting of a public directory overlay with many
private profile overlays. The directory overlay supports friend discovery and
verification and stores a lists of peers currently active in each profile
overlay. Profile overlays support message boards, private messages, and media
sharing.
\subsection{Finding and Verifying Friends}
In a traditional social network, a directory provides the ability to search
for users using public information, such as the user's full name, user ID,
e-mail address, group affiliations, and friends. The search results return zero
or more matching directory entries. Based upon the results, the user,
\textit{A}, can potentially make a friendship request. The request receiver,
\textit{B}, can review the public information of A to making a decision. If
\textit{B} accepts the request, \textit{A} and \textit{B} are given access to
each other's profiles. Once profile access has been enabled, the users can
learn more information, and if it turns out to be a mistake, the peers can
unilaterally end the relationship.
To map this to our proposed social overlay, the directory entries would be
inserted into the DHT of a public overlay. As discussed in previous work, the
DHT keys for these entries should consist of a subset of the user's
public information in lower-case format and hashed to an overlay address. The
value stored at these keys is the user's certificate, which consists of its public
information and an overlay address where the user expects to receive
notifications. This overlay address can be used for asynchronous offline
messaging, whose function we will explain shortly.
Because the users need a way to verify each other that involves social credentials,
we propose the use of a new form of certificate.
The main portion of the certificate is similar to a self-signed
x509 certificate with public information such as user's name, e-mail
address, and group affiliations embedded into the certificate. At the tail of
the certificate is a friend list represented by many friend entries. To do this
we propose employing a technique similar to PGP: users can acquire from their
friends a signed message consisting of a hash of the peer's certificate, the
time stamp, and the friend's certificate hash signed by the friend. Since PGP
does not provide a strong method for revocation, the time stamp provides
additional information to help decide whether or not a friendship link is still
active without accessing the profile overlay of either peers. Peers should
request a new friend list entry within a certain period of time or it will
appear that the friendship is no longer valid.
While looking for an individual, a peer may discover that many individuals have
overlapping public information components, such as the user's name. Assuming
all entries are legitimate, the overlay must have some method of supporting
multiple, distinct values at the same key, leaving the peer or the peer's DHT
client to parse the responses and determining the best match by reviewing the
contents of each certificate. Alternatively, a technique like
Sword~\cite{sword}, which supports distributing the data across a set of nodes
and using a bounded broadcast to discover peers that match all information,
could be used for searching.
If a peer, \textit{A}, desires a friendship with another peer, \textit{B},
\textit{A} issues a friendship request, which will be stored in the DHT
at the overlay address listed in \textit{B}'s certificate, as described earlier.
The friendship request consists of the self-signed certificate of
\textit{A}, the requesting peer; the public information of the request receiver,
\textit{B}; and a time stamp; all signed with the private key associated with
\textit{A}'s private key matched to their self-signed certificate.
Within a reasonable amount of time after a request has been inserted into the
DHT, \textit{B} can come online and check for outstanding requests. Upon
receiving a request, \textit{B} has three choices: a conditional accept, an
unconditional accept, or a reject. During an unconditional accept, \textit{B}
signs \textit{A}'s request and issues a request to befriend \textit{A}.
Alternatively in the case of a conditional accept, \textit{B} issues a friendship
request, waits for a reply, and investigates the profile prior to signing the
\textit{A}'s request. Once a user has received a signed certificate,
they may access the remote peer's profile overlay as discussed
in~\ref{profile_overlay}, which is also responsible for activities such as
revocation.
\subsection{The Profile Overlay}
\label{profile_overlay}
In a traditional social network, the profile or user-centric portion consists
of private messaging, data sharing, friendship maintenance, and a
public message board for status updates or public messages. In this
section, we explain how these components can be applied to a structured overlay
dedicated to an individual profile.
Using the techniques such as those described in~\cite{icdcs10}, it is feasible
to efficiently multiplex a P2P system across multiple, virtual private overlays enabling
each profile owner to have a profile overlay consisting of their online friends.
For access control, we employ a PKI, where each member uses the signed certificate
generated during the ``finding and verifying friends'' stage. All links are
encrypted using symmetric security algorithms established through the PKI,
thus preventing uninvited guests from gaining direct access to the overlay and
hence the profile. Because the profile owner also is the CA for all members of
the overlay, they can easily revoke users from access to the profile overlay.
In ~\cite{icdcs10}, we described mechanisms for overlay revocation through the
use of broadcasting for immediate revocation and the use of DHT for indirect
and permanent revocation.
The message board of a profile can be stored in two ways: distributed within the
profile overlay via a data store or stored on the profile owner's personal
computing devices. The distributed data store provide the profile when the
owner is offline and also distributes the load for popular profiles. For
higher availability, each peer should always be a provider for all data in their
profile when they are online. To ensure authenticity and integrity, all peers
should sign their messages and each peer's certificate should be available in
the overlay for verification. Messages that are unsigned should be ignored
by all members of the overlay. An ideal overlay for this purpose should
support complex queries~\cite{complex_queries} allowing easy access to data
stored chronologically, by content, by type, i.e., media, status updates,
or message board discussions.
Private messaging in the profile overlay is unidirectional meaning that only
the profile owner can receive private messages using their overlay. To
enforce this, a private message should be prepended with a symmetric key
encrypted by the profile owners public key, the message should be appended
by a hash of the message to ensure integrity and the entire message encrypted
by the symmetric key. This approach ensures that only the sender and the
profile owner can decrypt the private message. The contents of the private
message should include the sender, time sent, and the subject. Messages can
be stored in well known locations, so that the profile owner can either poll
the location or, alternatively, use an event based system to notify them of
the new message.
\subsection{Active Peers}
The directory overlay should be used to assist in finding currently active peers
in the profile overlays. By placing their node IDs at a well-known, unique
per-profile overlay keys in the DHT, active peers can bootstrap incoming peers
into the profile overlay. We implemented and evaluated this concept
in~\cite{icdcs10}. Because the profile overlay members all use PKI to ensure
membership, even if malicious peers insert their ID into the active list, it
would be useless as the peer would only form connections with peers who also
have a signed certificate.
\section{Challenges}
\label{outstanding}
While structured P2P overlays have been well-studied in a variety of applications,
their use in social profile overlays raises new interesting questions, including:
{\bf 1) Handling small overlay networks} - P2P overlay research typically focuses on
networks larger than the typical user's friend count (Facebook's average is
130\footnote{http://www.facebook.com/press/info.php?statistics}). Because social profile overlays are comparatively smaller, this can
impact the reliability of the overlay and availability of profile data. A user
can host their own profile; however when the user is disconnected it is important
that their profile remains available even under churn. It is thus important to
characterize churn in this application to understand how to best approach this
problem. An optional of per-user deployment of a virtual individual server (VIS)
and the use of replication schemes aware of a user's resources provide possible
directions to address this issue.
{\bf 2) Overlay support for low bandwidth, unconnected devices} - devices such as
smart phones cannot constantly be actively connected to the overlay and the
connection time necessary to retrieve something like a phone number may be
too much to make this approach useful. Similar to the previous challenge,
this approach could benefit from using a VIS enabling users access to their
social overlays by proxy without establishing a direct connection to the overlay
network.
{\bf 3) Reliability of the directory and profile overlay} - Overlays are
susceptible to attacks that can nullify their usefulness. While
the profile overlay does have point-to-point security, in the public,
directory overlay, the lack of any form centralization makes policing the system
a complicated procedure. While our approach of appending friends list can assist
users in making decisions on identity, it does not protect against denial of
service attacks. For example, users could attempt create many similar identities
in an attempt to overwhelm a user in their attempt to find a specific peer.
Previous work has proposed methods to ensure the usability of overlays even
while under attack. For the social overlay to be successful, we must identify
which methods should be used. A possible approach is to replicate public
information within a user's profile overlay thus providing an alternative
directory overlay for querying prior to using the public directory overlay.
\section{Conclusion}
\label{conclusion}
In this paper, we proposed methods by which a social network can be
decentralized through the use of structure P2P overlays. Our approach uses
multiple overlays where all users join a public directory overlay and a subset
of the individual profile overlays. The directory overlay enables
users to find and befriend other peers and bootstrap connections
into the secure profile overlays. Upon forming a friendship through the
directory overlay, peers are given CA signed certificates that allow them to
join each other's profile overlay. The owner of the profile overlay acts as
CA enabling unilateral dismissal of friendships via certificate revocation
using efficient and reliable methods.
For the purpose of storing profile information into the overlay, we cite
previous work that can be used to provide distributed data services and give
examples of how to store data securely in the overlay. Our proposed system
returns control of the social network and more importantly users' identity to
the users and eliminates the need for centralized social networks.
\bibliographystyle{abbrvnat}
\footnotesize{
|
\section{Introduction}
Investigation of the magnetic properties of the newly discovered layered
Fe-based pnictide supercondutors \cite{kami} with a microscopic probe such
as M\"ossbauer spectroscopy is of interest for (at least) two reasons:
i) to determine the characteristics of the expected spin-density wave (SDW)
order and ii) to give experimental evidence concerning the problem of the
interaction and/or local coexistence between superconductivity and magnetic
ordering of the Fe moments. In the so-called 122 family with Ba, the parent
(non superconducting) compound is BaFe$_2$As$_2$ and doping with Co
substituted
for Fe has been shown to induce superconductivity at low doping levels
\cite{sefat}. In the Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ series, the phase diagram
as a function of Co concentration $x$ is ``bell-shaped'', with $T_c(x)$
presenting a maximum of 26\,K for $x$=0.07 \cite{chu,flo}. Below this optimal
concentration, the material shows
apparently the persistence of an antiferromagnetic (AF) structure in the
superconducting phase, the AF or spin density wave (SDW) transition
temperature decreasing rapidly as $x$ increases. Neutron diffraction in the
undoped compound BaFe$_2$As$_2$ \cite{huang} has determined the AF magnetic
structure below $T_N \simeq$143\,K to be {\it a priori} commensurate.
$^{57}$Fe M\"ossbauer spectroscopy \cite{rotter0,rotter} and $^{75}$As NMR
\cite{kita} in BaFe$_2$As$_2$ have confirmed the commensurability of the
magnetic structure.
\begin{figure}
\includegraphics[width=11cm]{rho.eps}
\caption{[color online] Resistivity {\it vs.} temperature for BaFe$_2$As$_2$
with Co dopings $x$=0.014, 0.03 and
0.045 from Ref.\cite{flo}. The data for $x$=0.014 have been shifted by
100\,$\mu\Omega$.cm for clarity. The full arrows mark the inflexion point of
$\rho(T)$, attributed to the onset of magnetic ordering ($T_{\rm SDW}$), and
the dash-dotted arrows the superconducting transition temperature.}
\label{rhoT}
\end{figure}
In this work, we focus on the first problem evoked above: we performed a
$^{57}$Fe M\"ossbauer study of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ in the
underdoped region in order to investigate the SDW state. In our samples with
Co doping $x$=0.014, 0.03 and 0.045, the resistivity data show an inflexion
point \cite{flo}
respectively around 115\,K, 92\,K and 64\,K (see Fig.\ref{rhoT}). According to
neutron diffraction data in materials with similar
Co dopings \cite{pratt,christianson,lester}, this temperature corresponds
to the SDW-paramagnetic transition. We find that the spectra in the magnetic
phase have an unusual shape, which we interpret as due to an incommensurate
modulation of the AF structure, i.e. to an incommensurate (IC) spin density
wave.
We also find that the magnetic hyperfine subspectrum persists far above
$T_{\rm SDW}$, i.e. that AF correlations are present up to $T_{\rm AF} \sim
1.5 \ T_{\rm SDW}$.
\section{Experimental and sample characterisation}
The samples are single crystals prepared by the self-flux method, as described
in Ref.\cite{flo}. It was checked that Co doping is uniform and that
superconductivity occurs in the bulk \cite{lapl,lapl2}. For the M\"ossbauer
measurements, the crystals were
finely ground. Powder samples made by solid state reaction were also prepared,
with $x$=0 and 0.03, for comparison purposes. The spectra were recorded using
a commercial
$^{57}$Co$^*$:Rh $\gamma$-ray source mounted on an electromagnetic drive with
linear velocity signal.
Room temperature spectra ensured that no trace of FeAs (with a quadrupole
splitting $\Delta E_Q$=0.69\,mm/s) was present.
In the high temperature tetragonal phase (space group $I4/mmm$), the $4d$ Fe
site has fourfold symmetry ($\bar 4$m2). At low temperature, slightly above
$T_{\rm SDW}$, a
small orthorhombic distortion takes place (space group $Fmmm$), and the
local symmetry at the $8f$ Fe site is lowered (222).
\section{The $^{57}$Fe M\"ossbauer spectra at 4.2\,K}
The spectra at 4.2\,K in the three investigated compounds and in undoped
BaFe$_2$As$_2$, are represented in
Fig.~\ref{spe4k2Co}. In BaFe$_2$As$_2$, one observes a single magnetic
hyperfine pattern, in agreement with Ref.\cite{rotter}, with a hyperfine field
$H_{\rm hf} = 5.4(1)$\,T. By contrast, the spectral shapes in the Co-doped
compounds are rather far from this standard 6-line pattern.
Saturation effects can be discarded since the resonant
absorption is modest; for $x$=0.03, where it amounts to 7\%, we checked
that an absorber with 2\% absorption yields exactly the same spectrum.
Texture effects are also negligible since a powder absorber (for $x$=0.03)
yields spectra quasi-identical with those of the single crystal sample. The
high statistics reached in these experiments allows us to distinguish fine
details
of the spectral shape, which is a prerequisite for the fits to be described
below.
\begin{table}
\caption{In Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$: Fourier coefficients $h_{2k+1}$
and maximum value $h_m$ (in T) of the hyperfine field modulation at 4.2\,K .}
\label{coeff}
\begin{center}
\begin{tabular}{||c|c|c|c|c|c|c|c||} \hline
$x$ & $h_1$ & $h_3$ & $h_5$ & $h_7$ & $h_9$ & $h_{11}$ & $h_m$ \\ \hline
0.014 & 5.45 & 2.41 & 1.75 & 0.24 & 1.24 & 0.55 & 6.06\\ \hline
0.03 & 5.24 & 0.44 & 0.45 & 0.21 & 0.25 & -- & 5.29 \\ \hline
0.045 & 4.06 & -0.26 & 0.23 & -- & -- & -- & 4.55 \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\includegraphics[width=6cm]{spe4k2n_Co.eps}
\caption{M\"ossbauer absorption spectra on $^{57}$Fe at 4.2\,K
in Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ for $x$=0, 0.014, 0.03 and 0.045. Except for
$x$=0, the solid
lines, which are masked by the data due to the goodness of the fits,
are simulations with an incommensurate modulation of hyperfine fields (see
text).}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{modmom_4k2.eps}
\label{spe4k2Co}
\caption{[color online] Modulation of the hyperfine field along the direction
of {\bf q} at 4.2\,K in
Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ for $x$=0.014, 0.03 and
0.045, derived from the M\"ossbauer spectra. Insert: variation with the Co
concentration $x$ of the maximum hyperfine field value (circles); for $x$=0,
hyperfine field in BaFe$_2$As$_2$ (square). The dashed line is a guide for
the eye.}
\label{mod4k2}
\end{figure}
We interpret these unusual spectral shapes as due to a distribution of
hyperfine fields arising from an IC modulated magnetic structure.
Indeed, in this case, the $^{57}$Fe nuclei are submitted to hyperfine field
values ranging from zero to a maximum value, and although the M\"ossbauer
data cannot give access to the propagation vector {\bf q}, the shape of the
distribution reflects the particular shape of the modulation. In order to
describe the influence of the moment modulation on the spectral shape, we
assume a collinear magnetic structure, with no ferromagnetic component,
like in undoped BaFe$_2$As$_2$ where the Fe moments lie along the orthorhombic
{\bf a} axis \cite{huang}. We also assume the hyperfine field is collinear
with the Fe moment (see discussion below).
Then, along the direction OX parallel to the propagation vector {\bf q}, the
modulation of the hyperfine field is defined in terms of a Fourier series as:
\begin{equation}
H_{hf}(qX) = \sum_{k=0,n} h_{2k+1}\ \sin[(2k+1)qX],
\label{eqmod}
\end{equation}
where the $h_i$ are the odd Fourier coefficients of the modulation. The
spectra are then fitted to a superposition of individual hyperfine
6-line patterns with discrete $H_{hf}$ values according to Eq.(\ref{eqmod})
with a linear mesh along a period (0$\le qX\le 2\pi$). The number of Fourier
components in the fit is increased until a good reproduction of the lineshape
is obtained. At 4.2\,K, the very good fits shown in Fig.\ref{spe4k2Co} were
performed with 6, 5 and 3 Fourier components for $x$=0.014, 0.03 and 0.045
respectively. The resulting
hyperfine field modulations are shown in Fig.\ref{mod4k2}, and the values
of the Fourier coefficents are given in Table \ref{coeff}.
The isomer shift at 4.2\,K with respect to $\alpha$-Fe is 0.54(1)\,mm/s and
a very small quadrupolar interaction with $\Delta E_Q \simeq -$0.04(2)\,mm/s
is needed to reproduce the slight asymmetry of the spectra.
The 4.2\,K modulation is seen to depart more and more from ``squaredness''
(which occurs for $x$=0) as $x$ increases; for $x$=0.045, the modulation is
close to a pure
sine-wave. This is reflected in the number of Fourier coefficients needed
to reproduce the lineshape (see Table \ref{coeff}), which decreases as $x$
increases. The maximum value of the hyperfine field at 4.2\,K decreases
as $x$ increases (see insert in Fig.\ref{mod4k2}). For $x$=0.045, it is
75\% of the $x$=0.014 value.
\section{Thermal variation of the M\"ossbauer spectra}
The thermal variation of the spectra for $x$=0.014, for which $T_{\rm SDW}
\simeq$115\,K (Fig.\ref{rhoT}), is shown in Fig.\ref{spe1.4CoT}. No important
change
occurs up to about 80K, then the modulation changes shape, tending towards
a pure sine-wave as temperature is further increased. A striking
feature is observed when comparing Fig.\ref{spe4k2Co} and Fig.\ref{spe1.4CoT}:
increasing the Co doping has the same effect on the modulation at 4.2\,K as
increasing the temperature for the $x$=0.014 sample.
The maximum value of the hyperfine field shows little thermal
variation, as well as the main Fourier coefficient $h_1$. Above
115\,K, a single
line, characteristic of the paramagnetic phase, is present in the spectra and
it grows on heating at the expense of the magnetic subspectrum (see red line
in the 130\,K spectrum in Fig.\ref{spe1.4CoT}). The latter actually
persists in a rather large temperature range above $T_{\rm SDW}$, up to
150\,K at least.
\begin{figure}
\includegraphics[width=6cm]{spe_1.4CoT.eps}
\caption{[color on line] M\"ossbauer absorption spectra on $^{57}$Fe at
selected temperatures
in BaFe$_2$As$_2$ with $x$=0.014 Co doping. The lines are fits to an
incommensurate modulation of hyperfine fields; the red
subspectrum at 130\,K is a single Lorentzian-shaped line.}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{pourc_1.4Co.eps}
\label{spe1.4CoT}
\caption{[color online] In BaFe$_2$As$_2$ with $x$=0.014 Co doping, thermal
variation of the
fraction of the single line component in the M\"ossbauer spectra, for
increasing and decreasing temperature runs (left scale), and of the
resistivity (right scale).}
\label{fracT}
\end{figure}
In this temperature range, very good fits of the spectra were obtained with
two components: a single Lorentzian-shaped line and a magnetic subspectrum
with a modulation identical to that of the 115\,K spectrum. For this latter
component, a common scaling factor of the Fourier coefficients was
introduced in order to allow for a decrease of the modulation amplitude on
heating; it remains above 0.9 up to the
highest temperature. The thermal variation of the fraction of the single line
derived from these fits is shown in Fig.\ref{fracT}: a small hysteresis is
observed and a unique line is recovered only above 150-160\,K.
For the sample with $x$=0.03, the thermal variation of the 5 Fourier
components was studied in more detail (see Fig.\ref{pourc3} top) and a similar
trend is observed. The main component $h_1$ is an order
of magnitude larger than the higher harmonics, and its variation with
temperature is weak: it falls by about 20\% between 4.2 and 90\,K, which is
close to $T_{\rm SDW}$ (see Fig.\ref{rhoT}).
\begin{figure}
\includegraphics[width=7cm]{coefpourc3co.eps}
\caption{[color online] In BaFe$_2$As$_2$ with $x$=0.03 Co doping: {\bf Top}
thermal variation of the main Fourier component $h_1$ of the hyperfine field
modulation (the line is a guide for the eye) and of the 4
next components $h_3 - h_9$ (insert) below $T_{\rm SDW}$ = 92\,K;
{\bf Bottom} thermal
variation (on heating) of the fraction of the single line (left scale) and
of the resistivity (right scale).}
\label{pourc3}
\end{figure}
The higher harmonics are constant up to 60\,K, then they drop and $h_3$
becomes negative and almost opposite to $h_5$, the other two harmonics being
smaller. In terms of the shape of the modulation, it does not change up to
about 60\,K, and on further heating it evolves smoothly towards an almost
pure sine-wave at 85\,K. The spectrum at 85\,K for $x$=0.03 is very
similar to the 4.2\,K spectrum for $x$=0.045, illustrating the above mentioned
trend that increasing $x$ at 4.2\,K has the same effect on the shape of the
modulation as increasing $T$
for a given $x$. Above 90\,K, i.e. above $T_{\rm SDW}$, a coexistence
region is present, like in the sample with $x$=0.014. The thermal variations
of the fraction of the single line and of the resistivity are represented in
Fig.\ref{pourc3} bottom. Above 115\,K, the two components of the M\"ossbauer
spectra cannot be distinguished, but the coexistence region probably extends
up to 130-140\,K.
For $x$=0.045, the modulation remains unchanged up to about 60\,K, close to
$T_{\rm SDW}$, then a single line grows in the spectrum on further heating.
However, due to the lack of resolution of the spectra, no quantitative
assessment of the fraction of single line can be done.
For this doping level,
the spectrum at 4.2\,K belongs to the superconducting phase, but we postpone
a discussion of the mixed phase (superconducting/SDW)
to a future publication.
According to the neutron diffraction data \cite{pratt,christianson,lester},
the transition to a long range ordered SDW state occurs at $T_{\rm SDW}$.
Actually, this transition is not ``seen'' in the spectra of
our underdoped samples; rather, {\it above $T_{\rm SDW}$}, a paramagnetic
component (single line) appears and coexists with the magnetic hyperfine
component. The latter progressively dwindles as temperature increases,
its (maximum) hyperfine field decreases slowly on heating, and so does the Fe
moment (see discussion below). We also observe a small hysteresis (for
$x$=0.014). This points to
the persistence of AF correlations, probably short range and dynamic, far
above $T_{\rm SDW}$. We estimate that a ``fully paramagnetic''
phase is recovered around $T_{\rm AF} \sim 1.5\ T_{\rm SDW}$.
This behaviour could be linked with the strong bidimensional character of the
magnetic interactions in the Fe layers \cite{wang,zhang}.
As to the structural transition which, in the Co doped compounds, occurs at
a temperature higher than $T_{\rm SDW}$ \cite{pratt,lester}, nothing can be
said from the M\"ossbauer data due to the very
weak quadrupolar interaction in these compounds.
\section{Discussion}
\subsection{Disorder induced by Co substitution}
Since these materials are (weakly) substituted compounds, one may wonder
whether the observed lineshapes, where a distribution of hyperfine
fields is clearly present, could be accounted for by some disorder induced
spread in the hyperfine field values.
\begin{figure}
\includegraphics[width=\linewidth]{disthf_3Co.eps}
\caption{[color online] Hyperfine field distribution associated with the
hyperfine field
modulation at 4.2\,K for the $x$=0.03 sample (see Table \ref{coeff}). The
positions of the peaks near 2.8 and 4.3\,T correspond to the quasi-plateaus
of the modulation, and the (unphysical) divergence to the maximum hyperfine
field where $dH_{hf}/d(qX)$=0 (see Fig.\ref{mod4k2}). Insert: integrated
weight.}
\label{disthf3}
\end{figure}
In the layered structure of the pnictides, a Fe atom has 4 in-plane nearest
neighbors, and for Co dopings $x$=0.03 and 0.045, it results that resp. 12\%
and 18\% Fe atoms have one Co atom as nearest neighbour. Assuming that the
hyperfine field is different for a Fe atom with 4 Fe and with 1 Co - 3 Fe as
nearest neighbors, then the corresponding M\"ossbauer spectrum would show a
two peak hyperfine field distribution with, for instance for $x$=0.03,
relative weights of 88 and 12\%. The distribution associated with the 4.2\,K
modulation for $x$=0.03 is shown in Fig.\ref{disthf3}: it presents
3 peaks, each with 25-30\% relative intensity, which therefore cannot
be due to the above hypothesis of different environments of a Fe atom.
The hyperfine fields of Fe atoms with 4 Fe or 1 Co - 3 Fe are probably very
close due to the delocalised nature of the 3$d$ electrons.
\subsection{Other evidences for incommensurability}
An incommensurability of the AF magnetic order in Co-underdoped BaFe$_2$As$_2$
has been inferred by NMR measurements \cite{ning,lapl} from the anisotropy of
the line broadening. The IC wave-vector was estimated in Ref.\cite{lapl} to
depart only slightly (by a few
percent) from the commensurate value {\bf q}=(101) in the orthorhombic cell.
$\mu$SR data in a sample with $x$=0.04 \cite{bernh} can also be
interpreted in terms of an IC magnetic structure.
The neutron diffraction measurements \cite{pratt,christianson,lester} did not
report any incommensurability, but the resolution in q-space is
probably not sufficient for this purpose; we note that the (a,b) orthorhombic
splitting could not be resolved in these experiments as well. In LaOFeAs, a
calculation of the Lindhardt response
function $\chi_0(q)$ was performed for the undoped system and for the system
with $x$=0.1 F doping \cite{dong}. It is shown that F doping shifts the
maximum of $\chi_0(q)$ from the M point of the Brillouin zone to an
IC {\bf q} vector, inducing thus an IC magnetic structure.
The appearance of an IC SDW upon
doping is also predicted in Ref.\cite{cvetko}
In view of the similarities between the 1111
and 122 families of pnictides, these results could also hold for the latter
family,
accounting for the fact that a very small doping level ($x$=0.014) is enough
to push the material towards an IC SDW.
\subsection{The hyperfine constant and the Fe moment}
The saturated hyperfine field in BaFe$_2$As$_2$ is 5.4\,T
\cite{rotter} and the maximum hyperfine field at 4.2\,K in the Co doped
compounds has comparable magnitude (see insert of Fig.\ref{mod4k2}). These are
unusually small values for Fe, even in an itinerant magnet.
\begin{table}
\caption{Saturated Fe moment $m$ and hyperfine field $H_{hf}$, and the
ratio $r=H_{\rm hf}/m$ in AFe$_2$As$_2$ for A= Ba, Ca and Sr.}
\label{tchf}
\begin{center}
\begin{tabular}{||c|c|c|c||} \hline
A & Ba & Ca & Sr \\ \hline
$m$($\mu_B$) & 0.87\cite{huang} & 0.80\cite{gold} & 1.0\cite{kan} \\ \hline
$H_{\rm hf}$(T)& 5.4\cite{rotter0} &10\cite{kumar} & 8.9\cite{tegel} \\ \hline
$r$(T/$\mu_B$) & 6.3 & 12.5 & 8.9 \\ \hline
\end{tabular}
\end{center}
\end{table}
In intermetallic
compounds, it is usually assumed that the Fermi contact interaction with
$s$-electrons is the dominant hyperfine coupling, implying that the Fe moment
{\bf m} is proportional to the hyperfine field {\bf H$_{hf}$}
\cite{freewat}. This is true in $\alpha$-Fe, with a hyperfine constant
$C_{\rm hf}$=$H_{\rm hf}/m \simeq$ 15\,T/$\mu_B$, and in a number of other
intermetallics with similar $C_{\rm hf}$ value. In the 122
family of isostructural pnictides, the ratios of hyperfine field and moment
are smaller and present rather scattered values (see Table
\ref{tchf}, where the moment is taken from neutron diffraction data).
A strict proportionality between moment and hyperfine field can be however
questioned in the pnictides. Indeed, the magnetic susceptibility has been
found to be anisotropic ($\chi_\perp/\chi_{//} \simeq$ 1.5-2) both
in undoped BaFe$_2$As$_2$ \cite{wang} and in Co-doped compounds
\cite{fra}. This is a hint to the
presence of an ``unquenched'' orbital moment at the Fe site. This implies a
probably small, but non-zero contribution of the $d$-orbitals to
the hyperfine field, with a sign opposite to that of the Fermi contact field.
Then, like in Fe$^{2+}$ compounds \cite{imbert}, the hyperfine field is no
longer strictly proportional to the moment. The hyperfine field modulation
then would not reflect exactly that of the Fe moment.
However, assuming as a first good approximation that proportionality holds in
the Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$
series with the same constant as in BaFe$_2$As$_2$, then the maximum Fe moment
of the modulation at 4.2\,K would be 0.96, 0.84 and 0.72\,$\mu_B$ for
$x$=0.014, 0.03 and $x$=0.045 respectively.
\subsection{Thermal variation of the moment and nature of the SDW transition}
In the neutron diffraction studies of the series
Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ \cite{huang,pratt,christianson,lester},
the integrated intensity of the magnetic Bragg peaks continuously drops
towards zero as temperature is increased towards $T_{\rm SDW}$.
This behaviour is in contrast with the $^{75}$As NMR work in BaFe$_2$As$_2$
\cite{kita}, where the transferred hyperfine field at the As nucleus shows a
weak thermal variation and it was concluded that the SDW transition is first
order. A similar discrepancy is observed in the present work in Co-doped
materials, where the main Fourier component $h_1$ of the modulation retains a
sizeable value at $T_{\rm SDW}$ (see Fig.\ref{pourc3} Top). However, a correct
comparison with the neutron data must take into
account the fact that the intensity of a magnetic Bragg peak is proportional
to the square of the magnetic moment. In case of a modulated structure with
Fourier coefficients $m_{2k+1}$, it should scale with $\langle m^2 \rangle$,
where the brackets denote an average over a period of the moment modulation:
\begin{equation}
\langle m^2(x,T) \rangle = \frac{1}{2} \sum_{k=0,n} m_{2k+1}^2(x,T) \propto
\frac{1}{2} \sum_{k=0,n} h_{2k+1}^2(x,T).
\label{mneutr}
\end{equation}
Figure \ref{intnorm} shows the thermal variations of
$\langle m^2(x,T) \rangle$ (from the $h_{2k+1}(x,T)$ values
determined by M\"ossbauer spectroscopy) for $x$=0.03 and of the intensity of
a magnetic Bragg peak for $x$=0.04 \cite{christianson}. The $\langle m^2
\rangle$ value is reduced by half at $T_{\rm SDW}$, in disagreement with the
neutron data. Similar behaviours occur for $x$=0.014 and 0.045. It can thus be
concluded that the SDW transition in the Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$
series, as observed with the M\"ossbauer local probe, also presents a strong
first order character.
Comparing the moment values obtained from
the neutron data \cite{lester} for $x$=0.025 and 0.045 with the M\"ossbauer
results is more difficult since, in the scaling of the neutron Bragg
intensity, it was assumed that the magnetic structure is identical to that
of BaFe$_2$As$_2$, i.e. commensurate with the lattice spacings.
The mean saturated Fe moment value $m_s$ obtained from neutron diffraction
should be compared with:
\begin{equation}
m_s(x) = \sqrt{\langle m^2(x,T=4.2\,K) \rangle}.
\label{ms}
\end{equation}
Assuming proportionality between hyperfine field and moment, and using the
M\"ossbauer derived values for $h_{2k+1}(x,T=4.2\,K)$, one gets: $m_s$=0.70,
0.59 and 0.45\,$\mu_B$ resp. for $x$=0.014, 0.03 and 0.045. This must be
compared with $m_s$=0.35 and 0.17\,$\mu_B$ resp. for $x$=0.025 and 0.045
derived from the neutron data in Ref.\cite{lester}.
Further work is needed to resolve the disagreement between
local probe and neutron diffraction data about the Fe moment variation
with temperature and Co doping.
\begin{figure}
\includegraphics[width=\linewidth]{int_norm.eps}
\caption{[color online] Variation of the normalised values of the Bragg peak
intensity for $x$=0.04 from Ref.\cite{christianson} (diamonds) and of the
M\"ossbauer derived $\langle m^2(x,T) \rangle$ for $x$=0.03 from
Eqn.\ref{mneutr} (squares), as a function of the reduced temperature
$T/T_{\rm SDW}$. The M\"ossbauer data above $T_{\rm SDW}$ belong to the
short range AF regions.}
\label{intnorm}
\end{figure}
\section{Conclusions}
$^{57}$Fe M\"ossbauer absorption spectra have been measured in the
Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ series for $x$=0.014, 0.03 and 0.045. In all
the samples, unusual magnetic hyperfine lineshapes are observed from 4.2\,K
to $T_{\rm SDW}$. They are interpreted in terms of a continuous
modulation of hyperfine fields at the Fe sites arising from an incommensurate
electronic SDW state. Incommensurability is present at the
lowest doping level, i.e. for $x$=0.014, while the magnetic structure in
BaFe$_2$As$_2$ is commensurate. Above $T_{\rm SDW}$, we do not observe a
purely paramagnetic signal, but a coexistence of magnetic
hyperfine and paramagnetic subspectra in a large temperature range, up
to $T_{\rm AF} \sim 1.5 \ T_{\rm SDW}$ approximately for all samples. We
believe this is due to the persistence of AF short range correlated regions,
probably dynamic, well above the SDW transition, which is found to have a
strong first order character. Our
analysis of the hyperfine field modulations as a function of $x$ shows that
magnetism is not strongly depleted as doping increases, i.e. the mean Fe
moment or the maximum of the modulation does not decrease with $x$ as rapidly
as inferred from neutron diffraction data. We believe the
incommensurability of the magnetic structure
has not been evidenced by neutron scattering up to now due to the probable
very small value of the incommensurate component(s) of the SDW wave vector.
This makes M\"ossbauer spectroscopy a valuable tool to investigate these
structures: although the propagation vector cannot be obtained with this
technique, the spectral shape is sensitive to minute incommensurabilities.
\acknowledgments
We thank J. Bobroff and H. Alloul for fruitful discussions.
|
\section{Introduction}\label{sec:intro}
For a knot $K \subset S^3$, let $N_K$ denote a tubular neighborhood of
$K$ and let $M_K:=\overline{ S^3\setminus N_K}$ denote the exterior of
$K$. Let $\langle \mu, \lambda \rangle$ be the canonical
meridian--longitude basis of $H_1 ({\partial} N_K)$. An element $p/q \in
{{\mathbb{Q}}}\cup \{ 1/0\}$ is called a \emph{boundary slope} of $K$ if there
is a properly embedded essential surface $(S, {\partial} S) \subset (M_K,
{\partial} N_K)$, such that every circle of ${\partial} S$ is homologous to $p \mu + q \lambda \in
H_1 ({\partial} N_K)$. Hatcher has shown that every knot $K \subset S^3$
has finitely many boundary slopes \cite{hatcher:boundary-slopes}.
The \emph{colored Jones function} of $K$ is a sequence of Laurent
polynomial invariants $J_K(n, q) \in {{\mathbb{Z}}}[q, \ q^{-1}]$, for $n=1,2,
\ldots$. For $n=2$, $J_K(2,q)$ is the ordinary Jones polynomial. We
will use the normalization that $J_{\rm unknot}(n, q)=1$, for every $n
\geq 1$. Let $j(n)$ denote the highest degree of $J_K(n, q)$ in $q$,
and let $j^{*}(n)$ denote the lowest degree. Consider the sequences
$$js_K:= \left\{{{ 4j(n)}\over {n^2} } \: : \: n > 0\right\} \quad
\mbox{and} \quad js^*_K:= \left\{ {{ 4j^{*}(n)}\over {n^2}} \: : \: n
> 0 \right\}.$$
Garoufalidis conjectured \cite{garoufalidis:jones-slopes} that for each knot $K$, every cluster point
(i.e., every limit of a subsequence) of $js_K$ or $js^*_K$ is a
boundary slope of $K$. Thus, if the conjecture holds, the colored Jones polynomials detect certain boundary slopes of $K$.
He verified
the conjecture for alternating knots, torus knots, pretzel knots of
type $(-2, 3, p)$, and several low crossing knots.
In this paper, we prove Garoufalidis' conjecture for the class of
adequate knots. The precise definition of \emph{adequate} appears in
Section \ref{sec:state-graphs}. For the moment, we note that the
family of adequate knots includes all alternating knots, most
Montesinos knots, and all knots that are Conway sums of two
\emph{strongly alternating} tangles. See \cite{lick-thistle:adequate-knots} and Section
\ref{sec:examples} for more examples.
\begin{theorem}\label{thm:main}
Let $D(K)$ be a knot diagram. Then
\begin{itemize}
\item[(a)] If $D$ is $A$--adequate, then $\displaystyle{ \lim_{n \to
\infty} 4\, n^{-2}j^*(n)}$ exists, and is a boundary slope of $K$.
\item[(b)] If $D$ is $B$--adequate, then $\displaystyle{\lim_{n \to
\infty} 4\, n^{-2}j(n)}$ exists, and is a boundary slope of $K$.
\end{itemize}
In particular, if $K$ is a non-trivial adequate knot, then the set
$js_K \cup js^*_K$ has exactly two cluster points, both of which are integer
boundary slopes of $K$.
\end{theorem}
The proof of Theorem \ref{thm:main} involves four steps:
\begin{enumerate}
\item Starting with an $A$--adequate diagram $D(K)$, construct a
\emph{state surface} $S_A$ whose boundary is $K$. This is a standard
construction, generalizing the construction of a checkerboard surface.
\item Verify that $S_A$ is an essential surface. This result, stated as
Theorem \ref{thm:essential} below, was first proved by Ozawa
\cite{ozawa:adequate}; an alternate proof is given by the authors in \cite{fkp:guts}.
\item Relate the boundary slope of $S_A$ to the number of positive and
negative crossings in the diagram $D$. This is carried out in Lemma
\ref{lemma:surface-slope}.
\item Relate the limit of $js^*_K$ to the the number of positive and
negative crossings in the diagram $D$. This is carried out in Lemma
\ref{lemma:adequate-limit}.
\end{enumerate}
Taken together, Theorem \ref{thm:essential} and Lemmas
\ref{lemma:surface-slope} and \ref{lemma:adequate-limit} immediately
imply Theorem \ref{thm:main}(a). Part (b) of the theorem follows by
considering the mirror image of the diagram $D$.
\section{State graphs and surfaces}\label{sec:state-graphs}
Let $D$ be a link diagram, and $x$ a crossing of $D$. Associated to
$D$ and $x$ are two link diagrams, each with one fewer crossing than
$D$, called the \emph{$A$--resolution} and \emph{$B$--resolution} of
the crossing. See Figure \ref{fig:splicing}.
\begin{figure}[h]
\centerline{\input{splicing.pstex_t}}
\caption{$A$-- and $B$--resolutions of a crossing.}
\label{fig:splicing}
\end{figure}
A Kauffman state $\sigma$ is a choice of $A$--resolution or
$B$--resolution at each crossing of $D$. Corresponding to every state
$\sigma$ is a crossing--free diagram $s_\sigma$: this is a collection
of circles in the projection plane. We can encode the choices that led
to the state $\sigma$ in a graph ${\mathbb{G}}_\sigma$, as follows. The vertices
of ${\mathbb{G}}_\sigma$ are in $1-1$ correspondence with the state circles of
$s_\sigma$. Every crossing $x$ of $D$ corresponds to a pair of arcs
that belong to circles of $s_\sigma$; this crossing gives rise to an
edge in ${\mathbb{G}}_\sigma$ whose endpoints are the state circles containing
those arcs.
Every Kauffman state $\sigma$ also gives rise to a surface $S_\sigma$,
as follows. Each state circle of $\sigma$ bounds a disk in $S^3$. This
collection of disks can be disjointly embedded in the ball below the
projection plane. At each crossing of $D$, we connect the pair of
neighboring disks by a half-twisted band to construct a surface
$S_\sigma \subset S^3$ whose boundary is $K$. See Figure
\ref{fig:statesurface} for an example where $\sigma$ is the all--$A$
state.
Well--known examples of state surfaces include Seifert surfaces (where
the corresponding state $\sigma$ is defined by following an
orientation on $K$) and checkerboard surfaces for alternating links
(where the corresponding state $\sigma$ is either the all--$A$ or
all--$B$ state). In this paper, we focus on the all--$A$ and all--$B$
states of a diagram, but we do not require our diagrams to be
alternating. Thus our surfaces are generalizations of checkerboard
surfaces.
\begin{define}
A link diagram $D$ is called \emph{$A$--adequate} if the state graph
${\mathbb{G}_A}$ corresponding to the all--$A$ state contains no 1--edge loops.
Similarly, $D$ is called $B$--adequate if the all--$B$ graph ${\mathbb{G}_B}$
contains no 1--edge loops. A link diagram is \emph{adequate} if it is
both $A$-- and $B$--adequate. A link that admits an adequate diagram
is also called \emph{adequate}.
\end{define}
\begin{figure}
\includegraphics{state-surface.eps}
\caption{Left to right: A diagram. The diagram with $A$--resolutions at each crossing. The
graph ${\mathbb{G}_A}$. The state
surface $S_A$.}
\label{fig:statesurface}
\end{figure}
Adequate diagrams are quite common. All reduced alternating diagrams
are adequate. Every $n$--string planar cable of an adequate diagram is
adequate. The standard diagram of a Montesinos link is either
$A$--adequate or $B$--adequate, and typically both. Finally, observe
that a diagram $D(K)$ is $A$--adequate if and only if its mirror image
is $B$--adequate. This observation is useful for the proofs below:
once a result is proved for the all--$A$ state surface $S_A$ of an
$A$--adequate diagram, the corresponding statement about $B$--adequate
diagrams follows by reflection.
\begin{theorem}[Ozawa \cite{ozawa:adequate}]\label{thm:essential}
Let $D$ be an $A$--adequate diagram of a knot $K$. Then the state
surface $S_A$ is incompressible and ${\partial}$--incompressible in the
complement $M_K$. Similarly, if $D$ is a $B$--adequate diagram of a
knot $K$, then $S_B$ is incompressible and
${\partial}$--incompressible.
\end{theorem}
Ozawa's original proof of this theorem relies on building up the surface $S_\sigma$ via Murasugi sums. An
alternate proof from the point of view of normal surface theory will be
given by the authors in \cite{fkp:guts}, where we will also relate the coefficients of
the colored Jones polynomials $J_K(n,q)$ to the size of the \emph{guts} of the surfaces $S_A$ and $S_B$. These guts can be viewed as the hyperbolic pieces in the geometric decomposition of $S^3 \setminus S_\sigma$. Thus, taken together, \cite{fkp:guts} and Theorem \ref{thm:main} of this paper establish two separate connections between the colored Jones polynomials and classical geometric topology.
\medskip
Recall from the Introduction that if $S \subset M_K$ is a surface such
that ${\partial} S$ represents the homology class $p \mu + q \lambda \in
H_1({\partial} M_K)$, we say the boundary slope of $S$ is $p/q \in {\mathbb{Q}} \cup
\{ \infty \}$. It turns out that the boundary slope of a state surface
$S_\sigma$ is easy to read from a diagram $D$.
Suppose that $D(K)$ is a diagram of an oriented knot $K$. Then every
crossing of $D$ can be classified as either positive or negative, as
in Figure \ref{fig:local-slope}. For a state $\sigma$ of $D$, let
$c_+^B(\sigma)$ be the number of positive crossings at which $\sigma$
chooses the $B$--resolution. Similarly, let $c_-^A(\sigma)$ be the
number of negative crossings at which $\sigma$ chooses the
$A$--resolution.
\begin{lemma}\label{lemma:state-slope}
Let $D(K)$ be a diagram of an oriented knot $K$, and let $\sigma$ be a
state of $D$. Then the state surface $S_\sigma$ has as its boundary
the slope $2c_+^B(\sigma) - 2c_-^A(\sigma)$.
\end{lemma}
This lemma was observed by Curtis and Taylor for checkerboard surfaces
of alternating knots \cite[Proposition 2.6]{curtis-taylor}. However, both
the statement and the proof hold in complete generality: $S_\sigma$ is
not even required to be an essential surface.
\begin{figure}[h]
\includegraphics{local-slopes.eps}
\caption{Left: a positive crossing, and a piece of state surface
$S_\sigma$ that chooses the $B$--resolution at this crossing. Locally,
this crossing contributes $+2$ to the slope of $S_\sigma$. Right: if
$\sigma$ chooses the $A$--resolution at a negative crossing, the slope
of $S_\sigma$ receives a local contribution of $-2$.}
\label{fig:local-slope}
\end{figure}
\begin{proof}
Suppose, first, that $\sigma$ is the Seifert state, and $S_\sigma$ is
an oriented Seifert surface constructed from the diagram $D$. To
follow an orientation of $K$, $\sigma$ must choose the $A$--resolution
at every positive crossing and the $B$--resolution at every negative
crossing (the opposite of the choices depicted in Figure
\ref{fig:local-slope}). Thus $c_+^B(\sigma) = c_-^A(\sigma) = 0$, by
definition. Also, because ${\partial} S_\sigma$ is the boundary of an
orientable surface, this curve is null-homologous in $M_K$ and has
slope $0$. This verifies the lemma for the Seifert state.
Next, let $\sigma$ be an arbitrary state. Then it is still the case
that $S_\sigma$ intersects a meridian of $K$ only once. Thus the
boundary slope of $S_\sigma$ is still an integer $p$. Let $L$ be the
simple closed curve of intersection between $S_\sigma$ and ${\partial} M_K$,
oriented in the same direction as $K$. Then the boundary slope $p$ of
$S_\sigma$ is the linking number $lk(K,L)$, or equivalently the
oriented intersection number between $L$ and a Seifert surface for
$K$.
The linking number $lk(K, L)$ can be computed by summing the local
contributions near each crossing. If $\sigma$ chooses the
$A$--resolution at a positive crossing or the $B$--resolution at a
negative crossing, $L$ is locally disjoint from the Seifert surface,
and the local contribution to the linking number is $0$. On the other
hand, for every positive crossing where $\sigma$ chooses the
$B$--resolution, the left panel of Figure \ref{fig:local-slope} shows
that the neighborhood of the crossing contributes $+2$ to the linking
number $lk(K, L)$. Similarly, the right panel of Figure
\ref{fig:local-slope} shows that a negative crossing where $\sigma$
chooses the $A$--resolution contributes $-2$ to $lk(K,L)$.
Adding up these contributions, we conclude that $lk(K, L) =
2c_+^B(\sigma) - 2c_-^A(\sigma)$.
\end{proof}
As a special case of Lemma \ref{lemma:state-slope}, we obtain the
boundary slopes of $S_A$ and $S_B$.
\begin{lemma}\label{lemma:surface-slope}
Let $D(K)$ be a diagram of an oriented knot $K$. Let $c_+$ be the
number of positive crossings in $D$, and $c_-$ the number of negative
crossings. Then the all--$A$ surface $S_A$ has boundary slope $-2c_-$,
and $S_B$ has boundary slope $2c_+$.
\end{lemma}
\begin{proof}
The all--$A$ state $\sigma$ chooses the $A$--resolution at every
crossing. Thus for the all--$A$ state, $c_+^B(\sigma) = 0$ and
$c_-^A(\sigma) = c_-$, hence ${\partial} S_A$ has slope $-2c_-$ by Lemma
\ref{lemma:state-slope}.
Similarly, for the all--$B$ state $\sigma$, $c_+^B(\sigma) = c_+$ and
$c_-^A(\sigma) = 0$, hence ${\partial} S_B$ has slope $2c_+$.
\end{proof}
\section{Colored Jones polynomials}\label{sec:colored}
In this section, we relate the degrees of colored Jones polynomials to
the numbers $c_+$ and $c_-$ of positive and negative crossings in a
diagram $D$. A good reference for the following discussion is
Lickorish's book \cite{lickorish:book}.
The colored Jones polynomials of a link $K$ have a convenient
expression in terms of \emph{Chebyshev polynomials}. For $n \geq 0$,
the polynomial $S_n(x)$ is defined recursively as follows:
\begin{equation}\label{eq:cheb-recursive}
S_{n+1} = x S_n - S_{n-1}, \qquad S_1(x) = x, \qquad S_0(x) = 1.
\end{equation}
Let $D$ be a diagram of a link $K$. For an integer $m > 0$, let $D^m$
denote the diagram obtained from $D$ by taking $m$ parallel copies of
$K$. This is the $m$--cable of $D$ using the blackboard framing; if
$m=1$ then $D^1=D$. Let $\langle D^m \rangle$ denote the Kauffman
bracket of $D^m$: this is a Laurent polynomial over the integers in a
variable $A$. Let $w=w(D) = c_+ - c_-$ denote the writhe of
$D$. Then we may define the function
\begin{equation}\label{eq:unreduced}
G(n+1, A):= \left((-1)^n A^{n^2+2n} \right)^{-w} (-1)^{n-1}
\left(\frac{A^4 - A^{-4}}{A^{2n} - A^{-2n}} \right) \langle S_n( D)
\rangle,
\end{equation}
where $S_n(D)$ is a linear combination of blackboard cablings of $D$,
obtained via equation (\ref{eq:cheb-recursive}), and the notation
$\langle S_n(D) \rangle$ means extend the Kauffman bracket linearly.
That is, for diagrams $D_1$ and $D_2$ and scalars $a_1$ and $a_2$,
$\langle a_1 D_1 + a_2 D_2 \rangle = a_1 \langle D_1\rangle +
a_2\langle D_2\rangle$.
For the results below, the important corollary of the recursive
formula for $S_n(x)$ is that
\begin{equation}\label{eq:chebyshev}
S_n(D) = D^n + (1-n)D^{n-2} + \mbox{lower degree cablings of } D.
\end{equation}
Finally, the reduced $(n+1)$--colored Jones polynomial of $K$, denoted
by $J_K(n+1, q)$, is obtained from $G(n+1, A)$ by substituting
$q:=A^{-4}$.
Recall from the Introduction that $j(n)$ denotes the highest degree of
$J_K(n, q)$ in $q$, and $j^{*}(n)$ denotes the lowest degree. Following big--$O$ notation, we let $O(n)$ denote a term that is at most linear in $n$.
\begin{lemma} \label{lemma:adequate-limit}
Let $D$ be a link diagram with $c_+$ positive crossings and $c_-$
negative crossings.
\begin{itemize}
\item[(a)] If $D$ is $A$--adequate, then
$\displaystyle{j^{*}(n)= - \frac{c_-}{2}n^2+ O(n).}$
\item[(b)] If $D$ is $B$--adequate, then
$\displaystyle{j(n)= \frac{c_+}{2}n^2+ O(n).}$
\end{itemize}
\end{lemma}
\begin{proof}
For part (a), let $D$ be an $A$--adequate diagram with $c = c(D) = c_+
+ c_-$ crossings. Let $v_A = v_A(D)$ be the number of vertices of
${\mathbb{G}_A}$, which is equal to the number of state circles in the all--$A$
state. Then, for every $m > 0$, the link diagram $D^m$ is also
$A$--adequate with $c(D^m)=m^2 c$ and $v_A(D^m)=m v_A$.
Let $\deg ( P(A) )$ denote the highest degree of a polynomial $P$ in
$A$. Then the highest degree of $\langle D \rangle$ is $\deg \langle D \rangle = c+2v_A-2$; see \cite[Lemma
5.4]{lickorish:book} or \cite[Lemma 7.1]{dessin-jones} for a proof. From equation (\ref{eq:unreduced}),
one can see that $\deg( G(n, \ A) )$ comes from the highest--degree
term of $\langle S_{n-1}(D) \rangle$. Furthermore, by equation
(\ref{eq:chebyshev}) and the previous paragraph,
$$\deg \, \langle S_{n-1}(D) \rangle \: = \: \deg \, \langle D^{n-1}
\rangle \: = \: (n-1)^2 c + 2(n-1) v_A - 2.$$ Thus
\begin{eqnarray*}
\deg G(n, A) &=& -w(n^2+2n) + (4-2n) + \deg \, \langle S_{n-1}(D) \rangle \\
&=& -w(n^2+2n) + (4-2n) + (n-1)^2 c + 2(n-1) v_A - 2 \\
&=& (c-w) \, n^2 + O(n) \\
&=& ((c_+ + c_-) - (c_+ - c_-)) n^2 + O(n) \\
&=& 2 c_- n^2 + O(n).
\end{eqnarray*}
Finally, since $J_K(n, q)$, is obtained from $G(n, A)$ by substituting
$q:=A^{-4}$, we conclude that the lowest degree of $J_K(n, q)$ in $q$
is $j^{*}(n)= -c_- n^2/2+ O(n)$. This proves (a).
For part (b), it suffices to observe that the mirror image $D^{*}$ of
a $B$--adequate diagram $D$ will be $A$--adequate. Taking the mirror
image also interchanges positive and negative crossings, and replaces
$q$ with $q^{-1}$ in the colored Jones polynomials. Thus the result
follows from (a).
\end{proof}
We can now complete the proof of Theorem \ref{thm:main}.
\begin{proof}[Proof of Theorem \ref{thm:main}]
Let $D(K)$ be an $A$--adequate diagram. Then, by Theorem
\ref{thm:essential}, $S_A$ is an essential surface for $M_K$. By
Lemmas \ref{lemma:surface-slope} and \ref{lemma:adequate-limit}, the
boundary slope of $S_A$ is
$$-2c_- = \lim_{n \to \infty} 4\, n^{-2}j^*(n).$$
Similarly, if $D(K)$ is $B$--adequate, $S_B$ is an essential surface
with boundary slope
$$2c_+ = \lim_{n \to \infty} 4\, n^{-2}j(n).$$
In particular, if $D(K)$ is adequate, then both sequences $\{ 4\,
n^{-2}j^*(n) \}$ and $\{ 4\, n^{-2}j(n) \}$ converge to boundary
slopes of $K$. For a non-trivial knot $K$, the slopes $-2c_-$ and
$2c_+$ are distinct, because at least one of the integers $c_-$ and
$c_+$ is strictly positive.
\end{proof}
\section{Examples}\label{sec:examples}
\begin{example}
Let $p$ be an odd integer, and let $K_p$ denote the $(-2, 3, p)$ pretzel knot, with a standard pretzel diagram $D_p$. It is easy to check that $D_p$ is $A$--adequate iff $p>0$ and $B$--adequate iff $p<0$. Furthermore, for all values of $p$ except $p=-1$ (when $D_p$ is an unusual diagram of the $5_2$ knot), the knot $K_p$ does not admit an adequate diagram. This classical fact can also be seen via Theorem \ref{thm:main}, because all Jones slopes of adequate knots are integers.
Garoufalidis computed \cite{garoufalidis:jones-slopes} that
$$ \lim_{n \to \infty} \frac{4j(n)}{n^2} = \frac{2(p^2-p-5)}{p-3} \quad \mbox{if } p \geq 5, \qquad
\lim_{n \to \infty} \frac{4j^*(n)}{n^2} = \frac{2(p+1)^2}{p} \quad \mbox{if } p \leq -3.$$
From the work of Hatcher and Oertel
\cite{hatcher-oertel:montesinos-slopes} and
Dunfield \cite{dunfield:boundary-slopes}, it follows that these limiting numbers are indeed boundary slopes of $K_p$. Hence, since all Jones slopes of adequate knots are integers by Theorem \ref{thm:main}, we recover the classical fact that these knots are not adequate.
\end{example}
\begin{example}
Following \cite{lick-thistle:adequate-knots}, a $(2,2)$--tangle $T$ is called \emph{strongly
alternating} if each of the closures of $T$ is a reduced alternating
link diagram. Any knot obtained as a Conway sum of two
strongly alternating tangles is then adequate (see also \cite{fkp:conway}).
For example, any non-alternating pretzel knot $K(a_1, \ldots, a_r, b_1, \ldots,
b_k)$ with with $a_i, b_j, r, k\geq 2$ is adequate. Similarly, as
explained in \cite{lick-thistle:adequate-knots}, a reduced diagram of any Montesinos knot with
at least two positive rational tangles and at least two negative rational tangles will be
adequate.
Theorem \ref{thm:main} implies that these knots satisfy Garoufalidis'
conjecture.
\end{example}
\begin{example}
Let $B_n$ denote the braid group on $n$ strings, and let $\sigma_1, \cdots,
\sigma_{n-1}$ be the elementary braid generators. Let $D_b$ denote the closed braid diagram
obtained from the braid
$b=\sigma_{i_1}^{r_1}\sigma_{i_2}^{r_2} \cdots \sigma_{i_k}^{r_k}$. If $r_j > 0$ for all $j$, the positive braid diagram $D_b$ will be $A$--adequate. Since all crossings in this braid are positive, $c_-=0$. Thus, by Lemma \ref{lemma:adequate-limit},
$$ \lim_{n \to \infty} 4\, n^{-2}j^*(n) \:=\: -2c_- \:=\: 0.$$
Furthermore, the essential surface $S_A$ whose boundary is this slope will be a fiber in $S^3 \setminus K$.
Under the stronger hypothesis that $r_j \geq 3$ for all $j$, the diagram $D_b$ is not only $A$--adequate but also $B$--adequate. Thus Theorem
\ref{thm:main} applies. The other boundary slope detected by the colored Jones polynomials is
$$ \lim_{n \to \infty} 4\, n^{-2}j(n) \:=\: 2c_+ \:=\: 2 \sum_{j=1}^k r_j.$$
\end{example}
\bibliographystyle{hamsplain}
|
\section{Introduction}
Recent detailed Dalitz plot analyzes for different $B\to K\pi^+\pi^- $
decays were performed by BaBar and Belle collaborations
\cite{Garmash:2005rv,Garmash2007,Aubert:2008bj,Aubert:2007bs,Abe2005}.
New and rich data on $K\pi$ effective mass ($m_{K\pi}$) and helicity angle distributions,
branching fractions and CP asymmetries have been delivered.
The effective mass distributions show distinct accumulation of events below 2 GeV around the scalar
$K^*_0(1430)$ and the vector $K^*(892)$ resonances.
In experimental analyzes the event distributions have been studied using the isobar model with
decay amplitudes parameterized by sums of Breit-Wigner terms and a background.
In \cite{Aubert:2008bj,Aubert:2007bs}
an effective range component has been introduced in the $K\pi$ $S$-wave amplitude.
In this short report we present main results of our work on $B$ decays into three mesons,
$B\to K\pi^+\pi^-$, with final state interactions between kaons and pions in the $S$- and $P$-waves \cite{BFKLLM}.
In our analysis we restrict ourselves to the low effective $K\pi$ mass region $m_{K\pi} < 2$ GeV
where most of the $K\pi$ resonant structures are seen.
It is reasonable to expect that in such a kinematical configuration
three-body interactions are suppressed.
This allows us to grant the validity of quasi two-body factorization approach.
We also assume that the $K \pi$ pairs originate from a quark-antiquark state.
Main assumption in our approach is, however, that
the corrections to factorization can be absorbed into
$m_{K\pi}$ independent modifications of the QCD Wilson coefficients.
We take parts of these corrections from quasi-two-body
calculations and append a phenomenological part.
In the factorization approach the decay amplitudes can be expressed as a product
of the effective QCD part and the two matrix elements of the vector currents
which involve two functions of $m_{K\pi}^2$: the strange scalar and the vector form factors.
The introduction of form factors, constrained by theory and experiments other than $B$ decays,
is an alternative to the isobar model often used in experimental analyzes.
This latter approximation violates unitarity what can, for example, lead to a distortion of
information on resonances present in the final states.
\section{Theoretical model}
The weak amplitudes for the $B^{\pm}$, $B^0$ and $\bar B^0$ decays, have one term derived in the QCD factorization,
and the second one being a phenomenological contribution to the penguin amplitudes.
For example, the $K \pi$ $S$-wave contribution to the $B^- \to K^- \pi^+\pi^-$ decay amplitude reads
(expressions for other $B$ decay amplitudes can be found in \cite{BFKLLM}):
\begin{eqnarray}
\nonumber
\hspace{-9.cm}\mathcal{A}_S^-(m_{K\pi}^2) =
\frac{G_F}{\sqrt{2}} (M_B^2-m_{\pi}^2)\frac{m_{K}^2-m_{\pi}^2}{m_{K\pi}^2}
f_0^{B^-\pi^-}(m_{K\pi}^2) \ {f_0^{K^-\pi^+}(m_{K\pi}^2)} \\ \nonumber
\times
\bigg\{
\lambda_u\left(a_4^{u}(S)-\frac{a_{10}^{u}(S)}{2}+{ c_4^{u}}\right)
+\lambda_c\left(a_4^c(S)-\frac{a_{10}^c(S)}{2}+{ c_4^{c}}\right) \\
- \frac{2m_{K\pi}^2}{(m_b-m_d)(m_s-m_d)}
\times \left[
\lambda_u\left(a_6^{u}(S)-\frac{a_8^{u}(S)}{2}+{ c_6^{u}}\right)
+\lambda_c\left(a_6^c(S)-\frac{a_8^c(S)}{2}+{ c_6^{c}}\right)
\right]\bigg\},
\end{eqnarray}
where the $a_i$ ($i=4,6,8,10$) are the leading order factorization coefficients with
$O(\alpha_s)$ vertex and penguin corrections
added in the quasi-two-body approximation of pseudoscalar-scalar or pseudoscalar-vector final states.
The $c_i$ coefficients are additional phenomenological complex parameters fitted to the experimental data.
They can be interpreted as contributions from
charming penguins, hard spectator interaction in the perturbative regime and as annihilation terms.
$G_F$ denotes the Fermi coupling constant, $M_B,\ m_K$ and $m_\pi$ are the
masses of the charged $B$ mesons, kaons and pions, $f_0^{B^-\pi^-}(m_{K\pi}^2)$ and $f_0^{K^-\pi^+}(m_{K\pi}^2)$
are the $B^-$ to $\pi^-$ and the $K\pi$ scalar form factors respectively and $m_b,\ m_s$ and $\ m_d$ are $b,~ s$ and
$d$ quark masses.
The $\lambda_u= V_{ub} V^*_{us}$ and $\lambda_c= V_{cb} V^*_{cs}$ are products of the
Cabibbo-Kobayashi-Maskawa quark-mixing matrix elements $V_{qq'}$.
As we have shown in \cite{BFKLLM}, the $m_{K\pi}^2$ dependence of the $S$-wave amplitude
(3rd line of the Eq. (1)) plays an important role in
the description of the sizable enhancement in the $K\pi$ mass distributions below 1 GeV.
In the isobar model the corresponding $m_{K\pi}$ dependence is approximated by one fitted constant parameter.
Such approximation may not be very good for the wide $K^*_0(1430)$ \hspace{-0.07cm} resonance since the value of $m_{K\pi}^2$ varies by a factor almost
equal to 10 from the $K\pi$ threshold to the limit of our calculations about 1.8 GeV.
The scalar and vector $K\pi$ form factors are connected to the scattering amplitudes
in the $S$- and $P$-waves via unitarity relations. In the $S$-wave
we treat two coupled $ K\pi$ and $K\eta '$ channels and in the $P$-wave three ones:
$ K\pi$, $K^*\pi$ and $K\rho$.
The corresponding scattering amplitudes are constrained by experimental data,
especially by the LASS results obtained at SLAC, and by theoretical constraints at low effective mass.
The Muskhelishvili-Omnes equations are used to calculate the $K \pi$ form factors.
\section{Numerical results}
The $ K\pi$ effective mass and helicity angle distributions,
branching fractions, CP asymmetries and the phase difference between the $B^0$ and
$\bar B^0$ decay amplitudes to $K^*(892)\pi$ are calculated from our model and fitted,
using the minimization program with four phenomenological free parameters $c_{4,6}^{u,c}$,
to 319 data obtained by BaBar and Belle collaborations.
Branching fractions play a special role as they determine the absolute size of the decay amplitudes.
As an input
we use the branching fractions for the ${B} \to K^*(892)\pi$ decays which are well determined since
the $K\pi$ $P$-wave amplitude below 1 GeV is entirely dominated by the
$K^*(892)$ resonance.
Contrary to the $P$-wave amplitude, the $S$-wave amplitude contains not only the resonant part
related to the wide $K^*_0(1430)$ but also a very broad $K^*_0(800)$ below 1 GeV.
This complexity of the $S$-wave results in different parameterizations of the amplitudes
in the BaBar and Belle analyzes.
Therefore we do not use the experimental branching fractions for the
${B} \to K^*_0(1430)\pi$, which are not well determined, being largely model dependent.
Some of our model predictions for branching fractions in both waves are given in Table~1.
These numbers have been calculated by integration of the double differential branching fractions
in limited ranges of $m_{K\pi}$ around the $K^*_0(1430)$ \hspace{-0.07cm} and $K^*(892)$ \hspace{-0.07cm} resonances.
The model branching fractions for the $B \to K^*(892) \pi$ decays agree
quite well with the corresponding experimental values within their errors.
If the phenomenological parameters $c_i^p\ (i=4,6$ and $p=u,c$)
are equal to zero, than
the theoretical branching fractions for the $P$-wave are 4-5 times smaller than the experimental ones
and those for the $S$-wave differ by a factor of 2.
One can notice that the theoretical results for the $S$-wave
are lower than the experimental data, being substantially smaller than the
Belle results but closer to the BaBar numbers.
\begin{table}[h!]
\begin{center}
\caption{Branching fractions averaged over charge conjugate $B\to K\pi\pi$ decays
in units of $10^{-6}$. $(K^+\pi^-)_P$ and $(K^+\pi^-)_S$ denote the
$(K^+\pi^-)$ pair in $P$-wave and $S$-wave, respectively.
The values of the model, calculated by the integration over the given
$m_{K\pi}$ range,
are compared to the corresponding Belle ~\cite{Garmash:2005rv,Garmash2007} and
BaBar ~\cite{Aubert:2008bj,Aubert:2007bs} results.
The model errors are the phenomenological parameter uncertainties
found in the minimization procedure.
}
\begin{tabular}[]{|c|c|c|c|c|}
\hline
decay mode & $m_{K\pi}$ range (GeV) & Belle & BaBar & our model\\
\hline
$B^+\to (K^+\pi^-)_P\ \pi^-$ &$(0.82,0.97)$ & $5.35\pm0.59$ & $5.98\pm0.75$ &
$5.73\pm0.14$ \\
\hline
$B^0\to (K^0\pi^+)_P\ \pi^-$ &$(0.82,0.97)$ & $4.65\pm0.77$ & $6.47\pm0.75$ &
$5.42\pm0.16$ \\
\hline
$B^+\to (K^+\pi^-)_S\ \pi^-$ &$(0.64,1.76)$ & $27.0\pm2.5$ & $22.5\pm4.6$ &
$16.5\pm0.8$ \\
\hline
$B^0\to (K^0\pi^+)_S\ \pi^-$ &$(0.64,1.76)$ & $26.0\pm3.4$ & $17.3\pm4.6$ &
$15.8\pm0.7$ \\
\hline
\end{tabular}
\end{center}
\end{table}
In Table~2 we present comparison of our results for the CP asymmetries with experimental ones.
For $B^- \to (K^-\pi^+)_P\pi^-$ decays our asymmetries lie between those of Belle and BaBar.
The results for $B^-\to (K^-\pi^+)_S\ \pi^-$ and for $\bar B^0\to (\bar K^0\pi^-)_S\ \pi^+$
decays agree with the experimental values of both collaborations.
\begin{table}[h!]
\begin{center}
\caption{Direct $CP$ asymmetries averaged over charge conjugate reactions.
The values of the model, calculated over the indicated $m_{K\pi}$ range,
are compared to the Belle \cite{Garmash:2005rv} and BaBar \cite{Aubert:2008bj,Aubert:2007bs} results.
The model errors are the phenomenological parameter uncertainties
found in the minimization procedure.
}
\begin{tabular}{|c|c|c|c|c|}
\hline
Decay mode & $m_{K\pi}$ range (GeV) & Belle & BaBar & our model \\
\hline
$B^-\to (K^-\pi^+)_P \pi^-$ & $(0.82,0.97)$ & $-14.9\pm 6.8$ & $3.2\pm 5.4$ & $-2.5\pm 1.3$ \\
\hline
$\bar B^0\to (\bar K^0\pi^-)_P\ \pi^+$ & $(0.82,0.97)$ & $----$ & $14 \pm 12$ & $-19.6 \pm 3.0$\\
\hline
$B^-\to (K^-\pi^+)_S \pi^-$ & $(1.0,1.76)$ &$7.6\pm 4.6$ & $3.2\pm 4.6$ & $5.4 \pm 1.0$ \\
\hline
$\bar B^0\to (\bar K^0\pi^-)_S \pi^+$ & $(1.0,1.76)$ & $----$ & $17\pm 26$ & $-0.2 \pm 1.3$ \\
\hline
\end{tabular}
\end{center}
\end{table}
The fit on the model parameters leads to a good agreement with the experimental data, particularly
for the kaon-pion effective mass and helicity angle distributions.
Examples of these results are presented in
Figs \ref{Fig.MassDistr} and \ref{Fig.Helicity}.
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,scale=0.44]{Fig5.ps}
\caption{The $K^\pm \pi^\mp$ effective mass distributions in the
$B^\pm \to K^\pm \pi^\mp \pi^\pm$ decays from the fit to the experimental data \cite{Garmash:2005rv}.
The
dashed line corresponds to the $S$-wave contribution of our model,
the dotted line that of the $P$-wave and the histogram represents the coherent sum of the $S$- and $P$-wave contributions.
}
\label{Fig.MassDistr}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,scale=0.40]{Fig7a.ps}
\includegraphics[angle=0,scale=0.40]{Fig7b.ps}\\
\caption{Helicity angle distributions for $B^\pm \to K^\pm \pi^\mp \pi^\pm$ decays calculated
from the averaged double differential distribution integrated over $m_{K^\pm \pi^\mp}$ mass from
0.82 to 0.97 GeV in the $K^*(892)$ \hspace{-0.07cm} case a) and from 1.0 to 1.76 GeV in the $K^*_0(1430)$ \hspace{-0.07cm} case b).
Data points are from Ref.~\cite{Abe2005}.
Dashed lines represent the $S$-wave contribution of our model, dotted lines that of the $P$-wave and
the dot-dashed lines that of the interference term.
The histograms correspond to the sum of these three contributions.
}
\label{Fig.Helicity}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,scale=0.43]{Fig13a.ps}
\includegraphics[angle=0,scale=0.43]{Fig13b.ps}\\
\caption{Different components of the averaged $m_{K\pi}$ distributions of the $B^\pm \to (K^\pm\pi^\mp)_S
\pi^\pm$ decays for: a) our model and b) BaBar's LASS parameterization~\cite{Aubert:2008bj}.
The dashed lines correspond to the resonant $K^*_0(1430)$ contributions, the dotted-dashed lines to the background, dotted lines to the interference and the solid lines to their sum.
}
\label{Fig.components}
\end{center}
\end{figure}
In Fig. \ref{Fig.components} we show decomposition of the averaged $m_{K\pi}$ distributions
of the charged B meson decays into a resonant, background and interference terms.
We compare these components calculated using our amplitudes and the BaBar's LASS
parameterization.
The background part in our approach is several times smaller than that effective range term
in experimental analyzes.
Worthy of note is the fact that interference parts have very different sizes and even
different signs after integration over $m_{K\pi}$.
\section{Conclusions}
We have shortly presented main results of our analysis of the $K\pi$ interactions in
the $B \to K\pi^+\pi^-$ decays with $K\pi$ pairs in the $S$- and $P$-waves.
In our approach sufficiently good description of experimental data can be achieved after introduction
of four complex phenomenological coefficients into amplitudes.
These additional parameters are added to the QCD factorization coefficients and
can be interpreted as coming from long-distance charming penguin amplitudes, hard
spectator and weak annihilation contributions.
We have shown that the $K\pi$ $P$-wave amplitude is entirely dominated by
$K^*(892)$ whereas the $S$-wave with the wide resonance $K^*_0(1430)$ and the very wide
enhancement below 1 GeV (often interpreted as the $K^*_0(800)$ meson) is more complex.
In order to correctly account for these resonances
we use scalar and vector form factors in the $S$- and $P$-waves
instead of the Breit-Wigner amplitudes, often employed in the isobar model.
This leads to the background term in the effective mass distributions
which is several times smaller than the one obtained in each experimental analysis.
We therefore propose to use in the future
experimental analyzes the following parameterization of the $S$-wave
$B \to (K\pi)_S\pi$ decay amplitude:
\begin{equation}
\mathcal{A}_S^- = f_0^{K\pi}(m_{K\pi}^2)(c_0/m_{K\pi}^2 + c_1),
\nonumber
\end{equation}
where $f_0^{K\pi}$ is the scalar $K\pi$ form factor, while $c_0$ and $c_1$ are
complex numbers to be fitted from the data.
Numerical values of the complex scalar form factor can be provided on request.
Let us emphasize at the end that the amplitudes derived in our analysis \cite{BFKLLM} go
beyond the two-body approach applied to such decays as $B\to K^*\pi$.
We have proved that they correctly incorporate the two body final state interactions between
$K$ and $\pi$ if the effective $K\pi$ mass in the three-body $B \to K \pi^+\pi^-$ decays is limited to about 2 GeV.
|
\section{Introduction}
\subsection{Coagulation models}
In this paper we deal with the problem of uniqueness of post-gelation solutions of
several models of coagulation, namely Smoluchowski's and Flory's classical models, and the corresponding
models with limited aggregations recently introduced by Bertoin \cite{BertoinTSS}.
Smoluchowski's coagulation equations
describe the evolution of the concentrations of particles in a system where particles
can perform pairwise coalescence, see e.g. \cite{AldousReview, LauMi,NorrisReview}. In the original model of Smoluchowski \cite{Smolu}, a pair of particles
of mass, respectively, $m$ and $m'$, coalesce at rate
$\k(m,m')$ and produce a particle of mass $m+m'$.
In the discrete setting, the evolution of the concentration $c_t(m)$ of particles of
mass $m\in{\ensuremath{\mathbb N}} ^*$ at time $t \geq 0$ is given by the following system
\begin{equation}\label{coagu}
\frac{\mathrm{d}}{\mathrm{d}t} c_t(m) = \frac12 \sum_{m'=1}^{m-1} \k(m,m')\, c_t(m') \, c_t(m-m') - \sum_{m'\geq 1} \k(m,m') c_t(m)c_t(m').
\end{equation}
Norris considered in \cite{NorrisCC} far more general models of {\it cluster coagulation},
where the rate of coalescence does not depend only on the mass of the particles but
also on other parameters. In this general setting, most results on existence and uniqueness are obtained before a critical time, known as the {\it gelation time}, while the global behavior of the solutions after gelation, and in particular uniqueness, is not known.
An example of a solvable cluster coagulation model is Bertoin's {\it model with limited aggregations} \cite{BertoinTSS}, which we shall simply call the {\it model with arms}. In this
case, particles have a mass but also carry a certain number of potential links, called {\it arms}. Two particles of mass $m$ and $m'$ may coagulate only if they have a positive number of arms, say $a$ and $a'$. When they coagulate, an arm of each is used to create the bond and both arms are then deactivated, hence creating a particle with $a+a'-2$ arms and mass $m+m'$. The coagulation rate of these two particles is $aa'$.
Therefore, if $c_t(a,m)$ is the concentration of particles with $a\in{\ensuremath{\mathbb N}} =\{0,1,\dots\}$ arms and mass $m\in{\ensuremath{\mathbb N}} ^*=\{1,2,\dots\}$, then
the coagulation equation reads
\begin{equation} \label{smolubras0}
\begin{split}
\frac{\mathrm{d}}{\mathrm{d}t} c_t(a,m) & = \frac12 \sum_{a'=1}^{a+1} \sum_{m'=1}^{m-1} a'(a+2-a') c_t(a',m')c_t(a+2-a',m-m') \\
& \quad - \sum_{a' \geq 1} \sum_{m'\geq 1} aa'c_t(a,m) c_t(a',m').
\end{split}
\end{equation}
For monodisperse initial concentrations, i.e. $c_0(a,m)=\un{\{m=1\}}\mu(a)$,
with $\mu=(\mu(a))_{a\in{\ensuremath{\mathbb N}} }$ a measure on ${\ensuremath{\mathbb N}} $ with unit mean,
it is proved in \cite{BertoinTSS} that this equation has a unique solution on some interval $[0,T)$, where $ T = + \infty$ if and only if $K \leq 1$, where
\begin{equation} \label{K}
K:=\sum_{a\geq 1} a(a-1)\mu(a).
\end{equation}
In other words, if particles at time 0 have, on average, few arms, equation \eqref{smolubras0} has a unique solution defined for all $t \geq 0$.
When this is the case, as time passes, all available arms are used to create bonds and only
particles with no arms remain in the system. The limit concentrations $c_{\infty}(0,m)$
as $t\to+\infty$ of such particles turn out to be related
to the distribution of the total population generated by a sub-critical Galton-Watson
branching process (see e.g. \cite{AtNey}) started from two ancestors: see \cite{BertoinTSS,BertoinSido}
and section \ref{limconc} below.
\subsection{The gelation phase transition}
A formal computation shows that solutions
of \eqref{coagu} with multiplicative kernel $\k(m,m')=mm'$ should have constant mass
\begin{equation} \label{Mt0}
M_t := \sum_{m \geq 1} m c_t(m)
, \qquad t\geq 0,
\end{equation}
i.e. $\frac d{dt}M_t=0$. It is however well-known that if large particles can coagulate sufficiently fast, then
one may observe in finite time a phenomenon called {\it gelation}, namely the formation of particles of infinite mass, the {\it gel}. These particles do not count in the computation of the mass
so from the gelation time on, $M_t$ starts to decrease.
The reason why \eqref{smolubras0} can be solved, is that it can be transformed
into a solvable PDE involving the generating function of $(c_t)_{t\geq 0}$.
In Equation \eqref{coagu}, this transformation is also possible for several
particular choices of the kernel $\k(m,m')$, namely
when $\k$ is constant, additive or multiplicative: see e.g. \cite{DeaconuTanre}. In the multiplicative case $\k(m,m')=mm'$, which is our main concern here,
the total mass
is a parameter of \eqref{coagu} and of the associated PDE, which is therefore easy to solve only when $(M_t)_{t\geq 0}$ is known. Existence and uniqueness of solutions of \eqref{coagu} are thus
easy up to gelation, since in this regime, the total mass $M_t$ is constant.
After gelation, the gel may or may not interact with the other particles. If it does, Equation \eqref{coagu} has to be modified into Flory's equation \eqref{flory}. Else, the gel is inert, in which case Smoluchowski's equation continues to hold. Obviously, they are identical before gelation.
Occurrence of gelation depends heavily on the
choice of the coagulation rate $\k(m,m')$, and in the multiplicative case, gelation always occurs \cite{EscobedoGaMC,FouLauWP,LauCCC}. After gelation, the mass is not known, so $M_t$ itself becomes an unknown of the equation, and well-posedness of the equation is then much less trivial. The multiplicative kernel is therefore particularly interesting, since it exhibits a non-trivial behavior but can still be studied in detail by means of explicit computations.
\medskip
The same phenomenon of gelation has been observed in \cite{BertoinTSS} for \eqref{smolubras0} for monodisperse initial concentrations $c_0$.
A formal computation shows that the the mean number of arms $A_t$
\[
A_t:=\sum_{a,m \geq 1} a\, c_t(a,m), \qquad t\geq 0,
\]
satisfies the equation $\frac d{dt}A_t=-A_t^2$ and should therefore be equal to
$\frac{1}{1+t}$ for all $t\geq 0$. In fact, this explicit expression
holds only until a critical time, which is shown to be equal
to $1/(K-1)$ if $K>1$ and to $+\infty$ if $K\leq 1$, where $K$ is defined in \eqref{K}.
Again, the associated PDE is easy to solve before gelation since then, $A_t$ is known, while afterwards, the PDE contains
the unknown parameter $A_t$.
\subsection{Main result}
In this paper we investigate the global behavior of Smoluchow\-ski's equation with arms \eqref{smolubras0} before, at and
after the gelation phase transition, proving existence and uniqueness of global solutions
for a large class of initial conditions.
The technique used, as in \cite{BertoinTSS}, is to transform the equation into a PDE. Since the total number of arms $(A_t)_{t\geq 0}$ is not a priori known, this PDE is non local, unlike the one obtained in the regime before gelation. This is the main difficulty we have to deal with. We use a modification of the classical method of characteristics to show uniqueness of solutions to this PDE, and hence to \eqref{smolubras0}. We can consider initial conditions $(c_0(a,m), a \in {\ensuremath{\mathbb N}} , m \in {\ensuremath{\mathbb N}} ^*)$ with an initial infinite number of arms, that is, such that
\[
A_0 := \sum_{a, m \geq 1} a c_0(a,m)
\]
is infinite, and show that there is a unique solution ``coming down from infinity
sufficiently fast'', i.e. such that, for positive $t$,
\[
\int_0^t A_s^2 \: \mathrm{d}s < + \infty.
\]
Note however that this is no technical condition, but a mere assumption to ensure that the equation is well-defined.
We also consider a modification of this model which corresponds to Flory's equation for the model with arms. In this setting, the infinite mass particles, that is, the gel, interact with the other particles. We also prove existence, uniqueness and study the behavior of the solutions for this model.
In both cases, our technique provides a representation formula allowing to compute various quantities, as the mean number of arms in the system and the limiting concentrations. In Flory's case, we extend to all possible initial concentrations the computations done in \cite{BertoinTSS} in absence of gelation. In the first model, a slight modification appears which calls for a probabilistic interpretation; see
section \ref{limconc} below.
This seems to be the first case of a cluster coagulation model for which global well-posedness
in presence of gelation can be proven. Another setting to which these techniques could be applied
is the {\it coagulation model with mating} introduced in \cite{NorMFC}.
\subsection{Limiting concentrations}\label{limconc}
In \cite{BertoinTSS}, explicit solutions to \eqref{smolubras0} are given for monodisperse initial conditions $c_0(a,m) = \mu(a) \un{\{m=1\}}$ for some measure $\mu$ on ${\ensuremath{\mathbb N}} $ with unit first moment. In particular, when there is no gelation, i.e. $K\leq 1$ where $K$ is as in \eqref{K}, and $\mu \neq \frac12\delta_2$, there are limiting concentrations
\[
c_{\infty}(a,m) = \frac{1}{m(m-1)} \, \nu^{*m}(m-2)\, \un{\{a=0\}}, \qquad m \geq 2,
\]
where $\nu(m) = (m+1) \mu(m+1)$ is a probability measure on ${\ensuremath{\mathbb N}} $ different from $\delta_1$. This formula clearly resembles the well-known formula of Dwass \cite{Dwass}, which provides the law of the total progeny $T$ of a Galton-Watson process with reproduction law $\nu$, started from two ancestors:
\[
{\ensuremath{\mathbb P}} (T=m)=\frac2m \, \nu^{*m}(m-2), \quad m \geq 2.
\]
The similarity between the two formulas is no coincidence and is explained in \cite{BertoinSido} by means of the configuration model. For basics on Galton-Watson processes, see e.g. \cite{AtNey}.
Let us briefly explain the result of \cite{BertoinSido}, referring e.g. to \cite{vdHofstad} for more results on general random graphs. The configuration model aims at producing a random graph whose vertices have a prescribed degree. To this end, consider a number $n$ of vertices, each being given independently a number of arms (that is, half-edges) distributed according to $\mu$. Then, two arms in the system are chosen uniformly and independently, and form an edge between the corresponding vertices. This procedure is repeated until there are no more available arms. Hence, one arrives to a final state which can be described as a collection of random graphs. Then Corollary 2 in \cite{BertoinSido} and the discussion below show that, when there is no gelation, the proportion of trees of size $m$ tends to $c_{\infty}(0,m)$ when the number $n$ of vertices tends to infinity. Hence, the final states in the configuration model and in Smoluchowski's equation with arms coincide. This shows that the former is a good discrete model for coagulation.
Interestingly, the absence-of-gelation condition $K\leq 1$ is equivalent to (sub)-criticality
of the Galton-Watson branching process with reproduction law $\nu$, i.e. to almost sure extinction of the progeny, while $K>1$ and gelation at finite time are equivalent to super-criticality of the GW process.
In this paper, we obtain the limiting concentrations for \eqref{smolubras0} and its modified version when there is gelation. Let us start with the modified model, which is the counterpart of Flory's equation for the model with arms. In this case, and with the same notations as above, we obtain the limit concentrations
\[
c_{\infty}(a,m) = \frac{1}{m(m-1)} \, \nu^{*m}(m-2)\, \un{\{a=0\}}, \qquad m \geq 2,
\]
that is, the same explicit form as the one obtained in absence of gelation.
Again, this formula can be interpreted both in terms of a configuration model and
of a super-critical Galton-Watson branching process. The relation between
Flory's equation with arms and the configuration model is natural, since
in both cases all particles interact with each other, no matter what their size is.
It is worth noticing that, even though the limit concentrations have the same form with or without gelation, still some mass is eventually lost in presence of gelation, see \eqref{Minfty} below.
We also obtain the limiting concentrations for Smoluchowski's equation with arms, namely
\[
c_{\infty}(a,m) = \frac{1}{m(m-1)} \, \beta_{\infty}^{m-1} \nu^{*m}(m-2)\, \un{\{a=0\}},
\]
where $\beta_{\infty}$ is some constant, which is $1$ when there is no gelation, and is greater than $1$ otherwise, see Section \ref{seccinfsm}. However, the probabilistic interpretation of $\beta_{\infty}$ is unclear. One can recover Smoluchowski's equation with arms from discrete models by preventing big particles from coagulating, as is done in \cite{FouLauMLP} for the standard Smoluchowski equation, but the precise meaning of $\beta_{\infty}$ still seems to require some labor.
\subsection{Bibliographical comments}
Smoluchowski's equation \eqref{coagu} has been extensively studied; we refer to the reviews \cite{AldousReview,LauMi,NorrisReview}. Conditions on the kernel $\k$ are know for absence or presence of gelation, though this requires a precise definition of gelation, see e.g. \cite{EscobedoGiC}, or \cite{Jeon} in a probabilistic setting. For a general class of kernels Smoluchowski's solution has a unique solution before gelation \cite{NorrisReview,DuboStew,FouLauWP,LauMi}, and in the multiplicative case gelation always occurs \cite{EscobedoGaMC,FouLauWP,LauCCC}.
For the monodisperse initial condition $c_0(m)=\un{\{m=1\}}$, the first proof of existence and uniqueness to \eqref{coagu} before gelation is given in \cite{McLeod}, and a proof of global existence and uniqueness can be found in \cite{Kokholm}. The case of general nonzero initial conditions has been considered by several papers in the Physics literature \cite{ErnstKoG, ErnstCP,LeyvTschu,StellZiff,ZiffKoP}, and by at least one mathematical paper \cite{ShirRoesSR}, which however treats in full details only the regime before gelation, see Remark \ref{see} below. The same authors also provide in \cite{ShirRoesPGM} an exact formula for the post-gelation mass of \eqref{coagu}, but with no rigorous proof.
Thus, a clear statement about well-posedness of \eqref{coagu} for the most general initial conditions still seems to be missing, and our paper tries to fill this gap. We adapt the classical method of characteristics for generating functions, see \cite{DeaconuTanre,BertoinTSS}, which yields easily uniqueness before gelation for a multiplicative kernel \cite{MenonPego}. We can in particular consider initial concentrations with infinite total mass, i.e. such that
\[
M_0:=\int_{(0,+\infty)} m\, c_0(\mathrm{d}m)=+\infty,
\]
as long as $\int_{(0,+\infty)} (m\wedge 1)\, c_0(\mathrm{d}m)<+\infty$.
This covers for instance initial conditions of the type $c_0(\mathrm{d}m)=C_p\, m^{-p}\,
\mathrm{d}m$ with $p\in[1,2)$.
Our main concern is uniqueness, since existence of solutions has been obtained in a much more general setting by analytic \cite{LauGS,LauCCC,NorrisCC} or probabilistic \cite{FouLauMLP,Jeon} means. However, the case of an infinite initial mass seems to have been considered only in \cite{LauGS} in the discrete case, so we refer to Section \ref{exist} below for a proof.
\subsection{Plan of the article}
We start off in Section 2 by considering existence, uniqueness and representation formulas for global solutions of \eqref{coagu}, introducing and
exploiting all main techniques which are needed afterwards to tackle the same issues in the
case of \eqref{smolubras0}. We prove that for the most general initial conditions $\mu_0(\mathrm{d}m)$, a positive measure on
$(0,+\infty)$, Smoluchowski's equation with a multiplicative kernel has a unique solution
before and after gelation. We also show existence and uniqueness for the modified version of Smoluchowski's model, namely Flory's equation, in Section 3. The techniques used are generalized in Section 4 and 5, where we prove analogous results for the models with arms. We compute the limiting concentrations in Section 6, which are not trivial, in comparison with the standard Smoluchowski and Flory cases, for which they are always zero.
\section{Smoluchowski's equation}
In this section we develop our method in the case of equation \eqref{coagu},
proving existence, uniqueness and representation formulas for global solutions.
Let us first fix some notations.
\begin{itemize}
\item $\mathcal{M}_f^+$ is the set of all non-negative finite measures on $(0,+ \infty)$.
\item $\mathcal{M}_c^+$ is the set of all non-negative Radon measures on $(0,+ \infty)$.
\item For $\mu \in \mathcal{M}_c^+$ and $f \in L^1(\mu)$ or $f \geq 0$,
\[
\langle \mu, f \rangle = \int_{(0,+ \infty)} f(m) \: \mu(\mathrm{d}m).
\]
We will write $m$ for the function $m \mapsto m$, $m^2$ for $m \mapsto m^2$, etc.
\item For $\phi : (0,+ \infty) \to {\ensuremath{\mathbb R}} $ and $m,m' > 0$, $\Delta \phi(m,m') = \phi(m+m')-\phi(m)-\phi(m')$.
\item $C_c(0,\pinf)$ is the space of continuous functions on $(0,+ \infty)$ with compact support.
\item For a function $(t,x) \mapsto \phi_t(x)$, $\phi_t'(x)$ is the partial derivative of $\phi$ with respect to $x$.
\item $\dplus{}{t}$ or $\dplusdt{}$ denotes the right partial derivative with respect to $t$.
\end{itemize}
We are interested in Smoluchowski's equation \eqref{coagu} with multiplicative
coagulation kernel $\kappa(m,m')=mm'$. Note that the second requirement in the following definition is only present for the equation to make sense.
\begin{definition}\label{defsolsmolu}
Let $\mu_0\in\mathcal{M}_c^+$.
We say that a family $(\mu_t)_{t\geq 0}\subset \mathcal{M}_c^+$ solves Smoluchowski's equation if
\begin{itemize}
\item for every $t>0$, $\int_0^t \langle \mu_s(\mathrm{d}m), m\rangle^2 \: \mathrm{d}s < + \infty$,
\item for all $\phi \in C_c(0,\pinf)$ and $t>0$
\begin{equation}
\langle \mu_t , \phi \rangle = \langle \mu_0, \phi \rangle + \frac12 \int_0^t \langle \mu_s(\mathrm{d}m) \mu_s(\mathrm{d}m'), mm'\, \Delta \phi(m,m') \rangle \: \mathrm{d}s,
\label{smolu2}
\end{equation}
\item if $\langle \mu_0 , m^2 \rangle < + \infty$, then $t \mapsto \langle \mu_t , m^2 \rangle$ is bounded in a right neighborhood of 0.
\end{itemize}
\end{definition}
The global behavior of this equation has been studied first for monodisperse initial conditions (i.e. $\mu_0=\delta_1$), in which case it can be proven that there is a unique solution $(\mu_t)_{t\geq 0}$ on ${\ensuremath{\mathbb R}} ^+$, which is also explicit, see \cite{McLeod, Kokholm}. This solution clearly exhibits the gelation phase transition. Up to the gelation time $T_{\rm gel}=1$, the total mass $\langle \mu_t, m \rangle$ is constant and equal to 1, and then it decreases: $\langle \mu_t,m \rangle = 1/t$ for $t\geq1$. Moreover, the second moment $\langle \mu_t,m^2 \rangle$ is finite before time $1$, and then infinite on $[1,+\infty)$.
It is also known in the literature that for any nonzero initial conditions, there is a gelation time $0 < T_{\rm gel} < + \infty$, such that there is a unique solution to \eqref{smolu2} on $[0,T_{\rm gel} )$, and $\langle \mu_t , m^2 \rangle \to + \infty$ when $t \to T_{\rm gel} ^-$: see e.g. \cite{FouLauWP}.
\begin{theorem}\label{thsmolu}
Let $\mu_0\in\mathcal{M}_c^+$ a non-null measure such that
\begin{equation}\label{condition}
\langle \mu_0 , m\wedge 1 \rangle = \int_{(0,+ \infty)} (m\wedge 1) \, \mu_0(\mathrm{d}m)<+\infty.
\end{equation}
We can then define
\[
M_0:=\langle \mu_0 , m \rangle \in (0,+ \infty], \qquad K := \langle \mu_0 , m^2 \rangle \in (0,+ \infty],
\]
and the function
\begin{equation}\label{g_0}
g_0(x):=\langle \mu_0, mx^m \rangle = \int_{(0,+ \infty)} mx^m \: \mu_0(\mathrm{d}m), \qquad x\in[0,1]
\end{equation}
with $g_0(1)=M_0\in(0,+\infty]$. Let
\begin{equation} \label{Tgel}
T_{\rm gel} := 1/K \in [0,+\infty).
\end{equation}
Then Smoluchowski's equation \eqref{smolu2} has a unique solution on ${\ensuremath{\mathbb R}} ^+$.
It has the following properties.
\begin{enumerate}
\item The total mass $M_t=\langle \mu_t, m \rangle$ is continuous on $[0,+\infty)$. It is constant on $[0,T_{\rm gel}]$ and strictly decreasing on $[T_{\rm gel},+ \infty)$. It is analytic on ${\ensuremath{\mathbb R}} ^+ \backslash \{ T_{\rm gel} \}$.
\item If the following limit exists
\[
\nu:= - \lim_{x \to 1^-} \frac{\left(g_0'(x)\right)^3}{g_0'(x)+xg_0''(x)} \in[-\infty,0],
\]
then the right derivative $\dot{M}_{T_{\rm gel}}$ of $M$ at $t=T_{\rm gel}$ is equal to $\nu$.
\item Let $m_0 = \inf \mathrm{supp}\ \mu_0\in[0,+\infty)$. When $t \to + \infty$,
\[
\frac{1}{tM_t} \to m_{0}.
\]
\item The second moment $\langle \mu_t, m^2 \rangle$ is finite for $t\in[0,T_{\rm gel})$ and infinite for $t\in[T_{\rm gel},+\infty)$.
\end{enumerate}
\end{theorem}
\begin{rem}{\rm
\begin{itemize}
\item This result allows to recover the pre- and post-gelation formulas obtained with no rigorous proof in some earlier papers \cite{ErnstCP,ErnstKoG,Kokholm,LeyvTschu,ShirRoesSR,ShirRoesPGM,StellZiff}. The decrease of the mass in $1/t$ when $m_0 > 0$ was also observed in these papers. Also, some upper bounds in $1/t$ for the mass were proven in \cite{EscobedoGiC,LauCCC}.
\item If $m=0$, the mass tends to $0$ more slowly than $1/t$: small particles need to coagulate before any big particle can appear, and they coagulate really slowly. For instance, a straightforward computation shows that if $\mu_0(\mathrm{d}m)=e^{-m} \mathrm{d}m$, then $M_t \sim t^{-2/3}$. More generally, the explicit formula in Proposition \eqref{uniuni} allows to compute $M_t$ for any initial conditions.
\item With this formula, it is easy to check that $\dot{M}_{T_{\rm gel}+}$ can be anything from $- \infty$ to $0$. For instance, $\dot{M}_{T_{\rm gel}+}=0$ for $g_0(x)=(1-x) \log(1-x)+x$, $\dot{M}_0=- \infty$ for $g_0(x)=\sqrt{1-x} \log(1-x)+x$, and for $0 < \a < + \infty$, $\dot{M}_0=- \a$ for $g_0(x)=1 - \sqrt{1-x^{2 \a}}$. In particular, $M$ need not be convex on $[T_{\rm gel},+\infty)$.
\end{itemize}
}
\end{rem}
\subsection{Preliminaries}
Let $\mu_0$ be defined as in the previous statement.
We shall prove that, starting from $\mu_0$, there is a unique solution to \eqref{smolu2} on ${\ensuremath{\mathbb R}} ^+$, and give a representation formula for this solution. This allows to study the behavior of the moments. Let us start with some easy lemmas. So take a solution $(\mu_t)_{t\geq 0}$ to \eqref{smolu2} and set
\begin{equation}
M_t=\langle \mu_t , m \rangle.
\label{Mt}
\end{equation}
The two following lemmas are easy to prove, using monotone and dominated convergence.
\begin{lemma}\label{lemM}
$(M_t)_{t\geq 0}$ is monotone non-increasing and right-conti\-nuous. Moreover, $M_t<+\infty$ for all $t>0$.
\end{lemma}
\begin{proof}
Take $\phi^K(m)=m$ for $m \in [0,K]$, $\phi^K(m)=2K - m$ for $m \in [K,2K]$, and $\phi^K(m)=0$ for $m \geq 2K$, so that $\phi^K \in C_c$. Plugging $\phi^K$ in Smoluchowski's equation \eqref{smolu2}, letting $K \to + \infty$ and using Fatou's lemma readily shows that $(M_t)_{t\geq 0}$ is monotone non-increasing. Note also that $t\mapsto M_t = \sup_K \langle \mu_t , \phi^K \rangle$ is the supremum of a sequence of continuous functions and so is lower semi-continuous, which implies, for a monotone non-increasing function, right-continuity. Finiteness of $M_t$
is now obvious since $s \mapsto M_s^2$, and hence $s \mapsto M_s$, are integrable by Definition \ref{defsolsmolu}.
\end{proof}
\begin{lemma}
Assume that $t \mapsto \langle \mu_t , m^2 \rangle$ is bounded on some interval $[0,T_0]$. Then $M_t = M_0$ for $t \in [0,T_0]$.
\label{lemmt}
\end{lemma}
By Lemma \ref{lemM}, $\langle \mu_t , m \rangle<+\infty$ for $t>0$, so that we can define
\begin{equation}\label{g}
g_t(x)=\langle \mu_t, mx^m \rangle = \int_{(0,+ \infty)} mx^m \: \mu_t(\mathrm{d}m), \qquad x\in[0,1], \ t>0,
\end{equation}
which is the generating function of $m\mu_t(\mathrm{d}m)$.
Then, using a standard approximation procedure, it is easy to see that $g$ satisfies
\begin{equation}
\left \{
\begin{array}{ll}
{\displaystyle g_t(x) =g_0(x)+\int_0^t x(g_s(x)-M_s)\dplus{g_s}{x}(x)\: \mathrm{d}s, \quad t \geq 0, \ x\in(0,1)} \\
g_t(1) = M_t, \quad t\geq 0. \rule{0cm}{0.5cm}
\end{array}
\right.
\label{PDE}
\end{equation}
It is well-known, and will be proven again below, that $M_t=M_0$ for all $t\leq T_{\rm gel}$, since then, the PDE \eqref{PDE}
can be solved by the method of characteristics: the function $\phi_t(x):[0,1]\mapsto[0,1]$
\[
\phi_t(x)=xe^{t (M_0-g_0(x))}, \qquad x\in [0,1], \, t\leq T_{\rm gel}
\]
is one-to-one and onto, has an inverse $h_t:[0,1]\mapsto[0,1]$ and we find
\[
g_t(x) = g_0(h_t(x)), \qquad x\in[0,1], \, t \leq T_{\rm gel}.
\]
However $M_t$ is not necessary constant for $t>T_{\rm gel}$ and the form
of $\phi_t$ has to be modified; we thus define
\begin{equation}
\phi_t(x)=x \a_t e^{-t g_0(x)}, \qquad x\in [0,1], \, t>0
\label{phit}
\end{equation}
where
\begin{equation}
\a_t:=\exp \left(\int_0^t M_s \: \mathrm{d}s\right), \qquad t\geq 0.
\label{alphat}
\end{equation}
For $t>T_{\rm gel}$, $M_t$ is possibly less than $M_0$ and $\phi_t$, which depends
explicitly on $(M_s)_{s\in[0,t]}$, is possibly neither injective nor surjective. We shall prove
that it is indeed possible to find $\ell_t\in(0,1)$ such that $\phi_t(x):[0,1]\mapsto[0,\ell_t]$ is one-to-one
and $\ell_t$ is uniquely determined by $g_0$.
\subsection{Uniqueness of solutions}
Using an adaptation of the method of characteristics,
we are going to prove the following result. Note that in \cite{ShirRoesSR}, this properties are claimed to be true but a proof seems to lack. We will use the same techniques in the proof of Theorem \ref{thsmolubras} for the model with arms, but they are easier to understand in the present case.
\begin{prop}\label{uniuni} Let $(\mu_t)_{t\geq 0}$ be a solution of Smoluchowski's equation
\eqref{smolu2}.
\begin{enumerate}
\item For all $t\in[0,T_{\rm gel}]$, $M_t=M_0=g_0(\ell_t)$, where $\ell_t:=1$. For all $t>T_{\rm gel}$,
$M_t=g_0(\ell_t)$ where $\ell_t\in(0,1)$ is uniquely defined by
\begin{equation}\label{ell}
\ell_t \, g'_0(\ell_t)= \frac 1t.
\end{equation}
Moreover $\ell_t$ and $\phi_t(\cdot)$ satisfy
\begin{equation}\label{uuuu}
\phi_t'(\ell_t)=0, \qquad
\phi_t(\ell_t)=1>\phi_t(x), \quad \forall \, x\in(0,1).
\end{equation}
\item For all $t>0$, the function $\phi_t(\cdot)$ defined in \eqref{phit} has a right inverse
\begin{equation}
h_t:[0,1]\mapsto[0,\ell_t], \qquad \phi_t(h_t(x))=x, \quad x\in[0,1],
\label{ht}
\end{equation}
and
\begin{equation}\label{uuu}
g_t(x) = g_0(h_t(x)), \qquad t>0, \ x\in[0,1].
\end{equation}
\item The functions $(\ell_t)_{t\geq 0}$ and $(M_t)_{t\geq 0}$ are continuous.
\item $(\mu_t)_{t\geq 0}$ is uniquely defined by $\mu_0$.
\end{enumerate}
\end{prop}
\begin{figure}[htb]
\begin{tabular}{cc}
\includegraphics[width=7.1cm]{phitpregel.png}& \includegraphics[width=7.1cm]{phitpostgel.png}
\end{tabular}
\caption{$\phi_t$ before and after gelation. The dotted lines represent what $\phi_t$ may look like. The solid one is the actual $\phi_t$.}
\end{figure}
\begin{rem}\label{see}{\rm
\begin{itemize}
\item For all $t\leq T_{\rm gel}$, $M_t=M_0$, $\ell_t=1$ and $\phi_t:[0,1]\mapsto[0,1]$ is
one-to-one and onto. The first thing one needs to prove
is that for all $t>T_{\rm gel}$, $\ell_t<1$, i.e. there is indeed $x\in[0,1]$
such that $\phi_t(x)=1$, see Lemma \ref{ti}; the second one, is that $\ell_t=m_t$, i.e. $\phi_t(\cdot)$ has an absolute maximum
at $\ell_t$, see Lemma \ref{lemlt2}. In other words, one has to exclude the dotted lines as possible profiles of
$\phi_t(\cdot)$ in Figure 1. These properties are not obvious, since
$\phi_t$ depends on $(M_s)_{s\in[0,t]}$ which is, at this point, unknown.
All other properties are derived from these two.
\item In \cite[Section 6]{ShirRoesSR} one finds a discussion
of post-gelation solutions, in particular of the results of our Proposition \ref{uniuni}.
However this discussion falls short of a complete proof, since the two above-mentioned properties
are not proven. In particular, no precise statement about
what initial conditions can be considered is given.
\end{itemize}
}
\end{rem}
\noindent
The following lemma is a list of obvious but useful properties satisfied by $g$ and $\phi$.
\begin{lemma}
\label{defsolpde}
The function $g$ defined in \eqref{g} satisfies the following properties.
\begin{itemize}
\item[(a1)] $(t,x) \mapsto g_t(x)$ is finite and continuous on $[0,+\infty)\times [0,1)$;
\item[(a2)] For all $x \in [0,1)$, $t \mapsto g_t(x)$ is right differentiable on $(0,+\infty)$;
\item[(a3)] For all $t \geq 0$, $x \mapsto g_t(x)$ is analytic on $(0,1)$ and monotone non-decreasing;
\item[(a4)] For all $t > 0$, $x \mapsto g_t(x)\in[0,+\infty]$ is continuous on $[0,1]$.
\end{itemize}
The function $\phi$ defined in \eqref{phit} satisfies the following properties.
\begin{itemize}
\item[(b1)] $\phi_t$ is continuous on $[0,1]$ and analytic on $(0,1)$;
\item[(b2)] $\phi_t(0)=0$, $\phi_t(1)=e^{-\int_0^t(M_0-M_s) \: \mathrm{d}s}\in[0,1]$;
\item[(b3)] $\phi_t'(x)=\a_t e^{-t g_0(x)} (1-txg_0'(x))$ for $x \in (0,1)$;
\item[(b4)] For $t \leq T_{\rm gel}$, $\phi_t$ is increasing. For $t > T_{\rm gel}$, $x \mapsto x g_0'(x)$ is increasing, $\phi_t'(0)>0$ and $\phi_t'(1) < 0$. In particular, for $t > T_{\rm gel}$, there is precisely one point $m_t \in (0,1)$ such that
\begin{equation}\label{mmm}
\phi_t'(m_t)=0.
\end{equation}
\item[(b5)] For $t > T_{\rm gel}$, $\phi_t$ is increasing on $[0,m_t]$ and decreasing on $[m_t,1]$.
\end{itemize}
Moreover,
\begin{itemize}
\item[(c1)] The map $(t,x) \mapsto \phi_t(x)$ is continuous on ${\ensuremath{\mathbb R}} ^+ \times [0,1)$;
\item[(c2)] The map $(t,x) \mapsto \phi_t'(x)$ is continuous on ${\ensuremath{\mathbb R}} ^+ \times (0,1)$;
\item[(c3)] For every $x \in [0,1)$, $t \mapsto \phi_t(x)$ is right differentiable and
\begin{equation}\label{phi}
\dplus{\phi_t}{t} = \phi_t(x) (M_t - g_0(x)) \qquad x\in[0,1), \, t\geq 0.
\end{equation}
\end{itemize}
\end{lemma}
Property (b5) implies that there are at most two points in $(0,1)$ where $\phi_t$ equals 1. Take $\ell_t$ to be the smallest, if any, i.e.
\begin{equation}\label{lll}
\ell_t = \inf \{ x \geq 0: \, \phi_t(x)=1 \} \qquad (\inf\emptyset:=1).
\end{equation}
\begin{lemma}\label{ti}
\begin{enumerate}
\item For every $t \geq 0$ and every $x \in [0,\ell_t]$
\begin{equation}
g_t ( \phi_t(x) ) = g_0(x).
\label{gtphit}
\end{equation}
\item For all $t \in [0,T_{\rm gel}]$, $\ell_t=1$, and for $t > T_{\rm gel}$, $0<\ell_t <1$. In particular,
for all $t>0$, $\phi_t(\ell_t)=1$ and
\begin{equation}
g_0(\ell_t)=g_t(1)=M_t.
\label{g0lt}
\end{equation}
\item Finally, $t\mapsto \ell_t$ is monotone non-increasing and continuous on ${\ensuremath{\mathbb R}} ^+$.
\end{enumerate}
\label{lemlt1}
\end{lemma}
\begin{proof} \begin{enumerate}
\item Let us first prove that there exists $\tau>0$ such that \eqref{gtphit} holds for $t\in[0,\tau[$. Fix $0<a<b<1$. Since $0 < \min_{[a,b]}\phi_0 < \max_{[a,b]}\phi_0 < 1$, then by property (c1) there is $\tau>0$ such that
\[
0 < \min_{[a,b]}\phi_t < \max_{[a,b]}\phi_t < 1, \qquad \forall\, t \in [0,\tau).
\]
So, for a fixed $x \in [a,b]$, the function
\[
u_t:=g_t(\phi_t(x))-g_0(x)
\]
is well-defined and using \eqref{PDE} and \eqref{phi}, we see that
\begin{align*}
u_t= & \int_0^t \left(\dplus{g_s}{s}(\phi_s(x)) + \d{g_s}{x} (\phi_s(x))
\dplus{\phi_s}{s}(x) \right)\: \mathrm{d}s = \int_0^t \gamma_s \, u_s \: \mathrm{d}s
\end{align*}
where
\[
\gamma_t := \d{g_t}{x} (\phi_t(x)) \, \phi_t(x), \qquad t> 0.
\]
Since $x\in[0,1)$, $\sup_{t\in[0,\tau)} |\gamma_t|<+\infty$ and therefore $u_t\equiv 0$. Hence \eqref{gtphit} holds for $x\in[a,b]$ and $t\in[0,\tau[$.
Since both terms of \eqref{gtphit} are analytic functions of $x$ on $(0,\ell_t)$, by analytic continuation, \eqref{gtphit} actually holds on $(0,\ell_t)$, and hence on $[0,\ell_t]$ by continuity.
\item Let us now extend this formula to $t\in {\ensuremath{\mathbb R}} ^+$. Let
\[
T= \sup \{t > 0: \ \forall \, s \in [0,t], \ \forall \, x \in [0,\ell_s], \ g_s(\phi_s(x))=g_0(x) \} \geq \tau > 0,
\]
assume $T < + \infty$, and denote by $\ell$ the left limit of $(\ell_t)_{t\geq 0}$ at $T$. First, $\ell$ cannot be $0$, since otherwise we would get when $s \to T^-$
\[
1 = \phi_s(\ell_s)=\ell_s \a_s e^{-s g_0(\ell_s)} \to 0.
\]
For every $t < T^-$, $0< \ell \leq \ell_t$, so for every $x \in (0, \ell)$, $g_t(\phi_t(x))=g_0(x)$ and $\phi_t(x) < 1$. Using the continuity property (c1) and passing to the limit when $t \to T^-$ in this equality, we get
\[
g_T(\phi_T(x))=g_0(x), \qquad \forall \, x \in (0, \ell).
\]
By the same reasoning as in point (i), we obtain a $T' > T$ such that $g_t(\phi_t(x))=g_0(x)$ for all $t \in [T,T')$ and $x$ in a non-empty open subset of $(0,\ell)$. By analyticity and continuity, the formula $g_t(\phi_t(x))=g_0(x)$ holds for every $t \in [T,T')$ and $x \in [0,\ell_t]$. This contradicts the definition of $T$, and so $T = + \infty$. This concludes the proof of point (1) of the Lemma.
\item For the statement (2) of the Lemma, let us show first that $\langle \mu_t , m^2 \rangle$ is bounded on $[0,T_0)$, for every $T_0 \in [0,T_{\rm gel})$. Let $T'$, the smallest time when this fails (provided of course that $T_{\rm gel} > 0$). By assumption (see Definition \ref{defsolsmolu}), $T' > 0$. Differentiating \eqref{gtphit} with respect to $x$ and having $x$ tend to $\ell_t=1$ gives, for $t < T'$,
\[
g_t'(1)= \langle \mu_t , m^2 \rangle = \frac{1}{1-tK}.
\]
This quantity explodes only when $t = T_{\rm gel}=1/K$, so $T' = T_{\rm gel}$.
\item The boundedness of $(\langle \mu_t , m^2 \rangle)_{t\in[0,T_0)}$ just proven for all $T_0 \in [0,T_{\rm gel})$ and Lemma \ref{lemmt} imply that for $t \in [0,T_{\rm gel})$, $M_t = M_0$. By
the definition \eqref{phit} of $\phi_t$, it follows that $\phi_t(1)=1$ for $t \in [0,T_{\rm gel})$.
But $\phi_t$ is increasing, so $\ell_t = 1$ for $t \in [0,T_{\rm gel})$. Assume now that for some $t > T_{\rm gel}$, $\ell_t = 1$. Then \eqref{gtphit} holds on $[0,1]$, and this is impossible since the right term is an increasing function of $x$, whereas the left one decreases in a left neighborhood of 1 since $\phi'_t(1)<0$. The fact that $\phi_t(\ell_t)=1$ follows then directly from the definition of $\ell_t$ and the continuity of $\phi_t(\cdot)$. Finally, the inequality $\ell_t > 0$ is obvious since $\phi_t(0)=0$, and computing \eqref{gtphit} at $x=\ell_t$ gives \eqref{g0lt}. This concludes the proof of (2).
\item We know that $\ell_t=1$ and $M_t=M_0$ for all $t< T_{\rm gel}$. Now, $g_0$ is strictly increasing and continuous. Since
$(M_t)_{t\geq 0}$ is monotone non-increasing and right-continuous by Lemma \ref{lemM}, so is $(\ell_t)_{t\geq 0}$ by \eqref{g0lt}. To get left-continuity of $(\ell_t)_{t>T_{\rm gel}}$, consider $t > T_{\rm gel}$, and let $\ell$ be the left limit of $\ell_s$ at $t$. We have $\ell \leq \ell_{t+(t-T_{\rm gel})/2} < 1$, so by the continuity property (c1) above,
\[
1 = \phi_s(\ell_s) \underset{s \to t^-}{\to} \phi_t(\ell).
\]
Hence $\phi_t(\ell)=1$. Assume $\ell > \ell_t$ (that is, $\ell$ is the second point where $\phi_t$ reaches $1$). Take $x \in (\ell_t,\ell)$. By property (b5), $\phi_t(x) > 1$. But on the other hand, $x < \ell \leq \ell_s$ for $s <t$, so $\phi_s(x) \leq 1$, and so $\phi_t(x) \leq 1$, and this is a contradiction. So $\ell=\ell_t$ and $(\ell_t)_{t\geq 0}$ is indeed continuous. This concludes the proof of (3) and of the Lemma.
\end{enumerate}
\end{proof}
Finally, we will see that for $t > T_{\rm gel}$, $\ell_t=m_t$, so that $\phi_t$ increases from 0 to 1, which is its maximum, and then decreases. To this end, recall that $(\ell_t)_{t\geq 0}$ is monotone non-increasing and that $(\ell_t)_{t\geq 0}$ and $(\phi_t)_{t\geq 0}$ are continuous, so the chain rule for Stieltjes integrals and \eqref{phi} give
\begin{align*}
1 = \phi_t(\ell_t) & = \phi_0(\ell_0) + \int_0^t \phi_s'(\ell_s) \: \mathrm{d}\ell_s + \int_0^t \dplus{\phi_s}{s}(\ell_s) \: \mathrm{d}s \\
& = 1 + \int_0^t \phi_s'(\ell_s) \: \mathrm{d}\ell_s + \int_0^t \phi_s(\ell_s)(M_s-g_0(\ell_s)) \: \mathrm{d}s
\end{align*}
that is, with \eqref{g0lt},
\begin{equation}
\phi_t'(\ell_t) \: \mathrm{d}\ell_t = 0.
\label{phitprimlt}
\end{equation}
Hence, $\: \mathrm{d}\ell_t$-a.e $\phi_t'(\ell_t)=0$, i.e. $\ell_t=m_t$. This is actually true for all $t > T_{\rm gel}$, as we shall now prove. This result also has its counterpart in the model with arms, namely part 3 of the proof of Theorem \ref{thsmolubras}.
\begin{lemma}\label{lemlt2}
For every $t > T_{\rm gel}$, $\phi_t'(\ell_t)=0$, i.e. $\ell_t=m_t$, the point where $\phi_t$ attains its maximum.
In particular,
\begin{equation}
\ell_t \, g_0'(\ell_t)=\frac1t, \qquad \forall \, t>T_{\rm gel}.
\label{deflt1}
\end{equation}
\end{lemma}
\begin{proof}
First, recall that $\phi_t$ is increasing on $[0,\ell_t]$, so that $\phi_t'(\ell_t) \geq 0$, that is
\begin{equation}
\ell_t g_0'(\ell_t) \leq \frac1t.
\label{ineg}
\end{equation}
Assume now that there is a $t > T_{\rm gel}$ such that $\phi_t'(\ell_t)>0$, and consider
\[
s = \sup \{ r \in (T_{\rm gel},t) : \, \phi_r'(\ell_r) = 0 \}.
\]
As noted before, $t \mapsto \ell_t$ is strictly decreasing for $t > T_{\rm gel}$ for any $t > T_{\rm gel}$, so $\: \mathrm{d}\ell_t([T_{\rm gel},T_{\rm gel}+\varepsilon[)>0$ for all $\varepsilon>0$. Hence there are points $r<t$ where $\phi_r'(\ell_r)=0$, and thus the definition of $s$ does make sense.
Take now $(r_n)$ a sequence of points such that $T < r_n < t$, $\phi_r'(\ell_r)=0$ and $(r_n)$ converges to $s$. Since $0 < \ell_s < 1$, by property (c2) above, we get
\[
0 = \phi'_{r_n} (\ell_{r_n}) \to \phi'_s(\ell_s)
\]
so that $\phi'_s(\ell_s)=0$. This shows that $s < t$, and that for $r\in(s,t)$, $\phi'_r(\ell_r) >0$. Hence, by continuity of $(\ell_r)_{r\geq 0}$ and by \eqref{phitprimlt}, $(\ell_r)_{r\in[s,t]}$ is constant. This gives
\[
\frac1s = \ell_s \, g_0'(\ell_s) = \ell_t \, g_0'(\ell_t) \leq \frac1t
\]
which is a contradiction since $s< t$. In particular, $\phi'_t(\ell_t)=0$ implies \eqref{deflt1}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{uniuni}]
By Lemma \ref{lemlt2}, necessarily $M_t = M_0$ on $[0,T_{\rm gel}]$ and for $t> T_{\rm gel}$, $M_t := g_t(1) = g_0(\ell_t)$, where
\begin{equation}
\ell_tg_0'(\ell_t)=\frac1t.
\label{deflt}
\end{equation}
Since $x \mapsto x g_0'(x)$ is strictly increasing from $[0,1]$ to $[0,K]$, where $K=\langle \mu_0 , m^2 \rangle =1/T_{\rm gel}$, this equation has a unique solution for $t > T_{\rm gel}$. Hence $M_t$ is uniquely defined.
Therefore $\alpha_t$ and $\phi_t$ are uniquely determined by $g_0$,
so we can define $\phi_t$ as in \eqref{phit}, and
Lemma \ref{lemlt1} shows that $g_t(\phi_t(x))=g_0(x)$ for $x \in [0,\ell_t]$, and that $\phi_t$ is a bijection from $[0,\ell_t]$ to $[0,1]$. So it has a right inverse $h_t$, and compounding by $h_t$ in the previous formula gives
\begin{equation}
g_t(x)=g_0(h_t(x))
\label{gt}
\end{equation}
for all $x \in [0,1]$, $t \geq 0$. Thus $g_t$ can be expressed by a formula involving only $g_0$ and in particular, $(\mu_t)_{t\geq 0}$ depends only on $\mu_0$. This shows the uniqueness of a solution to Smoluchowski's equation \eqref{smolu2}.
\end{proof}
\subsection{Behavior of the moments}
In this paragraph, we will study the behavior of the first and second moment of $(\mu_t)_{t\geq 0}$ as time passes,
showing how to prove rigorously and recover the results of \cite{ErnstCP}. For more general
coagulation rates, one can obtain upper bounds of the same nature, see \cite{LauCCC}.
First consider the mass $M_t = \langle \mu_t, m \rangle$. We will always assume that $T_{\rm gel} < + \infty$. Let us start with the following lemma.
\begin{lemma}
Let $\nu \in \mathcal{M}_c^+$ be a measure which integrates $x \mapsto y^x$ for small enough $y > 0$. Let $m_0$ be the infimum of its support. Then
\[
\lim_{y \to 0^+} \frac{\langle \nu, x y^x \rangle}{\langle \nu, y^x \rangle} = m_0.
\]
\label{lemprob}
\end{lemma}
\begin{proof}
First, note that $x y^x \geq m y^x$ $\nu$-a.e. so
\[
\liminf_{y \to 0} \frac{\langle \nu, x y^x \rangle}{\langle \nu, y^x \rangle} \geq m_0.
\]
Let us prove now that
\[
\limsup_{y \to 0} \frac{\langle \nu, x y^x \rangle}{\langle \nu, y^x \rangle} \leq m_0.
\]
Assume this is not true. Then, up to extraction of a subsequence, we may assume that there exists $\a > 0$ such that for arbitrary small $y \in (0,1)$, $\langle \nu, x y^x \rangle \geq (m_0+ \a) \langle \nu, y^x \rangle$. Hence $\langle \nu, (x-m_0 - \a)y^x \rangle \geq 0$, so
\begin{equation}
\langle \nu, (x-m_0 - \a) y^x \un{\{x > m_0+\a\}} \rangle \geq \langle \nu, (m_0+\a-x) y^x \un{\{m_0 \leq x \leq m_0+\a\}} \rangle.
\label{eqlem}
\end{equation}
But
\begin{align*}
\langle \nu, (m_0+\a-x) y^x \un{\{m_0 \leq x \leq m_0+\a\}} \rangle & \geq \langle \nu, (m_0+\a-x) y^x \un{\{m_0 \leq x \leq m_0+\a/2\}} \rangle \\
& \geq \langle \nu, (m_0+\a-x) \un{\{m_0 \leq x \leq m_0+\a/2\}} \rangle y^{m_0 + \a/2}
\end{align*}
and
\[
\langle \nu, (x-m_0 - \a) y^x \un{\{x > m_0+\a\}} \rangle \leq \langle \nu, (x-m_0 - \a) \un{\{x > m_0+\a\}} \rangle y^{m_0 + \a}.
\]
With \eqref{eqlem}, this shows that
\[
\langle \nu, (x-m_0 - \a) \un{\{x > m_0+\a\}} \rangle y^{\a/2} \geq \langle \nu, (m_0+\a-x) \un{\{m_0 \leq x \leq m_0+\a/2\}} \rangle
\]
and having $y$ tend to zero gives
\[
0 \geq \langle \nu, (m_0+\a-x) \un{\{m_0 \leq x \leq m_0+\a/2\}} \rangle
\]
which is a contradiction since $\nu([m_0,m_0+ \a/2]) > 0$.
\end{proof}
\begin{cor}
The mass of the system is continuous and positive. It is strictly decreasing on $[T_{\rm gel},+ \infty)$. Moreover, denote $m_0 = \inf \mathrm{supp}\ \mu_0$. Then
\[
\lim_{t \to + \infty}\frac{1}{tM_t} = m_0.
\]
\end{cor}
\begin{proof}
Recall that $M_t=g_0(\ell_t)$ so the first properties follow from Lemma \ref{ti}. Denote now $\nu(\mathrm{d}m)=m\mu_0(\mathrm{d}m)$. For $t > T_{\rm gel}$, $\ell_tg_0'(\ell_t)=1/t$, so
\[
\frac{1}{tM_t}=\frac{\langle \nu, x\ell_t^x \rangle}{\langle \nu, \ell_t^x \rangle}
\]
and since $\ell_t \to 0$ when $t \to + \infty$, this tends to $m_0$ by Lemma \ref{lemprob}.
\end{proof}
We can also study the behavior of the mass for small times. Recall that before gelation, the mass is constant at $1$. We have seen that it is continuous at the gelation time. We may then wonder if its derivative is continuous, that is if $\dot{M}_{T_{\rm gel}+}$ is zero or not.
\begin{lemma}
The right derivative of $M$ at $T_{\rm gel}$ is given by
\[
\dot{M}_{T_{\rm gel}+}= - \lim_{x \to 1^-} \frac{g_0'(x)^3}{g_0'(x)+xg_0''(x)} \in [- \infty,0]
\]
provided the limit exists.
\label{Mpoint}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{Mpoint}]
For $t > T_{\rm gel}$, $f(\ell_t)=1/t$ with $f(x)=xg_0'(x)$, and $0 < \ell_t < 1$. But $f'(\ell_t) \neq 0$, so by the inverse mapping theorem, $(\ell_t)_{t\geq 0}$ is differentiable and
\[
\dot{\ell}_t=- \frac{1}{t^2 f'(\ell_t)}.
\]
Using the fact that $M_t = g_0(\ell_t)$, it is then easy to see that
\[
\dot{M}_t=-\ell_t^2 \frac{g_0'(\ell_t)^3}{g_0'(\ell_t)+\ell_tg_0''(\ell_t)}.
\]
Since $(\ell_t)_{t\geq 0}$ is continuous at $T_{\rm gel}$ and $\ell_{T_{\rm gel}}=1$, the result follows.
\end{proof}
Recall that the gelation time is precisely the first time when the second moment $\langle \mu_t, m^2 \rangle$ of $(\mu_t)_{t\geq 0}$ becomes infinite. It actually remains infinite afterwards.
\begin{cor}
For all $t \geq T_{\rm gel}$, $\langle \mu_t, m^2 \rangle = + \infty$.
\end{cor}
\begin{proof}
Note that
\[
\langle \mu_t , m^2 \rangle = g_t'(1),
\]
this formula being understood as a monotone limit. By \eqref{gtphit}, for $x < \ell_t$
\[
\phi_t'(x) g_t'(\phi_t(x))=g_0'(x).
\]
When $x \to \ell_t^-$, $\phi_t'(x) \to 0$ by Lemma \ref{lemlt2}, and $g_0'(x) \to g_0'(\ell_t) \neq 0$ since $\ell_t > 0$. So
\[
g_t'(\phi_t(\ell_t))=g_t'(1)=+ \infty.
\]
\end{proof}
\subsection{Existence of solutions}\label{exist}
Existence of solutions of \eqref{smolu2} is a well-known topic, see e.g. \cite{FouLauMLP}. However, the case $M_0=+\infty$ is apparently new, so that
we give a short proof for the general case based on previous papers, mainly \cite{ShirRoesSR}.
Let now $\mu_0\in\mathcal{M}_f^+$ be as in the statement
of Theorem \ref{thsmolu} and let us set $g_0$ as in \eqref{g_0}, $\ell_t$ and
$M_t$ as in point (1) of Proposition \ref{uniuni}, $\alpha_t$ and $\phi_t$ as in
\eqref{alphat} and \eqref{phit}. Then it is easy to see that $\phi_t$ admits a right
inverse $h_t$ satisfying \eqref{ht}, and we can thus define
\[
g_t(x):= g_0(h_t(x)), \qquad t\geq 0, \ x\in[0,1].
\]
It is an easy but tedious task to check that $g_t$ satisfies \eqref{PDE} and all properties
(a1)-(a4) above.
In particular,
if $g_0(1)=+\infty$ then $h_t(1)<1$ and therefore
$g_t(1)<+\infty$ for all $t>0$.
Following \cite{ShirRoesSR}, we can now prove the following.
\begin{prop}\label{shir}
For all $t>0$ there exists $\mu_t\in\mathcal{M}_f^+$ such that
\[
g_t(x) =\langle \mu_t, mx^m \rangle = \int_{(0,+ \infty)} mx^m \: \mu_t(\mathrm{d}m), \qquad x\in[0,1].
\]
\end{prop}
\begin{proof}
Let $t>0$ be fixed. We set for all $y\geq 0$
\[
\Phi(y) := g_0(e^{-y}), \quad
\Gamma(y) := tg_0(e^{-y}), \quad
G(y) := \Gamma(y)+y-\log\alpha_t= -\log\phi_t(e^{-y}).
\]
We recall
that $f:[0,+\infty)\mapsto[0,+\infty)$ is completely monotone if
$f$ is continuous on $[0,+\infty)$, infinitely many times differentiable on $(0,+\infty)$ and
\[
(-1)^k \frac{d^kf}{dy^k} (y) \geq 0,
\qquad \forall \, k\geq 0, \, y\in(0,+\infty).
\]
It is easy to see that $\Phi$ and $\Gamma$ are completely monotone. Moreover, $G$ has a right inverse
\[
G^{-1}:[0,+\infty)\mapsto[\log(1/\ell_t),+\infty), \quad G^{-1}(y)=-\log h_t(e^{-y}), \qquad y\geq 0,
\]
and therefore by the definitions
\[
g_0(h_t(e^{-y}))= \Phi(G^{-1}(y)), \qquad y\geq 0.
\]
By \cite[Thm. 3.2]{ShirRoesSR}, $\Phi\circ G^{-1}$ is completely monotone and therefore, by Bernstein's Theorem,
there exists a unique $\nu_t\in\mathcal{M}_f^+$ such that
\[
g_t(e^{-y})=g_0(h_t(e^{-y}))= \Phi(G^{-1}(y))=\int_{(0,+ \infty)} e^{-ym} \, \nu_t(\mathrm{d}m), \qquad y\geq 0.
\]
Since $g_t(1)<+\infty$ for all $t>0$, we obtain that $\langle \nu_t,m\rangle<+\infty$, so that
we can set $\mu_t(\mathrm{d}m):=m\, \nu_t(\mathrm{d}m)$, and we have found that there is a unique $\mu_t\in\mathcal{M}_f^+$ such that
\[
g_t(x)=g_0(h_t(x))= \int_{(0,+ \infty)} x^{m} \, m \, \mu_t(\mathrm{d}m), \qquad x\in(0,1].
\]
\end{proof}
In order to show that $(\mu_t)_{t\geq 0}$ is a solution of Smoluchowski's equation in the sense
of Definition \ref{defsolsmolu}, we have to check
that $\int_0^\varepsilon M_t^2 \, \mathrm{d}t<+\infty$ for all $\varepsilon>0$. This is the content of the next result.
\begin{lemma}\label{gutt}
If $(\mu_t)_{t\geq 0}$ is the family constructed in Proposition \ref{shir}, then for all $\varepsilon>0$,
$\int_0^\varepsilon \langle \mu_s, m\rangle^2 \: \mathrm{d}s< + \infty$.
\end{lemma}
\begin{proof}
If $M_0<+\infty$ then there is nothing to prove, since $(M_t)_{t\geq 0}$ is monotone non-increasing,
so let us consider the case $M_0=+\infty$ and thus $T_{\rm gel}=0$.
Since $M_t=g_0(\ell_t)$ is bounded and continuous for $t\in[\delta,\varepsilon]$ for all $\delta\in(0,\varepsilon)$,
we have by \eqref{ell} and \eqref{g_0}
\[
\begin{split}
\int_\delta^\varepsilon M_t^2\, \mathrm{d}t & =\int_\delta^\varepsilon g_0^2(\ell_t)\, \mathrm{d}t =
\varepsilon g_0^2(\ell_\varepsilon) -\delta g_0^2(\ell_\delta) - \int_\delta^\varepsilon 2tg_0(\ell_t)g_0'(\ell_t) \, \mathrm{d} \ell_t \\
& \leq \varepsilon g_0^2(\ell_\varepsilon) - \int_\delta^\varepsilon 2\, g_0(\ell_t)\, \frac{\mathrm{d} \ell_t}{\ell_t} =
\varepsilon g_0(\ell_\varepsilon) + 2\int_{\ell_\varepsilon}^{\ell_\delta} g_0(y)\, \frac{\mathrm{d} y}y \\
& \leq \varepsilon g_0(\ell_\varepsilon) +\frac{2}{\ell_\varepsilon} \, \langle \mu_0, \frac m{1+m}\rangle
\leq \varepsilon g_0(\ell_\varepsilon) +\frac{2}{\ell_\varepsilon} \, \langle \mu_0, m\wedge 1\rangle.
\end{split}
\]
Letting $\delta\downarrow 0$, by \eqref{condition} we obtain the desired result.
\end{proof}
We now finish the proof of existence of a solution by showing that $(\mu_t)_{t\geq 0}$
indeed solves \eqref{smolu2}. By choosing $x=e^{-y}$, $y\geq 0$, in \eqref{PDE}, we find
an equality between Laplace transforms. Since the Laplace transform is one-to-one,
then we obtain \eqref{smolu2}.
\begin{rem}
{\rm In the proof of uniqueness, we may only require that $\langle \mu_0, m y^m \rangle < + \infty$ for every $y \in [0,1)$. However, the same kind of computation as in Lemma \ref{gutt} shows that if this the case, but $\langle \mu_0, m \wedge 1 \rangle = + \infty$, then $\int_0^t M_s^2 \: \mathrm{d}s = + \infty$ for all $t > 0$, in contradiction with Definition \ref{defsolsmolu} of a solution.}
\end{rem}
\section{Flory's equation}
We will now consider the modified version of Smoluchowski's equation,
also known as {\it Flory's equation}, with a multiplicative kernel.
\begin{definition}\label{def2.1}
Let $\mu_0\in\mathcal{M}_c^+$.
We say that a family $(\mu_t)_{t\geq 0}\subset \mathcal{M}_c^+$ solves Flory's equation \eqref{smolu2} if
\begin{itemize}
\item for every $t>0$, $\int_0^t \langle \mu_s(\mathrm{d}m), m\rangle^2 \: \mathrm{d}s< + \infty$,
\item for all $\phi \in C_c(0,\pinf)$ and $t>0$
\begin{equation}\label{flory}
\begin{split}
\langle \mu_t , \phi \rangle & = \langle \mu_0 , \phi \rangle + \frac12 \int_0^t \langle \mu_s(\mathrm{d}m) \mu_s(\mathrm{d}m'), mm' \Delta \phi(m,m') \rangle \: \mathrm{d}s \\
& \quad - \int_0^t \langle \mu_s, \phi \rangle \langle \mu_0(\mathrm{d}m) - \mu_s(\mathrm{d}m), m \rangle \: \mathrm{d}s,
\end{split}
\end{equation}
\item if $\langle \mu_0 , m^2 \rangle < + \infty$, then $t \mapsto \langle \mu_t , m^2 \rangle$ is bounded in a right neighbor\-hood of 0.
\end{itemize}
\end{definition}
In equation \eqref{flory}, the mass that vanishes in the gel interacts with the other particles. It is a
modified Smoluchowski's equation, where a term has been added, representing
the interaction of the particles of mass $m$ with the gel, whose mass is
\[
\langle \mu_0 - \mu_s , m \rangle
\]
i.e. precisely the missing mass of the system. Notice that in this case the equation
makes sense only if $\langle \mu_0 , m \rangle<+\infty$.
The mass is expected to decrease faster in this case than for \eqref{smolu2}. This is actually true, as we can see in the following result.
%
\begin{theorem}\label{thflory}
Let $\mu_0\in\mathcal{M}_c^+$ a non-null measure such that
$\langle \mu_0 , m \rangle <+\infty$, and set
\[
M_0:=\langle \mu_0 , m \rangle \in (0,+ \infty), \qquad
K := \langle \mu_0 , m^2 \rangle \in (0,+ \infty].
\]
Let $T_{\rm gel} := 1/K \in [0,+\infty)$.
Then Flory's equation \eqref{flory}
has a unique solution $(\mu_t)_{t\geq 0}$ on ${\ensuremath{\mathbb R}} ^+$. It has the following properties.
\begin{enumerate}
\item We have $M_t=g_0(l_t)$, where $l_t=1$ for $t\leq T_{\rm gel}$ and, for $t> T_{\rm gel}$, $l_t$
is uniquely defined by
\[
l_t=e^{-t(M_0-g_0(l_t))}, \qquad l_t\in[0,1).
\]
Therefore $t\mapsto M_t$ is continuous on $[0,+\infty)$, constant on $[0,T_{\rm gel}]$, strictly decreasing on $[T_{\rm gel},+ \infty)$ and analytic on ${\ensuremath{\mathbb R}} ^+ \backslash \{ T_{\rm gel} \}$.
\item The function $\phi_t(x) = xe^{t(M_0-g_0(x))}$ has a right inverse $h_t \, : \, [0,1] \to [0,l_t]$. The generating function $g_t$ of $(\mu_t)_{t\geq 0}$ is given for $t \geq 0$ by
\[
g_t(x)=g_0(h_t(x)).
\]
\item Let $m_0 = \inf \mathrm{supp}\ \mu_0\geq 0$. Then, when $t \to + \infty$,
\[
M_t e^{m_0t} \to m_0 \mu_0(\{m_0\})
\]
and for every $\eps > 0$
\[
M_t e^{(m_0+\eps)t} \to + \infty.
\]
\item The second moment $\langle m^2,c_t \rangle$ is finite on ${\ensuremath{\mathbb R}} ^+ \backslash \{ T_{\rm gel} \}$ and infinite at $T_{\rm gel}$.
\end{enumerate}
\end{theorem}
\begin{rem}{\rm
\begin{itemize}
\item Norris \cite[Thm 2.8]{NorrisCC} has a proof of global uniqueness of Flory's equation \eqref{flory} for slightly less general initial
conditions ($\mu_0$ such that $\langle\mu_0,1+m\rangle<+\infty$), but for a much more general model.
\item When $m_0 > 0$, it was already observed (Proposition 5.3 in \cite{EscobedoGaMC}) that the mass decays (at least) exponentially fast (see also \cite{ErnstKoG,StellZiff,ZiffKoP}).
\end{itemize}
}
\end{rem}
\begin{proof}[Proof of Theorem \ref{thflory}] The proof is very similar to (and actually easier than) that of
Theorem \ref{thsmolu}.
\begin{enumerate}
\item Arguing as in the proof of Lemma \ref{lemM}, we obtain easily that $(M_t)_{t\geq 0}$ is
monotone non-increasing and right-continuous. As in Lemma \ref{lemmt}, if
$t \mapsto \langle \mu_t , m^2 \rangle$ is bounded on some interval $[0,T_0]$, then $M_t = M_0$ for $t \in [0,T_0]$ and
therefore $(\mu_t)_{t\geq 0}$ is a solution of Smoluchowski's equation \eqref{smolu2} on $[0,T_0]$.
\item Consider initial concentrations $\mu_0$ as in the statement, a solution $(\mu_t)_{t\geq 0}$ to Flory's equation and $g_t(x)$,
$x\in[0,1]$, generating function of $m\, \mu_t(dm)$. Then $g_t$ solves the PDE
\begin{equation}
\d{g_t}{t}=x(g_t-M_0)\d{g_t}{x}, \qquad \forall \, t>0, \, x\in[0,1],
\label{PDE2}
\end{equation}
the same as the one obtained for Smoluchowski's equation before gelation. It may be solved using the method of characteristics. Indeed, the mapping
\begin{equation}\label{phi2}
\phi_t(x)=x e^{t(M_0-g_0(x))}=x+\int_0^t (M_0-g_0(x))\, \phi_s(x) \: \mathrm{d}s,
\end{equation}
has the following properties
\begin{itemize}
\item[(d1)] $\phi_t(0)=0$, $\phi_t(1)=1$.
\item[(d2)] For all $t\geq 0$, $\phi_t'(x)=e^{t(M_0-g_0(x))}(1-txg'_0(x))$.
\item[(d3)] For $t \leq T_{\rm gel}$, $\phi_t(\cdot)$ is increasing; therefore,
$\phi_t(x)\in[0,1]$ for all $x\in[0,1]$ and $\phi_t(x)=1$ if and only if $x=1$
\item[(d4)] For $t > T_{\rm gel}$, $\phi_t(\cdot)$ is increasing on $[0,m_t]$ and decreasing on $[m_t,1]$, where $m_t$
is the unique $x\in(0,1)$ such that $\phi'_t(x)=0$, i.e. such that $txg_0'(x)=1$.
\item[(d5)] For $t>T_{\rm gel}$, $\phi_t(m_t)>1$, since $\phi_t(1)=1$ and $\phi_t'(1)<0$.
Therefore there is a unique $l_t\in(0,m_t)$ such that $\phi_t(l_t)=1$.
\item[(d6)] For $t > T_{\rm gel}$, $\phi_t'(l_t) \neq 0$, since $l_t<m_t$.
\end{itemize}
\begin{figure}[htb]
\begin{tabular}{cc}
\includegraphics[width=7.1cm]{phitpregel.png}& \includegraphics[width=7.1cm]{phitpostgelfl.png}
\end{tabular}
\caption{$\phi_t$ before and after gelation.}
\end{figure}
Setting $l_t:=1$ for $t\leq T_{\rm gel}$,
$\phi_t$ is thus a continuous bijection from $[0,l_t]$ to $[0,1]$, with continuous inverse function $h_t:[0,1]\mapsto
[0,l_t]$. By using \eqref{PDE2} and
\eqref{phi2} and arguing as in part (i) and (ii) of the proof of Lemma \ref{ti}, we can see that the function $u_t(x):=g_t(\phi_t(x))-g_0(x)$
satisfies $u_t(x)=u_0(x)=0$ for all $t\geq 0$ and $x\in[0,l_t]$. Therefore
the only solution of the PDE \eqref{PDE2} is given by
\begin{equation}
g_t(x)=g_0(h_t(x)), \qquad t\geq 0, \, x\in[0,1].
\label{gt2}
\end{equation}
Flory's equation has thus a unique solution on ${\ensuremath{\mathbb R}} ^+$, and its generating function is $g_t$.
\item We have seen in (d5) above that, for $t>T_{\rm gel}$, there is a unique
$l_t\in[0,1)$ such that $\phi_t(l_t)=1$. The relation $\phi_t(l_t)=1$ with $l_t\in[0,1)$ is equivalent
to $l_t=e^{-t(M_0-g_0(l_t))}$ with $l_t\in[0,1)$. This relation implies that $t\mapsto l_t$ is
analytic for $t>T_{\rm gel}$. A differentiation shows that
\[
\frac{dl_t}{dt} = -\frac{(M_0-g_0(l_t))l_t}{1-tg_0'(l_t)l_t}<0, \qquad t>T_{\rm gel},
\]
since $g_0'(l_t)l_t<g_0'(m_t)m_t=1/t$ and $g_0(l_t)<g_0(1)=M_0$. Let $\ell$ be the
limit of $l_t$ as $t\downarrow T_{\rm gel}$: then we obtain $\ell=e^{-T_{\rm gel}(M_0-g_0(\ell))}$,
i.e. $\phi_{T_{\rm gel}}(\ell)=1$. By (d3) above, this is equivalent to $\ell=1$.
\item Since $M_t=g_t(1)=g_0(h_t(1))=g_0(l_t)$,
the properties of $t\mapsto M_t=g_0(l_t)$ follow from those of $t\mapsto l_t$.
Recall now that $\phi_t(l_t)=1$, that is
\begin{equation}
\log(l_t)=t(g_0(l_t)-1).
\label{lnlt}
\end{equation}
If the limit $l$ of $l_t$ as $t\to+\infty$ were nonzero, then passing to the limit in this equality would give $\log(l)=- \infty$. So $l=0$ and
\begin{equation}
\log l_t \sim -t.
\label{equiv}
\end{equation}
\begin{itemize}
\item Assume $m > 0$. Now, obviously $g_0(x) \leq x^m$, so
\[
\log(tg_0(l_t))=\log l_t + \log g_0(l_t) \leq \log t + m \log l_t \to - \infty.
\]
Hence $t g_0(l_t) \to 0$ and \eqref{lnlt} yields $\log l_t + t \to 0$. Hence
$l_t^m \sim e^{-mt}$.
Finally
\[
\lim_{t \to + \infty} M_t e^{mt} = \lim_{t \to + \infty} \frac{g_0(l_t)}{l_t^m} = m \mu_0(\{m\})
\]
since by dominated convergence, $g_0(x) x^{-m} \to m \mu_0(\{m\})$ when $x \to 0$. Now, by monotone convergence, if $m' > m$, then $g_0(x)x^{-m'} \to + \infty$ when $x$ tends to 0, whence
\[
\lim_{t \to + \infty} M_t e^{m't} = \lim_{t \to + \infty} \frac{g_0(l_t)}{l_t^{m'}} = + \infty.
\]
\item Assume now $m = 0$ and let $\eps > 0$. By monotone convergence
$g_0(x)x^{-\eps} \to + \infty$ as $x\downarrow 0$,
so using \eqref{equiv} we see that
$g(l_t)e^{-\eps t} \to + \infty$ as $t\uparrow + \infty$, which is the desired result.
\end{itemize}
\item Finally, \eqref{gt2} gives for $x <1$ and $t > T_{\rm gel}$
\[
g_t'(x)=g_0'(h_t(x))\, h_t'(x) = \frac{g_0'(h_t(x))}{\phi_t'(h_t(x))}.
\]
When $x \uparrow 1$, $h_t(x) \uparrow l_t < 1$, and $\phi_t'(h_t(x)) \to \phi_t'(l_t) \neq 0$
by (d6) above. So $\langle \mu_t , m^2 \rangle = g_t'(1) < + \infty$.
\item Existence of a solution of \eqref{flory} follows arguing as in section \ref{exist}.
\end{enumerate}
\end{proof}
\begin{cor}\label{leq}
Let $\mu_0\in\mathcal{M}_c^+$ such that $\langle\mu_0,m\rangle<+\infty$ and let $(\mu_t^S)_{t\geq 0}$ and
$(\mu_t^F)_{t\geq 0}$ the solutions of \eqref{smolu2}, respectively, \eqref{flory}. Then
\begin{itemize}
\item $\mu_t^S\equiv \mu_t^F$ for all $t\leq T_{\rm gel}:=1/\langle\mu_0,m^2\rangle$;
\item $\langle\mu_t^F,m\rangle
<\langle\mu_t^S,m\rangle$ for all $t> T_{\rm gel}$.
\end{itemize}
\end{cor}
\begin{proof}
For all $t\leq T_{\rm gel}$, $\langle\mu_t^F,m\rangle=\langle\mu_0^F,m\rangle$ and therefore $\mu_t^F$ solves \eqref{smolu2},
so that by uniqueness of Smoluchowski's equation we have that $\mu_t^S= \mu_t^F$. For $t> T_{\rm gel}$
we have that $\langle\mu_t^F,m\rangle=g_0(l_t)$ while $\langle\mu_t^S,m\rangle=g_0(\ell_t)$, where $l_t$ and $\ell_t$
are defined respectively by
\[
l_t=e^{-t(M_0-g_0(l_t))}, \quad l_t\in[0,1)
\]
and
\[
\ell_tg_0'(\ell_t)=\frac1t.
\]
In points (d4) and (d5) of the proof of Theorem \ref{thflory}, we have shown that $l_t<m_t$ where
$tm_tg_0'(m_t)=1$, so that $m_t=\ell_t < l_t$. Hence $\langle \mu_t^F, m \rangle = g_0(\ell_t) < g_0(l_t) = \langle \mu_t^S, m \rangle$.
\end{proof}
As anticipated, the mass decreases faster in Flory's case than for Smoluchowski's equation. In particular, in
Flory's case $\langle \mu_t , m^2 \rangle$ becomes finite immediately after gelation, the mass remaining however continuous (we can think that the big particles, which have the biggest influence on this second moment, disappear into the gel). Moreover, if
$\inf \mathrm{supp}\ \mu_0>0$ then the mass decays exponentially fast, which is to be compared with the slow decrease in $1/t$ in Smoluchowski's equation.
\begin{rem}
{\rm The mass in Flory's equation may decrease slower if $\inf \mathrm{supp}\ \mu_0$ $=0$. For instance, if $\mu_0(\mathrm{d}m)=e^{-m} \mathrm{d}m$, then $M_t \sim t^{-2}$.}
\end{rem}
\section{The model with limited aggregation}
We now turn to our main interest, namely Equation \eqref{smolubras0}. We apply the same techniques as above in a slightly more
complicated setting. After giving all details in Smoluchowski's case, we will give a shorter proof and focus on the differences with the proof of Theorem \ref{thsmolu}. As above, we can transform the system \eqref{smolubras0} into a non-local PDE problem, which we are able to solve, thus obtaining existence and uniqueness to \eqref{smolubras0}. More
precisely, we consider the following system.
\begin{definition}\label{def3.1}
Let $c_0(a,m)\geq 0$, $a\in{\ensuremath{\mathbb N}} $, $m\in{\ensuremath{\mathbb N}} ^*$. We say that a family $(c_t(a,m))$, $t\geq 0$, $a\in{\ensuremath{\mathbb N}} $, $m\in{\ensuremath{\mathbb N}} ^*$, is a solution of Smoluchowski's equation \eqref{smolubras} if
\begin{itemize}
\item for every $t>0$, $\int_0^t \langle c_s, a\rangle^2 \: \mathrm{d}s< + \infty$,
\item for all $a\in{\ensuremath{\mathbb N}} $, $m\in{\ensuremath{\mathbb N}} ^*$ and $t>0$,
\begin{equation} \label{smolubras}
\begin{split}
c_t(a,m)& = \ c_0(a,m)+ \\
& \quad + \int_0^t\frac12 \sum_{a'=1}^{a+1} \sum_{m'=1}^{m-1} a'(a+2-a') c_s(a',m')c_s(a+2-a',m-m') \: \mathrm{d}s \\
& \quad - \int_0^t \sum_{a' \geq 1} \sum_{m'\geq 1} aa'c_s(a,m) c_s(a',m') \: \mathrm{d}s,
\end{split}
\end{equation}
\item if $\langle c_0 , a^2 \rangle < + \infty$, then $t \mapsto \langle c_t , a^2 \rangle$ is bounded in a right neighborhood of 0.
\end{itemize}
\end{definition}
Because of the interpretation of $a$ as a variable counting the number of arms a particle possesses, it is more
natural to state \eqref{smolubras} in the discrete setting, as in \cite{BertoinTSS}. In particular, since at each coagulation
two arms are removed from the system, a non-integer initial number of arms would lead to an ill-defined dynamics.
One could however with no difficulty consider an initial distribution of masses on $(0,+\infty)$.
It is easy to see that $(c_t)$ is a solution to this equation if and only if the function
\begin{equation}
k_t(x,y):= \sum_{a=1}^{+\infty} \sum_{m=1}^{+\infty} a\, c_t(a,m) \, x^{a-1}\, y^m,
\label{kt}
\end{equation}
defined for $t\geq 0$, $y \in [0,1]$ and $x \in [0,1)$, satisfies
\begin{equation}\label{PDE3}
\left \{
\begin{array}{l}
\displaystyle k_t(x,y)= k_0(x,y)+\int_0^t\left[
\left( k_s(x,y) - xA_s \right) \d{k_s}{x}(x,y) - A_sk_s(x,y)\right] \: \mathrm{d}s, \\
A_t:=k_t(1,1)=\langle c_t,a\rangle.
\end{array}
\right.
\end{equation}
We may solve this PDE with the same techniques as above and obtain the following result.
\begin{theorem}\label{thsmolubras}
Consider initial concentrations $c_0(a,m)\geq 0$, $a\in{\ensuremath{\mathbb N}} $, $m\in{\ensuremath{\mathbb N}} ^*$ such that $\langle c_0, 1 \rangle <+\infty$,
$A_0:=\langle c_0, a \rangle \in(0,+\infty]$
and with $K:=\langle c_0 , a(a-1) \rangle \in [0,+ \infty]$. Then $K=+\infty$ whenever $A_0=+\infty$. Let
\begin{equation}
T_{\rm gel}= \begin{cases}
\frac{1}{K-A_0} \quad & {\rm if} \ A_0<K<+\infty, \\ 0 \quad & {\rm if} \ K=+\infty,
\\ +\infty \quad & {\rm if} \ K\leq A_0<+\infty.
\end{cases}
\label{Tgelbras}
\end{equation}
Then equation \eqref{smolubras} has a unique solution defined on ${\ensuremath{\mathbb R}} ^+$.
When $T_{\rm gel} < + \infty$, this solution enjoys the following properties.
\begin{enumerate}
\item The number of arms $A_t := \langle c_t , a \rangle$ is continuous, strictly decreasing, and for all
$t>0$
\begin{equation}\label{to0}
A_t \leq \frac{A_0}{1+tA_0}\quad {\rm if} \ A_0<+\infty, \qquad A_t \leq \frac1t \quad {\rm if} \ A_0=+\infty.
\end{equation}
If we set
\[
\a_t = \exp\left(\int_0^t A_s \: \mathrm{d}s\right),
\]
then $\a_t$ is given by
\[
\a_t = 1+A_0t \quad \mathrm{for}\ t < T_{\rm gel}
\]
and for $t\geq T_{\rm gel}$
\begin{equation}
\a_t = \begin{cases}\Gamma^{-1}(1+A_0T_{\rm gel}+t-T_{\rm gel})\quad & {\rm if} \ A_0<+\infty, \\
\Gamma^{-1}(1+t) \quad & {\rm if} \ A_0=+\infty,\end{cases}
\label{alphat2}
\end{equation}
where
\[
\Gamma(x) =1+A_0T_{\rm gel}+\int_{1+A_0T_{\rm gel}}^x \frac{dr}{k_0(H(1/r))}, \quad x\geq1+A_0T_{\rm gel},
\]
and $H:[G(0),G(1))\mapsto[0,1)$ is the right inverse of the increasing function
\begin{equation}\label{G}
G:[0,1)\mapsto[G(0),G(1)), \qquad
G(x) := x - \frac{k_0(x,1)}{k_0'(x,1)}, \quad x\in[0,1),
\end{equation}
with $G(0):=G(0^+)\leq 0$, and
\[
0 < G(1):=G(1^-)=\begin{cases} \displaystyle 1-\frac{A_0}K\quad & {\rm if} \ A_0<+\infty \\ 1 \quad & {\rm if} \ A_0=+\infty. \end{cases}
\]
\item Let $k_0$ be defined as in \eqref{kt},
and
\begin{equation} \label{Atbt}
A_t = \langle c_t , a \rangle, \qquad \a_t = \exp\left(\int_0^t A_s \: \mathrm{d}s\right),
\qquad \b_t = \int_0^t \frac{1}{\a_s^2} \: \mathrm{d}s.
\end{equation}
Consider
\[
\phi_t(x,y) := \a_t(x - \b_t k_0(x,y)), \qquad t\geq 0, \, x,y\in[0,1].
\]
Then
\begin{itemize}
\item $\phi_t(\cdot,1)$ attains its maximum at a point $\ell_t$ such that $\phi_t(\ell_t,1)=1$. For $t \leq T_{\rm gel}$,
$\ell_t =1$, and for $t > T_{\rm gel}$, $0 < \ell_t < 1$ and
\begin{equation}
\d{\phi_t}{x}(\ell_t,1) = 0.
\label{phitprimlt2}
\end{equation}
In particular, for $t > T_{\rm gel}$, $\ell_t$ is given by
\begin{equation}\label{phitprimlt22}
\ell_t = H \left ( \frac{1}{\a_t} \right ),
\end{equation}
where $H$ is the right inverse of the function $G$ defined above.
\item For every $y \in [0,1]$, $\phi_t(\cdot,y)$ has a right inverse $h_t(\cdot,y):[0,1]\mapsto[0,1]$.
\end{itemize}
\item The generating function $k_t$ defined by \eqref{kt} is given by
\begin{equation}
k_t(x,y) = \frac{1}{\a_t} k_0(h_t(x,y),y)
\label{ktxy}
\end{equation}
for $y \in [0,1]$, $x \in [0,1]$. In particular, for $t>0$
\begin{equation}\label{k0lt}
\a_t A_t = \a_t k_t(1,1) = k_0(\ell_t,1), \qquad
A_t =\frac{k_0(\ell_t,1)}{1+\int_0^tk_0(\ell_s,1) \: \mathrm{d}s}.
\end{equation}
\item The second moment $\langle c_t, a^2 \rangle$ is finite on $[0,T_{\rm gel})$, infinite on $[T_{\rm gel},+ \infty)$.
\end{enumerate}
\end{theorem}
\subsection{Proof}
The only major difference with respect to the proof of Theorem \ref{thsmolu} is the
additional variable $y$ in the generating function $k_t(x,y)$. However, the variable
$y$ plays the role of a parameter in the PDE \eqref{PDE3}, and this allows to adapt
all above techniques.
\begin{proof}[Proof of Theorem \ref{thsmolubras}] The case $K\leq A_0<+\infty$, for which $T_{\rm gel}= + \infty$
has already been treated in \cite[Thm. 2]{BertoinTSS}, so
that we can restrict here to the cases where $T_{\rm gel}<+\infty$.
When $T_{\rm gel} > 0$, Thm. 2 in \cite{BertoinTSS} also shows that $\a_t = 1 + A_0 t$ on $[0,T_{\rm gel})$ (this however also requires that $\langle a^2 ,c_t \rangle$ be bounded in a neighborhood of 0: see point 3 of the proof of Lemma \ref{ti}).
\begin{enumerate}
\item First, by setting $u_t(x,y):=\alpha_tk_t(\phi_t(x,y),y)-k_0(x,y)$, we can see,
arguing as in points (i)-(ii) of the proof of Lemma \ref{ti}, that for all
$y\in(0,1]$ and $t>0$ there exists $\ell_t^0(y) < \ell_t(y)\in(0,1]$ such that
\begin{equation}\label{ktphit}
\alpha_tk_t(\phi_t(x,y),y)=k_0(x,y), \qquad \forall \, t\geq 0, \, y\in(0,1], \, x\in[\ell_t^0(y),\ell_t(y)]
\end{equation}
and $\phi_t(\cdot,y):[\ell_t^0(y),\ell_t(y)]\mapsto[0,1]$ is a continuous bijection and
has a continuous right inverse $h_t(\cdot,y):[0,1]\mapsto[\ell_t^0(y),\ell_t(y)]$.
\begin{figure}[htb]
\begin{tabular}{cc}
\includegraphics[width=7.1cm]{phitpregelbras.png}& \includegraphics[width=7.1cm]{phitpostgelbras.png}
\end{tabular}
\caption{$\phi_t(\cdot,1)$ before and after gelation. The dotted lines represent what $\phi_t$ may look like. The solid one is the actual $\phi_t$.}
\end{figure}
\item We denote for simplicity
\[
k_t(x):=k_t(x,1), \qquad \phi_t(x):=\phi_t(x,1), \qquad t\geq 0, \, x\in[0,1].
\]
For $y=1$, we set $\ell_t(1)=\ell_t$, i.e.
\[
1=\phi_t(\ell_t)=\alpha_t(\ell_t-\beta_tk_0(\ell_t)), \qquad t\geq 0.
\]
Arguing as in points (iv)-(v) of the proof of Lemma \ref{ti}, we can see that
$\ell_t=1$ for all $t\leq T_{\rm gel}$ and $\ell_t<1$ for all $t>T_{\rm gel}$. Moreover,
$t\mapsto \ell_t$ is continuous and monotone non-increasing.
Since $\phi_t$ is increasing on $[0,\ell_t]$, $\phi_t'(\ell_t)\geq 0$, i.e.
\[
\beta_t \leq \frac1{k_0'(\ell_t)},
\]
so that
\begin{equation}\label{4}
1 =\alpha_t(\ell_t-\beta_tk_0(\ell_t))\geq \alpha_t\, G(\ell_t),
\end{equation}
where we set $G(x):=x-\frac{k_0(x)}{k_0'(x)}$, $x\in[0,1)$. Notice that
\[
G'(x)= 1-\frac{(k_0'(x))^2-k_0(x)k_0''(x)}{(k_0'(x))^2}=\frac{k_0(x)k_0''(x)}{(k_0'(x))^2}>0,
\]
since $k_0$ is strictly convex (there is no gelation whenever $k_0'' \equiv 0$). Moreover $G(0)\leq 0$ and
\[
G(1)=1-\frac{A_0}K\quad {\rm if} \ A_0<+\infty, \qquad G(1)=1 \quad {\rm if} \ A_0=+\infty.
\]
Indeed, $k_0'(1)=K=\langle c_0,a(a-1)\rangle$ and, if $k_0(1)=A_0=+\infty$, then
\[
\lim_{x\uparrow 1} \frac{k_0(x)}{k_0'(x)} =0
\]
since, if $\liminf_{x\uparrow 1} \frac{k_0(x)}{k_0'(x)} > \varepsilon>0$, then
$k_0(1)\leq k_0(1-\delta)e^{\delta/\varepsilon}<+\infty$, for
some $\delta>0$, contradicting $k_0(1)=+\infty$.
In any case, $G$ has an inverse $H$, and $H(1/x)$ is defined for $x \in [1+A_0 T_{\rm gel},+ \infty)$.
\item
Computing \eqref{ktphit} at $(x,y)=(\ell_t,1)$ we obtain
\begin{equation}\label{2}
k_0(\ell_t)=\alpha_tk_t(1)=\alpha_tA_t=\dplusdt{\alpha_t}.
\end{equation}
Let us notice that
\[
\phi_t(x) = x+\int_0^t \left(A_s \phi_s(x)-\frac{k_0(x)}{\alpha_s}\right) \: \mathrm{d}s.
\]
Then by \eqref{2}, analogously to \eqref{phitprimlt} above,
\[
0 = \mathrm{d} \, \phi_t(\ell_t) = \left(A_t\phi_t(\ell_t)-\frac{k_0(\ell_t)}{\alpha_t}\right)\mathrm{d}t
+ \phi'_t(\ell_t)\,\: \mathrm{d}\ell_t= \phi'_t(\ell_t)\,\: \mathrm{d}\ell_t.
\]
In particular, for $\: \mathrm{d}\ell_t$-a.e. $t$, $\phi'_t(\ell_t)=0$, i.e. $\beta_t = 1/k_0'(\ell_t)$,
and therefore
\[
1 =\alpha_t(\ell_t-\beta_tk_0(\ell_t))= \alpha_t\, G(\ell_t), \qquad \: \mathrm{d}\ell_t-{\rm a.e.} \, t.
\]
Then, by \eqref{4}, we can write (note that $H$ is well-defined on the considered interval)
\[
\ell_t \leq H\left(\frac1{\alpha_t}\right), \quad \forall \, t>T_{\rm gel},
\qquad \ell_t = H\left(\frac1{\alpha_t}\right), \quad \: \mathrm{d}\ell_t-{\rm a.e.} \, t.
\]
Now, by \eqref{2}, setting $\Lambda:\,]1+A_0T_{\rm gel},+\infty[\,\mapsto\,]0,1[$, $\Lambda(z):=k_0\left(H\left(\frac1z\right)\right)$,
\[
\dplusdt{\alpha_t} \leq \Lambda(\alpha_t), \quad \forall \, t> T_{\rm gel},
\qquad \dplusdt{\alpha_t} = \Lambda(\alpha_t), \quad \: \mathrm{d}\ell_t-{\rm a.e.} \, t.
\]
Since $\alpha_t>1+A_0 T_{\rm gel}$ for any $t>T_{\rm gel}$, we obtain that $k_0(\ell_t)\leq\Lambda(\alpha_t)<1$ for
all $t>T_{\rm gel}$. In particular, $\: \mathrm{d}\ell_t$ is not identically equal to $0$. Suppose that for some
$t>T_{\rm gel}$ we have $\phi_t'(\ell_t)>0$. We set
\[
s:=\sup\{r<t: \phi_r'(\ell_r)=0\}=\max\{r<t: \phi_r'(\ell_r)=0\}.
\]
Then for all $r\in\,]s,t[$ we must have $\phi_r'(\ell_r)>0$. Then for all
$r\in\,]s,t[$ we have $\ell_r=\ell_s$. But, by definition of $\beta$,
\[
\beta_r>\beta_s=\frac1{k_0'(\ell_s)}=\frac1{k_0'(\ell_r)}
\]
and this is a contradiction.
Therefore for all $t>T_{\rm gel},$ we have $\dot{\alpha}_t = \Lambda(\alpha_t)$ for all
$t>T_{\rm gel}$ and the only solution of this equation with $\a_{T_{\rm gel}}=1+A_0 T_{\rm gel}$ is
given by \eqref{alphat2}.
\item In order to prove \eqref{k0lt}, let us note that by the preceding results
\[
\ddt{\a_t}=\a_t A_t = \a_t k_t(1,1) = k_0(\ell_t,1),
\]
\[
A_t = \ddt{}\log\a_t=\ddt{}\log\left(1+\int_0^tk_0(\ell_s,1) \: \mathrm{d}s\right)
=\frac{k_0(\ell_t,1)}{1+\int_0^tk_0(\ell_s,1) \: \mathrm{d}s}.
\]
\end{enumerate}
The rest of the proof follows the same line as that of Theorem \ref{thsmolu}.
\end{proof}
\section{The modified version}
Let us finally consider Flory's version of the model with arms. As in the case of Flory's equation
\eqref{flory}, we can consider only initial concentrations $c_0$ such that $A_0=\langle c_0,a\rangle<+\infty$.
Then, the equation we are interested in is
\begin{equation}
\begin{split}
\frac{\mathrm{d}}{\mathrm{d}t} c_t(a,m) & = \frac12 \sum_{a'=1}^{a+1} \sum_{m'=1}^{m-1} a'(a+2-a') c_t(a',m')c_t(a+2-a',m-m') \\
& \quad - \sum_{a' \geq 1} \sum_{m'\geq 1} aa'c_t(a,m) c_t(a',m') \\
& \quad - \left ( \frac{A_0}{1+tA_0} - \sum_{a',m' \geq 1} a' c_t(a',m') \right ) a c_t(a,m).
\end{split}
\label{florybras}
\end{equation}
With the same techniques as above, we can prove the following result.
\begin{theorem}\label{thflorybras}
Consider initial concentrations $c_0(a,m)\geq 0$, $a\in{\ensuremath{\mathbb N}} $, $m\in{\ensuremath{\mathbb N}} ^*$ such that $A_0:=\langle c_0, a \rangle \in(0,+\infty)$
and with $K:=\langle c_0 , a(a-1) \rangle \in [0,+ \infty]$. Let $T_{\rm gel}$ be defined as in \eqref{Tgelbras}.
Then equation \eqref{florybras} has a unique solution defined on ${\ensuremath{\mathbb R}} ^+$.
When $T_{\rm gel} < + \infty$, this solution enjoys the following properties.
\begin{enumerate}
\item We have
\begin{equation}
A_t=\frac{1}{1+tA_0} k_0(l_t)
\label{Atfl}
\end{equation}
where $l_t=1$ for $t\leq T_{\rm gel}$ and, for $t> T_{\rm gel}$, $l_t$ is uniquely defined by
\[
l_t = \frac{t}{1+tA_0} k_0(l_t), \qquad l_t\in[0,1).
\]
Therefore $t\mapsto A_t$ is continuous and strictly decreasing on $[0,+\infty)$ and analytic on ${\ensuremath{\mathbb R}} ^+ \backslash \{ T_{\rm gel} \}$.
\item The function $\phi_t(x,y)=(1+tA_0)x - t k_0(x,y)$ has, for every $y \in [0,1]$, a right inverse $h_t(\cdot,y) \, : \, [0,1] \to [0,l_t]$. The generating function $k_t$ defined in \eqref{kt} is given for $t \geq 0$ by
\begin{equation}
k_t(x,y) = \frac{1}{1+t A_0} k_0(h_t(x,y),y).
\label{ktfl}
\end{equation}
\item The second moment $\langle a^2,c_t \rangle$ is finite on ${\ensuremath{\mathbb R}} ^+ \backslash \{ T_{\rm gel} \}$ and infinite at $T_{\rm gel}$.
\end{enumerate}
\end{theorem}
\begin{proof}
The proof follows the same line of reasoning as the one of Theorem \ref{thflory}. First, for every $y \in [0,1]$, $\phi_t(\cdot,y)$, as defined in the statement, has the following properties:
\begin{itemize}
\item[(i)] $\phi_t(0,y) \leq 0$, $\phi_t(1,y) \geq \phi_t(1,1) = 1$;
\item[(ii)] For $t \leq T_{\rm gel}$, $\phi_t(\cdot,y)$ is increasing , and in particular, there are unique $0 \leq l_t^0(y)<l_t(y) \leq 1$ such that $\phi_t(l_t^0(y),y)=0$ and $\phi_t(l_t(y),y) =1$;
\item[(iii)] For $t > T_{\rm gel}$, $\phi_t(\cdot,y)$ is increasing then decreasing for, and in particular, there are unique $0 \leq l_t^0(y) < l_t(y) < 1$ such that $\phi_t(l_t^0(y),y)=0$ and $\phi_t(l_t(y),y) =1$.
\end{itemize}
\begin{figure}[htb]
\begin{tabular}{cc}
\includegraphics[width=7.1cm]{phitpregelbras.png}& \includegraphics[width=7.1cm]{phitpostgelbrasfl.png}
\end{tabular}
\caption{$\phi_t(\cdot,1)$ before and after gelation.}
\end{figure}
In any case, it is easy to check that for $x \in [l_t^0(y),l_t(y)]$,
\[
\exp \left ( \int_0^t A_s \: \mathrm{d}s \right ) k_t(\phi_t(x,y),y)=k_0(x,y)
\]
where $A_t$ is defined by \eqref{Atfl}. Then, the properties above show that $\phi_t(\cdot,y)$ has a right inverse $h_t$ defined on $[0,1]$, and compounding by $h_t$ in the previous equation shows that \eqref{ktfl} holds. The other properties then follow easily.
\end{proof}
\section{Limiting concentrations}
We compute here some explicit formulas for the concentrations and their limit for the two models above. In the standard Smoluchowski and Flory cases, particles keep coagulating, and they all eventually disappear into the gel: $c_t(m) \to 0$ for every $m \geq 1$. When the aggregations are limited, there may remain some particles in the system, since whenever a particle with no arms is created, it becomes inert, and so it will remain in the medium forever. In the following, we consider monodisperse initial conditions, i.e. $c_0(a,m) = \mu(a) \un{\{m=1\}}$ for a measure $\mu$ on ${\ensuremath{\mathbb N}} $. We also denote
\[
\nu(m) = (m+1) \mu(m+1).
\]
In \cite{BertoinTSS}, it is assumed that $\nu$ is a probability measure, what we do not require. The results of \cite{BertoinTSS} can hence be recovered by taking $A_0=1$ below. Now, note the two following facts.
\begin{itemize}
\item Equations \eqref{to0} and \eqref{Atfl} readily show that
\begin{equation} \label{cinf}
c_{\infty}(a,m) := \lim_{t \to + \infty} c_t(a,m) = 0, \quad a \geq 1,
\end{equation}
that is, only particles with no arms remain in the medium (else, a coagulation ``should'' occur).
\item There is an arbitrary concentration of particles with no arms at time 0, and they are the only particles with no arms and mass $1$ which will still be in the medium in the final state. Hence, the limit concentrations $c_{\infty}(0,1)=c_0(0,1)$ have no physical meaning. We will thus only consider $c_{\infty}(0,m)$ for $m \geq 2$.
\end{itemize}
Note now that if at time 0, each particle has zero or more than two arms, then obviously, this property still holds for any positive time. Rigorously, this is easy to check with the representation formula \eqref{ktxy} or \eqref{ktfl}. Then, because of \eqref{cinf},
\[
c_{\infty}(m)=0
\]
for each $m \geq 2$. We thus rule out this trivial case by assuming that
\begin{equation} \label{hypnu}
\nu(0) > 0.
\end{equation}
This is actually a technical assumption which is needed to apply Lagrange's inversion formula in the proof of the following corollaries. We will relate our results to a population model known as the Galton-Watson process. For some basics on this topic, see e.g. the classic book \cite{AtNey}. The formula providing the total progeny of these processes was first obtained by Dwass in \cite{Dwass}.
\subsection{Modified model} \label{seccinffl}
\begin{cor} \label{corcinffl}
Let $c_t(a,m)$ be the solution to Flory's equation with arms \eqref{florybras} and with initial conditions $c_0(a,m) = \mu(a) \un{\{m=1\}}$ with $\mu(1) > 0$.
\begin{itemize}
\item For all $t \geq 0$, $m \geq 2$, $a \geq 0$,
\[
c_t(a,m) = \frac{(a+m-2)!}{a!m!} \frac{t^{m-1}}{(1+tA_0)^{a+m-1}} \nu^{*m}(a+m-2).
\]
\item In particular, there are limiting concentrations $c_{\infty}(a,m) = c_{\infty}(m) \un{\{a=0\}}$ with
\begin{equation} \label{cinffl}
c_{\infty}(m)=\frac{1}{m(m-1)} \nu^{*m}(m-2).
\end{equation}
\end{itemize}
\end{cor}
\begin{proof}
With the notation of Theorem \ref{thflorybras}, we have
\[
(1+t A_0)h_t(x,y) - t y k_0(h_t(x,y)) = x, \quad k_t(x,y) = \frac{1}{1+t A_0} y k_0(h_t(x,y)).
\]
Up to some obvious changes (just replace $1+t$ by $1+tA_0$), these are precisely the equations solved in Section 3.2 of \cite{BertoinTSS} under the assumption \eqref{hypnu}. Theorem 2 and Corollary 2 therein hence give the desired result (with only $1+t$ replaced by $1+t A_0$).
\end{proof}
If $A_0=1$, which we may always assume up to a time-change, we observe as in \cite{BertoinTSS} that $2 (m-1) c_{\infty}(0,m)$ is the probability for a Galton-Watson process with reproduction law $\nu$, started from two ancestors, to have total progeny $m$. This Galton-Watson process is (sub)critical when $K:=\sum_{a\geq 1}a(a-1)\mu(a) \leq 1$, that is, by Theorem \ref{thflorybras}, when there is no gelation, and supercritical when $K > 1$. Denote by $p_{\nu}$ its extinction probability, i.e. the smallest root of $k_0(x)=x$, so $p_{\nu} = 1$ when $K \leq 1$ and $p_{\nu} < 1$ when $K > 1$. Let us compute the mass at infinity, as in \cite{BertoinTSS}, by writing
\begin{align*}
M_{\infty} := \sum_{m \geq 1} m c_{\infty}(m) & = c_{\infty}(1) + \sum_{m \geq 2} \frac{1}{m-1} \nu^{*m}(m-2) \\
& = c_{\infty}(1) + \sum_{a \geq 0} \nu(a) \sum_{m \geq a + 2} \frac{1}{m-1} \nu^{*m-1}(m-2-a) \\
& = c_{\infty}(1) + \sum_{a \geq 0} \nu(a) \sum_{n \geq a + 1} \frac{1}{n} \nu^{*n}(n-1-a).
\end{align*}
Now, the Lagrange inversion formula \cite{Genefunc} shows that
\[
\frac{a+1}{n}\nu^{*n}(n-1-a)
\]
is precisely the coefficient of $x^n$ in the analytic expansion of $\phi(x)$ around $0$, where $\phi$ is the unique solution to $\phi(x) = x k(\phi(x))$. Hence
\[
\sum_{n \geq a + 1} \frac{1}{n} \nu^{*n}(n-1-a) = p_{\nu},
\]
where $p_{\nu}$ is defined above. Note also that $c_{\infty}(1) = \mu(0)$, so finally
\begin{equation}\label{Minfty}
M_{\infty} = c_{\infty}(1) + \sum_{a \geq 0} \nu(a) \, \frac{1}{a+1} \, p_{\nu}^{a+1} = \sum_{a \geq 0} \mu(a) \, p_{\nu}^a.
\end{equation}
The mass at time 0 is $M_0 = \sum \mu(a)$, so when there is no gelation, $p_{\nu}=1$ and no mass is lost in the gel. When there is gelation, $p_{\nu} < 1$ and the mass $M_0 - M_{\infty} > 0$ is lost in the gel. By Dwass' formula \cite{Dwass}, $M_{\infty}$ is also the probability that a Galton-Watson process, with reproduction law $\mu$ for the ancestor and $\nu$ for the others, has a finite progeny.
\subsection{Non-modified model} \label{seccinfsm}
\begin{cor}
Let $c_t(a,m)$ be the solution to Smoluchowski's equation with arms \eqref{smolubras} and with initial conditions $c_0(a,m) = \mu(a) \un{\{m=1\}}$ with $\mu(1) > 0$.
\begin{itemize}
\item For all $t \geq 0$, $m \geq 2$, $a \geq 0$,
\[
c_t(a,m) = \frac{(a+m-2)!}{a!m!} \frac{\beta_t^{m-1}}{\alpha_t^a} \nu^{*m}(a+m-2)
\]
where $\a_t$ and $\b_t$ are defined in Theorem \ref{thsmolubras}.
\item In particular, there are limiting concentrations $c_{\infty}(a,m) = c_{\infty}(m) \un{\{a=0\}}$ with
\begin{equation}\label{cinfflb}
c_{\infty}(m)=\frac{1}{m(m-1)} \beta_{\infty}^{m-1} \nu^{*m}(m-2)
\end{equation}
where $\beta_{\infty}$ is defined by
\[
\beta_{\infty} = \frac{1}{k_0'(c)} = \frac{c}{k_0(c)}
\]
and $c$ is the unique solution to $k_0'(c) = k_0(c)/c$. Moreover, $\beta_{\infty} = 1$ when there is no gelation, and $\beta_{\infty} > 1$ otherwise.
\end{itemize}
\end{cor}
\begin{proof}
As for Corollary \ref{corcinffl}, the proof of the formula for $c_t(a,m)$ is the same as in \cite[Section 3.2]{BertoinTSS}, just replacing $1+t A_0$ by $\a_t$ and $t$ by $\a_t \b_t$. So we just have to find the limit of $\b_t$. First \eqref{alphat2} shows that $\a_t \to + \infty$,
hence, by \eqref{phitprimlt22}, $\ell_t \to \ell_{\infty} = H(0)$. Now, \eqref{phitprimlt2} gives $\b_t = 1/k_0'(\ell_t)$, so $\b_t$ tends to
\[
\b_{\infty} = \frac{1}{k_0'(H(0))}
\]
where by definition $c := H(0)$ is the unique solution to $k_0'(c)=k_0(c)/c$. Finally, when there is gelation, $\a_t < 1+t$ after gelation because of \eqref{alphat2}, so by \eqref{Atbt}, $\beta_{\infty} > 1$.
\end{proof}
By a similar computation as above, we may also compute the mass at infinity in this case and get
\[
M_{\infty} = \sum_{a \geq 0} \mu(a) \, c^a
\]
where $c$ is defined in the corollary. Note that $c$ is the slope of the straight line passing by $0$ and tangent to the graph of $k$, so $c > p_{\nu}$. In particular, less mass is lost than in Flory's case.
A final remark is that despite the striking resemblance between Formulas \eqref{cinfflb} and \eqref{cinffl}, the meaning of the factor $\beta_{\infty}$ is unclear. A probabilistic interpretation using the configuration model may explain its appearance.
\medskip
\noindent \textbf{Acknowledgements} We thank Jean Bertoin for useful discussions and advice.
|
\section{Introduction}
In recent years, the concept of entanglement \cite{hor09} has
received much attention in quantum information and quantum
computation \cite{nielson} due to its fascinating features in
developing the idea of quantum communication and quantum computer.
Entanglement is one kind of quantum correlations \cite{hor09} and
offers support for lots of quantum tasks such as quantum
teleportation \cite{bennett93,bou97}, quantum dense code
\cite{bennett92} and quantum computing \cite{shor99,gro97}. However,
it is not the only part of quantum correlations \cite{maz09}. So
what is the quantum correlation without entanglement? An answer to
this question is the notion of quantum discord (QD) given by
Ollivier and Zurek \cite{zurek01}. It introduces a measure of the
quantumness of correlations (including entanglement). This notion
has attracted much attention nowadays
\cite{sarandy09,dil08,luo08,wer09,werlang09,maz09,chen09}. It can be
treated as an important quantum resource and even used for the
quantum computation task \cite{dat08,lanyon08,kni98}.
As a kind of quantum resource, entanglement can be generated,
maintained and controlled in so many condensed matter fields
especially solid state systems such as quantum dots
\cite{tra07,hanson07}, Josephson junctions \cite{wei06}, trapped
ions \cite{por04} and cavity-QED systems \cite{zheng00}, etc. Among
systems above, owing to its concise form of Hamiltonian and ability
to be realized in real physical systems, there has been plenty of
work discussing entanglement and thermal entanglement qualitatively
and quantitatively in spin-chain systems
\cite{wang02,sun03,zhang05,arn01,aso05,gun01}.
Since much work discuss the spin-spin interaction in spin-chain
systems such as Heisenberg model, more and more researches study the
influence of Dzyaloshinskii-Moriya (DM) interaction
\cite{dzy58,mo60} which is arising from spin-orbit interaction in
spin-chain systems \cite{aristov00,kargarian09,jafari08}. Recently,
there is an interesting paper \cite{li08} investigating the thermal
entanglement in an anisotropic XXZ model with DM interaction. The
characteristic of thermal entanglement in that model is notable
especially when you control the entanglement via tuning
$x$-component DM interaction.
Since QD is considered as quantum correlation and quantum resource
as well as entanglement. It is curious to ask what the
characteristic of QD is at finite temperature, and what the
differences are between thermal QD and entanglement in this system.
In our paper we introduce this two-qubit anisotropic Heisenberg XXZ
model with DM interation. We discuss the dependence of thermal QD in
this system on temperature $T$ and some other tunable parameters
such as DM interaction parameters $D_z$, $D_x$. We also compare the
QD with entanglement measured by concurrence
\cite{1bennett96,2bennett96,wootters98} between two qubits in the
model and find remarkable differences between them. QD is more
robust than entanglement concurrence versus temperature $T$. (This
property can also be found in Ref. \cite{werlang09} where a
Heisenberg XYZ model with a magnetic term was considered.) We find
that the characteristic of QD is special in this system since QD
decreases with certain tunable parameters such as $D_z$ and $D_x$
when entanglement increases. And this possibly offers a potential
solution to enhance entanglement of a system. Furthermore, we
analyze the influence of different DM interaction parameters $D_z$
and $D_x$ on QD and reveal that tunable parameter $D_x$ is more
efficient than parameter $D_z$ in most regions for controlling the
QD.
Our paper is organized as follows. In Sec. II we review the concept
of quantum discord. In Sec. III we investigate the thermal quantum
discord in two XXZ Heisenberg model with different DM interaction
parameters. We show the main results and compare thermal QD with
thermal entanglement. In Sec. IV we discuss the opposite tendency of
thermal QD and entanglement. Conclusion are then presented in Sec.
V.
\section{Quantum Discord}
In classical information theory, the correlation between two random
variables $A$ and $B$ are described by mutual information which
reads in two expressions.
\begin{equation}
\mathcal{I}(A:B) = H(A)+H(B)-H(A,B) \label{i1}\end{equation}
\begin{equation}
\mathcal{J}(A:B) = H(A)-H(A|B) \label{i2}\end{equation} Here,
$H(A)=-\sum_{a}p_{a}\log_{2}p_{a}$ is Shannon entropy. $H(A,B)$ is
the joint Shannon entropy which is defined by
$H(A,B)=-\sum_{a,b}p_{a,b}\log_{2}p_{a,b}$. $H(A|B)$ is conditional
entropy and introduced for quantifying the ignorance (on average)
about the value of A given B is known. This two expressions are
equivalent using Bayes' rule:
$p(a_{i},b_{j})=p(a_{i}|b_{j})p(b_{j})=p(b_{j}|a_{i})p(a_{i})$,
where $p(a_{i}|b_{j})$ is the conditional probability.
In order to generalize these expressions above to quantum case, we
replace classical probability distributions by density matrices and
the Shannon entropy by von Neumann entropy
$S(\rho)=-\textrm{tr}(\rho\log\rho)$. Since we can denote the
density matrix of a composite system $AB$ by $\rho^{AB}$ and the
density matrices of parts $A$ and $B$ by $\rho^A$ and $\rho^B$,
respectively, we can rewrite Eq. (\ref{i1}) and Eq. (\ref{i2}) as
\begin{equation}
\mathcal{I}(\rho^A:\rho^B) = S(\rho^A)+S(\rho^B)-S(\rho^{AB})
\label{q1}\end{equation}
\begin{equation}
\mathcal{J}(\rho^A:\rho^B) = S(\rho^A)-S(\rho^A|\rho^B),
\label{q2}\end{equation} where $S(\rho^{AB})$ is quantum joint
entropy and $S(\rho^A|\rho^B)$ is quantum condition entropy.
However, here we can not just obtain quantum condition entropy
simply via replacing Shannon entropy by von Neumann entropy. Due to
the definition of condition entropy, $S(\rho^A|\rho^B)$ depends on
the choice of measurement. To get quantum condition entropy, we
should choose a measurement which is described by a set of
projectors $\{\hat{\Pi}_{i}^{B}\}$ performed locally on B. After the
measurement, the state of system is
$\rho_{i}=\hat{\Pi}_{i}^{B}\rho^{AB}\hat{\Pi}_{i}^{B}/p_{i}$ where
$p_{i}=\textrm{tr}(\hat{\Pi}_{i}^{B}\rho^{AB}\hat{\Pi}_{i}^{B})$.
Now we can get the definition of quantum condition entropy.
Following the Eq. (\ref{q2}), another expression of quantum mutual
information is
\begin{equation}
\mathcal{J}(\rho^{AB}:\{\hat{\Pi}_{i}^{B}\}) =
S(\rho^{A})-S(\rho^{AB}|\{\hat{\Pi}_{i}^{B}\}) \label{mut}
\end{equation}
Since different choices of $\{\hat{\Pi}_{i}^{B}\}$ decide different
$\mathcal{J}(\rho^{AB}:\{\hat{\Pi}_{i}^{B}\})$, Eq. (\ref{q1}) and
Eq. (\ref{q2}) may not be equal any more. According to the original
definition \cite{zurek01}, quantum discord is the minimum of
difference between Eq. (\ref{q1}) and Eq. (\ref{mut}), i.e.
\begin{equation}
D(\rho^{AB})=\textrm{min}\left[\mathcal{I}(\rho^A:\rho^B)-\mathcal{J}(\rho^{AB}:\{\hat{\Pi}_{i}^{B}\})\right]\end{equation}
It is a measure of the quantumness of a pairwise correlation. With
the expression of quantum mutual information of Eq. (\ref{mut}),
following Ref. \cite{zurek01,henderson01}, we can define classical
correlation between $A$ and $B$ as
\begin{equation}
CC(\rho^{AB})=\underset{{\{\hat{\Pi}_{i}^{B}\}}}{max}(\mathcal{J}(\rho^{AB}:\{\hat{\Pi}_{i}^{B}\}))\label{classical}
\end{equation}
Eventually, we get quantum discord in another expression:
\begin{equation}
Q(\rho^{AB})=\mathcal{I}(\rho^A:\rho^B)-CC(\rho^{AB})\label{discord}
\end{equation} Here the quantum mutual information is used as a measure of total
correlations \cite{verdral02,gro05}.
Quantum discord is a measure to quantify all the quantum
correlations including entanglement of a pairwise correlation. For
two subsystems $A$ and $B$ are correlated classically only, the QD
is zero. Moreover, for some pairwise system such as Werner states
\cite{zurek01}, QD is non-zero although the system is separable. In
this sense, some two-body states that don't have entanglement do not
mean there is no quantum correlations between them. We can even use
this kind of quantum correlation power to do some quantum
computation tasks \cite{dat08,lanyon08,kni98}.
\section{XXZ Heisenberg model with different DM interaction parameters}
In this section, we introduce an anisotropic XXZ Heisenberg model
with different DM interaction parameters. We show the dependence of
thermal QD on temperature $T$ and some other tunable parameters such
as $J_z$, $D_z$ and $D_x$ in this model. We compare the thermal QD
with thermal entanglement concurrence in the same model and find
some attractive results.
\subsection{XXZ Heisenberg model with DM interaction parameter $D_z$}
The Hamiltonian $H$ of a two-qubit anisotropic Heisenberg XXZ chain
with DM interaction parameter $D_z$ is
\begin{equation}
H=J\sigma_1^x\sigma_2^x+J\sigma_1^y\sigma_2^y+J_z\sigma_1^z\sigma_2^z+D_z(\sigma_1^x\sigma_2^y-\sigma_1^y\sigma_2^x).
\label{h1}\end{equation} where $J$ and $J_z$ are the real coupling
coefficients, $D_z$ is the $z$-component parameter of the DM
interaction. Here the so-called DM interaction is a supplemented
magnetic term arising from the spin-orbit coupling
\cite{dzy58,mo60}. $\sigma^i$ ($i = x, y, z$) are Pauli matrices. In
our paper, we consider the antiferromagnetic case ($J>0$, $J_z>0$).
This model is reduced to the isotropic XX model when $J_z=0$ and to
the isotropic XXX model when $J_z=J$. Parameters $J$, $J_z$ and
$D_z$ are dimensionless here.
First we show the matrix form of this $H$ in the standard basis
$\{|00\rangle,|01\rangle,|10\rangle,|11\rangle\}$, and get
\begin{equation}
H = \left(
\begin{array}{cccc}
J_z&0&0&0\\
0&-J_z&2J+2iD_z&0\\
0&2J-2iD_z&-J_z&0\\
0&0&0&J_z
\end{array}
\right)\label{h1m}
\end{equation}
Now we want to consider the state of the system as thermal
equilibrium state in a canonical ensemble. The state of spin chain
system at thermal equilibrium is given by the Gibb's density
operator $\rho(T)=\frac{exp(-\beta H)}{Z}$ where $Z=tr[\exp(-\beta
H)]$ is the partition function of the system, $H$ is the system
Hamiltonian and $\beta=\frac{1}{K_BT}$ with T temperature. $K_B$ is
the Boltzmann's constant which we take equal to 1 for simplicity.
Here $\rho(T)$ presents a thermal state. Plenty of paper have
discussed about entanglement in thermal state
\cite{wang02,sun03,zhang05,arn01,aso05,gun01} which is called
thermal entanglement \cite{nielsen00}. In our paper, we discuss
another kind of quantum correlation-quantum discord at finite
temperature i.e., thermal quantum discord.
By some routine calculations, we can get expression of $\rho(T)$:
\begin{equation}
\rho(T) =\frac{1}{Z} \left(
\begin{array}{cccc}
e^{-\beta J_z}&0&0&0\\
0&u&ve^{i\theta}&0\\
0&ve^{-i\theta}&u&0\\
0&0&0&e^{-\beta J_z}
\end{array}
\right)
\end{equation}
where $u=\frac{1}{2}(1+e^{4\beta w})e^{\beta(J_z-2w)}$,
$v=\frac{1}{2}(1-e^{4\beta w})e^{\beta(J_z-2w)}$, $Z=2e^{-\beta
J_z}[1+e^{2\beta J_z}\cosh(2\beta w)]$ and $w=\sqrt{J^2+D_z^2}$,
$\theta=\arctan(\frac{D_z}{J})$.
According to Eq. (\ref{discord}), to obtain QD of a two-qubit
system, we should first get $\rho^A$, $\rho^B$ for subsystems $A$
and $B$ respectively and $\rho^{AB}$ for the composite system. Since
we have $\rho^{AB}$ already, we need to evaluate $\rho^A$, $\rho^B$.
Choosing $|0\rangle$, $|1\rangle$ as a basis, we can trace out the
rest of the subsystem to get reduced density matrix of $A$ and $B$
employing formulas below.
\begin{equation}
\rho^A\equiv tr_B\rho^{AB}\equiv
\sum_{j=1}^{N_B}(I_A\bigotimes\langle\phi_j|)\rho^{AB}(I_A\bigotimes|\phi_j\rangle)
\end{equation}
\begin{equation}
\rho^B\equiv tr_A\rho^{AB}\equiv
\sum_{j=1}^{N_A}(\langle\psi_j|\bigotimes
I_B)\rho^{AB}(|\psi_j\rangle\bigotimes I_B)
\end{equation} where $I_A$ and $I_B$ are the identity operators, $|\phi_j\rangle$ ($j=1, 2, ..., N_B$) and $|\psi_j\rangle$ ($j=1, 2,
..., N_A$) are orthonormal bases in $\mathcal{H}_B$ and
$\mathcal{H}_A$ respectively.
Now we can get $\rho_A=\rho_B=\frac{1}{Z}(e^{-\beta J_z}+u)I$ where
$I$ is the identity operator. Next we evaluate quantum discord
numerically according to the following instruction. Choose the set
of projectors
$\{|\psi_1\rangle\langle\psi_1|,|\psi_2\rangle\langle\psi_2|\}$,
where
$|\psi_1\rangle=\cos{\theta}|0\rangle+e^{i\phi}\sin{\theta|1\rangle}$
and
$|\psi_2\rangle=-\cos{\theta}|1\rangle+e^{-i\phi}\sin{\theta|0\rangle}$
, to measure one of the subsystems. The classical correlation
$CC(\rho^{AB})$ is obtained numerically by varying the angles
$\theta$ and $\phi$ from $0$ to $2\pi$ and eventually according to
Eq. (\ref{discord}) we can get QD as well.
\begin{figure}
[ptbh](a)
\begin{minipage}{2.4in}
\includegraphics[
width=2.4in,height=1.4in
{fig.1.eps}
\end{minipage}
\hspace{0.0in}
(b)
\begin{minipage}{2.4in}
\includegraphics[
width=2.4in,height=1.4in
]
{fig.2.eps}%
\end{minipage}
\hspace{0.0in}
(c)
\begin{minipage}{2.4in}
\includegraphics[
width=2.4in,height=1.4in
]
{fig.3.eps}%
\end{minipage}
\caption{The QD (a), (c) and concurrence (b) versus T in the system with $z$-component parameter $D_z$. Here $J=1$.
(a) Here $J_z=0.2$ and for the thick line $D_z=1.0$, normal line $D_z=0.7$, dashed line $D_z=0.5$.
(b) Here $J_z=0.2$ and for the thick line $D_z=1.0$, normal line $D_z=0.7$, dashed line $D_z=0.5$.
(c) Here $J_z=1$ and for the thick line $D_z=1.0$, normal line $D_z=0.7$, dashed line $D_z=0.5$.}
\end{figure}
Using the method in Ref. \cite{1bennett96,wootters98,2bennett96},
and results in \cite{li08}, we can also simply get the concurrence
(In this paper, we use concurrence as the measure of entanglement)
of this model at finite temperature. $C(\rho(T))=\frac{\beta
J_z}{Z}(e^{2\beta w}-e^{-2\beta J_z}-e^{-2\beta w}-e^{-2\beta J_z})$
if $J_z>-w$, and $C(\rho(T))=0$ if $J_z<-w$.
Now by fixing some parameters we can see the variations of thermal
QD and entanglement under the influence of other parameters. Looking
at Fig. 1, where $J=1$, we can find that the QD decreases with
increasing of temperature $T$. This tendency is similar to the
situation of entanglement as shown in panel (b) where the conditions
are the same. Differently, we can see from panel (b) that thermal
entanglement will vanish when $T$ reaches certain point. And the
larger parameter $D_z$ is the larger critical point is. The point
denotes so-called critical temperature of the system. This behavior
is caused by mixing of the maximally entanglement state with other
states. In other words, it means that with the increasing of
temperature $T$, thermal entanglement of the system will be
disappeared at critical temperature and there is no entanglement
resource to use then. However, as we see in panels (a) and (c),
thermal QD will not reach zero point. It acts asymptotically near
zero when the temperature is high. And in the temperature regions
where thermal entanglement is zero while QD is still alive. We even
find that we adjust $T$ very high (not shown in the plots), QD
becomes very small but nonzero. This interesting phenomenon can be
explained as follows. With the temperature grows, the role of
thermal fluctuations exceeds quantum ones. However, it cannot kill
the quantumness. As we know, entanglement is not the only part of
quantum correlations. Based on Ollivier and Zurek's argument
\cite{zurek01}, it shows that absence of entanglement does not imply
classcality. If you want to destroy the quantumness of a system, you
should use the process of decoherence. Decoherence will eventually
make the quantumness of a system to be zero and then the system will
be in totally classical state. Since QD measures total quantum
correlations, it is obvious that even temperature is high, without
introducing the decoherence, quantum correlations will not vanish.
So in this sense, QD is robust than entanglement at a finite
temperature.
Focusing now at panel (a) of Fig. 1, when temperature $T$ increases,
thermal fluctuation takes control of the system in this region, and
the role of coupling constant such as $J$, $J_z$ and $D_z$ become
weaker. So we can see the lines in panel (a) of Fig. 1 become
concentrated together. Now we enhance the parameter $J_z$ to 1.0 as
we can see in panel (c) of Fig. 1. We find that at the same
temperature, lines become more separate than the case of panel (a).
This effect is because of the influence of coupling constant $J_z$
makes the quantumness considerable. Next, let us take a look at
panels (a) and (c) again and consider the influence of parameter
$D_z$ to thermal QD. Due to the symmetry of DM interaction
parameter, we can only take the strength of $D_z$ into account. So
here only consider the $D_z>0$ case. Apparently, we can see a
remarkable phenomenon (Also shown in Fig. 3) in certain region that
when the parameter $D_z$ increases, thermal QD decreases while this
trend of QD is in contrast to the behavior of entanglement. Since QD
and entanglement are both kinds of measure of quantum correlations.
This is very wired. Worth to be mentioned, this phenomenon appears
in $D_x$ case as well, and we will analyze this in detail in Sec.
IV.
Moreover, take a look at Fig. 2, where we fix $J=1$ and $D_z=1$.
Asymptotical phenomenon of QD also appears at finite temperature for
different $J_z$ in panel (a). However, we can see that QD increases
with increasing of $J_z$. This is corresponding to the case of
entanglement shown in panel (b). Here, $J_z$ as a anisotropic
coupling constant of Heisenberg model, plays the same role in
thermal QD as in the case of entanglement.
\begin{figure}
[ptbh](a)
\begin{minipage}{2.4in}
\includegraphics[
width=2.4in,height=1.4in
{fig.4.eps}
\end{minipage}
\hspace{0.0in}
(b)
\begin{minipage}{2.4in}
\includegraphics[
width=2.4in,height=1.4in
]
{fig.5.eps}%
\end{minipage}
\caption{The QD (a) and concurrence (b) versus T in the system with $z$-component parameter $D_z$. Here $J=1$, $D_z=1$.
(a) Here for the thick line $J_z=3.0$, normal line $J_z=2.0$ and dashed line $J_z=1.0$.
(b) Here for the thick line $J_z=3.0$, normal line $J_z=2.0$ and dashed line $J_z=1.0$.}
\end{figure}
\begin{figure}
\includegraphics[width=7cm]{fig.6.eps}
\caption{(Color online) The quantum discord is plotted versus $T$
and $D_z$ where $J=1$ and $J_z=0.2$.}
\end{figure}
\subsection{XXZ Heisenberg model with DM interaction parameter $D_x$}
Here we consider the case of the two-qubit anisotropic Heisenberg
XXZ chain with DM interaction parameter $D_x$. The Hamiltonian $H'$
of this model reads
\begin{equation}
H'=J\sigma_1^x\sigma_2^x+J\sigma_1^y\sigma_2^y+J_z\sigma_1^z\sigma_2^z+D_x(\sigma_1^y\sigma_2^z-\sigma_1^z\sigma_2^y).
\end{equation}
where $D_x$ is the $x$-component parameter of the DM interaction and
other parameters are the same as in part A of this section.
Following the exact same approach in part A, we can get
\begin{equation}
H' = \left(
\begin{array}{cccc}
J_z&iD_x&-iD_x&0\\
-iD_x&-J_z&2J&iD_x\\
iD_x&2J&-J_z&-iD_x\\
0&-iD_x&iD_x&J_z
\end{array}
\right)
\end{equation}
\begin{equation}
\rho'(T) =\frac{1}{2Z'} \left(
\begin{array}{cccc}
\mu_+ &-\xi &\xi &\mu_-\\
\xi &\nu_+ &\nu_- &-\xi\\
-\xi &\nu_- &\nu_+ &\xi\\
\mu_- &\xi &-\xi &\mu_+
\end{array}
\right)
\end{equation}
where $w'=\sqrt{(J+J_z)^2+4D_x^2}$
$\phi=\arctan(\frac{2D_x}{J+J_z-w'})$
$\varphi=\arctan(\frac{2D_x}{J+J_z+w'})$, and
$\mu_\pm=e^{-\beta
J_z}\pm(e^{\beta(J-w')}\sin^2{\phi}+e^{\beta(J+w')}\sin^2{\varphi})$
$\nu_\pm=e^{\beta
(J_z-2J)}\pm(e^{\beta(J-w')}\cos^2{\phi}+e^{\beta(J+w')}\cos^2{\varphi})$
$\xi=ie^{\beta(J-w')}\sin{\phi}\cos{\phi}+ie^{\beta(J+w')}\sin{\varphi}\cos{\varphi}$
$Z'=2e^{-\beta J}\cosh[\beta(J-J_z)]+2e^{\beta J}\cosh(\beta w')$.
Also we can get $\rho'^A$ and $\rho'^B$ via straightforward
calculation: $\rho'_A=\rho'_B=\frac{1}{2Z'}(\mu_++\nu_+)I$ where $I$
is the identity operator.
Now, first we look at Fig. 4, where $J=1$ and $D_x=1$. We can see
that when $J_z$ increases, the thermal QD becomes higher. This trend
of QD is the same like entanglement case. (For better comparison we
choose the exact the same value of parameters in Fig. 4 as in Ref.
\cite{li08}) And also, this fits with the behavior of $J_z$ in the
system with $z$-component parameter $D_z$.
While looking at Fig. 5, we find that besides the asymptotical
phenomenon mentioned above, we can also find the opposite trend of
thermal QD and entanglement. So here we can derive that different DM
parameters can both influence the behavior of QD and entanglement
and make them move towards different directions in certain region.
We will address this in Sec. IV.
In addition, back to the Fig. 1 and Fig. 5 which demonstrate the
behaviors of thermal QD in aniosotropic Heisenberg model with
parameters $D_z$ and $D_x$. It is obvious there are some differences
between two different cases. First, let us look at panel (a) of Fig.
1 and panel (b) of Fig. 5, we find that in the situation where other
parameters are exactly the same, $D_x$ is more reliable than $D_z$
when T becomes higher. This indicates that $D_z$ is easy to
influence by the increasing of $T$ (This also shown in Fig. 3 and
Fig. 6). And from Fig. 1 and Fig. 5 we can find that if we tune the
parameter $J_z$ larger, thermal QD depending on different $D_x$
change more obviously than the case of parameter $D_z$. This means
that parameter $D_x$ is more sensitive in certain condition along
with change of parameters such as $J_z$. So we can derive from this
results that it is more efficient to use parameter $D_x$ to control
the QD. In addition, we can also find that when tune parameter $J_z$
to 1.0, the two plots (panel (c) of Fig. 1 and panel (a) of Fig. 5)
become similar, which is mainly because the DM interactions now are
not the dominating parameters controlling the behavior of the
system.
\begin{figure}
\includegraphics[width=7cm]{fig.7.eps}
\caption{The QD versus T in the system with $x$-component parameter
$D_x$. Here $J=1$, $D_x=1$ and for the thick line $J_z=0.9$, normal
line $J_z=0.4$, dashed line $J_z=0$.}
\end{figure}
\begin{figure}
[ptbh](a)
\begin{minipage}{2.4in}
\includegraphics[
width=2.4in,height=1.4in
{fig.8.eps}
\end{minipage}
\hspace{0.0in}
(b)
\begin{minipage}{2.4in}
\includegraphics[
width=2.4in,height=1.4in
]
{fig.9.eps}%
\end{minipage}
\caption{The QD versus T in the system with $x$-component parameter $D_x$. Here $J=1$.
(a) Here $J_z=1.0$ and for the thick line $D_z=1.0$, normal line $D_z=0.7$, dashed line $D_z=0.5$.
(b) Here $J_z=0.2$ and for the thick line $D_z=1.0$, normal line $D_z=0.7$, dashed line $D_z=0.5$. }
\end{figure}
\begin{figure}
\includegraphics[width=7cm]{fig.10.eps}
\caption{(Color online) The quantum discord is plotted versus $T$
and $D_x$ where $J=1$ and $J_z=0.2$.}
\end{figure}
\section{Discussion about opposite tendency of thermal QD and entanglement}
In this section, we will discuss the opposite tendency of thermal QD
and entanglement appeared in Sec. III.
In previous section, we find that, in certain regions, when
parameter $D_z$ increases, thermal QD decreases while at the same
time thermal entanglement increases. This also appears in the case
of parameter $D_x$. So we say that when we tune the parameters of DM
interaction, entanglement and QD have the opposite trend of
movement. Obviously, this notable phenomenon is very wired since
both QD and entanglement quantify the quantumness of correlations.
There are some underlying reasons about that. Although QD and
entanglement both kinds of measure of quantum correlations.
Entanglement is a non-local correlation while QD figures a measure
of whole quantumness of correlations for a pairwise system. There
are some other quantum correlations without entanglement which can
be characterized by QD but not any kinds of entanglement measure
such as concurrence. And we reveal that these other quantum
correlations have remarkably different characteristic from the
entanglement. Recent researches \cite{dat08,lanyon08,kni98} also
suggest that these other quantum correlations can also be an
important resource and maybe take respondence in mixed state quantum
computation. Secondly, DM interaction is a magnetic term arising
from the spin-orbit coupling. The staggered DM interactions bring on
helical magnetic structures due to its form
$\mathbf{D}\cdot[\mathbf{S_1}\times\mathbf{S_2}]$, where
$\mathbf{S_1}$, $\mathbf{S_2}$ denote the spins and $\mathbf{D}$ is
DM vector. DM interaction can contribute the strong planar quantum
fluctuations which pose the alignment ordering to both qubit-1 and
qubit-2 in two-qubit system. Some researches
\cite{jafari08,kargarian09} have shown that in certain condition, DM
interaction can build entanglement. So it is acceptable that
entanglement increase with increasing of DM interaction in this
paper.
Now we can see that since QD figures the whole quantum correlations
including entanglement, the phenomenon here indicate that, the other
quantum correlations without entanglement we mentioned above
decrease rapidly (more rapidly than the increase of entanglement at
the same time) with increasing of parameters of DM interaction. So
we can see from Fig. 1 and Fig. 5. that when QD decreases while the
entanglement increases with the increasing of DM interactions. Here,
this phenomenon can coincide what we have raised above that the
characteristic of QD is exotic. Furthermore, using this
characteristic, under certain conditions, we can tune the parameters
of DM interactions to increase the entanglement of a system and
meanwhile decrease the other quantum correlations of the same
system. The advantage of the method is obvious. If you want to take
entanglement as a quantum resource to do quantum tasks, you want to
enhance your entanglement resource and make the other quantum
correlations of system as less as possible, this approach would
offer a possible solution.
\section{Conclusion}
The thermal QD of a two-qubit anisotropic Heisenberg XXZ chain with
two distinct DM interaction parameters is investigated. We find that
thermal QD is more robust than thermal entanglement versus
temperature $T$. And numerical results indicate that it is more
efficient to control QD in the model with parameter $D_x$ than with
parameter $D_z$. Furthermore, with the increasing of parameters
$D_z$ and $D_x$, thermal QD varies oppositely to the case of
entanglement. This phenomenon reveals that quantum correlations
without entanglement have quite exotic characteristic from
entanglement. In addition, this effect also offers a solution to
make the entanglement of a system enhanced.
\begin{acknowledgments}
The work is supported in part by the NNSF of China Grant No.
90503009, No. 10775116, and 973 Program Grant No. 2005CB724508.
\end{acknowledgments}
|
\section{Introduction}
In recent years, the ADS/CFT correspondence \cite{adscft_Maldacena,adscft_gubser,adscft_witten} has become a powerful tool to study various properties of the strong coupling limit of conformal field theories, with applications to QCD and also more recently to some aspects of condensed matter physics. In principle, one has to differentiate between top-down setups that are constructed within string theory and imply consistency -- and bottom-up setups, where the gravitational duals are constructed from a phenomenological point of view. In this paper, we use the former approach as we would like to explore what happens to a particular consistent theory.
In condensed matter applications, there has been particular interest in 2+1 dimensional systems that can be typically constructed using M2 branes \cite{pavel}, or as a defect in a 3+1 dimensional background using D3-D5 \cite{fancycon,conpaper} and D3-D7 \cite{fancycon,conpaper,rey} intersections -- and also as bottom-up setups in various contexts such as superconductivity. As our world is 3+1 dimensional, the defect setup may be more realistic, even though there are some problems with the consistency of the D3-D7 systems \cite{conpaper,fancycon}.
Recently, there has also arisen significant interest in fermi-liquid-like aspects at low temperature compared to the density, for example in a 3+1 dimensional D3-D7 setup \cite{zerosound} or a 1+1 dimensional D3-D3 configuration \cite{luttinger}. This low-temperature limit will also be in the focus of this paper.
A common way how to introduce fundamental matter in ADS/CFT is in probe brane configurations, where one considers an $AdS$ blackhole background. For example one considers the well-known $AdS_5\times S^5$ (black hole) solution (above the deconfinement phase transition) from the decoupling limit of a stack of $N_c\gg 1$ D3 branes, that is dual to a (thermal) $SU(N_c)$ $\mathcal{N}=4$ supersymmetric Yang-Mills theory \cite{adscft_gubser}. Then one inserts $N_f \ll N_c$ intersecting Dp branes, giving $N_f$ families of (charged) fields in the fundamental representation of the $SU(N_c)$, living along the directions of the intersection \cite{winters,robfirst} -- such as the above-mentioned D3-Dp intersections.
In these setups, the mass $M_q$ of the fundamental matter is usually given by a scalar describing the size of the compact sphere of the probe brane geometry inside the $S^5$ \cite{winters}. Typically large masses or small temperatures correspond to a vanishing size of the compact sphere away from asymptotic infinity, i.e. to a ``narrow funnel'' that depends for example on the baryon number density $\rho$.
However it was noted in \cite{fancytherm}
that in
$d +1$ dimensional probe brane setups
there is a coincidence in the scalings at finite $\rho$ and $M_q$: The relevant parameters are dimensionless with appropriate powers of the temperature and at small temperatures the dimensionless mass scales as
$\frac{M_q}{T} \sim \alpha \left(\frac{\rho}{T^{d}}\right)^{1/d}$, where the proportionality constant $\alpha$ depends on the ``size of the funnel''.
Hence, the exponents of $T$ cancel and the size of the compact sphere remains finite and independent of temperature.
This can be visualized in the usual picture by the extra tension or "stiffness" of the probe branes from the increasing $\frac{\rho}{T^d}$, as $T$ decreases, which will be illustrated in figure \ref{physembed}.
This scaling coincidence suggests that there may be some interesting low-temperature physics in this limit, and hence we will explore the details of the embeddings and their thermodynamics and related physical properties. We will mostly consider the 2+1 dimensional defect setup, but eventually we will also briefly look at the results in 3+1 and 1+1 dimensions, and point out some important differences.
This paper is organized as follows: First, we review the string theory setup and the necessary ADS/CFT dictionary of the D3-D5 defect in section \ref{setup}. Then, we will demonstrate the scaling solution for the probe-brane embedding that is central to this paper in section \ref{limit} where we also comment on the relevant thermodynamic variables to consider in the massive case. In section \ref{thermo}, we first obtain the leading terms of the free energy, then discuss the thermodynamics/statistical mechanics in 2+1 dimensions and finally study the subleading terms that give rise to the heat capacity. Throughout this section, we will see how the system interpolates between a mass-dominated limit and a density-dominated limit. Finally, we will briefly discuss the generalized 3+1 and 1+1 dimensional systems in section \ref{gencas};
and then we will discuss the results and conclude in section \ref{conclu}.
\section{Setup}\label{setup}
We start with the supergravity background of a planar black hole in $AdS_5$,
\begin{eqnarray}
ds^2 & =& \frac{r^2}{L^2} \left( -(1-r_0^4/r^4)dt^2 +d\vec{x}_3^2\right)
+ \frac{L^2}{r^2} \left( \frac{dr^2}{1-r_0^4/r^4} +r^2 d\Omega_5^2\right) \
, \ \ \ C^{(4)}_{txyz}=-\frac{r^4}{L^4} \labell{D3geom} \ .
\end{eqnarray}
This
corresponds to the decoupling limit of $N_c$ black D3-branes
dual to ${\cal N}=4$ $SU(N_c)$ super-Yang-Mills theory at finite temperature $T$, living along the flat directions of the AdS, \cite{adscft_witten}.
The temperature $T$ is given by the Hawking temperature $T = \frac{r_0}{\pi L^2}$ and the Yang-Mills coupling by $g_{ym}^2 = 4 \pi g_s$. Since the curvature $L$ is
given in terms of the string coupling $g_s$ and
string length $l_s$ as $L^4 = 4\pi\, g_s N_c \, l_s^4$, the 't Hooft
coupling $\lambda=g_{ym}^2 N_c$ can be written as $\lambda=\frac{L^4}{l_s^4}$. Hence the ``supergravity limit'' $L \gg l_s$ in which the type IIB supergravity action and the solution \reef{D3geom} are valid corresponds to strong coupling $\lambda \gg 1$.
In practice, however, we will use coordinates that are made dimensionless with factors of $\frac{L^2}{r_0} =\frac{1}{\pi T}$, denoted by $\tilde{(\cdot)}$, such as $\tlt: = \frac{r_0 \, t}{L^2} = t\, (\pi T)$, and the inverse dimensionless radius $u := \frac{r_0}{r}$, giving us
\begin{equation}\label{branemetric}
ds^2 \, = \, L^2 u^2\left(-(1-u^4)d\tlt^2 + d\vec{{\tilde{x}}}_3^2 + \frac{du^2}{1-u^4} + d\Omega_5^2 \right) \ .
\end{equation}
Because in this setup all the fields transform in the adjoint representation of the $SU(N_c)$ its use in QCD or condensed matter physics is very much limited as one would like to consider also matter that is charged under this symmetry, i.e. that transforms in the fundamental representation.
To introduce the fundamental matter one then creates an intersection of ``probe'' Dp branes with the D3 branes, such that at the string theory side there are fields at the massless level of field theory at the intersection. From the point of view of the probe branes, they correspond to endpoints of D3-Dp strings, and in the gravity side they correspond to fundamental fields in the (defect) field theory.
Here, we use the well- known D3-D5 defect setup (see e.g. \cite{lisa,hirosi,johannadef}):
\begin{equation}
\begin{array}{rccccc|c|cccccl}
& & 0 & 1 & 2 & 3 & 4& 5 & 6 & 7 & 8 & 9 &\\
& & t & x & y & z & r& & & & & \theta &\\
\mathrm{background\,:}& D3 & \times & \times & \times & \times & & & & & & & \\
\mathrm{probe\,:}& D5 & \times & \times & \times & & \times & \times & \times & & & & \ \ \ .
\end{array}
\labell{array}
\end{equation}
The dual field
theory is now the SYM gauge theory coupled to $N_f$ fundamental
hypermultiplets, which are confined to a (2+1)-dimensional defect.
This construction is still supersymmetric, but the supersymmetry has been reduced from ${\cal N}=4$ to
${\cal N}=2$ by the introduction of the defect. In the limit $N_f
\ll N_c$, the D5-branes may be treated as probes in the
supergravity background, i.e. we may ignore their gravitational
back-reaction.
Here we want to turn on only the overall $U(1)$ factor of the world-volume gauge field, so the the probe branes are governed by the DBI action
\begin{equation}\label{braneaction}
S \, = \, - T_5 N_f \int_{D5} \sqrt{- det(P[G] + 2\pi l_s^2 F )} \ ,
\end{equation}
evaluated in the D3 background \reef{D3geom}.
We assume also translational invariance along the flat directions and rotational invariance on the sphere. Hence, the pullback in \reef{braneaction} gives us one scalar field corresponding to the position in the $z$ direction, which was extensively studied in \cite{conpaper,fancycon}, and another scalar which describes the size of the compact sphere and corresponds to turning on the mass of the fundamental matter, studied in \cite{fancytherm,fancycon} and more extensively in the similar D3-D7 system in \cite{long,johanna,ingo,robdens,robchem}. Parametrizing the $S^5$ as
$d\Omega_5^2 = d\theta^2 + \sin^2 \theta\, d\Omega_2^2 + \cos^2 \theta\, d\Omega_2^2$ and putting the brane on the first $S^2$ of the $S^5$,
the induced metric on the probe branes is given by
\begin{equation}\labell{branemet}
ds^2 \, = \, \frac{L^2}{u^2} \left( -(1- u^4)d\tlt^2 + d\vec{{\tilde{x}}}_2^2 + \left(\frac{1}{1-u^4}+ \frac{u^2 \Psi'(u)^2}{1-\Psi(u)^2} \right) du^2 \, + \, u^2 (1-\Psi(u)^2) d\Omega_2^2\right) \ ,
\end{equation}
where we defined the scalar as $\Psi(u) = \sin \, \theta(u)$.
We also choose to turn on the world-volume $U(1)$ gauge field which is dual to the $U(1)$ current operator of the $U(N_f)$ that gives rise to the ``flavor symmetry''. In particular, the flux of the electric field
\begin{equation}
F \ = \ \partial_u A_{t}(u) \, du \wedge dt \ ,
\end{equation}
corresponds in the field theory side to the baryon number density (see e.g. \cite{bigrev})
\begin{equation}\label{rhodic}
\rho \ = \ \big<\!\ J^t \big> \ = \ - \frac{1}{N_c}\frac{\delta S}{\delta A_{t}^{bdy.}} \ = \ 4\pi N_f \frac{L^2 T_5}{N_c r_0} \lim_{u\rightarrow 0} \partial_u A_t(u) \ .
\end{equation}
In this parametrization, the DBI action becomes
\begin{equation}\labell{backact}
S \ = \ 4 \pi L^2 T_5 N_f \int\! d\sigma^4 \sqrt{-\det P[G]_4}\sqrt{1-( 2\pi l_s^2 \partial_u {A}_t(u))^2P[G]^{tt}P[G]^{uu} }\, (1-\Psi^2)
\end{equation}
where the integral is taken over the $AdS_4$ part of the world-volume.
Using this action, it is straightforward to obtain the solution for the gauge field
\begin{equation}\labell{backgd}
\ \partial_u {A}_t(u) \ = \ \sqrt{\lambda} T \frac{{\tilde{\rho}}\sqrt{1- \Psi^2(u) + u^2 h(u) \Psi'(u)^2}}{\sqrt{1-\Psi(u)^2}\sqrt{ {\tilde{\rho}}^2 u^4 + (1-\Psi(u)^2)^2}} \nonumber,
\end{equation}
where ${\tilde{\rho}} := \frac{\rho}{ N_f T^2}$.
The equation of motion for $\Psi(u)$ becomes
\begin{eqnarray}\labell{psieom}
\!\!\!\!\!\!\!\!\!\!\frac{2(1-\Psi^2)^3 +u^2 (1-u^4)\big({\tilde{\rho}}^2 u^4+(1-\Psi^2)^2 \big)\Psi'^2}{u^4 (1\!-\!\Psi^2)\sqrt{(1\!-\!\Psi^2)\big(1\!-\!\Psi^2+(u^2\!-u^6)\Psi'^2 \big)\big(1+{\tilde{\rho}}^2 u^4 + \Psi^2(\Psi^2\!-2)\big)}} \nonumber \\
\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\! = \
\partial_u \left(\Psi' \frac{1-u^4}{u^2}\sqrt{\frac{{\tilde{\rho}}^2 u^4 + (1-\Psi^2)^2}{(1-\Psi^2)(1-\Psi^2+(u^2-u^6)\Psi'^2)}} \right)~~~~~~~~~~~~~
\ ,
\end{eqnarray}
and on the horizon $u=1$, the equations reduce to
\begin{equation}\labell{psibdy}
\lim_{u\rightarrow 1^-} \partial_u \Psi\, = \, \frac{1}{2}\frac{\Psi_0(1-\Psi_0^2)^2}{{\tilde{\rho}}^2 + (1-\Psi_0^2)^2} \ , \ \mathrm{where} \ \Psi_0 \, = \, \lim_{u\rightarrow 1^-} \Psi\ .
\end{equation}
The asymptotic solution at $u \rightarrow 0$ is
\begin{equation}\labell{psiasym}
\Psi \, \sim \, {\tilde{m}} \, u \, + \, {\tilde{c}} \, u^2 \ ,
\end{equation}
where ${\tilde{m}}$ and ${\tilde{c}}$ are dimensionless free parameters that are determined by the value on the horizon $\Psi_0$ and \reef{psibdy} that acts as a second boundary condition on the horizon. Following arguments of the T-dual $(3+1)$ dimensional D3-D7 setup \cite{long,johanna,ingo,robdens,robchem}, the quark mass $M_q$ and dual condensate $C$ are given by
\begin{equation}
M_q \, = \, \frac{r_0 \,{\tilde{m}}}{2 \pi l_s^2} \, = \, \frac{\sqrt{\lambda}}{2} T {\tilde{m}}\ \ \ \mathrm{and} \ \ \ C\, = \, T^2 N_f N_c{\tilde{c}}\ .
\end{equation}
One can understand this identification of the mass from the separation between the D3 and D5 branes in flat space, such that $M_q$ is the mass of a stretched D3-D5 string and the condensate is just the thermodynamic dual of the mass.
At vanishing density, there is a critical temperature-mass ratio below which the probe branes do not extend down to the horizon \cite{johannaphase,long}. At finite densities, and unless one turns on the scalar in the $z$ direction considered in \cite{fancycon,fancytherm}, the brane always extends down to the horizon even though a phase transition may still be observed at small densities \cite{fancytherm}. In the limit considered in this paper, however, this phase transition is of no concern.
In general, equation \reef{psieom} has no analytical solution and \reef{psibdy} implies that one has to start integrating the equation from the horizon to obtain the mass and the related condensate at the asymptotic boundary, rather than setting either of them first. We show the full numerical solutions for $\Psi(u)$ in figure \ref{physembed}, where we illustrate the ``physical'' embedding $\left(\frac{1}{u} \cos\theta ,\frac{1}{u} \sin\theta\right)$ of the probe brane for both fixed $\rho/T^2$ and fixed $\rho/M_q^2$ to support the physical interpretation outlined already in the introduction.
\begin{figure}
\includegraphics[width = 0.99\textwidth]{fullembed.pdf}
\caption{The full ``physical embedding'' of the probe brane, represented by the ``radial'' direction of the bulk spacetime $\frac{1}{u}\cos\, \theta$ and the ``separation'' $\frac{1}{u}\sin\, \theta$ for fixed ${\tilde{\rho}} = 1$ (at fixed temperature and varying mass ${\tilde{m}}$, black) and fixed ${\bar{\rho}} = \frac{{\tilde{\rho}}}{{\tilde{m}}^2} = 1$ (at fixed mass and varying temperature ${\bar{T}} = \frac{1}{{\tilde{m}}}$, gray). The corresponding curves of both cases should meet at asymptotic infinity, $\frac{1}{u}\cos\theta \rightarrow \infty$}\label{physembed}
\end{figure}
In the following section, however, we will demonstrate a particular case in which the equation of motion for $\Psi$ simplifies dramatically.
\section{Taking the limit}\label{limit}
In this section, we obtain the scaling solution in a ``blow-up region'' near asymptotic infinity $u\rightarrow 0$ at small temperatures and finite density and mass, i.e. ${\tilde{\rho}} \gg 1$, ${\tilde{m}} \gg 1$.
Before we do so, we need to first consider an approximate solution at finite values of $u$ in order to match the asymptotic solution to the black hole horizon.
Hence, let us consider a linear expansion around an approximately finite size funnel, $\Psi(u) \ = \ \Psi_0 + \psi(u)$, which reduces \reef{psieom} to
\begin{equation}
2\Psi_0(1- \Psi_0^2)\ =\ {\tilde{\rho}}^2 u^6 \partial_u \left((1-u^4) \partial_u \psi(u)\right) \ +\ {\mathcal{O}}(\psi^2)\ .
\end{equation}
Enforcing the boundary condition \reef{psibdy} and $\psi = 0$ at the horizon gives the solution
\begin{equation}\labell{psilogsol}
\psi(u) \, = \, \frac{-\Psi_0(1-\Psi_0^2)}{10\, {\tilde{\rho}}^2 u^4}\left(1-u^4 \, - 2 u^4\tan^{\! -1}\frac{1-u}{1+u} + u^4 \ln\frac{(1+u^2)(1+u)^2}{8 \, u^4} \right) \,+ {\mathcal{O}}({\tilde{\rho}}^{-4}) \ ,
\end{equation}
which is negative over $u\in ]0,1[$ and becomes finite at small $u\sim {\mathcal{O}}(1/\sqrt{{\tilde{\rho}}})$ as it should.
Now we wish to find some overlap of this solution with a solution in the blow-up region at small $u$ that we can match around $\frac{1}{\sqrt{{\tilde{\rho}}}} \ll u \ll 1$. To do so, we define the ``blow-up'' coordinate $\xi:=u \sqrt{{\tilde{\rho}}}$, which maps the horizon to large $\xi = \sqrt{{\tilde{\rho}}} \sim T^{-1} \gg 1$, and the scale over which most of the change in $\Psi$ takes place to $\xi \in ]0,{\mathcal{O}}(1)]$. Then, we consider all of $\xi\in \mathbb{R^+}$ and expand for large ${\tilde{\rho}}$.
The leading term (up to order ${\tilde{\rho}}^{-2}$) is written in the most compact form as
\begin{eqnarray}
0 & = & \partial_\xi\left(\sqrt{1-\Psi(\xi)^2}\sqrt{\xi^4 + (1-\Psi(\xi)^2)^2} \partial_\xi \left(\frac{\Psi(\xi)}{\xi}\right) \right) \\ \nonumber & & + \
\frac{\partial_\xi\Psi(\xi)}{\Psi(\xi)} \left(\xi \partial_\xi \left(\frac{\Psi(\xi)}{\xi}\right)\right)^2
\partial_{\Psi(\xi)}\left(\sqrt{1-\Psi(\xi)^2}\sqrt{\xi^4 + (1-\Psi(\xi)^2)^2} \right) \ ,
\end{eqnarray}
where the erratic solution $\Psi = m \, \xi$ has to be ignored.
It is straightforward to show that the expansion for large $\xi$ is indeed consistent with $\Psi = \Psi_0 - \frac{\Psi_0 (1-\Psi_0^2)^2}{10 \xi^4}$, i.e. with the small-$u$ limit of \reef{psilogsol}. In practice for the numerics, we have to choose $\xi^{-4}$ as a coordinate and implement this matching as a boundary condition as $\xi^{-4} \rightarrow 0$.
Because we want to consider fixed mass, but vary the temperature and density, the suitable physical quantities to consider at finite mass are
\begin{eqnarray}
{\bar{T}} & =& \frac{1}{{\tilde{m}}} \ = \ \sqrt{\lambda} \frac{T}{2 M_q} \ , \\
{\bar{\rho}} & = & \frac{{\tilde{\rho}}}{{\tilde{m}}^2} \, =\, \frac{ \rho}{ N_f}\frac{\lambda}{4 M_q^2} \ \mathrm{and} \\
{\bar{c}} & =& \frac{{\tilde{c}}}{{\tilde{m}}^2} \ =\ \frac{C}{ N_c N_f}\frac{\lambda}{4 M_q^2} \ ,
\end{eqnarray}
where the numerical coefficients are chosen to give
us parameters that are straightforwardly related to the embedding.
Using \reef{rhodic}, \reef{psiasym} and the definition of $\xi$, they are now related to the scalar by
\begin{equation}
{\bar{\rho}} \, = \, (\partial_\xi \Psi |_{\xi\rightarrow 0})^{-2} \, , \ \ \ {\bar{c}}\, =\, {\bar{\rho}} \frac{1}{2}\partial_\xi^2 \Psi |_{\xi\rightarrow 0} \ \ \mathrm{and} \ \ \ {\bar{T}}\, =\, \sqrt{{\bar{\rho}}/{\tilde{\rho}}} \ .
\end{equation}
It turns out that there seems to be an exact relation $\partial_\xi^2 \Psi |_{\xi\rightarrow 0} = - 2 \Psi_0$, such that ${\bar{c}} =- {\bar{\rho}}\, \Psi_0$.
For small $\Psi_0$, we can expand the equations of motion to linear order,
\begin{equation}
0 \ = \ 2\Psi \, - \, 2 \xi \partial_\xi \Psi \, + \, \xi^2(1+\xi^4)\partial_\xi^2\Psi \ .
\end{equation}
This has an analytic solution in terms of the elliptic integral of the first kind, $\mathcal{F}(\phi|k) = \int_0^\phi \frac{d\varphi}{\sqrt{1-k^2\sin^2\varphi}}$ and the complete elliptic integral $\mathcal{K}(k)=\mathcal{F}\!\left(\frac{\pi}{2}\big|k\right)$
\begin{equation}\labell{psiK}
\Psi \ = \ \Psi_0 \xi \left( e^{i\pi/4} \mathcal{F} \left(i \sinh^{-1}(e^{i\pi/4} \xi) \Big| i\right) \ + \ \mathcal{K}(1/\sqrt{2})\right) \ ,
\end{equation}
which is itself not particularly interesting because it just interpolates between the linear behavior at small $\xi$ and the constant $\Psi_0$ at large $\xi$; and since it is the linear expansion in $\Psi_0$ it tells us that at large densities the relevant scales in $\xi$ do not depend on the density. The profile can be seen in the curves with large ${\bar{\rho}}$ on the right in fig. \ref{emplot}.
The important point is that because $\forall k: \, \mathcal{F}(x|k) \sim x + {\mathcal{O}}(x^2)$, we can straightforwardly read off the first and second derivatives at $\xi \rightarrow 0$ to find that
\begin{equation}
\Psi_0 \, \sim\, \frac{1}{\sqrt{{\bar{\rho}}} \, \mathcal{K}(1/\sqrt{2})}\ \ \mathrm{and} \ \ \ {\bar{c}} \, \sim\, -\Psi_0\, {\bar{\rho}}\, \sim \, \frac{\sqrt{{\bar{\rho}}}}{ \mathcal{K}(1/\sqrt{2})} \ .
\end{equation}
In the regime of $\Psi -1 \ll 1$, however, one cannot find such a simple expression. By studying the equations of motion, one can only argue that the transition from the large-$\xi$ solution to the asymptotic solution occurs at $\xi\sim{\mathcal{O}}(1-\Psi_0)^{(1/4)}$, such that $1-\Psi \propto {\bar{\rho}}^2$.
The corresponding value of the condensate is ${\bar{c}} = - {\bar{\rho}} + {\mathcal{O}} ({\bar{\rho}}^3)$ for small values of ${\bar{\rho}}$.
The full numerical dependence $\Psi_0({\bar{\rho}})$ and these limiting cases are shown in figure \mref{emplot}, where we also show the corresponding embeddings $\Psi_{{\bar{\rho}}}(\xi)$.
\begin{figure}
\includegraphics[width = 0.49\textwidth]{psiofrho.pdf}
\hspace{0.02\textwidth}
\includegraphics[width = 0.49\textwidth]{embeddings.pdf}
\caption{Left: $\Psi_0({\bar{\rho}})$ with the approximations at small and large ${\bar{\rho}}$. Right: $\Psi(\xi)$ for various values of ${\bar{\rho}}$.}\label{emplot}
\end{figure}
We further illustrate\footnote{The color scheme is an artist's impression to illustrate the similarity with the flowers of Angel's Trumpets ({\it Brugmansia}).}
the ``size'' of the sphere, $\cos \theta = \sqrt{1-\Psi^2}$ in figure \ref{fancyplot}.
\begin{figure}
\includegraphics[width = 0.49\textwidth]{angelc1.jpg}
\hspace{0.02\textwidth}
\includegraphics[width = 0.49\textwidth]{angelc2.jpg}
\caption{Illustration
of the size of the temperature-independent size $\cos\,\theta = \sqrt{1-\Psi^2}$ of the compact $S^2$ (only $S^1$ shown here) in the asymptotic blow-up region as a function of the inverse radial coordinate $\xi$, for $\Psi_0 = 0.95$. The color scheme is an artist's impression to illustrate the similarity with the flowers of Angel's Trumpets ({\it Brugmansia}).}\label{fancyplot}
\end{figure}
\section{Thermodynamics}\label{thermo}
In order to study the thermodynamic quantities, we follow the standard procedure of obtaining the free energy at fixed quark mass as a function of density and temperature, $F=F_{M_q}(\rho,T)$, and then computing its derivatives.
The extrinsic thermodynamic quantities of the defect should be considered as a contribution to the overall 3+1 dimensional system.
In \cite{fancytherm}, it was argued, however,
that
they can be discussed independently, since the asymptotic properties of the bulk do not depend on the details of the defect.
As usual, we consider all extrinsic quantities densitized and will comment on obtaining the pressure separately.
\subsection{Free Energy}
In \cite{fancytherm}, we computed the free energy density for this system via the standard procedure from the Euclidean action, $I_e$, \cite{long} using
\begin{equation}\mlabel{freeendef}
F \ = \ T \tilde{I}_e \ ,
\ \ \ \tilde{I}_e \ = \ \int_{u_{min}}^{u_{max}} \mathcal{L}_e \ + \ I_{bdy} \ + \ \rho A_t\ ,
\end{equation}
where the boundary terms
\begin{equation}\label{ibdy}
I_{bdy.} \ = \ - \frac{1}{3} \sqrt{\gamma} + \frac{1}{2}\Psi^2\sqrt{\gamma}
\end{equation}
are dictated to us by consistency \cite{skenderis}. The Legendre transformation needs to be done in order to obtain $\rho$ as a thermodynamic variable, rather than the chemical potential (see e.g. \cite{robdens,fancytherm} for a more detailed discussion).
At large temperatures, i.e. ${\tilde{m}},{\tilde{\rho}} \ll 1$, this evaluates to
\begin{equation}
\bar{F} \ := \ F\frac{\lambda}{ 4N_c N_f M_q^3} \ = \ -\frac{1}{3} {\bar{T}}^3 \ ,
\end{equation}
with the corresponding entropy $\bar{S} = {\bar{T}}^2$ and heat capacity $\bar{c}_V = 2 {\bar{T}}^2$. Note that the unusual factor $\lambda$ arises because of the definition of ${\bar{T}}$.
To compute the value of the free energy in our case, we can again use $\xi=u \sqrt{{\tilde{\rho}}}$ and expand for large ${\tilde{\rho}}$ to obtain
\begin{equation}\labell{keyeq}
\bar{F} \ = \ {\bar{\rho}}^{3/2}\left( \int \frac{d\,\xi}{\xi^4}\sqrt{\xi^4+(1-\Psi^2)^2}\sqrt{1+\frac{\xi^2\Psi'^2}{1-\Psi^2}} \ - \ \left.\frac{2-3\Psi^2}{6 \xi^3}\right|_{bdy.}\right) \ + \ {\mathcal{O}}({\tilde{\rho}}^{-1/2}) \ .
\end{equation}
This has to be evaluated over $\xi \in ]0,\sqrt{{\tilde{\rho}}}]$, and we will try to split it into an integral over $\xi \in ]0,\infty [$ and $\xi \in ]\sqrt{{\tilde{\rho}}},\infty [$, for the latter of which we can consider a simple expansion of the integrand.
At large $\xi$, the integral becomes just $\int d\,\xi\,(\xi^{-2} \, + \, {\mathcal{O}}(\xi^{-6}))$.
To see that
this is consistent at $\xi\sim{\mathcal{O}}(\sqrt{{\tilde{\rho}}})$ with the full expression for the integral found in \cite{fancycon}, we substitute the approximate the solution \reef{psilogsol} into the full expression and expand for large ${\tilde{\rho}}$ -- giving us $\int d u\,(\frac{{\tilde{\rho}}}{u^2} \, + \, {\mathcal{O}}({\tilde{\rho}}^{-1}))$.
In contrast to the procedure in \cite{fancytherm} and the ``main'' part of the integral $\xi \in ]0,\infty [$, we did not add the boundary term to this part of the integral. This is because it vanishes in the full solution on the horizon and in our approximate solution as $\xi \rightarrow \infty$, but not sufficiently fast near $\xi \sim \sqrt{{\tilde{\rho}}}$.
Putting both parts of the integral together, we find the result
\begin{equation}\labell{fformula}
\bar{F} \ = \ {\bar{\rho}}^{3/2} \Phi({\bar{\rho}}) \, - \, {\bar{T}} {\bar{\rho}} \, + \, {\mathcal{O}}({\bar{T}}^4/\sqrt{ {\bar{\rho}}}) \ ,
\end{equation}
where
the term ${\bar{T}} {\bar{\rho}}$ comes from the integral over $\xi \in ]\sqrt{{\tilde{\rho}}},\infty [$
and
the integral
\begin{equation}\label{keyint}
\Phi({\bar{\rho}}) \ = \ \lim_{\varepsilon \rightarrow 0}\int_{\varepsilon}^{\infty}\frac{d\, \xi}{2 \xi^4}\left( 2 \sqrt{\xi^4+(1-\Psi^2)^2}\sqrt{1+\frac{\xi^2\Psi'^2}{1-\Psi^2}} -2 +3\Psi^2-2\xi\Psi\Psi'\right) \
\end{equation}
is a function of ${\bar{\rho}}$ only. An important observation is that the subleading term is of higher order in temperature than the $T^3$ behavior \ref{keyint} in the high-temperature regime.
At $\Psi = 0$, i.e. at ${\bar{\rho}} \rightarrow \infty$, the integral simplifies to $\int_0^\infty d\xi \frac{\sqrt{1+\xi^4} -1}{\xi^4} \, = \, \frac{2}{3}\mathcal{K}(1/\sqrt{2})$. We can also expand the integral to order $\Psi^2$ using the approximate solution for $\Psi$, equation \reef{psiK}, but the resulting integral can only be evaluated numerically and yields
\begin{equation}
\Phi({\bar{\rho}}) \ = \ \frac{2}{3}\mathcal{K}(1/\sqrt{2})+ \frac{0.2697}{{\bar{\rho}}} + {\mathcal{O}}({\bar{\rho}}^{-4}) \ .
\end{equation}
In the opposite limit at ${\bar{\rho}} \ll 1$,
we approximate the solution by two integrals, below and above the ``singular'' point around $\sqrt{{\bar{\rho}}}$, which is the ``kink'' in the profiles shown in figure \ref{emplot}.
The integral from $0$ to $\sqrt{{\bar{\rho}}}$ with $\Psi \sim \xi/\sqrt{{\bar{\rho}}} $ gives a contribution $ \frac{1}{6 {\bar{\rho}}^{3/2}} + {\mathcal{O}}(1)$ and the integral from $1/\sqrt{{\bar{\rho}}}$ to $\infty$ gives $ \frac{-1}{6 {\bar{\rho}}^{3/2}}+\frac{1}{\sqrt{{\bar{\rho}}}}$ to leading order in $1-\Psi_0$. Hence the leading behavior at small ${\bar{\rho}}$ is
$\Phi\sim \frac{1}{\sqrt{{\bar{\rho}}}}$ or $\bar{F} \sim {\bar{\rho}}(1-{\bar{T}})$. The next order in $\Phi(\rho)$ is numerically approximately $1.95 {\bar{\rho}}^{3/2}$, i.e. $1.95 {\bar{\rho}}^{3}$ in $\bar{F}$.
We show the full solution for $\Phi ({\bar{\rho}})$ together with the approximations in figure \ref{fofrhoplot}.
\begin{figure}
\includegraphics[width = 0.49\textwidth]{fofrho.pdf}\hspace{0.02\textwidth}\includegraphics[width = 0.49\textwidth]{fofrholog.pdf}
\caption{$\Phi({\bar{\rho}})$ as a linear plot (left) and log-log plot (right), illustrating the asymptotic limits.}\label{fofrhoplot}
\end{figure}
Reinstating the dimensionful coefficients, the free energy density becomes
\begin{equation}
F \, = \, \frac{\sqrt{\lambda}}{2}N_c \rho \left( \sqrt{\frac{\rho}{N_f}}\Phi({\bar{\rho}})\, - \, T\right)\, \sim \, \left\{\begin{array}{ll}
N_crho\, M_q - \frac{\sqrt{\lambda}}{2}N_c rho\, T & \ \ : \frac{\rho}{ N_f} \ll \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \\
\frac{1}{3}\mathcal{K}(\frac{1}{\sqrt{2}})\, N_c\sqrt{\frac{\lambda}{ N_f}}\rho^{3/2}- \frac{\sqrt{\lambda}}{2}N_c rho\, T & \ \ :
\frac{\rho}{ N_f} \gg \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \ .
\end{array}\right.
\end{equation}
\subsection{Thermodynamic quantities}\label{thermquan}
The entropy and internal energy densities are given by $S = -\left.\frac{\partial\, F}{\partial\, T}\right|_{\rho }$ and $U = F +TS$, which evaluates in our case to
\begin{eqnarray}
\!\!\!\!\!\!\!\! S & = & \frac{\sqrt{\lambda}}{2}N_c \rho \, + \, {\mathcal{O}}({\bar{T}}^3/\sqrt{ {\bar{\rho}}}) \ \ \mathrm{and} \\
\!\!\!\!\!\!\!\! U & = & \sqrt{\frac{\lambda}{N_f}}\frac{N_c}{2}\rho^{3/2}\Phi({\bar{\rho}})\, + {\mathcal{O}}({\bar{T}}^4/\sqrt{ {\bar{\rho}}})\, \sim \left\{\begin{array}{ll}
N_crho\, M_q & \ \ : \frac{\rho}{ N_f} \ll \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \\
\frac{1}{3}\mathcal{K}(\frac{1}{\sqrt{2}})\, N_c\sqrt{\frac{\lambda}{ N_f}}\rho^{3/2} & \ \ :
\frac{\rho}{ N_f} \gg \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \ .
\end{array}\right. ~~{ }
\end{eqnarray}
Correspondingly the specific heat capacity $c_V = \left.\frac{\partial \, U}{\partial\, T}\right|_{\rho }$ is
\begin{equation}
c_v \ = \ 0\, + \, {\mathcal{O}}({\bar{T}}^3/\sqrt{ {\bar{\rho}}}) \ .
\end{equation}
We can also obtain the chemical potential $\mu = \left.\frac{\partial \, F}{\partial\, \rho}\right|_{T }$ ,
\begin{equation}
\mu \, = \,
\sqrt{\frac{\lambda}{N_f}}\frac{N_c}{2}\,
\partial_{\rho}
\left(\rho^{3/2} \Phi({\bar{\rho}})\right)
\, - \,\frac{\sqrt{\lambda}}{2} T
\, \sim \,
\left\{\begin{array}{ll}
N_c M_q - \frac{\sqrt{\lambda}}{2}N_c T & \ \ : \frac{\rho}{ N_f} \ll \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \\
\frac{1}{2}\mathcal{K}(\frac{1}{\sqrt{2}})\, N_c\sqrt{\frac{\lambda\, \rho}{ N_f}}- \frac{\sqrt{\lambda}}{2}N_c T & \ \ :
\frac{\rho}{ N_f} \gg \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \ .
\end{array}\right.
\end{equation}
and the ``inverse density'' of states $\varepsilon^{-1} = \left.\frac{\partial^2 \, F}{\partial\, \rho^2}\right|_{T }$,
\begin{equation}
\epsilon^{-1} \, = \,
\sqrt{\frac{\lambda}{N_f}}\frac{N_c}{2}\,
\partial_{\rho}^2
\left(\rho^{3/2} \Phi({\bar{\rho}})\right)
\, \sim \, \left\{\begin{array}{ll}
0 + 0.73 \frac{N_c \lambda^2 \rho}{N_f^2 M_q^2}& \ \ : \frac{\rho}{ N_f} \ll \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \\
\frac{1}{4}\mathcal{K}(\frac{1}{\sqrt{2}})\, N_c\sqrt{\frac{\lambda }{\rho N_f}} & \ \ :
\frac{\rho}{ N_f} \gg \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \ .
\end{array}\right.
\end{equation}
where one has to keep in mind that ${\bar{\rho}} = \frac{ \rho}{ N_f}\frac{\lambda}{4 M_q^2}$ and the subleading term in these expressions is of order ${\bar{T}}^3$.
The density scaling of the chemical potential in the large-$\rho$ limit is in principle characteristic for a 2 dimensional Fermi liquid.
Interpreting this chemical potential however as a Fermi-sea -type chemical potential may be misleading, as the results of \cite{fancycon} suggest that the finite density gives rise to some finite length scale and to a quasiparticle spectrum. Furthermore, for a Fermi sea one would expect a heat capacity proportional to the temperature. Hence, it should be rather interpreted as a ground state energy that depends on the density.
Somewhat similarly, the entropy and the $-\sqrt{\lambda}N_c T/2$ term in the chemical potential suggest a ground state degeneracy proportional to the total baryon number, or $\sqrt{\lambda}/2$ times the quark number.
It may be an interesting exercise to write down a distribution and density of states that reproduces this.
The interpretation of the system as being in a ground state is also consistent with the large-mass limit. There, the energy of the system is approximately the baryon number times the baryon mass scale $N_c M_q$, or
precisely the quark number times the quark mass. The cross-over between the high and low density limits occurrs as the induced length scale $\sqrt{\frac{N_f}{\lambda\, \rho}}$ is of the order of the quark mass.
To obtain the pressure, we have to keep in mind that $\bar{F}$ is densitized. Using $\partial_V G(\rho)|_N = -\frac{\rho}{V} \partial_\rho G(\rho) $ for some function $G$ that only depends on $\rho$, we can obtain $\bar{P} = -\bar{F} + {\bar{\rho}} \partial_{\bar{\rho}} \bar{F}$:
\begin{equation}
P \ = \ \sqrt{\frac{\lambda}{N_f}}\frac{N_c}{2}\rho^{3/2} \left( {\textstyle \frac{1}{2}} \Phi({\bar{\rho}}) \, -\, \rho \partial_\rho \Phi({\bar{\rho}}) \right) \, \sim \,
\left\{\begin{array}{ll}
0 + 0.59 \frac{N_c \lambda^2 \rho^3}{N_f^2 M_q^2}& \ \ : \frac{\rho}{ N_f} \ll \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \\
\frac{1}{6}\mathcal{K}(\frac{1}{\sqrt{2}})\, N_c \rho^{3/2}\sqrt{\frac{\lambda }{ N_f}} & \ \ :
\frac{\rho}{ N_f} \gg \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \ .
\end{array}\right.
\end{equation}
From the pressure, we can then obtain the isothermal and adiabatic bulk moduli $K_T$ and $K_S$. Since the pressure is an intrinsic quantity, we can compute the volume derivative as $\bar{K}_T = - V \left. \frac{\partial \bar{P}}{\partial_V}\right|_{T,N} = {\bar{\rho}} \partial_{\bar{\rho}}\left(-\bar{F} + {\bar{\rho}} \partial_{\bar{\rho}} \bar{F} \right)= {\bar{\rho}}^2 \partial_{\bar{\rho}}^2 \bar{F}$:
\begin{equation}
K_T \ = \
\sqrt{\frac{\lambda}{N_f}}\frac{N_c}{2}\rho^{2}
\left(\rho^{3/2} \Phi({\bar{\rho}})\right)
\, \sim \,
\left\{\begin{array}{ll}
0 + 1.76 \frac{N_c \lambda^2 \rho^3}{N_f^2 M_q^2}& \ \ : \frac{\rho}{ N_f} \ll \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \\
\frac{1}{4}\mathcal{K}(\frac{1}{\sqrt{2}})\, N_c \rho^{3/2}\sqrt{\frac{\lambda }{ N_f}} & \ \ :
\frac{\rho}{ N_f} \gg \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \ .
\end{array}\right.
\end{equation}
and because the full, extrinsic, entropy is just given by the baryon number, we have $K_T=K_S$.
Overall, these results suggest that at small densities compared to the mass, we are dealing with some type of pressureless gas of quarks, where the pressure scales proportional to $T^4$ and $\rho^3$, compared to $\rho T$ for an ideal classical gas. As the density increases above the mass scale, the zero-temperature pressure and bulk modulus become large. This is consistent with the density dependence of the ground state energy and with a length scale induced by the finite density.
We can also obtain the speed of sound. For the small-density limit this is just the non-relativistic ${v}_s^2 = \frac{K}{M_q N_c \rho} $, which vanishes because the bulk modulus vanishes. In general we can use the fact that the energy density depends only on the density and obtain $v_s^2 = \left.\frac{\partial P}{\partial U}\right|_S = \left(\frac{\partial U}{\partial \rho}\right)^{-1} \frac{\partial P}{\partial \rho} = \frac{K}{\rho \partial_\rho U}$:
\begin{equation}
v_s^2 \, = \,
\rho \partial_\rho \ln \left( \partial_\rho
\left(\rho^{3/2} \Phi({\bar{\rho}})\right) \right)
\, \sim \,
\left\{\begin{array}{ll}
0 + 1.76 \frac{N_c \lambda^2 \rho^3}{N_f^2 M_q^2}& \ \ : \frac{\rho}{ N_f} \ll \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \\
\frac{1}{2} & \ \ :
\frac{\rho}{ N_f} \gg \left(\frac{2M_q \!}{\sqrt{\lambda}}\right)^{\! 2} \ .
\end{array}\right.
\end{equation}
The density-independent result in the large-density limit is dictated by causality.
It would be very interesting to compare these results with the ones obtained from the normal modes in the gravity side as in \cite{zerosound,luttinger}.
\subsection{Subleading Terms}
To study the subleading terms, we obtain the difference between the low-temperature limit for $F$ from equation \reef{keyeq} and the full numerical result that we can obtain from the methods in \cite{fancytherm}.
To do so, we proceed as follows:
We fix $\Psi_0$ and then vary $\tilde{\rho}$. Beyond the dominant constant term, there is some small temperature dependence in ${\bar{\rho}}$
that appears in the numerical results at ${\mathcal{O}}({\bar{T}}^3)$.
This is compensated for by using the actual numerical value ${\bar{\rho}}(\Psi_0,{\bar{T}})$
in \reef{fformula} and compare this to the numeric result for the same $\Psi_0$ and ${\bar{T}}$. From that point on, we ignore the small variation in ${\bar{\rho}}$ and consider it fixed. It turns out to be most convenient to consider a temperature range over which the significant subleading term in $\bar{F}$ is suppressed by at least $10^{-4}$ and ${\bar{\rho}}$ varies at most by a factor of $10^{-3}$. Because the numerical accuracy in $\bar{F}$ is approximately $10^{-7}$, the systematic errors from ignoring the variation in ${\bar{\rho}}$ appear then roughly at the same level as the numerical noise.
It turns out that this subleading term is always negative and interpolates from approximately $\delta \bar{F} \approx - \frac{{\bar{T}}^4}{10}$, i.e. $\bar{c}_v \approx \frac{6{\bar{T}}^3}{5}$, at small density to precisely $\delta \bar{F} = - \frac{T^5}{10 {\bar{\rho}}}$ at large density. The latter is consistent with the results in the massless case in \cite{fancytherm} and \cite{zerosound}, from which we expect at large densities $\bar{c}_V = 2 \frac{T^4}{{\bar{\rho}}}$.
The most convenient ways to parametrize the numeric results are $\delta \bar{F} = - a {\bar{T}}^4 - b {\bar{T}}^5$ and $\delta \bar{F} = - \alpha {\bar{T}}^\beta$, which are not distinguishable at the level of the straightforwardly achievable numerical accuracy and are shown in figure \ref{tempexpplot}.
\begin{figure}
\includegraphics[width = 0.49\textwidth]{tempcoeff.pdf}\hspace{0.02\textwidth} \includegraphics[width = 0.49\textwidth]{tempexp.pdf}
\caption{The coefficients in the subleading term in the free energy parametrized as $\delta \bar{F} = - a {\bar{T}}^4 - b {\bar{T}}^5$ (left) and $\delta \bar{F} = - \alpha {\bar{T}}^\beta$ (right). On the left, the dots indicate a regime in which the fits for the smaller term are not numerically reliable anymore.}\label{tempexpplot}
\end{figure}
An exponential suppression of the subleading term, that would be suggestive of an energy gap, is however ruled out.
\section{General cases}\label{gencas}
Even though it is not the focus of this paper, let us briefly comment about what happens in other cases than the 2+1 dimensional defect. In principle, the methods discussed in sections \ref{limit} and \ref{thermo} are straightforwardly generalized to fundamental matter supported in (d+1) dimensions. In the following $\tilde{(\cdot)}_d$ and $\bar{(\cdot)}_d$ will denote dimensionless quantities analogous to the definitions in sections \ref{setup} and \ref{limit}.
The setups that come most straightforwardly to mind are the usual supersymmetric D3-D7 and D3-D3 setups that are T-dual to the D3-D5 case,
\begin{equation}
\begin{array}{rccccc|c|cccccl}
& & 0 & 1 & 2 & 3 & 4& 5 & 6 & 7 & 8 & 9 &\\
& & t & x & y & z & r& & & & & \theta &\\
\mathrm{background\,:}& D3 & \times & \times & \times & \times & & & & & & & \\
\mathrm{probe\ (1+1)\,:}& D3 & \times & \times & & & \times & \times & & & & & \ \ \mathrm{or} \\
\mathrm{probe\ (3+1)\,:}& D7 & \times & \times & \times & \times & \times & \times & \times & \times & & & \ \ \
\end{array}
\labell{arraymany}
\end{equation}
with the induced metric and the action given analogous to equation \reef{branemetric} and \reef{braneaction}, respectively. For the action in the form of equation \reef{backact}, we have to keep in mind that in a D3-Dp embedding, the factor $(1-\Psi^2)$ is replaced by $(1-\Psi^2)^{(p-d-1)/2}$, which is in our supersymmetric cases $(1-\Psi^2)^{d/2}$.
The solutions for the gauge field are
\begin{equation}\label{shita}
\partial_u \tilde{A}_{d\,t}(u) \ = \ \frac{{\tilde{\rho}}\sqrt{1 + u^2 h(u) \frac{\Psi'(u)^2}{1- \Psi^2(u)}}}{\sqrt{ {\tilde{\rho}}_d^2 u^{2d} + (1-\Psi(u)^2)^{(p-d-1)}}}
\end{equation}
but we will not review further details of the backgrounds as they are readily found in the literature \cite{winters,robdens,luttinger} and they are not essential here.
The most important property that we need to know for the the embeddings $\Psi$ at this point is that near $u\rightarrow 0$, in the D3-D7 system $\Psi\sim {\tilde{m}}_3 u + {\tilde{c}}_3 u^3$ and $\Psi\sim {\tilde{m}}_1 u + {\tilde{c}}_1 u\, \ln\, u$ in the D3-D3 system.
To obtain the brane profiles, the radial coordinate has to be defined by $\xi_d = \frac{u}{\tilde{\rho}_d^{1/d}}$, and the large-$\xi_d$ expansions of $\Psi$ are $\Psi_0 - \frac{1}{14 \xi_3^6} \Psi_0(1-\Psi_0^2)^3$ and $\Psi_0 - \frac{1}{6 \xi_1^2} \Psi_0(1-\Psi_0^2)$. We show the corresponding $\Psi_0$ as a function of the density ${\bar{\rho}}_d = \left. \partial_{\xi_d} \Psi\right|_{\xi_d \rightarrow 0}$ in figure \ref{lastgoddamnplot}.
We find that at large densities, $\Psi_0$ scales proportional to ${\bar{\rho}}_d^{-1/d}$. As in the 2+1 case, there seems to be an exact relation ${\bar{c}}_3 = {\bar{\rho}}_3 \frac{\Psi_0}{ 2}$ and ${\bar{c}}_1 = {\bar{\rho}}_1 \Psi_0$.
To compute the thermodynamics, we first note that the overall factor ${\bar{\rho}}^{3/2}$ in the free energy in equation \reef{keyeq} becomes ${\bar{\rho}}^{(d+1)/d}$.
In analogy to \reef{keyeq}, the factor in the integral coming from the AdS structure is $\int\! \frac{d\xi_d}{\xi_d^{d+2}}$, and in the large-$\xi$ limit, the solution \reef{shita} gives us a term $\xi_d^2$ in the Legendre-transformed action such that we can reproduce the $\xi \in ]\sqrt{{\tilde{\rho}}},\infty[$ part of the integral \reef{keyeq}. Hence, the free energy becomes
\begin{equation}
\bar{F}_d \ = \ {\bar{\rho}}_d^{(d+1)/d} \Phi_d({\bar{\rho}}_d) \, - \, {\bar{\rho}}_d {\bar{T}} \ + \ {\mathcal{O}}({\tilde{\rho}}^{(1-d)/d}) \
\end{equation}
and the entropy is just the baryon number as in the 2+1 dimensional case.
\begin{figure}
\includegraphics[width = 0.49\textwidth]{manydensX.pdf}
\hspace{0.02\textwidth}
\includegraphics[width = 0.49\textwidth]{lastgoddamnplotX.pdf}
\caption{Left: The boundary value $\Psi_0$ as a function of the density ${\bar{\rho}}$ for various setups. The large-density approximations are shown dashed. Right: $\Phi_d$ as a function of the density ${\bar{\rho}}_d$, illustrating also the ${\bar{\rho}} \ll 1$ limit.}\label{lastgoddamnplot}
\end{figure}
\subsection{3+1 Dimensions}
To discuss $\Phi_d({\bar{\rho}}_d)$ let us start with the 3+1 dimensional system. In this case, the counter terms from \cite{skenderis} are structurally the same as in the D3-D5 case, and hence the integral is similar to \reef{keyint}, up to the appropriate powers in the expression. The result is shown in figure \ref{lastgoddamnplot} in comparison to the 2+1 defect.
We find that the power scaling at small densities approaches $\Phi_3({\bar{\rho}}_3) \sim \frac{1}{4 {\bar{\rho}}_3^{4/3}}$. This means that the leading behavior in ${\bar{\rho}}_3$ is $\bar{F}_3 = const. - {\bar{\rho}}_3 {\bar{T}}_3$, i.e. the free energy (and also the internal energy) per unit volume (given by the meson mass scale) is proportional to the mass -- independent of the baryon number density. Essentially, this implies that the configuration is unstable in this regime, with a negative pressure proportional to $-M_{meson}^4$. As the chemical potential is monotonously increasing (i.e. we have a positive ``density of states'') this instability is different from the one studied in \cite{yunseok}. Furthermore, there the ``funnel'' on the gravity side is becoming very narrow, whereas here it has narrow but finite width and is becoming very long. Still, $\Psi'$ is becoming very small and hence the tension of the brane that falls into the horizon is close to the tension of an appropriate number of fundamental strings attached to a brane that does not extend down to the horizon.
The subleading term in $\bar{F}_3$ is just ${\bar{\rho}}_3^{-1/3}$, giving again just the baryon number times the meson mass scale as we observed in the 2+1 system.
At large densities, $\Phi_3$ becomes constant, close to 1, so the free energy density is approximately $\bar{F} \sim {\bar{\rho}} (\rho^{1/3} - {\bar{T}})$ which is consistent with the picture of an induced length scale discussed in section \ref{thermquan} and mimics again a Fermi surface. The sound speed in this limit is now $c_V^2 = \frac{1}{3}$ as in \cite{isospin,responses}.
\subsection{1+1 Dimensions}
The D3-D3 system is profoundly different. While the overall prefactor of the integral in \reef{keyint} and first counter term in \reef{ibdy} just follow dimensional arguments, there are a few logarithmic divergences. Hence, the appropriate counter term from \cite{skenderis} for the scalar is now $\frac{1}{2}\sqrt{\gamma}\Psi(\epsilon)^2(1+1/\ln(\epsilon))$.
In contrast to the 3+1 and 2+1 dimensional systems, where the boundary term gave us the appropriate variation of the action w.r.t. the scalar, ${\tilde{c}}\,\delta {\tilde{m}}$, the variation is now $-{\tilde{m}}\, \delta {\tilde{c}}$. Hence, we have to do a ``Legendre transformation'' and add the extra boundary term ${\tilde{c}}\, {\tilde{m}}$ to the action in order to consider a fixed mass.
This counter term cures the divergence coming from the scalar but there remains one divergence
coming from the logarithmically divergent $A_t \sim {\tilde{\mu}} + \ln \epsilon$.
While the variation of the action is finite, and also the Legendre transformation can be trivially made finite by using ${\tilde{\rho}} (A_t - {\tilde{\rho}}\ln\epsilon)$, the action itself is divergent.
In principle, one can add extra boundary terms to cure this divergence.
For example in \cite{luttinger}, the authors use a counter term $\frac{A_\mu A_\nu \gamma^{\mu\nu} \sqrt{\gamma}}{2\, \ln\, \epsilon}$ which has, however, a non-trivial variation $\frac{A_\mu A_\nu \gamma^{\mu\nu} \sqrt{\gamma}}{ \ln\, \epsilon}\delta A_\nu$.
In the limit $\epsilon \rightarrow 0$ and for a finite variation $\delta A_t$, this counter term is identical to the Hawking-Ross boundary term \cite{hawkingross}, and cancels precisely the boundary term in the variation of the gauge field. This is because, just as with the scalar, implicitly this term also includes a Legendre transformation, and in terms of the electric flux (or density) the asymptotic variation of the gauge field is $\delta A_t \sim {\tilde{\rho}} \ln \epsilon$, so it diverges at the boundary and results in a finite variation $\mu \, \delta {\tilde{\rho}}$. Coincidentally, this counter term is identical to adding $\frac{{\tilde{\rho}}^2}{2} \ln\epsilon$ and then doing the finite Legendre transformation -- or first doing a (diverging) Legendre transformation ${\tilde{\rho}} A_t$ and then adding $-\frac{{\tilde{\rho}}^2}{2} \ln\epsilon$.
For practical aspects, the integral for $\Phi$ does converge only logarithmically in the UV limit, so we have to take care of this using approximate analytical solutions to extend the numerical result to even smaller $x$.
We show the result for $\Phi_1$ in figure \ref{lastgoddamnplot}. There, we see that $\Phi$ converges to approximately $0.596$ at large densities ${\bar{\rho}}_1 \gg 1$, giving us physics with an induced length scale similar to the other cases, and now a speed of sound of $v_s^2 = 1$, which is again the ``conformal'' result.
At small densities or large masses ${\bar{\rho}}_1 \ll 1$, $\Phi_1$ approaches $\Phi_1\sim \frac{1}{{\bar{\rho}}_1}$, which implies again that the energy of the system is described by a pressureless gas of mesons. Similarly, the subleading term giving rise to the pressure implies that that the pressure is at least of order ${\bar{\rho}}^4$, but numerical accuracy implies that we cannot give a more detailed result and also cannot exclude a very small but finite (positive or negative) pressure at vanishing density.
One might worry in how far the redefinition $u\, {\tilde{\rho}}_1 = \xi_1$ in the logarithmic terms causes some mixing between the density and the temperature. As the final action however is free of logarithmic divergences, there should be no such remaining terms. Furthermore, we see that our results for the free energy rely certainly heavily on using the correct boundary terms. Further discussions of peculiarities of this 1+1 dimensional system can be found in \cite{luttinger}, which focuses on this case, albeit not in this limit.
\section{Conclusions}\label{conclu}
In this paper, we used holography to explore the low-temperature regime of fundamental matter with finite mass and baryon number density coupled to the usual $N_c \gg 1$, $\mathcal{N}=4$ SYM theory above the deconfinement phase transition. The fact that we used a top-down approach guarantees us that the field theory is consistent and well defined, giving more relevance to our findings. Using $N_f$ Dp probe-branes in a background of $N_c \gg N_f$ D3 branes, we found a new temperature-independent scaling solution of the probe brane embeddings. On the field theory side, this manifests in the fact that the low-temperature physics is only governed by the density to mass ratio, $\rho / M_q^2$ and interpolates between a mass-dominated low-density regime and a density-dominated large-density regime. On the gravity side the radius-dependent size of the compact sphere of the probe brane geometry forms a long ``funnel'' of finite size, with the scaling solution in the asymptotic ``opening''.
First, we concentrate on a defect supported on 2+1 dimensions that is dual to a D3-D5 setup.
From the usual thermodynamic relations, we obtained the properties of this matter, which suggest that there is a new kind of quantum liquid. The entropy is just given by the baryon number times $N_c \sqrt{\lambda}/2$ or the quark number times $ \sqrt{\lambda}/2$, which implies a large ground state degeneracy. Consistent with this, the chemical potential has a negative term $-T N_c \sqrt{\lambda}/2$, that is consistent with a Bose-Einstein distribution with a degeneracy proportional to the baryon number.
The heat capacity however vanishes completely in the scaling solution, which means that it is of higher order in temperature than a classical Fermi or Bose gas. To explore the possibility of a gap, we explored the subleading terms in the free energy numerically. It turns out that the heat capacity is not exponentially suppressed -- so there is no energy gap -- but rather interpolates between $c_V\propto T^3$ at small densities and $c_V \propto T^4$ at large densities; in the latter case the coefficient matches with the analytic results of the massless case in \cite{zerosound} and \cite{fancycon}.
In the small-density limit, the internal energy is just given by the baryon number times the baryon mass scale, or precisely the overall quark number times the quark mass.
At the same time, the pressure, bulk modulus and sound speed vanish; suggesting that the system becomes a pressureless gas of quarks. This is, however, still not a classical non-interacting gas, as the leading terms in the pressure are of higher order in temperature and pressure than for a classical gas, i.e. they vanish more quickly as $T\rightarrow 0$ and $\rho \rightarrow 0$.
At large densities, the internal energy is given by the quark number times an energy scale that is proportional to $\sqrt{\rho}$ with a corresponding leading term in the chemical potential proportional to $\sqrt{\rho}$. This is, however, not to be interpreted as a Fermi surface, but rather as a length scale of the order of $\sqrt{\frac{ N_f}{\lambda\, \rho}}$
that is induced by the finite density at strong coupling and determines the ground state energy. Note that this is much shorter than just the geometrical separation of quarks. Consistent with this, the pressure and bulk modulus are proportional to $N_c \rho \sqrt{\frac{\lambda\, \rho}{ N_f}}$
and the sound speed becomes $\frac{1}{\sqrt{2}}$ times the speed of light.
The cross-over between the high and low density regimes occurrs when the induced length scale is of the order of the quark mass.
We also looked at the cases of fundamental matter supported in 3+1 (D3-D7) and 1+1 (D3-D3) dimensions. The embeddings and the general structure of the thermodynamics are very similar, but there are also some important differences. In the 3+1 case, the system exhibits a new thermodynamic instability characterized by a negative pressure at small density, and in the 1+1 case it turns out that there are some non-trivial logarithmic divergences. Applying our analysis to this case and adding additional
counter terms, such as the one discussed in \cite{luttinger}, yields qualitatively similar results to the 2+1 dimensional case.
To study the properties of this quantum liquid a bit further -- in particular in the 2+1 defect case -- it would be interesting to obtain properties that are obtained from linear response theory and two point functions, such as the viscosity, thermal and electric conductivity, quasiparticle spectrum and also the speed of sound to confirm consistency of the results.
On the gravity side, these are straightforwardly obtained from the equations
of motion for perturbations of the world-volume gravitational and gauge fields. Certainly, there may be some mixing between the different modes, but dealing with coupled systems in probe-brane backgrounds both analytically and numerically is straightforward \cite{fancycon,luttinger,recenttwist}. Furthermore, the equations should simplify by following the limiting procedure considered in section \ref{limit}.
In the 3+1 dimensional case, it would be interesting to further explore this new instability, and it would be also interesting in both the 3+1 and 1+1 cases to study the physics and subleading terms more thoroughly and to study dimensional peculiarities in the 1+1 case.
Certainly, it would also be interesting to see in how far some of the properties that we find are realized in experimentally studied systems.
\acknowledgments The author would like to thank Bum-Hoon Lee for useful comments on the draft of this paper, Janet Hung and Aninda Sinha for helpful clarifications regarding \cite{luttinger} and
Sang-Jin Sin, Yun-Seok Seo and Yumi Ko
for useful discussions.
This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MEST)
through the Center for Quantum Spacetime(CQUeST) of Sogang University with grant number 2005-0049409.
|
\section{Introduction and Motivation}
\label{sec:introduction}
Wireless sensor networks are composed of huge numbers of small
physical devices capable of sensing the environment and connected
using ad-hoc networking protocols over radio links~\cite{survey:akyildiz:etal:02}.
These platforms have several unique characteristics when compared with
other ad-hoc networks.
First, sensor networks are often designed for specific applications or
application domains making software re-usability and portability an
issue.
Sensor devices have very limited processing power (CPU), available
memory, and battery lifetime, and are often deployed at remote
locations making physical access to the devices (\emph{e.g.}\; for maintenance)
difficult or even impossible.
For these reasons, programming such large scale distributed systems
can be daunting. Programs must be \emph{lightweight}, produce a
\emph{small memory footprint}, be \emph{power conservative}, be
\emph{self-reconfigurable} (\emph{i.e.}\; may be reprogrammed dynamically
without physical intervention on the devices) and, we argue, be
\emph{(type-)safe}.
To date several programming languages and run-time systems have been
proposed for wireless sensor networks (see~\cite{bookchapter} and
references therein) that address some of the above issues, but few
tackle the \emph{safety} issue. Regiment~\cite{regiment2}, a strongly
typed functional \emph{macroprogramming} language, is the closest to
achieve this goal by providing a type-safe compiler.
However, Regiment is then compiled into a low-level \emph{token
machine language} that is not type-safe. This intermediate language
is itself compiled into a nesC implementation of the run-time based on
the \emph{distributed token machine} model, for which no safety
properties are available.
In fact, in general, an underlying model with well-studied operational
semantics for sensor networks seems to be lacking. The absence of such
a model reveals itself as a considerable semantic gap between the
semantics of the (sometimes high-level) programming languages and their
respective implementations.
In this paper we propose Callas, a calculus for programming sensor
networks, based on the formalism of process
calculi~\cite{async-pi:honda:tokoro:91,pi:milner:parrow:walker:92},
that aims to establish a basic computational model for sensor
networks. The goal is to diminish the above mentioned semantic gap
by proceeding bottom-up, using Callas as a basic assembly language
upon which high-level programming abstractions may be encoded as
semantics preserving, derived constructs.
Callas is an evolution from a previous proposal~\cite{sensorcomm07} by
the authors, which unlike its original sibling provides: (a)
decoupled semantics for in-sensor computation (associated with the
\emph{application layer}) and networking (associated with the
\emph{data-link} and \emph{network} layers); b) support for a form of
timed \emph{events}; and c) \emph{event-driven} semantics.
\section{Overview of Callas}
\label{sec:calculus}
The syntax of Callas is provided by the grammar in
Figure~\ref{fig:syntax}. Let $\vec \alpha$ denote a possibly empty
sequence $\alpha_1 \dots \alpha_n$ of elements of some syntactic
category $\alpha$. We let~$l$ range over a countable set of
\emph{labels} representing function names, and let $x$ range
over a countable set of \emph{variables}. These sets are pairwise
disjoint.
\begin{figure}
\begin{equation*}
\begin{aligned}
S & \Space{::=} & & \text{\emph{Sensors}}\\
& \quad \; \textbf{\textsf{0}} & & \text{empty network}\\
& \;\mid\; S \,\vert\, S & & \text{composition}\\
& \;\mid\; \sensor {P,R} {M,T} I O p t & & \text{sensor}\\[0.5cm]
%
v & \Space{::=} & & \text{\emph{Values}}\\
& \quad \; b & & \text{built-in value}\\
& \;\mid\; x & & \text{variable}\\
& \;\mid\; M & & \text{module}\\
& \;\mid\; \keyw{sensor} & & \text{installed functions}\\[0.4cm]
%
m & \Space{::=} \langle l(\vec v)\rangle & & \text{messages}\\[0.4cm]
R & \Space{::=} P_1::\dots::P_n & & \text{run-queue} \\
I,O & \Space{::=} m_1::\dots::m_n & & \text{message queues}
\end{aligned}
\qquad
\begin{aligned}
M & \Space{::=} & & \text{\emph{Modules}}\\
& \quad \; \{ {l_i} = {\abstr {\vec x_i}{P_i}}\}_{i\in I} & & \text{module}\\[0.5cm]
%
P & \Space{::=} & & \text{\emph{Processes}}\\
& \quad \; v & & \text{value}\\
& \;\mid\; \invk {v}{l}{\vec v} & & \text{function call}\\
& \;\mid\; \system {\msg l {\vec v}} & & \text{external call}\\
& \;\mid\; \fire {\msg l {\vec v}} v v & & \text{timed call}\\
& \;\mid\; \send {\msg l {\vec v}} & & \text{communication}\\
& \;\mid\; \receivek & & \text{communication}\\
& \;\mid\; \install v v & & \text{install code}\\
& \;\mid\; \Let x P P & & \text{sequence}\\[0.3cm]
T & \Space{::=} \{(l_i(\vec v_i),v_i,v_i,v_i)\}_{i \in I} & & \text{timed calls}
\end{aligned}
\end{equation*}
\caption{The syntax of Callas.}
\label{fig:syntax}
\end{figure}
A network $S$ is an abstraction for a network of real-world sensors
connected via radio links. We write it as a flat, unstructured
collection of sensors combined using the parallel composition
operator. The empty network is represented by symbol $\textbf{\textsf{0}}$.
A sensor $\sensor {P,R} {M, T} I O t p$ is an abstraction for a sensor device. It
features a running process~($P$) and a double-ended
queue of processes scheduled for execution~($R$). Its memory stores
both the installed code for the application~($M$) and a table of
timers for function calls~($T$). These components represent the
application layer of the protocol stack for the sensor. The interface
with the lower level networking and data-link layers is modeled using
incoming~($I$) and outgoing~($O$) queues of messages.
The sensors have a measurable position~($p$) and their own
clocks~($t$), and are able to measure some physical property
(e.g.\@ temperature, humidity) by calling appropriate \emph{external}
functions.
The code in $M$ consists of a set of named functions.
The syntax $l = (\vec x) P$ represents a function, where $l$ is the name,
$(\vec x)$ the parameters, and $P$ the body.
The $I$ ($O$) queue buffers messages received from (sent to) the
network. Messages are just packaged function calls
$\langle l(\vec v)\rangle$.
Finally, $T$ is a set that keeps information on timers for function
calls. For each timer, a tuple is maintained with the following
information: the call to be triggered, the timer period, the
time after which the timer expires and, the time of the next
call.
A process $P$ can be one of the following:
(a) a value $v$ that represents
the data exchanged between sensors. It can be a basic value ($b$) that
can intuitively be seen as the primitive data types supported by the
sensor's hardware or a module ($M$). The special value $\keyw{sensor}$
represents the module that holds the functions installed at the
sensor;
(b) a synchronous call $\invk {v}{l}{\vec v}$ to a function $l$ in a module $v$;
(c) a synchronous external call -- $\system {\msg l {\vec v}}$;
(d) a timer -- $\fire {\msg l {\vec v}} v v$, that
calls an installed function $l(\vec v)$ periodically, controlled by a
timer;
(e) an asynchronous remote call -- $\send {\msg l {\vec v}}$, that adds a message
$\langle l(\vec v)\rangle$ to the outgoing queue ($O$);
(f) a receptor --
$\receivek$, that gets a message from the incoming queue ($I$);
(g) a module installation -- $\install v{v'}$, that adds the set of
functions in $v'$ to $v$; and, finally
(h) a $\keyw{let}$ construct that
allows the processing of intermediate values in computations.
The latter is also useful to derive a basic sequential composition
construct (in fact, $\keyw{let}\ x = P\ \keyw{in}\ P' \equiv P~;~P'$ with $x
\not\in \operatorname{fv}(P')$). We make frequent use of this construct to impose a
more imperative style of programming.
Each function in a module has as the first parameter the variable
\text{self}{} that is, as usual, a reference to the current module, i.e.,
the one the function belongs to. Each call to a function $v.l(\vec v)$
passes $v$ as the first argument in $l(\vec v)$.
In the sequel we present two small examples of programs written in
Callas. Both examples have two components: the code to be run at a
base-station (\emph{sink}) and the code to be run at each of the
other nodes (\emph{sensor}).
\textbf{Streaming data.} The program that runs on the \emph{sink}
starts by installing, in the local memory ($M$), a module with a
\lstinline{receiver} function and a \lstinline{gather} function. The
former just listens for messages from the network on the incoming
queue. The latter simply logs the arguments using a built-in external
call. Then, it starts a timer for the \lstinline{receiver} function
with a period of 5 milliseconds for 10 seconds. Finally, the sink
broadcasts a \lstinline{setup} message with a period of 100
milliseconds and a duration of 10 seconds. The call is placed in the
outgoing queue of the sink ($O$). In these examples we write
\lstinline{install} as a compact form for \lstinline{sensor.install}.\\
\begin{lstlisting}
// sink
install {
receiver = (self)
receive
gather = (self,x,y)
external log(x,y) };
timer receiver() every 5 expire 10000;
send setup(100,10000)
// sensor
install {
receiver = (self)
receive
setup = (self,x,y)
timer sample() every x expire y
sample = (self)
let x = external time() in
let y = external data() in
send gather(x,y) };
timer receiver() every 5 expire 10000;
\end{lstlisting}
Each \emph{sensor} starts by installing a module with a
\lstinline{receiver} function, similar to that on the sink, and
\lstinline{setup} and \lstinline{sample} functions. Then it starts a
timer on \lstinline{receiver} and waits for incoming messages. When a
sensor receives a \lstinline{setup} message from the network, it sets
up another timer to periodically call \lstinline{sample} in the same
module. When this function is executed the local time and the desired
data are read with external calls and a \lstinline{gather} message is
sent to the network carrying those values.
Note that the routing of messages is transparent at this level. It is
controlled at the network and data-link layers and we model this by
having an extra semantic layer for the network
(\emph{c.f.}\; Figure~\ref{fig:reduction-net}). In this example, the messages
from the sink are delivered to every sensor that carries a
\lstinline{setup} function. The information originating in the
sensors, in the form of \lstinline{gather} messages, on the other
hand, is successively relayed up to the sink (since sensors have no
\lstinline{gather} functions implemented).
\textbf{The maximum value of a data attribute and the MAC address of
the sensor that reads it.} This example follows much the same
principles of the above, except that it is a single shot
request. Instead of computing the maximum value of the data attribute
only at the sink, we optimize the program so that each sensor has two
attributes \lstinline{max_data} and \lstinline{max_mac} that keep,
respectively, the maximum value for the data that passed through the
sensor, and the associated MAC address.\\
\begin{lstlisting}
// sink
install {
receiver = (self)
receive
gather = (self,x,y)
external log(x,y) };
timer receiver() every 5 expire 10000;
send setup()
// sensors
install {
receiver = (self)
receive
setup = (self)
let x = external data() in
let y = external mac() in
self.install { max_data = (self) x
max_mac = (self) y };
send gather (x,y)
gather = (self,x,y)
let val = self.max_data() in
if x > val then
self.install { max_data = (self) x
max_mac = (self) y };
send gather(x,y); };
timer receiver() every 5 expire 10000;
\end{lstlisting}
The program that runs on the \emph{sink} is very similar to that of
the previous example. After installing the \lstinline{receiver} and
the \lstinline{gather} functions, it starts the receiver and
broadcasts a \lstinline{setup} message to the network.
The \emph{sensors} get the call from the network using their receivers
and execute \lstinline{setup}. The data and MAC address are obtained
by calling external functions and sent to the network in
\lstinline{gather} messages. Each time such a message is relayed by a
sensor on its way to the sink, the relaying sensor checks whether it
is worth to send the data forward by comparing it with the local
maximum. This strategy manages to substantially reduce the required
bandwidth at the sensors closest to the sink. The sink implementation
of \lstinline{gather} stops the relaying and logs the data. Note that
in this example, to simplify, more than one maximum value may be
recorded at the sink. Also, we use an \lstinline{if-then} construct
that is not provided in the base calculus but that can easily be added
for convenience.
Unlike the previous example, here every sensor will relay
\lstinline{gather} messages only after some internal processing, by
its own version of the homonym function.
\myparagraph{Semantics.}
\label{sec:semantics}
The calculus has two variable binders: the $\keyw{let}$ and the function
constructs, inducing the usual definition for free and bound variables.
The displayed occurrence of variable $x$ is a \emph{binding} with
\emph{scope}~$P$ both in $\keyw{let}\ x=P'\ \keyw{in}\ P$ and in
$l = (\dots, x, \dots) P$.
An occurrence of a variable is \emph{free} if it is not in the scope of a
binding.
Otherwise, the occurrence of the variable is \emph{bound}.
The set of free variables of a sensor $S$ is referred to as $\operatorname{fv}(S)$.
We present the reduction relation with the help of a structural
congruence, as it is usual~\cite{ccs:milner:80},
given in Figure~\ref{fig:congruence}.
Here, $\mkSrule{init-Send}$ is the only non-standard rule and provides a sensor
with a conceptual \emph{membrane} that engulfs neighboring sensors as
they become engaged in communication. This prevents the reception of
duplicate copies of the \emph{same} message from the source sensor
during a transmission.
\begin{figure}
\begin{gather*}
\tag{\mkSrule{monoid-Sensor}} S_1 \,\vert\, S_2 \equiv S_2 \,\vert\, S_1, \qquad S \,\vert\, \textbf{\textsf{0}} \equiv S, \qquad S_1 \,\vert\, (S_2 \,\vert\, S_3) \equiv (S_1 \,\vert\, S_2) \,\vert\, S_3 \\ \\
\sensor {P,R} {M,T} I O p t \equiv \tagsensor {P,R} {M,T} I O p
t {\textbf{\textsf{0}}} \qquad \tag{\mkSrule{init-Send}}
\end{gather*}
\caption{Structural congruence for sensors.}
\label{fig:congruence}
\end{figure}
The reduction relation is inductively defined by the rules in
Figures~\ref{fig:reduction} and~\ref{fig:reduction-net}.
Since processes evaluate to values, we allow for reduction within the
$\keyw{let}$ construct and therefore present the reduction relation using
the following reduction contexts: $\context{\cdot} \Space{::=} [\;]
\;\mid\; \keyw{let}\ x=\context{\cdot} \;\keyw{in}\ P$.
The reduction in a sensor is driven by running process~$P$.
\begin{figure}
\begin{gather*}
\tag{\mkRrule{extern}}
\frac{
\noEvent T t
}{
\sensor {\context{\system {\msg l {\vec v}}},R} {M,T} I O p t
\rightarrow
\sensor {\context{v},R} {M, T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{install-sensor}}
\frac{
\noEvent T t
}
{
\sensor {\context{\install {\keyw{sensor}} {M'}},R} {M, T} I O p t
\rightarrow
\sensor {\context{\obj{}},R} {M+M', T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{install-module}}
\frac{
\noEvent T t
}
{
\sensor {\context{\install {M'} {M''}},R} {M, T} I O p t
\rightarrow
\sensor {\context{M' + M''},R} {M, T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{send}}
\frac{
\noEvent T t
}
{
\sensor {\context{\send {\msg l {\vec v}}},R} {M, T} I O p t
\rightarrow
\sensor {\context{\obj{}},R} {M, T} I {O::\langle \msg l {\vec v} \rangle} p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{receive}}
\frac{
\noEvent T t
}
{
\sensor {\context{\receivek},R} {M, T} {\langle \msg l {\vec v} \rangle :: I} O p t
\rightarrow
\sensor {\context{\obj{}},R:: \invk \keyw{sensor} l {\vec v}} {M, T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{no-message},\mkRrule{idle}}
\frac{
\noEvent T t
}
{
\sensor {\context{\receivek},R} {M, T} {\varepsilon} O p t
\rightarrow
\sensor {\context{\obj{}},R} {M, T} I O p {t+1}
}
\qquad
\frac{
\noEvent T t
}
{
\sensor {v,\varepsilon} {M, T} I O p t
\rightarrow
\sensor {v,\varepsilon} {M, T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{next},\mkRrule{move}}
\frac{
\noEvent T t
}
{
\sensor {v,P::R} {M, T} I O p t
\rightarrow
\sensor {P,R} {M, T} I O p {t+1}
}
\qquad
\frac{
\noEvent T t
}
{
\sensor {P,R} {M, T} I O p t
\rightarrow
\sensor {P,R} {M, T} I O {p'} t
}
\\[0.2cm]
\tag{\mkRrule{let}}
\frac{
\noEvent T t
}
{
\sensor {\context{\Let x v P},R} {M, T} I O p t
\rightarrow
\sensor {\context{P\subs v x},R} {M, T} I O p t
}
\\[0.2cm]
\tag{\mkRrule{call-sensor}}
\frac{
M(l) = (\text{self}\ \vec x) P
\quad
\noEvent T t
}{
\sensor {\context{\invk \keyw{sensor} l {\vec v}},R} {M,T} I O p t
\rightarrow
\sensor {\context{P\subs {M\ \vec v} {\text{self}\ \vec x}},R} {M, T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{call-module}}
\frac{
M'(l) = (\text{self}\ \vec x) P \quad \noEvent T t
}{
\sensor {\context{\invk {M'} l {\vec v}},R} {M,T} I O p t
\rightarrow
\sensor {\context{P\subs {M'\ \vec v} {\text{self}\ \vec x}},R} {M, T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{no-function}}
\frac{
l \not\in \operatorname{dom}(M) \quad \noEvent T t
}{
\sensor {\context{\invk \keyw{sensor} l {\vec v}},R} {M,T} I O p t
\rightarrow
\sensor {\obj{},R::\context{\invk \keyw{sensor} l {\vec v}}} {M, T} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{timer}}
\frac{
T' = T \uplus (\msg l {\vec v}, v, t + v', t + v)
\quad \noEvent T t
}{
\sensor {\context{\fire \msg l {\vec v} {v} {v'}},R} {M,T} I O p t
\rightarrow
\sensor {\context{\obj{}},\invk \keyw{sensor} l {\vec v}::R} {M,T'} I O p {t+1}
}
\\[0.2cm]
\tag{\mkRrule{trigger}}
\frac{
t \le v' \quad T' = T \uplus (\msg l {\vec v}, v, v', t + v)
}{
\sensor {P,R} {M,T \uplus (\msg l {\vec v}, v, v', t)} I O p t
\rightarrow
\sensor {P,\invk \keyw{sensor} l {\vec v}::R} {M, T'} I O p t
}
\\[0.2cm]
\tag{\mkRrule{expire}}
\frac{
t > v'
}{
\sensor {P,R} {M,T \uplus (\msg l {\vec v}, v, v', t)} I O p t
\rightarrow
\sensor {P,R} {M, T} I O p t
}
\end{gather*}
\center See Definition~\ref{def:plus} for the formal meaning of operator $+$.
\caption{Reduction semantics for sensors.}
\label{fig:reduction}
\end{figure}
Within sensors reduction proceeds without obstacle while the internal
clock $t$ is not such that a timed call must be triggered. This is
controlled by the predicate $\underline{\mathop{\mathrm{noEvent}}}$
that checks the time of the next activation for every timed call
against the current time.
There is no special reason
why the increments in the clock are unitary. One could easily assume
that each instruction consumes a different number of processor cycles
and reflect that scenario in the rules. Some rules (\emph{e.g.}\; \mkRrule{let}{})
simply re-structure a process and thus we assume that no cycles are
consumed.
Rule \mkRrule{extern}{} calls a synchronous external function and receives a
value as the result. The rules \mkRrule{install-sensor}{} and \mkRrule{install-module}{} handle
module updates. The former takes the module with the code installed at
the sensor and updates it with the code of another module $M'$. The
resulting new module is installed in the sensor. The latter applies
only to volatile anonymous modules and therefore the resulting module
is not installed in the sensor. The rule \mkRrule{send}{} (\mkRrule{receive}{})
handles the interaction with the network by putting (getting) messages
in (from) the outgoing (incoming) queue. Notice that receiving a
message is non-blocking (\mkRrule{no-message}).
The rules \mkRrule{call-sensor}{}, \mkRrule{call-module}{} and \mkRrule{no-function}{} handle calls to
functions in modules. \mkRrule{call-sensor}{} selects the function in the
sensor's module, gets its code and replaces the parameters with the
arguments passing the sensor's module $M$ as the first argument
in variable \text{self}. \mkRrule{call-module}{} is similar to \mkRrule{call-sensor}{} but uses
module $M'$ instead of the sensor's module $M$.
Rule \mkRrule{no-function}{} handles the case of a call to a
function that is not yet installed. The call is deferred to the end of
the run-queue. The idea is that the module containing the function
may not have arrived at the sensor to be installed and so we postpone
the execution of the function.
When a value of $t$ is reached such that it implies the triggering of
a call, the rules \mkRrule{trigger}{} and \mkRrule{expire}{} come into action. Rule
\mkRrule{trigger}{} places a timed function call $l(\vec v)$ at the front of
the run-queue. The execution of the call is delegated to rule
\mkRrule{call-sensor}{}. Note that only calls to functions installed in the
sensor~(in $M$) are allowed. Other calls are deferred to the end of the
run-queue by the rule \mkRrule{no-function}{}. If the timer has expired, rule
\mkRrule{expire}{} removes the corresponding tuple from $T$.
\begin{figure}
\begin{gather*}
\\[0.2cm]
\tag{\mkRrule{network}, \mkRrule{congr}}
\frac{
S \rightarrow S'
}
{
S \,\vert\, S''
\rightarrow
S' \,\vert\, S''
}
\qquad
\frac{
S_1 \equiv S_2
\qquad
S_2 \rightarrow S_3
\qquad
S_3 \equiv S_4
}{
S_1 \rightarrow S_4
}
\\[0.2cm]
\\[0.2cm]
\tag{\mkRrule{broadcast}}
\frac{
\operatorname{inRange}(p,p')
\qquad (I'',O'') = \operatorname{networkRoute}(m, I', O')
}
{
\tagsensor {P,R}
{M, T} I {m::O} p t S \,\vert\,
\sensor{P',R'}{M', T'} {I'} {O'} {p'}{t'}
\rightarrow
\tagsensor {P,R}
{M, T} I {m::O} p t
{S \,\vert\, \sensor{P', R'}
{M', T'} {I''} {O''} {p'}{t'}}
}
\\[0.2cm]
\\[0.2cm]
\tag{\mkRrule{release}}
{
\tagsensor {P,R} {M, T} I {m::O} p t S
\rightarrow
\sensor {P,R} {M, T} I O p t \,\vert\, S
}
\end{gather*}
\caption{Reduction semantics for sensor networks.}
\label{fig:reduction-net}
\end{figure}
Network level reduction proceeds concurrently with in-sensor
processing. It handles the distribution of messages placed by the
sensors in their outgoing queues.
A message broadcast starts with the creation of an empty membrane for
the broadcasting sensor (rule \mkSrule{init-Send}{} from the structural
congruence). Then, each time a new sensor is added to the membrane of
a broadcasting sensor (rule \mkRrule{broadcast}), a function
$\underline\operatorname{networkRoute}$ decides where the message in the $O$ queue of the
broadcasting sensor should be copied into the new sensor. The function
can be thought off as implementing the routing protocol for the sensor
network.
The message broadcast ends with the destruction of the membrane, the
captive sensors becoming again free to engage in communication (rule
\mkRrule{release}).
\section{The Type System}
\label{sec:types}
In this section we present a simple type system for Callas, discuss
run-time errors, and prove a type safety result guaranteeing that a
well-typed sensor network does not get ``stuck'' while computing.
\paragraph{Type checking.}
\label{sec:type-checking}
The syntax for types is depicted in Figure~\ref{fig:syntax-types}.
Types $\tau$ are built from the built-in type~$\beta$, the types
for functions $\arrowType {\vec \tau} {\tau}$, where~$\vec \tau$ is
the type for parameters of the function and $\tau$ is its return
type, the types for the sensor code module $\anyObj {l_i \colon \vec \tau_i \rightarrow \tau_i}_{i \in I}$ that is a record
type gathering type information for each function of the code module,
the types for anonymous code modules $\obj {l_i \colon \vec \tau_i \rightarrow \tau_i}_{i \in I}$, recursive types,
and type variables. The need for distinct code module types comes from
the fact that we need to distinguish from installing code in the sensor module
or in an anonymous module.
The $\mu$ operator is a binder, giving rise, in the standard way, to
notions of bound and free variables and alpha-equivalence. We do not
distinguish between alpha-convertible types. Furthermore, we take an
equi-recursive view of
types~\cite{pierce:types-programming-languages}, not distinguishing
between a type $\recType{\alpha} {\tau}$ and its unfolding $\tau\subs\recType{\alpha} {\tau} X$.%
\begin{figure}
\begin{align*}
& \tau \Space{::=} & & \text{\emph{Types}}
\\
& \qquad \; \beta & & \text{built-in type}
\\
& \quad \;\mid\; \arrowType {\vec \tau} {\tau} & & \text{recursive function type}
\\
& \quad \;\mid\; \anyObj {l_i \colon \vec \tau_i \rightarrow \tau_i}_{i \in I} & & \text{sensor code type}
\\
& \quad \;\mid\; \obj {l_i \colon \vec \tau_i \rightarrow \tau_i}_{i \in I} & & \text{anonymous code type}
\\
& \quad \;\mid\; \recType{\alpha} {\tau} & & \text{recursive type}
\\
& \quad \;\mid\; \alpha & & \text{type variable}
\end{align*}
\caption{The syntax of types.}
\label{fig:syntax-types}
\end{figure}
\begin{definition}
\label{def:plus}
The $+$ operator is defined (overloaded) for modules, code types, and
type environment as follows:
\begin{itemize}
\item $\obj{l_i = (\vec x_i) P_i}_{i \in I} + \obj{l'_j = (\vec x_j) P'_j}_{j \in J} =
\obj{\l_i = (\vec x_i) P_i, l'_j = (\vec x_j) P'_j}_{i \in (I \setminus J), j \in J}$
\item $\obj{l_i \colon \tau_i}_{i \in I} + \obj{l'_j \colon \tau'_j}_{j \in J} =
\obj{\l_i \colon \tau_i, l'_j \colon \tau'_j}_{i \in (I \setminus J), j \in J}$
\item $\Gamma_1 + \Gamma_2 = (\Gamma_1 \setminus \Gamma_2) \cup
\Gamma_2$.
\end{itemize}
\end{definition}
The typing rules for values, processes, sensors and queues are
presented in Figures~\ref{fig:type-system-values}
to~\ref{fig:type-system-queues}.
Type judgments for values are of the form $\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v \colon
\tau$, where $\tau_S$ and $\tau_M$ are code module types
representing the types for the built-in functions of the sensor ($\tau_S$) and
for functions installed in the sensor memory $\tau_M$, and $\Gamma$ is a
typing environment mapping variables to types.
The rules are straightforward, but notice that rule \mkTrule{sensor}{} assigns the
sensor code type $\tau_M$ to $\keyw{sensor}$ value.
The judgments for processes are the same as for values.
Rule~\mkTrule{extern}{} ensures that no user-defined function is executed as a
system call and that a system call always belongs to a predefined
type $\tau_s$ ($\tau_S \vdash l \colon \functionType {\vec \tau} {\tau}$).
Broadcasting a call (Rule~\mkTrule{send}{}) is only possible if the call can
be made locally ($\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \invk \keyw{sensor} l {\vec v} \colon \obj{}$)
and for functions that return the empty module, since it is an
asynchronous remote call and no value is going to be returned
(cf.\@ the return value of a system or a local call, which is synchronous).
Notice that the type system does not distinguish between
local and remote functions, however such refinement may be
interesting and can easily be added.
Installing code in the sensor's code module~(Rule~\mkTrule{sInstall})
implies that the module is entirely replaced and that its type is preserved.
On the other hand, installation over an anonymous
module~(Rule~\mkTrule{mInstall}) is more flexible and only requires that
functions common to both code modules should agree
on their type (\textit{vide} the definition of~$+$ operation).
When calling a local function the type of the first parameter
($\tau_1$) corresponds to the type of module containing the function
being called (\textit{vide} operation semantic Rules~\mkRrule{call-sensor}{} and~\mkRrule{call-module}{}
in Figure~\ref{fig:reduction}).
The rules for $\keyw{let}$ and $\keyw{receive}$ are straightforward.
Finally, firing an event (Rule~\mkTrule{timer}{}) amounts to calling a user-defined
function locally.
Typing judgments for sensor networks are of the form $\tau_S;
\tau_M \vdash S$. We only comment the rule for typing a sensor
(Rule~\mkTrule{sensor}{}), in particular, that the type of each function
in the sensor's code module~($M$) must agree with predefined
sensor's type interface~($\tau_M$), apart from the self parameter.
Typing the run-queue (Rule~\mkTrule{run-queue},
Figure~\ref{fig:type-system-queues}), the incoming and outgoing
queues, and the event table is equivalent to typing each element of
the structure individually (Rules~\mkTrule{comm-queue}{} and~\mkTrule{event-queue}{}).
Notice that each element of the incoming (outgoing) queue is typable if it
can be called as a local sensor function. The same holds for timed calls ($\msg l {\vec v}$).
\begin{figure}
\begin{gather*}
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash b \colon \beta
\qquad
\seqSets {\tau_S} {\tau_M} {\Gamma} \operatorname{,} x \colon \tau \vdash x \colon \tau
\qquad
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \keyw{sensor} \colon \tau_M
\tag{\mkTrule{built-in}, \mkTrule{var}, \mkTrule{sensor}}
\\[0.2cm]
\frac{
j \in I
}
{
\senObj{l_i \colon \functionType {\vec \tau_i} {\tau_i}}_{i \in I}
\vdash l_j \colon \functionType {\vec \tau_j} {\tau_j}
}
\qquad
\frac{
\forall i. \seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v_i \colon \tau_i
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \vec v \colon \vec \tau
}
\tag{\mkTrule{label}, \mkTrule{seq}}
\\[0.2cm]
\frac{
\forall i \in I.\seqSets {\tau_S} {\tau_M} {\Gamma}
\operatorname{,} s_i \colon \tau_{M'} \operatorname{,} \vec x_i \colon \vec \tau_i \vdash P_i \colon \tau_i
\qquad
\tau_{M'} = \recType \alpha {\obj{l_i \colon
\functionType{\alpha \vec{\tau_i}} \tau_i}_{i \in I}}
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \obj{l_i = {\abstr {s_i,\vec x_i}{P_i}}}_{i \in I} \colon \tau_{M'}
}
\tag{\mkTrule{code}}
\end{gather*}
\center where $\senObj{l_i \colon \functionType {\vec \tau_i} {\tau_i}}_{i
\in I} $ means either a sensor or an anonymous code type.
\caption{Typing rules
for values.}
\label{fig:type-system-values}
\end{figure}
\begin{figure}
\begin{gather*}
\tag{\mkTrule{extern},\mkTrule{send}}
\frac{
\tau_S \vdash l \colon \functionType {\vec \tau} {\tau}
\qquad
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \vec v \colon \vec \tau
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \system {\msg l {\vec v}} \colon \tau
}
\qquad
\frac{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \invk \keyw{sensor} l {\vec v} \colon \obj{}
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \send {\msg l {\vec v}} \colon \obj{}
}
\\[0.2cm]
\tag{\mkTrule{sInstall}, \mkTrule{mInstall}}
\frac{
\begin{array}{c}
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v_1 \colon
\recType \alpha {\anyObj {l_i \colon \alpha \vec \tau_i \rightarrow \tau_i}_{i \in I}}\\
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v_2 \colon
\recType \alpha {\obj {l_i \colon \alpha \vec \tau_i \rightarrow \tau_i}_{i \in I}}
\end{array}
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \install {v_1} {v_2} \colon \obj{}
}
\qquad
\frac{
\begin{array}{c}
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v_1 \colon \tau_1
\qquad
\tau_1 = \obj{l_i \colon \vec \tau_i \rightarrow \tau_i}_{i \in I}\\
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v_2 \colon \tau_2
\qquad
\tau_2 = \obj{l_j \colon \vec \tau_j \rightarrow \tau_j}_{j \in J}
\end{array}
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \install {v_1} {v_2} \colon \tau_1 + \tau_2
}
\\[0.2cm]
\tag{\mkTrule{call},\mkTrule{let}}
\frac{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v_1 \colon \tau_1
\quad
\tau_1 \vdash l \colon \functionType {\tau_1 \vec \tau}{\tau_2}
\quad
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \vec v_2 \colon \vec \tau
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \invk {v_1} {l} {\vec v_2} \colon \tau_2
}
\qquad
\frac{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash P_1 \colon \tau_1
\quad
\seqSets {\tau_S} {\tau_M} {\Gamma} \operatorname{,} x \colon \tau_1 \vdash P_2 \colon \tau_2
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \Let x {P_1} {P_2} \colon \tau_2
}
\\[0.2cm]
\tag{\mkTrule{receive},\mkTrule{timer}}
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \receivek \colon \obj{}
\qquad
\frac{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \invk \keyw{sensor} l {\vec v} \colon \obj{}
\qquad
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash v_1 v_2 \colon \beta \beta
}
{
\seqSets {\tau_S} {\tau_M} {\Gamma} \vdash \fire {\msg l {\vec v}} {v_1} {v_2} \colon \obj{}
}
\end{gather*}
\caption{Typing rules for processes.}
\label{fig:type-system-processes}
\end{figure}
\begin{figure}
\begin{gather*}
\tau_S;\tau_M \vdash \textbf{\textsf{0}}
\qquad
\frac{
\begin{array}{c}
\tau_S;\tau_M;\emptyset \vdash P \colon \_
\quad
\tau_S;\tau_M \vdash R
\quad
\tau_S;\tau_M;\emptyset \vdash M \colon
\recType \alpha
{\obj{l_i \colon \functionType {\alpha \vec \tau_i}
{\tau_i}}_{i \in I}}
\\
\forall i \in I.\tau_M \vdash l_i \colon
\functionType {\tau_M \vec \tau_i} {\tau_i}
\qquad
\tau_S;\tau_M; \emptyset \vdash t p \colon \beta \beta
\\
\tau_S;\tau_M \vdash T
\qquad
\tau_S;\tau_M \vdash I
\qquad
\tau_S;\tau_M \vdash O
\end{array}
}
{
\tau_S;\tau_M \vdash \sensor {P,R} {M, T} I O t p
}
\tag{\mkTrule{off}, \mkTrule{sensor}}
\\[0.2cm]
\frac{
\tau_S;\tau_M \vdash \sensor {P,R} {M, T} I O t p
\qquad
\tau_S;\tau_M \vdash S
}
{
\tau_S;\tau_M \vdash \tagsensor {P,R} {M,T} I O p t S
}
\qquad
\frac{
\tau_S;\tau_M \vdash S_1
\qquad
\tau_S;\tau_M \vdash S_2
}
{
\tau_S;\tau_M \vdash S_1 \,\vert\, S_2
}
\tag{\mkTrule{bSensor}, \mkTrule{par}}
\end{gather*}
\caption{Typing rules for sensor networks.}
\label{fig:type-system-sensors}
\end{figure}
\begin{figure}
\begin{gather*}
\frac{
\tau_s;\tau_M;\emptyset \vdash P \colon \_
\qquad
\tau_s;\tau_M \vdash R
}
{
\tau_S;\tau_M \vdash P::R
}
\qquad \qquad
\frac{
\tau_S;\tau_M;\emptyset \vdash \invk \keyw{sensor} l {\vec v} \colon \_
\qquad
\tau_S;\tau_M \vdash I
}
{
\tau_S;\tau_M \vdash \langle \msg l {\vec v} \rangle :: I
}
\tag{\mkTrule{run-queue}, \mkTrule{comm-queue}}
\\[0.2cm]
\frac{
\tau_S;\tau_M;\emptyset \vdash \invk \keyw{sensor} l {\vec v} \colon \_
\qquad
\tau_S;\tau_M;\emptyset \vdash v_1 v_2 v_3 \colon \beta_1 \beta_2 \beta_3
\qquad
\tau_S;\tau_M \vdash T
}
{
\tau_S;\tau_M \vdash T \uplus (\msg l {\vec v}, v_1, v_2, v_3)
}
\tag{\mkTrule{event-queue}}
\end{gather*}
\caption{Typing rules for queues of messages, processes, and events.}
\label{fig:type-system-queues}
\end{figure}
The proofs for our main results (Theorem~\ref{teo:subject-reduction},
Theorem~\ref{teo:type-safety}, and Corollary~\ref{cor:type-safety})
are based on the following auxiliary results.
We call \textit{context process}, denoted $\context{P}$, the
processes resulting from filling its context hole.
Informally, Lemma~\ref{lemma:context-process} states that if a context
process is well typed, then the same also holds for the process that
fills its hole, although not necessarily with an identical type.
Lemma~\ref{lemma:process-context} states that the typability of a
context process holds and its type is preserved if we fill the context's
hole with processes of the same type.
Lemma~\ref{lemma:module-substitution} handles module's substitution.
Lemmas~\ref{lemma:substitution-lemma} and~\ref{lemma:congruence} are
discussed below.
\begin{lemma}
\label{lemma:context-process}
If $\tau_S; \tau_M;\Gamma \vdash \context P \colon \tau$, then $\tau_S;
\tau_M;\Gamma' \vdash P \colon \tau'$.
\end{lemma}
\begin{proof}
The proof proceeds by induction on the contexts' structure and
both cases are straightforward.
\end{proof}
\begin{lemma}
\label{lemma:process-context}
If $\tau_S; \tau_M;\Gamma \vdash \context P \colon \tau$, $\tau_S;
\tau_M; \Gamma' \vdash P \colon \tau'$, and $\tau_S; \tau_M;\Gamma' \vdash P' \colon
\tau'$, then $\tau_S; \tau_M;\Gamma \vdash \context {P'} \colon \tau$.
\end{lemma}
\begin{proof}
We proceed by induction on the contexts' structure analysing each
definition case. Both cases follow easily.
\end{proof}
\begin{lemma}
\label{lemma:module-substitution}
If $\tau_S; \tau_M;\Gamma \vdash M_1 \colon \tau_1$ and $\tau_S; \tau_M;\Gamma'
\vdash M_2 \colon \tau_2$,
then $\tau_S; \tau_M; \Gamma + \Gamma' \vdash M_1 +
M_2 \colon \tau_1 + \tau_2$.
\end{lemma}
\begin{proof}
Directly from the definition of $+$ and using Rule~\mkTrule{code}.
\end{proof}
The Substitution Lemma is used in the proof of the Subject Reduction
Theorem, to show the cases that involve the replacement of formal by
actual parameters, specifically for function call and for the let
construct.
The proof is standard, so we omit it, but the interested reader may
find similar proofs in the literature, for instance
in~\cite[Section 6.3]{sangiorgi.walker:theory-mobile}.
\begin{lemma}[Substitution Lemma]
\label{lemma:substitution-lemma}
If $\tau_S; \tau_M;\Gamma \vdash v \colon \tau'$ and $\tau_S; \tau_M;\Gamma
\operatorname{,} x \colon \tau' \vdash P \colon \tau$, then $\tau_S; \tau_M;\Gamma \vdash
P \subs v x \colon \tau$.
\end{lemma}
The following results state type invariance during reduction.
\begin{lemma}[Congruence Lemma]
\label{lemma:congruence}
If $\tau_S; \tau_M \vdash S$ and $S \equiv S'$, then $\tau_S;
\tau_M \vdash S'$.
\end{lemma}
\begin{proof}
We proceed by induction on the derivation tree for $S \equiv S'$.
The proof is straightforward.
\end{proof}
\begin{theorem}[Subject Reduction]
\label{teo:subject-reduction}
If $\tau_S; \tau_M \vdash S$ and $S \rightarrow S'$, then $\tau_S;
\tau_M \vdash S'$.
\end{theorem}
\begin{proof}
By induction on the derivation tree for $S \rightarrow S'$.
In each case, we proceed by case analysis on the last typing rule of
the inference tree for $\tau_S; \tau_M \vdash S$.
\end{proof}
\paragraph{Type safety.}
\label{sec:type-safety}
Our claim is that well-typed sensor networks are free from run-time errors.
The unary relation $S \overset{\text{err}}{\longmapsto}$, defined as the least relation on
networks closed under the rules in Figure
\ref{fig:runtime-errors}, identifies processes that would get ``stuck''
during computation (reduction).
We write $S \overset{\text{err}}{\,\,\arrownot\!\!\longmapsto}$ for $\lnot (S \overset{\text{err}}{\longmapsto})$.
Our Sensor Networks may exhibit two kinds of failures upon
computing: when calling a function or when installing a module.
In the former, the call may result in a run-time error when the target
of the call is neither \lstinline{sensor}, nor an anonymous module (Rule
\mkErule{call});
or when the function name is unknown or there is a mismatch
between the number of arguments ($v_1 \dots v_n$) and the number
parameters ($x_1 \dots x_m$) (Rule \mkErule{cFunction}).
In the latter, an error may occur if we are installing some value
that is not a module (Rule \mkErule{install}).
As an example, recall the \lstinline{gather} function from the
streaming data example that we sketched below.
\begin{lstlisting}[frame=none]
{ gather = (self,x,y) ...}
\end{lstlisting}
The process
\begin{lstlisting}[frame=none]
let t = extern getTime() in send gather(t)
\end{lstlisting}
exhibits a run-time error, since function
\lstinline|gather| is being called with two arguments instead of three.
In fact, the above network may reduce using Rules~\mkRrule{extern}{} and~\mkRrule{let}, but
then we cannot apply Rule \mkRrule{function}, since the substitution is not
defined.
Run-time error Rule \mkErule{cFunction}{} captures this kind of failure.
\begin{figure}
\begin{align*}
\sensor {\context{\invk v l {\vec v}}, R} {M,T} I O p t & \overset{\text{err}}{\longmapsto} & &
\text{if } v \text{ is not \lstinline{sensor}, nor } M'
\tag{\mkErule{call}}\\[0.07cm]
\sensor {\context{\invk {M'} l {v_1 \dots v_n}}, R} {M,T} I O p t & \overset{\text{err}}{\longmapsto} & &
\text{if } l \not \in \operatorname{dom}(M') \text{ or }
\\
& & & (M'(l) = \abstr {x_1 \dots x_m} {P} \text{ and }\\
& & & n \neq m)
\tag{\mkErule{cFunction}}\\[0.07cm]
\sensor {\context {\install v v}, R} {M, T} I O p t & \overset{\text{err}}{\longmapsto} & &
\text{if } v \text{ is not \lstinline{sensor}, nor } M'
\tag{\mkErule{install}}\\[-1cm]
\end{align*}
\begin{align*}
\frac{
S \overset{\text{err}}{\longmapsto}
}
{
S \,\vert\, S' \overset{\text{err}}{\longmapsto}
}
\qquad \qquad
\frac{
S \equiv S' \quad
S \overset{\text{err}}{\longmapsto}
}
{
S' \overset{\text{err}}{\longmapsto}
}
\tag{\mkErule{par}, \mkErule{str}}
\end{align*}
\caption{Run-time errors for sensors.}
\label{fig:runtime-errors}
\end{figure}
The Type Safety result states that well-typed networks do not incur in
run-time errors.
\begin{theorem}[\bf Type Safety]
\label{teo:type-safety}
If $\tau_S;\tau_M \vdash S$, then $S \overset{\text{err}}{\,\,\arrownot\!\!\longmapsto}$.
\end{theorem}
\begin{proof}
We prove the contra-positive result, namely $S \overset{\text{err}}{\longmapsto}$ implies that
$\tau_S;\tau_M \not \vdash S$, proceeding by induction on the
definition of $S \overset{\text{err}}{\longmapsto}$ relation.
\end{proof}
Finally, a well-typed network is free of flaws, at any time during
reduction.
\begin{corollary}[\bf Absence of Runtime Errors]
\label{cor:type-safety}
If $\tau_S;\tau_M \vdash S$ and $S \reduces^{*} S'$, then $S' \overset{\text{err}}{\,\,\arrownot\!\!\longmapsto}$.
\end{corollary}
\begin{proof}
By hypothesis $\tau_S;\tau_M \vdash S$, then, since types are
preserved during reduction (Theorem \ref{teo:subject-reduction}), by
induction on the length of $\reduces^{*}$ we obtain $\tau_S;\tau_M
\vdash S'$.
%
Using the Type Safety theorem (Theorem \ref{teo:type-safety}) we
conclude that $S' \overset{\text{err}}{\,\,\arrownot\!\!\longmapsto}$.
\end{proof}
\section{Related work}
\label{sec:related}
The majority of available programming tools for
sensor networks are based on rather low-level programming languages,
most notably the module-based idiom nesC~\cite{nesc:gay:levis:etal},
which promotes a system level programming style on top of a
small-scale operating system such as TinyOS~\cite{tinyos}. Other
examples include C and Prothothreads for Contiki~\cite{contiki} and at
the extreme Pushpin~\cite{pushpin}.
Moving away from the hardware and system level programming we have
virtual machines like Mat\'{e}~\cite{mate:levis:culler:02} and its
associated core language TinyScript that provide programmers with a
suitable abstraction layer for the hardware.
Middleware platforms such as Deluge~\cite{deluge:hui:culler:04} and
Agilla~\cite{agilla:fok:roman:lu:05} enable higher level control of
sensor networks for critical operations such as massive code
deployment.
True high-level programming languages such as Regiment~\cite{regiment},
Cougar~\cite{cougar}, and TinyDB~\cite{tinydb} abstract away from the
physical network by viewing sensor networks as time varying data
streams or as data repositories.
Regiment, for instance, adopts a data-centric view of sensor networks
and provides the programmer with abstractions to manipulate data
streams and to manage network regions.
Although Regiment is a strongly typed language --- an essential
characteristic to enable the scalable development of applications
--- its construction is not based on a formal calculus and it is not
clear that the semantics is amenable to proving correctness results
for the system and applications.
In fact, the state-of-the-art in the design of sensor network
programming languages~\cite{bookchapter} follows, invariably, a
\emph{top-down} approach, in which system engineers start by
identifying useful patterns and abstractions based on case studies of
applications and then attempt to provide the programmer with language
constructs and system features that reflect these patterns. These
building blocks must then be compiled into nesC/TinyOS code or some
other API that interacts with the low-level operating system.
The problem with such approaches is that the semantic gap between
the original language specification and the actual implementation
inevitably precludes a thorough analysis of the correctness of
the envisioned sensor networking application.
Seeking a fundamentally sound path towards the development of
programming languages for sensor networks, we propose
a somewhat disruptive \emph{bottom-up} approach.
Inspired by process calculi
theory~\cite{async-pi:honda:tokoro:91,pi:milner:parrow:walker:92}, our
basic idea is to start by constructing a fundamental programming
model, which (a) captures the specific computing and communication
aspects of sensor networks and (b) enables us to reason about their
fundamental operations.
This approach is justified by the fact that most high-level languages,
even those that fully abstract from the networking aspects and view
sensor networks as time varying data streams or data repositories,
ultimately map their high-level constructs into a lower level
communication-centric language and run-time system.
Previous work on process calculi for wireless systems is scarce.
Prasad~\cite{broadcast:prasad:91} established the first process
calculus approach to modeling broadcast based systems.
Later work by Ostrovsk\'y, Prasad, and
Taha~\cite{broadcast-high-order:ostrovsky:prasad:taha:02} established
the basis for a higher-order calculus for broadcasting systems.
More recently, Mezzetti and
Sangiorgi~\cite{wireless:mezzetti:sangiorgi:06} discuss the use of
process calculi to model wireless systems, again focusing on the
details of the lower layers of the protocol stack
(\emph{e.g.\@}~collision avoidance) and establishing an operational
semantics for the networks.
\section{Conclusions}
\label{sec:conclusions}
\paragraph{Discussion.}
\label{sec:discussion}
Programming languages based on type safe specifications are
fundamental for applications where development and debugging can be
complex. Sensor networks are one such case. Difficulties in physically
accessing deployed sensors, resource limitations of the devices, and
dynamic ad-hoc routing protocols, all conspire to make the programming
and debugging of these infrastructures a difficult task.
In this paper we present a strongly typed calculus for programming
sensor networks. Sensor network applications are built by plugging
together components called \emph{modules}. Dynamic reprogramming
is supported by making modules first class entities that
can be exchanged between sensors and by allowing modules to be
installed locally upon reception on a sensor. A type system provides a
static verification tool, which allows for premature detection of
protocol errors in the usage of modules. This feature is of utmost
importance when programming large-scale applications for sensor
networks, since it eliminates many errors that would have to be
corrected online, at run-time. We prove two fundamental
properties of the operational semantics and of the type system,
namely, \emph{subject reduction} and \emph{type safety}. Together,
these results establish the calculus as a sound framework for
developing programming languages for sensor networks.
\paragraph{Future work.}
As part of our ongoing work, we are pursuing two different lines of
research. First, we are exploring the theoretical properties of the
calculus. By applying techniques from process calculi theory we hope
to be able to prove fundamental properties of sensor networking
applications and protocols (\emph{e.g.}\; protocol correctness). Second, we
designed a core programming language based on the calculus and
implemented the corresponding compiler and virtual machine. We expect
to prove the correctness of the virtual machine relative to the base
calculus. This will provide an unequivocal link between the semantics
of the calculus (and core language) and the semantics of higher-level
programming languages that we implement on top of it.
~\vspace{-0.5cm}
\paragraph{Acknowledgments.}
The authors are partially supported by project CALLAS of the
Funda\c{c}\~ao para a Ci\^encia e Tecnologia (contract
PTDC/EIA/71462/2006).\\[-0.8cm]
|
\section{Introduction.} Given a closed smooth $n-$manifold $W$,
consider the collection ${\mathcal M}_W$ of connected smooth codimension zero
submanifolds of $W$.
\begin{definition} \label{def: arbiter}
A {\em topological arbiter} on $W$ is a function ${\mathcal A}\co
{\mathcal{M}}_W\longrightarrow \{0,1\}$ satisfying axioms (1) -- (3):
(1) \parbox[t]{14.25cm}{{\em ``${\mathcal A}$ is topological''}: If $M, M'\in {\mathcal M}_W$ and $M$ is ambiently isotopic
to $M'$ in $W$ then ${\mathcal A}(M)={\mathcal A}(M')$.}
(2) \parbox[t]{14.25cm}{{\em ``Greedy axiom''}: If $M\subset M'$
and ${\mathcal A}(M)=1$ then ${\mathcal A}(M')=1$.}
(3) \parbox[t]{14.25cm}{{\em ``Duality''}: Suppose $A,B\in{\mathcal M}_W$ are such that $W=A\cup B$, with $A\cap B
=\partial A\cap \partial B$.
Then ${\mathcal
A}(A)+{\mathcal A}(B)=1$.}
\end{definition}
The notion of a topological arbiter was introduced by the first author in \cite{F0} (the term {\em exclusive property} was used
in that reference) in connection with percolation theory. The purpose of this paper is to investigate further the structure of these
invariants in $4$ and higher dimensions.
A prototypical example of an arbiter is the homological arbiter ${\mathcal A}_h$ on ${\mathbb R}P^2$. Specifically,
given $M\subset {\mathbb R}P^2$ define ${\mathcal A}_h(M)=1$ if $M$ carries the non-trivial
first homology class of ${\mathbb R}P^2$, i.e. if $$H_1(M; {\mathbb Z}/2)\longrightarrow
H_1({\mathbb R}P^2; {\mathbb Z}/2)\cong {\mathbb Z}/2$$ is onto, and set ${\mathcal A}_h(M)=0$ otherwise. The first two axioms
are immediate, and (3) easily follows from Poincar\'{e} duality, see lemma \ref{projective space} below.
In fact, ${\mathcal A}_h$ is the {\em unique} topological arbiter on the projective plane. Indeed, suppose there is a different
arbiter ${\mathcal A}$ on ${\mathbb R}P^2$, so there is $A\subset {\mathbb R}P^2$ such that
${\mathcal A}(A)=1$ but $H_1(A; {\mathbb Z}/2)\longrightarrow
H_1({\mathbb R}P^2; {\mathbb Z}/2)$ is the trivial map. Consider the complement
$B={\mathbb R}P^2\smallsetminus A$, then by Poincar\'{e}-Lefschetz duality $B$ carries the first homology of ${\mathbb R}P^2$.
Since $A$ lies in the complement of a non-trivial cycle, it is contained in a $2$-cell $D^2\subset {\mathbb R}P^2$. It follows from the axioms (2), (3) that ${\mathcal A}(D^2)=0$, so by (2) ${\mathcal A}(A)=0$, a contradiction proving that ${\mathcal A}_h$ is the unique
arbiter on the projective plane.
It is easy to see that for any $n$, the $n$-sphere $S^n$ does not admit a topological arbiter:
consider a decomposition $S^n=A\cup B$ where both $A,B$ are diffeomorphic to $D^n$. According to axiom (1), ${\mathcal A}(A),
{\mathcal A}(B)$ are either both $0$ or both $1$, contradicting the duality axiom (3).
More generally, note that any manifold $W$ admitting an open book
decomposition does not support a topological arbiter. (However $W$ may admit a {\em multi-}arbiter, see the Appendix.)
Indeed, in this case the complement
of the binding fibers over $S^1$, where the binding is a specified codimension $2$ submanifold of $W$.
Decompose $S^1$ into a union of two intervals and consider the corresponding decomposition
$W=A\cup B$. $A$ and $B$ are diffeomorphic (in fact, isotopic), so again axioms (1), (3)
contradict the existence of an arbiter. Note that odd-dimensional manifolds admit open book decompositions \cite{Lawson, Winkelnkemper}.
Simply-connected even-dimensional
manifolds (of dimension $\geq 6$) are open books, provided that the signature is zero \cite{Winkelnkemper}.
In the general even dimensional case there is a further obstruction given by nonsingular bilinear forms \cite{Quinn}.
Generalizing the discussion of the projective plane above, consider the ``homological'' arbiter on even-dimensional
real and complex projective spaces:
\begin{lemma} \label{projective space} \nl \sl
(1) Let $M\subset{\mathbb R}P^{2n}$.
Define ${\mathcal A}_h(M)=0$ if $H_n(M;{\mathbb Z}/2)\longrightarrow
H_n({\mathbb R}P^{2n}; {\mathbb Z}/2)\cong {\mathbb Z}/2$ is the trivial map and
${\mathcal A}_h(M)=1$ otherwise.
(2) Fix a field $F$ and let $M\subset{\mathbb C}P^{n}$, where $n$ is even. Define ${\mathcal
A}_F(M)=0$ if $H_n(M;F)\longrightarrow H_n({\mathbb C}P^{n}; F)\cong F$ is
the trivial map and ${\mathcal A}_F(M)=1$ otherwise.
Then ${\mathcal A}_h$ is a topological arbiter on
${\mathbb R}P^{2n}$, and for any $F$, ${\mathcal A}_F$ is a
topological arbiter on ${\mathbb C}P^{n}$.
\end{lemma}
{\em Proof}. The first two axioms are immediate. To prove axiom (3), let $W={\mathbb R}P^{2n}$ or ${\mathbb
C}P^{n}$, let the coefficients $F={\mathbb Z}/2$ in the first case
and an arbitrary field in the second case. Consider the long exact
sequences
\begin{equation} \label{comm diagram}
\xymatrix{
H_n(A) \ar[r]
\ar[d] &
H_n(W)
\ar[d] \ar[r] &
H_n(W,A) \ar[d]
\\
H^n(W,B) \ar[r]
&
H^n(W)
\ar[r] &
H^n(B)
}
\end{equation}
where the vertical maps are isomorphisms given by
Poincar\'{e}-Lefschetz duality (cf \cite{Hatcher}). Since $H_n(W;F)\cong F$ it is clear
that precisely one of the two maps $H_n(A;F)\longrightarrow H_n(W;F)$,
$H_n(B;F)\longrightarrow H_n(W;F)$ is non-trivial. \qed
There are further examples of manifolds and arbiters induced by homology.
For example, generalizing lemma \ref{projective space} one may define homological arbiters on any
even-dimensional manifold $W^{2n}$ whose middle-dimensional homology group $H_n(W;F)$ has odd rank.
The arbiter ${\mathcal A}_F$ associates $1$ to submanifolds representing more than half the rank of
$H_n(W;F)$.
The results in this paper concern arbiters on even-dimensional balls $D^{2n}$, the local version
of topological arbiters discussed above. {\em Local arbiters} (see definition \ref{def: local arbiter}
in section \ref{sec: Local arbiters} below) associate a value $0$ or $1$ to codimension zero submanifolds of $D^{2n}$
with a prescribed boundary condition: $M\cap \partial D^{2n}$ is a regular neighborhood of the standard
$(n-1)$-sphere $S^{n-1}$ in $S^{2n-1}=\partial D^{2n}$, $\partial M\cap \partial D^{2n}
=\partial D^n\times D^n$, writing $D^{2n}=D^n\times D^n$. The duality axiom satisfied by local arbiters is modeled on
the Alexander duality satisfied by submanifolds $D^{2n}=A\cup B$ bounding the Hopf link $S^{n-1}\cup S^{n-1}
\subset S^{2n-1}$. The reader may wish to glance ahead to absorb definition \ref{def: local arbiter}.
All arbiters discussed so far are induced by homology (with various coefficients).
It follows from the universal coefficient theorem that
the arbiter ${\mathcal A}_F$ on ${\mathbb C}P^n$ (and the analogous local arbiter on $D^{2n}$)
depends only on the characteristic of the
field $F$, so $\{ {\mathcal A}_F \}$ is a countable collection.
Focusing on $D^4$, we construct
a collection of arbiters different from the homological ones:
\begin{theorem} \label{uncountable} \sl There exists an uncountable collection of local topological arbiters in dimension $4$.
\end{theorem}
The construction (discussed in more detail below and in section \ref{sec: Local arbiters})
relies on the existence of non-abelian embedding obstructions in $D^4$, reflected
in particular by the Milnor group. These obstructions, detecting in a certain sense the failure of the Whitney trick in $4$ dimensions,
have been extensively studied in the setting of link-homotopy theory \cite{Milnor}.
Since the Whitney trick is valid in dimensions $> 4$, our construction of (uncountably
many) arbiters on $D^4$ does not have an immediate analogue in higher dimensions.
In section \ref{sec: Higher dimensions} we give a rather different construction of families of local arbiters, not induced by homology, in
dimensions $6$ and higher. These higher-dimensional arbiters
are homotopy-theoretic in nature, the following theorem summarizes our construction:
\begin{theorem} \label{higher dim thm}\sl
To each non-trivial square in the stable homotopy ring of spheres
there is associated a local topological arbiter not induced by homology on
$D^{2n}$ for sufficiently large $n$. In particular, there exist local arbiters not induced by homology on $D^{2n}$ for each
$n>2$.
\end{theorem}
The main ingredient used in the construction of these arbiters in higher dimensions is the following homotopy-theoretic obstruction
to a generalized link-slicing problem: for each $n>2$ we show that there exist submanifolds $(M,S^{n-1})\subset (D^{2n},S^{2n-1})$,
where $S^{n-1}$ is a distinguished ``attaching'' sphere in the boundary of $M$, such that
(1) \parbox[t]{14.25cm}{for any coefficient ring $R$ the attaching $(n-1)$-sphere
of $M$ is non-trivial in $H_{n-1}(M; R)$, and}
(2) \parbox[t]{14.25cm}{The components of the Hopf link $S^{n-1}\cup S^{n-1}\subset S^{2n-1}$ do not bound
disjoint embeddings of two copies of $M$ into $D^{2n}$.}
Observe that if the attaching sphere $S^{n-1}$ were trivial in $H_{n-1}(M;F)$ for some field coefficients $F$, then Alexander
duality would immediately imply (2). The proof of (1) and (2), given in section \ref{sec: Higher dimensions},
can be carried out using a ``secondary'' obstruction to disjointness where
$n+1$-cells are attached to $S^{n-1}$ via the generator of ${\pi}_1^S$, $n>3$. A further refinement is necessary for $D^6$, see section \ref{D6}.
To prove theorem \ref{higher dim thm}, one defines a ``partial'' arbiter, setting it equal to $1$ on all submanifolds of $D^{2n}$ containing
$M$ and equal to zero on all submanifolds contained in $D^{2n}\smallsetminus M$. The condition (2) above implies that this partial arbiter
is consistently defined, and (1) shows that it is different from any homological arbiter. The proof of the theorem is completed
by extending this partial arbiter to all submanifolds of $D^{2n}$, the details are given in section \ref{sec: Higher dimensions}.
A brief outline of the rest of the paper is as follows.
The proof of theorem \ref{uncountable} begins in section \ref{sec: Local arbiters}
with the introduction of a dyadic collection of ``model'' submanifolds of $D^4$ (i.e.
a special collection of submanifolds, parametrized by vertices of a trivalent tree.)
A generalization of the
Milnor group is used to find obstructions to pairwise embeddings of these model submanifolds. The collection ${\mathcal M}_{D^4}$ of all submanifolds
of $D^4$ has a partial ordering induced by inclusion.
The non-embedding results mentioned above allow one to
define an uncountable collection of arbiters on those elements of ${\mathcal M}_{D^4}$ which are comparable
to model submanifolds. One then extends these arbiters to ${\mathcal M}_{D^4}$, concluding the proof of theorem
\ref{uncountable}, see section \ref{sec:proof}.
Section \ref{sec: embeddings} points out that the embeddings of the submanifolds $M\hookrightarrow W$ in definition \ref{def: arbiter}
of arbiters play a very important role. The result of \cite{K0} shows that if one considers topological arbiters as
invariants of manifolds, disregarding their embedding information, then in the context of $D^4$ whenever the attaching circle
of $M\subset D^4$ is trivial in $H_1(M;{\mathbb Z})$ the value of any arbiter on $M$ has to be $1$.
The proof of theorem \ref{uncountable} is valid only when the embedding data is taken into account. Indeed
the arbiters constructed in its proof assume value $1$ on manifolds where the attaching circle does not vanish in homology with any
coefficients.
Section \ref{sec: CP2} adapts
the local construction in the proof of theorem \ref{uncountable}
to the setting of ${\mathbb C}P^2$. It seems likely that this construction should yield an
uncountable collection of arbiters on ${\mathbb C}P^2$, however proving this would require a substantial
extension of the techniques of Milnor's theory of link homotopy.
In appendix \ref{sec: Surgery} we note that the existence of a local arbiter on $D^4$ satisfying an
additional ``Bing doubling'' axiom would give an obstruction to topological $4$-dimensional surgery for
the non-abelian free groups, using the A-B slice reformulation \cite{F, F1} of the surgery conjecture.
Appendix \ref{sec: multiarbiters} introduces two generalizations of topological arbiters, {\em multi-arbiters}.
We discuss their properties and applications, in particular a construction of non-homological multi-arbiters,
and an application to percolation theory.
\section{Local arbiters.} \label{sec: Local arbiters}
The purpose of this section is to prove theorem \ref{uncountable}.
We start by
defining the notion of a local topological arbiter on $D^4$ (more generally, the definition
makes sense for $D^{2n}$ for any even dimension $2n$.)
Consider
$\mathcal{M}=\mathcal{M}_{D^4}=\{(M,{\gamma})|\, M$ is a codimension zero, smooth,
compact submanifold of $D^4$, ${\gamma}\subset
\partial M$, and $M\cap\partial D^4$ is a tubular neighborhood of an
unknotted circle ${\gamma}\subset S^3 \}$. (For $n>2$, $M\cap \partial D^{2n}$
is a tubular neighborhood of a standard $S^{n-1}$, isotopic in $\partial D^{2n}
=\partial (D^n\times D^n)$ to $\partial D^n\times D^n$.)
Consider a local version of definition \ref{def: arbiter}:
\begin{definition} \label{def: local arbiter}
A {\em local topological arbiter} is a function ${\mathcal A}\co
{\mathcal{M}}\longrightarrow \{0,1\}$ satisfying axioms (1) -- (3):
(1) \parbox[t]{14.25cm}{If $(M,{\gamma}), (M',{\gamma}')\in {\mathcal M}$ and $(M,{\gamma})$ is ambiently isotopic
(not necessarily relative to the identity on the boundary) to $(M',{\gamma}')$ in $(D^4,S^3)$ then ${\mathcal A}(M,{\gamma})={\mathcal A}(M',{\gamma}')$.}
(2) \parbox[t]{14.25cm}{If $(M,{\gamma})\subset (M',{\gamma}')$ and
${\mathcal A}(M,{\gamma})=1$ then ${\mathcal A}(M',{\gamma}')=1$.}
(3) \parbox[t]{14.25cm}{Suppose $(A,{\alpha}),(B,{\beta})\in{\mathcal M}_{D^4}$ are such that $D^4=A\cup B$, with $A\cap B
=\partial A\cap \partial B$, and so that the attaching curves ${\alpha}, {\beta}$ of $A,B$
form the Hopf link in $\partial D^4$. (In other words, $D^4=A\cup B$ is a decomposition of
$D^4$ extending the standard genus $1$ Heegaard decomposition of the
$3-$sphere.) Then ${\mathcal
A}(A,{\alpha})+{\mathcal A}(B,{\beta})=1$. (For $n>2$ the ``Hopf link'' above is replaced with
$\partial D^n\times 0\sqcup 0\times \partial D^n\subset S^{2n-1}.$)}
\end{definition}
We summarize local versions of the properties discussed in the introduction.
In dimension $4$ (more generally, in any even dimension), given a field $F$ there is a ``homological''
local arbiter ${\mathcal A}_F$. This arbiter (by the universal coefficient theorem depending
only on the characteristic of $F$) is defined by ${\mathcal A}_F(M,{\gamma})=1$ if and only if ${\gamma}=0\in
H_1(M;F)$. The axioms (1), (2) are satisfied automatically, (3) is a
consequence of Alexander duality.
In dimension $2$ there is a unique local arbiter ${\mathcal A}$ (equal to ${\mathcal A}_F$ for any
field $F$.)
The difference between arbiters in dimensions $2$ and higher is explained by the divergence
between homology and homotopy: for $(A,{\alpha})\subset (D^2,S^1)$, ${\alpha}=0\in H_0(A)$ iff ${\alpha}=0\in {\pi}_0(A)$.
If this is the case, the complement $B$ of $A$ in $D^2$ may be pushed off an arc connecting $\alpha$ in $A$ into a collar
on its attaching curve $\beta$, therefore the value of any arbiter on $B$ is zero.
In higher dimensions, being trivial in homology is a weaker condition, and we show that there exist arbiters
whose value on $B$ may be $1$ even though $\beta$ is homologically non-trivial with any coefficients.
\subsection{Model decompositions.} \label{sec: Model decompositions}
The proof of theorem \ref{uncountable} starts with the construction of a special collection of submanifolds of $D^4$, which are shown
below to satisfy pairwise non-embedding properties. These codimension zero
submanifolds form complementary pairs $A_I, B_I$ parametrized by a dyadic multiindex $I$ discussed in more
detail below. For each index $I$,
the pair is a decomposition of the $4$-ball in the sense of definition \ref{def: local arbiter} (3) above: $D^4=A_I\cup B_I$, and
the distinguished curves ${\alpha}_I, {\beta}_I$ form the Hopf link in $\partial D^4$.
To start the construction, consider the genus one surface
$S$ with a single boundary component, and set $B_1=S\times D^2$.
Denote ${\beta}_1=\partial S\times 0$.
To specify an embedding $(B_1,{\beta}_1)\subset (D^4, S^3)$, embed the surface $S$ into $S^3$, push
it into the $4$-ball and take a regular neighborhood.
\begin{figure}[ht]
\includegraphics[width=2.6cm]{surface.eps} \hspace{3cm} \includegraphics[width=3.5cm]{Bingpair.eps}
{\small
\put(-290,37){$B_1$}
\put(-258,48){${\gamma}_1$}
\put(-212,42){${\gamma}_2$}
\put(-270,-6){${\beta}_1$}
\put(-80,-6){${\alpha}_1$}
\put(-116,73){$H_1$}
\put(3,80){$H_2$}
\put(-8,23){$A_1$}}
\caption{}
\label{fig:surface Bing double}
\end{figure}
\begin{proposition} \label{surface complement}
\sl The complement $A_1=D^4\smallsetminus B_1$ is obtained from the collar
$S^1\times D^2\times I$ by attaching a pair
of zero-framed $2$-handles to the Bing double of the core of the
solid torus $S^1\times D^2\times 1$, figures \ref{fig:surface Bing double}, \ref{fig:Kirby Bing double}.
\end{proposition}
\begin{figure}[ht]
\vspace{.3cm}
\includegraphics[width=4cm]{KirbySurface1.eps} \hspace{3cm} \includegraphics[width=4cm]{KirbyBing1.eps}
{\small
\put(-336,5){$B_1$}
\put(-128,5){$A_1$}}
{\scriptsize
\put(-304,20){$0$}
\put(-259,28){${\beta}_1$}
\put(-107,30){$0$}
\put(-17,22){$0$}
\put(-46,35){${\alpha}_1$}}
\caption{}
\label{fig:Kirby Bing double}
\end{figure}
The distinguished curve ${\alpha}_1\subset \partial A_1$ is formed by $S^1\times 0 \times 0$.
Figure \ref{fig:surface Bing double} is a schematic ``spine'' picture of
the manifolds $A_1, B_1$, a precise description in
terms of Kirby diagrams (drawn in a solid torus) is given in figure \ref{fig:Kirby Bing double}.
Note that a distinguished pair of curves ${\gamma}_1, {\gamma}_2$,
forming a symplectic basis in the surface $S$, is determined as the
meridians (linking circles) to the cores of the $2-$handles $H_1,
H_2$ in $D^4$. In other words, ${\gamma}_1$,
${\gamma}_2$ are fibers of the circle normal bundles over the cores
of $H_1, H_2$ in $D^4$.
The proof of proposition \ref{surface complement} is an exercise in Kirby calculus, see for example
\cite{FL}.
The manifolds $A_1, B_1$ are the basic building blocks in the inductive construction of model decompositions.
Consider the $2$-handle $H_1$ (one of the $2$-handles in the definition of $A_1$, figure \ref{fig:surface Bing double})
in place of the original
$4$-ball. The pair of curves (${\gamma}_1$, the attaching circle
${\delta}_1$ of $H_1$) forms the Hopf link in the boundary of $H_1$.
In $H_1$ consider the standard genus one surface bounded by
${\delta}_1$. As discussed above, its complement is given by two
zero-framed $2$-handles attached to the Bing double of ${\gamma}_1$.
Assembling this data, consider the new decomposition $D^4=A_0\cup
B_0$. The manifold $A_0$ and its Kirby diagram are shown in figure \ref{fig:A0} below, and both
$A_0, B_0$ are included in figure \ref{fig:model decompositions}.
We are now in a position to define the more general collection of special
elements $A_I, B_I$ of ${\mathcal M}$ (``model decompositions'',
$D^4=A_I\cup B_I$), figure \ref{fig:model decompositions}. These decompositions correspond to
vertices of a trivalent tree, and each $A_I, B_I$ is labeled by a
dyadic multiindex $I$. Each arrow in figure \ref{fig:model decompositions} represents a
an inclusion. These model decompositions are defined
inductively; the transition from $A_I$ to $A_{I,0}$ replaces one of
the two $2$-handles of $A_I$, introduced during the previous step,
by $A_0$. $A_{I,1}$ is obtained by replacing one of these two
``new'' $2$-handles of $A_I$ by $A_1$ (in other words, by Bing
doubling this $2$-handle.) Then $B_I$ is defined as the complement
$D^4\smallsetminus A_I$. Note that this is a subset of the collection of
decompositions of $D^4$ introduced in the context of the A-B slice problem in \cite{FL}.
Figure \ref{fig:model decompositions} shows only the ``spines'' of the submanifolds
$A_I, B_I$; we refer the reader to \cite{FL} for a precise description of
these $4$-manifolds in terms of
Kirby diagrams.
\begin{figure}[ht]
\includegraphics[height=8cm]{models2.eps}
{\footnotesize
\put(-200,209){$A$}
\put(-280,127){$A_0$}
\put(-103,127){$A_1$}
\put(-340,12){$A_{00}$}
\put(-250,12){$A_{01}$}
\put(-150,12){$A_{10}$}
\put(-58,12){$A_{11}$}
\put(-180,185){${\alpha}$}
\put(-238,113){${\alpha}_0$}
\put(-61,113){${\alpha}_1$}}
\includegraphics[height=8cm]{othermodels.eps}
{\footnotesize
\put(0,265){}
\put(-190,7){$B$}
\put(-274,63){$B_0$}
\put(-99,65){$B_1$}
\put(-323,165){$B_{00}$}
\put(-233,165){$B_{01}$}
\put(-140,165){$B_{10}$}
\put(-53,165){$B_{11}$}
\put(-146,-6){${\beta}$}
\put(-223,45){${\beta}_0$}
\put(-53,43){${\beta}_1$}}
\caption{The collection of model decompositions $A_I, B_I$. The
arrows represent inclusions.}
\label{fig:model decompositions}
\end{figure}
Consider the {\em lexicographic order} on the set of dyadic
multiindices: given $I=(i_1,\ldots,i_m)$, $J=(j_1,\ldots,j_n)$, set
$I<J$ if $i_1=j_1,\ldots, i_k=j_k$ and $i_{k+1}=0, j_{k+1}=1$ for
some $k$. Also set $I<J$ if $I$ is an initial segment of $J$:
$J=(I,j_{k+1},\ldots$)
Consider the trivalent tree $T$ encoding the submanifolds $A_I$. The root
of $T$ corresponds to $A=2$-handle. Let $r$ be a ``geodesic'' ray in $T$
originating at the root; $r$ is viewed as an infinite dyadic multiindex.
For a technical reason which will be apparent in the proof below, it is convenient to consider all rays $r$ passing
through $A_1$ (i.e. whose first coefficient is $1$),
and the lexicographic order is extended
to the set of such rays.
\begin{definition}
Let ${\mathcal M}_0\subset {\mathcal M}$ denote the collection of submanifolds
$\{ A_I, B_I\}$ defined above.
For each $r$, a ray as above, define a {\em partial arbiter}
${\mathcal A}_r\co{\mathcal M}_0\longrightarrow \{0,1\}$ by setting
$${\mathcal A}_r(A_I)=1 \,\, {\rm if} \, \, I<r, \,\, {\rm and}\, \,
{\mathcal A}_r(A_I)=0\, \, {\rm if}\,\, I>r.$$ Then the
duality axiom $(3)$ in definition \ref{def: local arbiter} forces one to define $${\mathcal A}_r(B_I)=0
\,\, {\rm if}\,\, I<r, \,\, {\rm and} \,\, {\mathcal A}_r(B_I)=1\,\,
{\rm if} \,\, I>r.$$
\end{definition}
\medskip
\begin{lemma} \label{greedy axiom} \sl
For each $r$, ${\mathcal A}_r\co{\mathcal M}_0 \longrightarrow \{0,1\}$ satisfies
the greedy axiom (2) in definition \ref{def: local arbiter}.
In other words, given $(M,{\gamma}),(M',{\gamma}')\in {\mathcal M}_0$ with $(M,{\gamma})\subset (M',{\gamma}')$, if
${\mathcal A}_r(M,{\gamma})=1$ then ${\mathcal A}_r(M',{\gamma}')=1$.
\end{lemma}
The proof follows from the following proposition.
Recall from the beginning of section \ref{sec: Local arbiters}
that each element $(M,{\gamma})\in {\mathcal M}$ comes equipped with an
embedding $(M,{\gamma})\subset (D^4,S^3)$.
\begin{proposition} \label{proposition} \sl \nl
(1) If $I<J$ then $(A_I,
{\alpha}_I)\not\subset (A_J, {\alpha}_J)$.
(2) For any $I, J,$ $(A_I, {\alpha}_I)\not\subset (B_J, {\beta}_J)$.
(3)
For any multiindices $I,J$ whose first entries equal $1$, there does not exist an embedding $(B_I,
{\beta}_I)\subset (A_J, {\alpha}_J)$.
\end{proposition}
We postpone the proof of proposition \ref{proposition} to complete the
proofs of lemma \ref{greedy axiom} and theorem \ref{uncountable}. One needs to check that for
any two elements $(M_1,{\gamma}_1), (M_2,{\gamma}_2)\in {\mathcal M}_0$ if ${\mathcal A}_r(M_1)=1$ and ${\mathcal
A}_r(M_2)=0$ then
$(M_1,{\gamma}_1)\not\subset (M_2,{\gamma}_2)$. Consider the list of all pairs $M_1, M_2$ where
${\mathcal A}_r(M_1)=1$ and ${\mathcal A}_r(M_2)=0$:
(a) $M_1=A_I$ where $I<r$, $M_2=A_J$ where $J>r$,
(b) $M_1=A_I$ where $I<r$, $M_2=B_J$ where $J<r$,
(c) $M_1=B_I$ where $I>r$, $M_2=A_J$ where $J>r$,
(d) $M_1=B_I$ where $I>r$, $M_2=B_J$ where $J<r$.
The cases (a), (b), (c) follow from the statements (1), (2), (3) of proposition \ref{proposition}
respectively. To eliminate (d), suppose
$B_I\subset B_J$ where $J<I$. Considering their complements, there
is an inclusion $A_J\subset A_I$, a contradiction with
proposition \ref{proposition} (1). This completes the proof of lemma \ref{greedy axiom}. \qed
\begin{figure}[ht]
\medskip
\includegraphics[height=6cm]{cones.eps}
{\scriptsize
\put(-70,177){$A=2-$handle}
\put(-145,108){$B_I$}
\put(-66,105){$A_J$}
\put(-106,44){$A_I$}
\put(-27,70){$B_J$}
\put(-145,64){${\mathcal A}_r\!=\!0$}
\put(-112,91){${\mathcal A}_r\!=\!1$}
\put(-73,-14){$B=$collar}}
\medskip
\caption{The vertices of the graph encode codimension zero
submanifolds of $D^4$ (elements of ${\mathcal M}$), and the edges
correspond to inclusions. This figure illustrates the case
$I<r$, $J>r$: then ${\mathcal A}_r(A_I)={\mathcal A}_r(B_J)=1,$ and
${\mathcal A}_r(B_I)={\mathcal A}_r(A_J)=0$. These requirements
force ${\mathcal A}_r(M)=1$ for any $M$ in the ``cones'' over $A_I,
B_J$ and ${\mathcal A}_r(M)=0$ for any $M$ in the ``cones'' under
$B_I, A_J$. The upward-facing cones are disjoint from the
down-facing cones by lemma \ref{greedy axiom}.}
\smallskip
\label{fig:partial ordering}
\end{figure}
\subsection{Proof of theorem \ref{uncountable}.} \label{sec:proof}
Given a ray $r$, by lemma \ref{greedy axiom} the partial
arbiter ${\mathcal A}_r$ is consistently defined on the collection ${\mathcal M}_0$
of model decompositions. Given $M_1, M_2\in {\mathcal M}_0$ with ${\mathcal
A}_r(M_1)=0$, ${\mathcal A}_r(M_2)=1$, axiom (2) in definition \ref{def: local arbiter} forces ${\mathcal
A}_r(M)=0$ for any $M$ isotopic to a submanifold of $M_1$ and ${\mathcal A}_r(M)=1$ for any
$M$ isotopic to $M'$ where $M_2\subset M'$. Graphically, ${\mathcal A}_r=0$ in the ``cone''
below $M_0$ and ${\mathcal A}_r=1$ in the ``cone'' above $M_1$,
figure \ref{fig:partial ordering}. The upward-facing cones are disjoint from the down-facing
cones by lemma \ref{greedy axiom}. This defines a consistent extension of ${\mathcal
A}_r$ to the union of all such cones. Duality axiom (3) in definition \ref{def: local arbiter}
for this extension follows from the fact that the upward-facing cone
over, say, $A_I$ consists precisely of the complements in $D^4$ of
the elements of the down-facing cone under $B_I$.
To extend ${\mathcal A}_r$ to ${\mathcal M}$, consider any local arbiter on $D^4$, for example a homological
arbiter ${\mathcal A}_F$ for some fixed field $F$. Set ${\mathcal A}_r(M)={\mathcal A}_F(M)$ for any $M$
in the complement of the union of the ``cones'' discussed above. It is straightforward to check that ${\mathcal A}_r$
satisfies the axioms of a local arbiter.
Note that if $r\neq r'$ then ${\mathcal A}_r\neq {\mathcal A}_{r'}$:
assume $r<r'$, then their difference is detected by a codimension zero submanifold $A_I$ where
$I$ is a multiindex with $r<I<r'$. This completes the proof of theorem \ref{uncountable}, assuming
proposition \ref{proposition}. \qed
{\bf Remarks.}\\
{\bf 1.} Because we use a definite procedure to extend ${\mathcal A}_r$ beyond the ``cones''
on which it is determined by $r$, the proof of theorem \ref{uncountable} does not depend on the axiom of choice.
Nor does any other assertion in this paper. To be complete, our argument actually has shown
that the cardinality of local arbiters on $D^4$ is precisely the continuum.
{\bf 2.} The argument extending partial arbiters ${\mathcal A}_r$ to arbiters defined
on all codimension zero submanifolds of $D^4$ in the proof of theorem \ref{uncountable} may also be used to define ``hybrids''
of homological arbiters of different characteristics. For example, suppose $(M, {\gamma})$ is an element of
${\mathcal M}$ such that ${\gamma}=0\in H_1(M;{\mathbb Z}/p)$ but ${\gamma}\neq 0\in H_1(M;{\mathbb Z}/q)$
for some primes $p$ and $q$. By Alexander duality with ${\mathbb Z}/p$ coefficients, two copies
of $M$ cannot be embedded disjointly into $D^4$ so that the distinguished curves form the Hopf link on the boundary.
Then a partial arbiter ${\mathcal A}$ may be defined by setting it equal to
${\mathcal A}_{{\mathbb Z}/p}$ in the cone over $M$ and also in the cone under $D^4\smallsetminus M$. Extend it by
${\mathcal A}_{{\mathbb Z}/q}$ to the rest of ${\mathcal M}$. The arbiters constructed in the proof of theorem
\ref{uncountable} assume value $1$ on elements $(M, {\gamma})\in {\mathcal M}$ such that
${\gamma}\neq 0\in H_1(M;R)$ for {\bf any} coefficients $R$, therefore they
are different from any such hybrid of homological arbiters.
\smallskip
{\bf Proof of part (1) of proposition \ref{proposition}.} Before giving the
general (inductive) proof consider several special cases which
clarify the inductive step. (Refer to figure \ref{fig:model decompositions}
and the beginning of section \ref{sec: Model decompositions} to recall the definition
of the submanifolds $A_I, B_I$.)
(i) {\sl $(A_0, {\alpha}_0)$ does not embed in $(A_1, {\alpha}_1)$.}
The proof is immediate since ${\alpha}_0=0\in H_1(A_0)$ and
${\alpha}_1\neq 0\in H_1(A_1)$.
(ii) {\sl $(A_{10}, {\alpha}_{10})\not\subset (A_{11}, {\alpha}_{11})$.}
Suppose such an embedding exists. Then consider the complements:
$(B_{11},{\beta}_{11})\subset (B_{10},{\beta}_{10})$.
The submanifold $B_{11}$ is (a thickening of) a $2$-stage {\em grope}, see \cite{FL} (the term used in that reference is
a {\em half-grope}), a geometric counterpart of the commutators in the lower central series ${\pi}_n$ of a group $\pi$. This series
is defined inductively by ${\pi}_n=[{\pi},{\pi}_{n-1}]$, ${\pi}_1={\pi}$. The attaching curve ${\beta}_{11}$
of the $2$-stage grope $B_{11}$ is in the
third term $({\pi}_1 (B_{11}))_3$.
This is a contradiction since ${\beta}_{11}$ maps to ${\beta}_{10}$, however
$${\beta}_{11}=1\in{\pi}_1(B_{11})/({\pi}_1(B_{11}))_3 \hspace{.25cm} {\rm
but} \hspace{.25cm} {\beta}_{10}\neq
1\in{\pi}_1(B_{10})/({\pi}_1(B_{10}))_3.$$
(iii) {\sl $(A_{00},
{\alpha}_{00})\not \subset (A_{01}, {\alpha}_{01})$.}
\begin{figure}[ht]
\vspace{.5cm}
\includegraphics[height=4cm]{nonembedding11.eps}
{\footnotesize
\put(-283,42){$A_{00}$}
\put(-137,48){$i$}
\put(-5,35){$A_{01}$}
\put(-270,15){${\Sigma}'$}
\put(-80,15){${\Sigma}''$}
\put(-250,-9){${\alpha}_{00}$}
\put(-55,-9){${\alpha}_{01}$}
\put(-210,90){${\gamma}_1$}
\put(-192,-2){${\gamma}_2$}
\put(-20,90){${\delta}_1$}
\put(-2,-2){${\delta}_2$}}
\vspace{.35cm} \caption{}
\label{fig:non-embedding1}
\end{figure}
\begin{figure}[ht]
\vspace{.5cm}
\includegraphics[height=4cm]{nonembedding2.eps}
{\footnotesize
\put(-280,17){$B_{00}$}
\put(-80,17){$B_{01}$}
\put(-129,39){$\overline i$}
\put(-250,-9){${\beta}_{00}$}
\put(-55,-9){${\beta}_{01}$}
\put(-265,40){${\gamma}_1$}
\put(-196,83){${\gamma}_2$}
\put(-80,40){${\delta}_1$}
\put(-6,79){${\delta}_2$}
\put(-239,68){$T'$}
\put(-182,58){$S'$}
\put(-102,73){$T''$}
\put(2,58){$S''$}}
\vspace{.35cm} \caption{}
\label{fig:non-embedding2}
\end{figure}
The argument in this example will serve as an ingredient in the proof in the general case further below.
Suppose the embedding $i$ exists as in figure \ref{fig:non-embedding1}. Considering the complements one gets
$\overline i\co (B_{01},{\beta}_{01})\longrightarrow
(B_{00},{\beta}_{00})$, figure \ref{fig:non-embedding2}. The argument will involve two steps,
the first one using the embedding $i$ and second step using $\overline i$. The connection between the two steps
is provided by the Alexander duality.
{\em Step 1.}
Denote by ${\Sigma}', {\Sigma}''$
the bottom surface stages of $A_{00}, A_{01}$. Since
${\alpha}_{00}$ maps to ${\alpha}_{01}$, $i$ ``restricts'' to a
degree $1$ map $({\Sigma}',{\alpha}_{00})\longrightarrow ({\Sigma}'',{\alpha}_{01})$. (More
precisely, consider the restriction $i|_{{\Sigma}'}$. $A_{01}$ is
homotopy equivalent to the wedge of ${\Sigma}''$ and three
$2-$spheres. The composition of $i|_{{\Sigma}'}$ with the
retraction onto ${\Sigma}''$ has degree $1$.)
Consider the classes in $H_2(B_{00})$, $H_2(B_{01})$ represented by
the tori $T'$, $T''$ and spheres $S'$, $S''$, figure \ref{fig:non-embedding2}. (These are
closed surfaces since they are capped off by the null-homotopies for
the Bing-double curves.) The punctured tori ${\Sigma}'\subset
A_{00}, {\Sigma}''\subset A_{01}$ have preferred symplectic bases of
curves: ${\gamma}_1, {\gamma}_2$ on ${\Sigma}'$, and ${\delta}_1, {\delta}_2$ on
${\Sigma}''$. The curves ${\gamma}_1, {\gamma}_2$ are Alexander dual
respectively to $T', S'$, and similarly ${\delta}_1, {\delta}_2$
are dual to $T'',S''$. Consider the matrix $M=\left(
\begin{smallmatrix} a & b\\ c & d \end{smallmatrix} \right)$ representing
the induced map $i_*\co H_1({\Sigma}')\longrightarrow
H_1({\Sigma}'')$ in these bases. Since the map
$({\Sigma}',\partial)\longrightarrow ({\Sigma}'',\partial)$ has
degree $1$, $M$ is an element of $SL_2({\mathbb Z})$.
{\em Step 2.} The matrix for the map $\overline i_*\co H_2(B_{01})\longrightarrow H_2(B_{00})$,
restricted to the $2$-dimensional subspaces spanned by $[S''],
[T'']$ and $[S'], [T']$, is the transpose of $M$, by Alexander
duality. Since $[S'']$ is a spherical class, the component of
$\overline i_*([S''])$ along $[T']$ is trivial, so $b=0$ and $M$
must be of the form $\left(
\begin{smallmatrix} 1 & 0\\ c & 1 \end{smallmatrix} \right)$.
Therefore the class $[T'']$ maps to $[T']$ plus a multiple $c$ of $[S']$,
but in particular one gets a degree $1$ map $T''\longrightarrow T'$.
This leads to a
contradiction when one considers the elements represented by $T'$, $T''$ in
the {\em Dwyer's filtration} (see \cite{Dwyer}) of the corresponding spaces.
A geometric description of the $n$th term of the Dwyer's filtration ${\phi}_n(X)
\subset H_2(X;{\mathbb Z})$ of a space $X$ is the subgroup consisting of those
elements in $H_2(X;{\mathbb Z})$ which may be represented by maps of closed gropes
of class $n$ into $X$.
Since $[T'']\in {\phi}_3(B_{01})$ but $[T']\not\in
{\phi}_3(B_{00})$, $[T'']$ cannot map to $[T']$, a contradiction.
\medskip
{\bf Proof of part (1) of proposition \ref{proposition} in the general case.} Suppose an
embedding $i\co (A_I, {\alpha}_I)$ $\longrightarrow (A_J,
{\alpha}_J)$ exists; let $\overline i\co
(B_J,{\beta}_J)\longrightarrow (B_I,{\beta}_I)$ be the corresponding
inclusion of the complements. Since $I<J$, $I=(i_1,\ldots, i_n, 0,
\ldots)$ and $J=(i_1, \ldots, i_n, 1, \ldots)$ for some $n$.
First consider the case $i_1=\ldots=i_n=0$, the proof for arbitrary $i_1,\ldots,i_n$ will be analogous.
Since all $i_j$ are assumed to be zero, $j=1\dots,n$, each of the first $n$ steps in the construction
of $A_I, A_J$ replaces an existing $2$-handle with (a thickening of) the genus $1$ surface, and then attaches
a pair of $2$-handles to the Bing double of an essential curve on this surface. Denote the
successive genus $1$ surface stages of $A_I$ by ${\Sigma}'_1,\ldots, {\Sigma}'_n$, and similarly
denote the genus $1$ surface stages of $A_J$ by ${\Sigma}''_1,\ldots, {\Sigma}''_n$.
The construction on the $B$-side is dual (in the sense of proposition \ref{surface complement}):
each of the first $n$ steps attaches a pair of $2$-handles
to a Bing double, and then replaces one of these $2$-handles with a punctured torus. Denote the
successive tori in $B_I$ by $T'_1,\ldots, T'_n$ and the corresponding tori in $B_J$ by $T''_1,\ldots, T''_n$.
(In the special case $I=(00), J=(01)$ considered in (iii) above, $n$ was equal to $1$.)
{\em The base of the induction}.
The embedding $i$ induces a degree $1$ map $({\Sigma}'_1,{\alpha}_I)
\longrightarrow ({\Sigma}''_1,{\alpha}_J)$.
{\em The inductive step}.
The inductive hypothesis is that the embedding $i$ induces a degree $1$ map ${\Sigma}'_k
\longrightarrow {\Sigma}''_k$ for some $1\leq k<n$. The argument in the special case (iii)
above proves that then the complementary embedding $\overline i$ induces a degree $1$ map $T'_k
\longrightarrow T''_k$. Since the construction of the first $n$ stages of the submanifold
$A_I$ is precisely dual to the construction of $B_I$ (and similarly for $A_J, B_J$, compare figures
\ref{fig:non-embedding1}, \ref{fig:non-embedding2}), exactly the same
argument now propagates the conclusion to the $A$-side: if $\overline i$ induces a degree $1$ map
$T''_k \longrightarrow T'_k$, then the embedding $i$ induces a degree $1$ map ${\Sigma}'_{k+1}
\longrightarrow {\Sigma}''_k$. This concludes the proof of the inductive step.
A contradiction is reached after $n-1$ applications of the inductive
step: the outcome is a degree $1$ map ${\Sigma}'_n \longrightarrow {\Sigma}''_n$;
run the argument one more time to get a degree $1$ map
$T''_n\longrightarrow T'_n$. However since the $n+1$-st coefficient of $J$ is $1$, it follows
from the construction of $B_J$ that $[T''_n]\in {\phi}_3(B_J)$. On the other hand
the $n+1$-st coefficient of $I$ is $0$, so $[T'_n]\not\in {\phi}_3(B_I)$. This
contradicts the fact that $\overline i$ induces a degree $1$ map $T''_n\longrightarrow T'_n$, concluding the
proof when $i_1=\ldots=i_n=0$.
A slight modification of the argument is necessary in the general case when $i_1=0$ but
$(i_1,\ldots,i_n)$ contains subsequences with one or more consecutive $1$'s.
Then the corresponding tori on the $B$-side are replaced by
gropes of height $h$ equal to ($1+$ the number of consecutive $1$'s in a subsequence),
all of whose surface stages have genus $1$. (Correspondingly, the $2$-handles on the $A$-side are attached
to higher iterated Bing doubles.) The inductive step discussed above applies without any changes on the $A$-side.
On the $B$-side a variation is necessary: Let $f$ be a map $f\co G\longrightarrow G$ where $G$ is a half-grope of
height $n$ with all surface stages of genus $1$. The second homology class of the bottom stage surface
of $G$ reads off the relation $[{\gamma}_1,[{\gamma}_2,\ldots,[{\gamma}_h,{\gamma}_{h+1}]\ldots]$, where
${\gamma}_i$ are the curves generating ${\pi}_1(G)$, corresponding to the ``tips'' of the grope.
Note that ${\gamma}_h,{\gamma}_{h+1}$ generate ${\pi}_1$ of the top stage surface of $G$.
Suppose $f_*$ carries the class of the bottom stage surface of $G$ in $H_2(G)$ to itself. Then the map
induced by $f$ on the $2$-dimensional subspace of $H_1(G)$ spanned by ${\gamma}_h,{\gamma}_{h+1}$ is in
$SL(2,{\mathbb Z})$. The rest of the argument in the inductive step goes through.
Finally, if $i_1=1$ then the first step of the induction is applied on the
$B$-side. This concludes the proof of \ref{proposition} (1).
\medskip
{\bf The proof of part (2) of proposition \ref{proposition}} involves a generalization of the Milnor group
from the context of link homotopy (a link in $S^3$ bounding disjoint maps of disks in $D^4$)
to a class of more general submanifolds of the $4$-ball.
We outline this construction below and refer the reader to \cite{Krushkal} for a detailed
exposition.
Suppose $(A_I,{\alpha}_I)\subset (B_J, {\beta}_J)$.
Then $A_I$ embeds in the complement of $A_J$, where the attaching curves
${\alpha}_I$, ${\alpha}_J$ form the Hopf link in $S^3=\partial D^4$.
As a warmup, consider the special case where both $I, J$ are dyadic multiindices with no zeros, for example
$A_{1}$ and $A_{11}$ are shown in figure \ref{fig:model decompositions}. For such multiindices $I, J$, the
handlebodies $A_I, A_J$ consist of $2$-handles attached to iterated Bing doubles of their respective
attaching curves ${\alpha}_I, {\alpha}_J$ which form the Hopf link in $\partial D^4$. It is straightforward
to prove that any link $L$ obtained by iterated Bing doubling of the Hopf link is a homotopically essential link \cite{Milnor},
by computing that it has a non-trivial Milnor's ${\bar\mu}$-invariant with non-repeating coefficients. Therefore the components of
$L$ do not even bound disjoint maps of disks into $D^4$, contradicting the assumption
$(A_I,{\alpha}_I)\subset (B_J, {\beta}_J)$.
\begin{figure}[ht]
\centering
\bigskip
\includegraphics[width=3.6cm]{prelimB1.eps}
\hspace{1.5cm} \includegraphics[width=5.6cm]{KirbyAprime1.eps}
{\small
\put(-309,36){$A_0$}
\put(-291,-3){${\alpha}_0$}
\put(-175,42){${\alpha}_0$}}
{\scriptsize
\put(-99,13){$0$}
\put(-82.5,42){$0$}
\put(-53,42){$0$}}
\smallskip
\caption{The submanifold $A_0$ and its Kirby diagram. (The handle diagram is
drawn in the solid torus which is the complement of the dashed circle in the $3$-sphere.)}
\label{fig:A0}
\end{figure}
Now consider a more interesting example $A_0$ (figure \ref{fig:A0}), illustrating the argument in the general case.
Consider a pair of pants $P$ with boundary components ${\gamma}, {\alpha}_1, {\alpha}_2$, and let
$C$ denote $P\times D^2\; \cup$ four zero-framed $2$-handles $h_1\ldots,h_4$ attached to the Bing doubles of the curves ${\alpha}_1, {\alpha}_2$,
figure \ref{cell figure}.
\begin{figure}[ht] \label{model figure1}
\centering
\bigskip
\includegraphics[width=5.75cm]{1.eps}\hspace{2cm} \includegraphics[bb=0 -55 250 100, width=5.5cm]{kirby.eps}
{\small
\put(-349,9){${\gamma}$}
\put(-387,73){$h_1$}
\put(-256,20){$C$}
\put(-360,122){$h_2$}
\put(-282,120){$h_3$}
\put(-227,65){$h_4$}}
{\scriptsize
\put(-164,58){$0$}
\put(-141,80){$0$}
\put(-18,79){$0$}
\put(7,63){$0$}}
{\small
\put(-75,25){${\gamma}$}}
\put(-118,73){\circle*{3}}
\put(-39,72){\circle*{3}}
\vspace{.3cm}
\caption{The handlebody $C$ (a Bing cell of height $1$): a schematic
picture and a Kirby diagram}
\label{cell figure}
\end{figure}
Observe that there is a map $C\longrightarrow A_0$ taking $\gamma$ to ${\alpha}_0$, such that the $2$-handle $h_1$ is disjoint from $h_2$,
and similarly $h_3$ is disjoint from $h_4$. On the level of spines, the proof amounts to cutting the punctured torus into
a pair of pants; the separation of parallel copies of the $2$-handles results in $h_1-h_3$ and $h_2-h_4$ double points.
Therefore to prove that the components of the Hopf link do not bound disjoint embeddings of two copies of $A_0$, it suffices to show that
its components do not bound disjoint maps of two copies $C', C''$ of $C$, subject to the disjointness condition on the $2$-handles discussed above.
Suppose the opposite is true. Consider meridians $m'_i$ to
the $2$-handles $\{h'_i\}$ of $C'$ and meridians $m''_i$ to the $2$-handles $\{ h''_i\}$ of $C''$ in $D^4$.
(By a meridian we mean a based loop in the complement representing the same homology class as a fiber
of the normal circle bundle over a $2$-handle.) Consider the quotient $\overline M {\pi}$
of ${\pi}:={\pi}_1(D^4\smallsetminus (C'\cup C''))$ by the relations $[(m'_i)^x,( m'_j)^y]$ and $[(m''_i)^x, (m''_j)^y]$, where
$\{i,j\}\neq \{1,2\}, \{3,4\}$ and the conjugating elements $x,y$ range over all elements of ${\pi}$.
These relations correspond to the Clifford tori linking the double points between
the various $2$-handles in $D^4$, and the fact that the pairs $\{1,2\}, \{3,4\}$ are excluded reflects the fact that these
handles are required to be disjoint. (This is a generalization of the Milnor group $M{\pi}$ (see \cite{Milnor}) which in this context is defined
as the quotient of ${\pi}$ by the relations $[(m'_i)^x,( m'_i)^y], [(m''_i)^x,( m''_i)^y], i=1,\ldots, 4$, which correspond to
{\em self}-intersections of the handles. These relations are included in the defining relations of $\overline M{\pi}$.)
The classical Milnor group $M{\pi}$ is nilpotent, and since $\overline M{\pi}$ is a quotient of $M{\pi}$, it is nilpotent
as well. It follows that $\overline M{\pi}$ is generated by the chosen meridians, and its relations may be read off from the
generators of $H_2(D^4\smallsetminus (C'\cup C''); {\mathbb Z})$, which is Alexander dual to the relative
group $H_1(C'\cup C'', \partial(C'\cup C'')\cap \partial D^4)$. There are two types of generators of this second homology
group: the Clifford tori that give rise to the commutation relations used in the definition of $\overline M{\pi}$, and
the generators dual to the relative first homology group of the pairs of pants defining $C', C''$. The latter relations
are of the form $[m'_1,m'_2]=[m'_3,m'_4], [m''_1,m''_2]=[m''_3,m''_4]$.
Let $m', m''$ denote the respective meridians to the components of the Hopf link in $\partial D^4$,
formed by the attaching curves ${\gamma}', {\gamma}''$. The relation $[m', m'']=1$ holds in ${\pi}_1(S^3\smallsetminus
({\gamma}'\cup {\gamma}''))$, therefore it also holds in ${\pi}_1(D^4\smallsetminus (C'\cup C''))$. On the other hand,
in terms of the generators of $\overline M{\pi}$, $m'=[m'_1,m'_2]=[m'_3,m'_4], m''=[m''_1,m''_2]=[m''_3,m''_4]$.
One then shows using the Magnus expansion that, based on the relations in $\overline M{\pi}$ discussed above,
the commutator $[m', m'']$ is in fact non-trivial in $\overline M{\pi}$: phrased differently, there are
``not enough'' relations in ${\pi}_1(D^4\smallsetminus (C'\cup C''))$ to imply the relation $[m', m'']=1$
which holds in ${\pi}_1(S^3\smallsetminus ({\gamma}'\cup {\gamma}''))$.
This contradiction shows that
the components of the Hopf link do not bound in $D^4$ disjoint copies of the manifold $A_0$.
The manifolds $A_0$, $A_1$ considered above are the building blocks in the definition of model decompositions
$D^4=A_I\cup B_I$, see section \ref{sec: Model decompositions}. There is a similar inductive construction of {\em Bing cells}
based on the manifold $C$ above,
where the inductive step replaces a $2$-handle with a copy of $C$. {\em Moreover, the $2$-handles in the more general definition
below may be attached to iterated} Bing doubles, since such links arise in the model submanifolds $A_I$ (see for example
$A_{01}$ in figure \ref{fig:model decompositions}.)
\begin{definition} \label{model cell}
A {\em model Bing cell of height $1$} is a smooth $4$-manifold $C$
with boundary and with a specified curve ${\gamma}\subset
\partial C$, defined as follows.
Consider a pair of pants $P$ with boundary components ${\gamma}, {\alpha}_1, {\alpha}_2$, and let
$L_i$ be a (possibly iterated) Bing double of the core of the solid torus ${\alpha}_i\times D^2$.
Then $C$ is obtained from
$P\times D^2$ by attaching zero-framed $2$-handles along the
components of $L_1\cup L_2$.
A {\em model Bing cell $C$ of height $h$} is obtained from a model
Bing cell of height $h-1$ by replacing some of its $2$-handles with Bing cells
of height one.
\end{definition}
{\em Bing cells in a $4$-manifold} $M$ are defined as maps of model
Bing cells into $M$, subject to certain crucial disjointness
requirements. For each link $L_i$ in the inductive definition above,
surfaces and more generally Bing cells attached to different components of
$L_i$ are required to be disjoint in $M$.
Generalizing the arguments above, one shows (see \cite{Krushkal}) that
(1) \parbox[t]{14.25cm}{If the components of a link $L\subset S^3=\partial D^4$ bound
disjoint Bing cells in $D^4$ then $L$ is homotopically trivial, and}
(2) For each index $I$ the attaching curve ${\alpha}_I$ bounds a Bing cell in $A_I$.
Therefore the components of the Hopf link do not bound disjoint embeddings of $A_I, A_J$, i.e. $(A_I,{\alpha}_I)\not
\subset (B_J, {\beta}_J)$. This concludes the proof of proposition \ref{proposition} (2).
{\em The proof of part (3) of proposition} \ref{proposition} is immediate since
${\beta}_I=0\in H_1(B_I)$ and ${\alpha}_J\neq 0\in H_1(A_J)$.
This completes the proof of
proposition \ref{proposition} and of theorem \ref{uncountable}. \qed
\subsection{Arbiters and embeddings.} \label{sec: embeddings}
In this section we note the important role played by the embeddings in definition
\ref{def: local arbiter} of arbiters on $D^4$.
According to \cite{K0} for any given $g$ there exist decompositions (in the sense of axiom (3) in definition
\ref{def: local arbiter}) $D^4=A_g\cup B_g$, where the attaching curves ${\alpha}, {\beta}$ form the Hopf link
in $\partial D^4$, such that the conditions (i), (ii) are satisfied:
(i) Let $S$ denote the genus $g$ surface with one boundary
component, ${\gamma}=\partial S$. Denote by $S_0$ its
untwisted $4-$dimensional thickening, $S_0=S\times D^2$, and
set ${\gamma}_0={\gamma}\times\{0\}$. Then there exists a
proper embedding $(A_g, {\alpha})\subset (S_0,
{\gamma}_0)$.
(ii) $B_g$ embeds in a collar on its attaching curve. More precisely, there
exists a proper embedding $(B_g, {\beta})\subset
(S^1\times D^2\times [0,1],\, S^1\times \{0\}\times\{0\})$
Suppose ${\mathcal A}$ is an invariant as in definition \ref{def: local arbiter}, where the embedding in the axiom (2) in \ref{def: local arbiter}
is assumed to be an arbitrary embedding of abstract pairs ($M,{\gamma}), (M', {\gamma}')$, rather than just the inclusion of submanifolds of $D^4$. Since the value of ${\mathcal A}$ on the collar is necessarily $0$, (ii) implies
that for each $g$, ${\mathcal A}(B_g)=0$, therefore ${\mathcal A}(A_g)=1$. (Note that the embeddings in (i), (ii) are indeed not isotopic to
the original embeddings into $D^4$.)
Now suppose $(M,{\gamma})$ is a submanifold of $(D^4, S^3)$ such that the attaching curve $\gamma$
of the $4$-manifold $M$ is trivial in $H_1(M;{\mathbb Z})$. Then $\gamma$ bounds an embedded surface $S$ in $M$ with the trivial
normal bundle. It follows that $(A_g, {\alpha})\subset (M,{\gamma})$, where $g$ is the genus of $S$. Again by axiom (2) which allows
arbitrary embeddings,
${\mathcal A}(M,{\gamma})=1$. That is, a function ${\mathcal A}$ which is insensitive to the embedding of submanifolds into $D^4$
is determined by homology, at least on those submanifolds whose attaching curve is trivial in the first homology with integer coefficients, and
on their complements in $D^4$.
The proof of theorem \ref{uncountable} is valid only when the embeddings are inclusions of submanifolds of $D^4$, indeed
the arbiters constructed in its proof assume value $1$ on manifolds where the attaching circle does not vanish in homology with any
coefficients.
\subsection{Submanifolds of ${\mathbb C}P^2$.} \label{sec: CP2}
This section outlines a construction of submanifolds of ${\mathbb C}P^2$ which may be viewed as
an ``absolute'' analogue of the relative model decompositions of $D^4$ defined in section \ref{sec: Model decompositions}.
Consider the usual handle structure on ${\mathbb C}P^2$ with one handle of each index $0,2,4$:
${\mathbb C}P^2=0H\cup2H\cup 4H$. Consider decompositions ${\mathbb C}P^2=\widehat A\cup \widehat B$, where
$\widehat A\cap \widehat B=\partial \widehat A\cap \partial \widehat B$. The decompositions are constructed of the form
$\widehat A=0H\cup A$, $\widehat B=4H\cup B$, and $2H=A\cup B$ is a decomposition of the $4$-ball
$2H=D^2\times D^2$ in the sense discussed in section \ref{sec: Local arbiters}: $A\cup B$
extends the standard genus $1$ decomposition of the $3$-sphere $D^2\times S^1\cup S^1\times D^2$.
Applying this construction to model decompositions $D^4=A_I\cup B_I$ of the $4$-ball in section \ref{sec: Model decompositions}
(see figure \ref{fig:model decompositions}), one gets a dyadic collection of decompositions ${\mathbb C}P^2=\widehat A_I\cup \widehat B_I$.
More concretely, $\widehat A_I$ is obtained by attaching a $1$-framed $2$-handle to $A_I$ along the attaching curve ${\alpha}_I$, and
similarly its complement $\widehat B_I$ is obtained by attaching a $1$-framed $2$-handle to $B_I$ along the attaching curve ${\beta}_I$.
It seems reasonable to conjecture that there are non-embedding results analogous to proposition \ref{proposition}, which should
enable one to construct an uncountable collection of topological arbiters on ${\mathbb C}P^2$. However proving this would require an
extension of the Milnor group techniques which is outside the scope of this paper. This question remains open.
\section{Squares in stable homotopy and arbiters in higher dimensions} \label{sec: Higher dimensions}
Recall that homology with field coefficients gives rise to topological arbiters on even dimensional projective spaces,
and similarly there are local homological arbiters on $D^{2n}$ for all $n$, see lemma \ref{projective space} and the
discussion following definition \ref{def: local arbiter}.
In this section we use stable homotopy theory to show that there exist local arbiters different from the homological ones
in even dimensions $\geq 6$.
Recall that the construction of an uncountable
collection of local arbiters on $D^4$ in section \ref{sec: Local arbiters}
was based on obstructions related to the failure of the Whitney trick in dimension $4$. More specifically, this is reflected in
the fact that the middle-dimensional (index $2$) handles of certain submanifolds of $D^4$ cannot be embedded disjointly. This argument
does not have an analogue in higher dimensions, since the Whitney trick is valid then.
Our construction in dimensions $6$ and higher, discussed below, is quite different. It is homotopy-theoretic in nature, and
the non-embedding results involve handles above the middle dimension. We point out that conversely, this higher-dimensional
approach does not work in dimension $4$ since the higher homotopy groups of the circle (the attaching region of the
submanifolds in $D^4$) are trivial.
Before outlining a general construction of families of non-homological arbiters on $D^{2n}$ for all $n\geq 4$, we discuss a specific
example in dimension $8$. (The construction in $D^6$ requires an additional step, discussed in section \ref{D6} below.) Consider codimension zero submanifolds $M\subset D^8$ such that $\partial M \cap \partial D^8$ is a regular
neighborhood of a standard $S^3\subset S^7=\partial D^8$. We will call this $S^3\subset \partial M$ the attaching $3$-sphere of $M$,
and consider {\em proper} embeddings $M\hookrightarrow D^8$ taking the attaching $3$-sphere to $\partial D^8$. Our construction
of a non-homological arbiter on $D^8$ relies on the following result.
\begin{lemma} \label{Hopf lemma} \sl
There exist submanifolds $(M,S^3)\subset (D^8,S^7)$ such that
(1) \parbox[t]{14.25cm}{for any coefficient ring $R$ the attaching $3$-sphere
of $M$ is non-trivial in $H_3(M; R)$, and}
(2) \parbox[t]{14.25cm}{The components of the Hopf link $S^3\cup S^3\subset S^7$ do not bound
disjoint embeddings of two copies of $M$ into $D^8$.}
\end{lemma}
{\em Proof.} Consider a standard $3$-sphere $X=S^3\subset \partial D^8$,
let $C=S^3\times D^4\times I\subset D^8$ be a collar on its regular neighborhood,
and denote by $Y$ a copy pushed into $D^8$, $Y=S^3\times\{0\}\times\{1\}$.
Let ${\Sigma}h\co S^4\longrightarrow Y$ be the suspension of the Hopf map
$h\co S^3\longrightarrow S^2$. Consider
a general position map $f\co D^5\longrightarrow D^8$ whose restriction $f|_{\partial D^5}$ is ${\Sigma}h$. Define $M$ to be the collar
$C$ union a regular neighborhood of the
image of $f$ in $D^8$.
To check (1), first note that the attaching $3$-sphere is non-trivial in homology of the abstract CW-complex
$Z:=S^3\cup_{{\Sigma}h} D^5$. The singular set $S$ of a general position map $f$ is $2$-dimensional. It follows
that the attaching $3$-sphere $X$ is non-trivial in $H_3(M)$, indeed since the third homology of the singular set
is trivial it follows from the Mayer-Vietoris sequence that $X\neq 0\in H_3(Z\cup_S cone(S))$.
Suppose the components $S^3_i$ of the Hopf link bound disjoint copies $M_i, i=1,2$ of $M$ in $D^8$.
The attaching $4$-sphere of the $5$-cell in the definition of $M$ is mapped by ${\Sigma}h$ to the copy $Y$ of $S^3$
pushed into $D^8$. Using the collar structure, it also may be viewed as being mapped by ${\Sigma}h$ to the
actual attaching sphere $X\subset \partial D^8$. Consider the resulting maps $S^4_i\longrightarrow X$.
It is convenient to consider $S^7=\partial D^8$ as ${\mathbb R}^7$ union the point at infinity, bounding the half space
${\mathbb R}^8_{+}$.
Let $A$ denote the
``arrow'' map $$A\co M_1\times M_2\longrightarrow S^7, \; A(x,y)=\frac{x-y}{|x-y|}.$$
Also consider the composition ${\phi}\co S^4_1\times S^4_2\longrightarrow S^3_1\times S^3_2\longrightarrow S^6$
of ${\Sigma} h\times {\Sigma} h$ with the arrow map in ${\mathbb R}^7$.
The $5$-cells $D^5_i$ attached to $S^4_i$ map to $M_i$, and composing with $A$ one has
a map ${\psi}\co D^5_1\times D^5_2\longrightarrow S^7$.
Pick a point $z\in S^6$ which is a regular value both of $\phi$ and $\psi$. The preimage of $z$ in $S^3_1\times S^3_2$
is a single point, and since the fiber of the Hopf map (and of its suspension) is a circle with the Lie group framing, ${\phi}^{-1}(z)$
is an Arf invariant $1$ torus $T$, the generator of
${\Omega}_2^{fr}\cong{\pi}_2^{S}$, cf \cite{Stong}. (Related stable homotopy invariants of higher dimensional links
have been considered by several authors in the context of link homotopy, see for example \cite{MR}, \cite{Kos}.)
However ${\phi}^{-1}(z)$ is the boundary
of the $3$-manifold $M^3={\psi}^{-1}(z)\subset D^5_1\times D^5_2$. Note that $T$ is codimension $6$ in
$S^4\times S^4$, while
$M^3$ is codimension $7$ in $D^5_1\times D^5_2$. Consider the pullback $\nu$ under $\psi$ of the normal line bundle
to $S^6$ in $S^7$. The $5$-cells $D^5_i$ may be assumed to be mapped near their boundary into the collar $C=S^3\times D^4\times I\subset D^8$
as a product ${\Sigma}h\times Id_{I}$. Using this product structure, one checks that there is a well-defined limit of $\nu$ at the boundary
$\partial M = T$, which amounts to a stabilization of the $6$-framing of $T$. This exhibits the Arf invariant $1$ torus $T$ as the framed boundary
of $M$, a contradiction.
\qed
Lemma \ref{Hopf lemma} easily yields a non-homological arbiter ${\mathcal A}$ on $D^8$. This step is analogous to
the corresponding part of the proof of theorem \ref{uncountable} in dimension $4$:
consider the decomposition
$D^8=M\cup N$, where $M$ is the manifold constructed in the proof of lemma \ref{Hopf lemma},
and $N$ is its complement. Given a submanifold $P\subset D^8$ such that $M$ is isotopic to a submanifold
of $P$, set ${\mathcal A}(P)=1$. On the other hand, if $P$ is isotopic to a submanifold of $N$,
set ${\mathcal A}(P)=0$. The condition (2) in lemma \ref{Hopf lemma} ensures that this consistently defines a
partial arbiter on the ``cone'' (with respect to
the partial ordering induced by inclusion) over $M$ and the corresponding cone under $N$, see figure \ref{fig:partial ordering}.
Now pick
any arbiter on $D^8$, for example the rational homological arbiter ${\mathcal A}_{\mathbb Q}$, and use it to extend
${\mathcal A}$ to all other submanifolds of $D^8$. The condition (1) in lemma \ref{Hopf lemma} shows that
${\mathcal A}$ is different from the homological arbiters on $D^8$.
Suspending the Hopf map further, one gets the analogue of lemma \ref{Hopf lemma} and a non-homological arbiter on $D^{2n}$ for any
$n>3$. In fact, the relevant structure used in the proof of the lemma is a non-trivial square in the stable homotopy ring of spheres.
As another illustration of this construction, consider the Hopf fibration $H\co S^7\longrightarrow S^4$. Following the proof of
lemma \ref{Hopf lemma} with ${\Sigma} h$ replaced by ${\Sigma}^3 H$ (and suspending ${\Sigma}^3 H$ further), one gets arbiters on $D^{2n}, n\geq 8$.
Here at least three suspensions are needed to ensure that the dimension of the singular set $S$ is smaller than the dimension of the attaching sphere of $M$, so that \ref{Hopf lemma} (1) holds.
The following theorem summarizes the results discussed above, the case $n=3$ is proved in section \ref{D6} below.
\begin{theorem} \sl
To each non-trivial square in the stable homotopy ring of spheres
there is associated a local topological arbiter not induced by homology on
$D^{2n}$ for sufficiently large $n$. In particular, there exist non-homological local arbiters on $D^{2n}$ for each
$n\geq 3$.
\end{theorem}
\subsection{Arbiters on $\mathbf{D^6}$} \label{D6} Constructing non-homological arbiters on $D^6$ requires a bit more care.
Indeed, note that a direct analogue of the proof of lemma \ref{Hopf lemma} (with the actual Hopf map, rather than its suspension) does not hold.
(A conceptual explanation for this is provided by Lemma \ref{trivial in homology} below.) Specifically, to recall the notation consider
a standard $2$-sphere $X=S^2\subset \partial D^6$, let $C=S^2\times D^3\times I\subset D^6$ be a collar on its regular neighborhood,
and denote by $Y$ a copy of the $2$-sphere pushed into $D^6$, $Y=S^2\times\{0\}\times\{1\}$.
Let $h\co S^3\longrightarrow Y$ be the Hopf map. Consider
a general position map $f\co D^4\longrightarrow D^6$ whose restriction $f|_{\partial D^4}$ is $h$, and define $M$ to be the collar
$C$ union a regular neighborhood of the
image of $f$ in $D^6$.
Analogously to the proof of \ref{Hopf lemma}, the attaching $2$-sphere is non-trivial in homology of
$Z:=S^2\cup_{h} D^4(={\mathbb C}P^2)$, and the singular set $S$ of a general position map $f$ is still $2$-dimensional. However, in contrast with
\ref{Hopf lemma}, we show in the
lemma below that the attaching $2$-sphere is in fact trivial in $H_2(M;{\mathbb Z}/2)$ for any map $f$.
To illustrate this point, we consider a specific map
$f\co ({\mathbb C}P^2, {\mathbb C}P^1)\longrightarrow (D^4, S^3)$. Take a Chern class one normal circle bundle $B$ over $Y$ in $D^6$.
(To be more specific, $B$ is considered as a subset of the boundary $Y\times S^3$ of a regular neighborhood of $Y$, and the fiber
of $B\longrightarrow Y$ over $y\in Y$ is the fiber $h^{-1}(y)\subset \{ y\}\times S^3$ of the Hopf fibration $h$.)
Cone $B$ to the origin of
$D^6$ and also consider the mapping cylinder of the fibration $B\longrightarrow Y$, that is the circle fiber over each
point $y\in Y$ is coned off to $y$. (This mapping cylinder gives a map of the Chern class $1$ disk bundle into $D^6$, bounded by $B$.) This construction yields a map of $CP^2$ into $D^6$ which meets $\partial D^6$ in the $2$-sphere $X$ and the singularities may be arranged to be contained in the suspension $\Sigma$ of a single fiber $F$ of the circle bundle $B$.
The suspension is formed by the disk fiber bounded by $F$ union with ($F$ coned off
to the center of $D^6$.)
Note that this suspension $\Sigma$ represents the same class in $H_2({\mathbb C}P^2)$ as $[CP^1]$. Moreover, when $\Sigma$
is mapped into $D^6$ it is flattened into a plane, thereby killing the attaching $2$-sphere $X$ in $H_2(M;{\mathbb Z})$.
The reader may find it useful to consider a construction analogous
to the above for $({\mathbb R}P^2, {\mathbb R}P^1)$ mapped into $(D^3, S^2)$ where the Chern class $1$ bundle
$B$ is replaced by the M\"obius band. The singularity in this case is formed by a circle, generating
${\pi}_1({\mathbb R}P^2)$, flattened in a line through the origin in $D^3$.
The following lemma shows that the behavior exhibited in the example above holds (at least with ${\mathbb Z}/2$ coefficients) for {\em any}
general position map $f$.
\begin{lemma} \label{trivial in homology} \sl
Let $f\co {\mathbb C}P^2\longrightarrow D^4$ be a general position map taking ${\mathbb C}P^1$ to a $2$-sphere in $\partial D^4$.
Then the class $[{\mathbb C}P^1]$ is trivial
in $H_2(f({\mathbb C}P^2);{\mathbb Z}/2)$.
\end{lemma}
Before proving this lemma, observe that as a consequence the construction given in lemma \ref{Hopf lemma} of an arbiter
on $D^8$ fails to give a {\em non-homological} arbiter on $D^6$: since the attaching $2$-sphere is trivial in
$H_2(M;{\mathbb Z}/2)$, there is a homological obstruction to embedding two copies of $M$ into $D^6$ forming the
Hopf link in $\partial D^6$. A non-homological arbiter on $D^6$ is constructed further below in section \ref{non-homological in six dimensions}.
{\em Proof of lemma \ref{trivial in homology}.} Let $M$ be a regular neighborhood of $f({\mathbb C}P^2)$ in $D^6$. Attaching another
$6$-cell to $D^6$, consider $M$ as a submanifold of $S^6$. Suppose the statement of the lemma is false for this map $f$, then in
particular $[{\mathbb C}P^1]\neq 0\in H_2(\partial M;{\mathbb Z}/2)$. $\partial M$ is a $5$-manifold,
and there exists \cite{Thom} a $3$-manifold $N^3\subset \partial M$ representing the Poincar\'{e} dual $PD([{\mathbb C}P^1])
\in H_3(\partial M;{\mathbb Z}/2)$. Push $N$ outward to get a copy $N'$ which links the $2$-sphere
$Y$ in $S^6$. Consider the arrow map $A\co S^2\times N\longrightarrow S^5$, and composing with the
Hopf map one has $A'\co S^3\times N\longrightarrow S^5$. Since the linking number between $Y=S^2$ and $N'$ is $1$ (mod $2$), the preimage
$A'^{-1}(y)$ is a collection of an odd number of circles, each one with the Lie group framing.
We will give an argument assuming the preimage $A'^{-1}(y)$ is a single such circle $C$, the general case is proved analogously.
Since $N$ is disjoint from $M$, $A'$ extends to a map $D^4\times N\longrightarrow S^5$, and therefore
$A'^{-1}(y)$ is null-cobordant: there exists a framed (perhaps non-orientable) surface $F\subset D^4\times N^3$, $\partial F=C$.
Consider $D^4$ as the unit ball in ${\mathbb R}^4$; $C$ is a circle with non-trivial framing in $\partial D^4\times N$.
Close off $F$ to get $\overline F\subset {\mathbb R}^4\times N^3$ whose normal bundle $\nu$ is trivialized over $F$ but
$w_2({\nu})=1$. The proof of the lemma concludes with the following proposition.
\begin{proposition} \label{Whitney class} \sl
Suppose $\overline F$ is a surface embedded in $N^3\times {\mathbb R}^4$, where $N$ is a (perhaps non-orientable)
$3$-manifold. Denoting by ${\nu}$ the normal bundle over $\overline F$, the following relation holds between the
Stiefel-Whitney classes of $\nu$: if $w_1({\nu})=0$ then $w_2({\nu})=0$.
\end{proposition}
{\em Proof.} Consider the composition $g$ of the embedding $e$ with the projection onto $N$,
$g\co {\Sigma}\hookrightarrow N\times {\mathbb R}^4 \longrightarrow N$, and homotope $e$ so that $g$ is an immersion.
Denoting by ${\nu}'$ the normal bundle of the immersion $g$, note that since $\overline F$ is codimension one in $N$,
$w_1({\nu}')=0$ implies $w_2({\nu}')=0$. By naturality the same conclusion holds for $\nu$ as well. \qed
\subsection{Non-homological arbiters on $D^6$.} \label{non-homological in six dimensions}
In this section we present a modification of the argument in lemma \ref{Hopf lemma} which works
in dimension $6$. That is, we prove the existence of submanifolds $(M,S^2)\subset (D^6,S^5)$ such that
(1) \parbox[t]{14.25cm}{for any coefficient ring $R$ the attaching $2$-sphere
of $M$ is non-trivial in $H_2(M; R)$, and}
(2) \parbox[t]{14.25cm}{The components of the Hopf link $S^2\cup S^2\subset S^5$ do not bound
disjoint embeddings of two copies of $M$ into $D^6$.}
The construction of $M$ relies on the map $f\co D^4\longrightarrow D^6$ introduced at the beginning of
section \ref{D6}. Recall that the singular set $S(f)$ of that map is a $2$-sphere in $D^4$, representing
the same class as $[{\mathbb C}P^1]$ (and the same as the class of the attaching $2$-sphere $[X]$)
in the second homology of ${\mathbb C}P^2=S^2\cup_h D^4$. We will show that
a suitable $M$ may be constructed as a regular neighborhood of a map $g$: punctured $S^2\times S^2\longrightarrow D^6$.
Consider a circle linking the $2$-sphere $S(f)$ in $D^4$ and perform a $1$-surgery on $D^4$ along this circle, the result
$D^4\smallsetminus (S^1\times D^3)\cup (D^2\times S^2)$ is a punctured $S^2\times S^2$. One checks that
the $5$-dimensional $2$-handle $D^2\times D^3$ (the trace of the surgery) may be embedded in the complement of
$f(D^4)$. Denoting by $g$ the resulting map: punctured $S^2\times S^2\longrightarrow D^6$, note that the singular set of
$g$ is the same $2$-sphere, $S(g)=S(f)$.
Let $M$ be a regular neighborhood of the image of $g$, and denote by $a, b$ the generators of $H_2(S^2\times S^2)$,
corresponding respectively to the core and the co-core of the $2$-handle.
The following relation holds in $H_2(M)$: $[X]=[S(g)]+b$, where $X$ is the attaching sphere of $M$. It follows
that $[X]\neq 0\in H_2(M)$. As in the beginning of section \ref{D6}, it may be helpful to follow this
construction in the half-dimensional real context, where the analogue of $g$ maps a punctured torus into $D^3$.
The final observation is that the proof of (2) in lemma \ref{Hopf lemma} still goes through when the $5$-cells $D^5_i$ are replaced by
an arbitrary orientable punctured $5$-manifold attached to the $3$-sphere by the suspension of the Hopf map. The same argument in
dimension $6$ establishes for $M$ the property (2) above. A non-homological arbiter ${\mathcal A}$ is then defined by
setting ${\mathcal A}(N)=1$ for any $N\supset M$, ${\mathcal A}(N)=0$ for $N\subset D^6\smallsetminus M$, and extending
it to an arbiter on all submanifolds of $D^6$.
|
\section{Introduction}\label{Introduction}
Dwarf galaxies are the most common galaxies in the Universe (e.g. \citet{1997AJ....113..185M}) and the Local Group sample is an excellent laboratory where to study these systems and their properties (\citet{1998ARA&A..36..435M, 1999IAUS..192...17G}). Our interest in these structures is related to many important mutually related topics: the growth of structure, the nature of the dark matter and dynamical interactions. In the last decades, new observational data have become available for many dwarf galaxies (in particular in our Local Group) especially regarding the velocity dispersion profiles of stars and the rotation curves of gas, providing clean measures of the dynamical mass at all radii where baryonic matter can be found. This is gradually permitting a deeper investigation on the equilibrium and stability of the dark matter profiles.
Such studies revealed a discrepancy between the prediction of the $\Lambda CDM$
models by numerical simulations, which require power law density profiles with cusps $\rho \propto r^{ - \gamma } $
where $\gamma = 1$
or $1.5$
(e.g. \citet{1999ApJ...524L..19M, 1997ApJ...490..493N, 2004MNRAS.349.1039N}), and the more flattened profiles derived from observations (e.g. \citet{1995ApJ...447L..25B, 2002A&A...385..816D, 2005ApJ...634L.145G, 2005ApJ...634..227D}).
Other problems strictly related to the nature of the dark matter and its clustering properties are, e.g., the missing satellites problem (e.g. \citet{1999ApJ...524L..19M, 1999ApJ...516..530K, 1999MNRAS.310.1147M, 1999ApJ...522...82K}), the triaxiality of the halos for clusters of galaxies inferred from gravitational lensing (e.g. \citet{1998ApJ...498L.107T}) and the angular momentum problem (e.g. \citet{2000ApJ...538..477N}).
Despite the many successes of the $\Lambda CDM$
paradigm, a large number of open questions remain. Among the large body of literature on the nature of dark matter and tests of different formulations of gravity, the work of, e.g., \citet{2001ApJ...556...93B, 2000PhRvL..84.3760S} and \citet{2004PhRvD..70h3509B} provide classical results.
Our goal in this paper is to extend these studies by examining the interplay between dynamics and star formation in the evolution of dwarf galaxies. Although including star formation makes the modeling more complex, it also allows for comparisons with dwarf galaxies in the Local Group.
Chemodynamical modeling of this type has been productive in the past, particularly in investigations of the effects of winds from high-mass stars and supernova explosions in the ISM (e.g. \citet{2004A&A...422...55P, 2005MNRAS.356..737S, 2005A&A...436..585D}). Given the small dynamical mass inferred for dwarf galaxies, it is surprising that the ISM can remain bound long enough to permit star formation episodes that last for a few Gyr (as found by, e.g., \citet{2005MNRAS.359..985B, 1986AJ.....92...23C, 2002MNRAS.332...91D, 1998ARA&A..36..435M,2003AJ....125.1926G}) giving a typical baryonic matter\footnote{Hereafter we identify baryonic matter with the stellar component; the dark matter will generally be called not-baryonic matter even if its nature could be baryonic as in the case of baryonic-dark-matter objects such as MACHOs or brown dwarfs.} binding energy of $ \sim 10^{53} \text{erg}$ with a total expected number of SN events of $10^3 \div 10^4 $.
We will focus our attention in this paper on the dwarf spheroidal (dSph) galaxies. These systems are the least massive galaxies known but their velocity dispersions suggest a mass-to-light ratio of up to $100M_ \odot /L_ \odot $, making them some of the most dark matter dominated objects in the universe (e.g., \citet{2007ApJ...663..948G}). For these galaxies, the star formation history and chemical composition are very environment dependent, making the realization of a consistent theory of their origin and evolution that combines dynamics, star formation and chemical enrichment more difficult. High-resolution spectroscopy of several dSphs shows a large spread in metallicity (e.g. \citet{2001ApJ...548..592S}) corroborated by photometric studies
(e.g. \citet{2001AJ....122.3092H,2001AJ....122.2524A, 2002AJ....124.3222B}). Several authors suggested that the ISM of dSph systems could be entirely removed by SNe explosions (\citet{1986ApJ...303...39D, 2004A&A...426...25H, 2002ApJ...571...40M, 2004ApJ...613L..97M, 1997ApJ...478L..21M, 1999MNRAS.309..161M, 2004MNRAS.351.1338L, 1999ApJ...513..142M}).
\citet{2006Natur.442..539M} discuss how star formation processes can successfully act as flattener for the central density cusps, within very short timescales (more details will be given in the following). \citet{2005MNRAS.356..107R} obtain the same result as effect of an external impulsive mass loss event. \citet{2002MNRAS.333..299G} investigate the influence of winds. \citet{1996MNRAS.283L..72N} were one of the first to point out the possibility that feedback mechanisms can turn the central dark-matter cusp into a cored one.
We start our analysis by demonstrating the gravitational stability of the isolated model. Then, once the stellar processes are included, we combine the dynamical and chemical analysis. In Section \ref{ModelApproach} we outline our research approach. In Section \ref{Settingup} we set up the initial model, in Section \ref{Settingup} we present the evolution of the models with and without stellar processes, in Section \ref{chemistry} we add the chemistry analysis and in Section \ref{conclusion}, we discuss our results.
\section{The modeling approach} \label{ModelApproach}
In this paper we are trying to address the problem of the dark matter profiles in the dwarf galaxies introduced in the previous Section. Despite the complicated processes acting on the true physical systems, different tools have been developed and presented in the literature to follow the chemodynamical evolution of an isolated spherical multi-component self-gravitating model of a dwarf galaxy system. Briefly, we will base our arguments on a chemodynamical code. Our code can follow the time evolution of dynamically interacting smoothed-particles that represent stars, gas and dark matter components of a dwarf galaxy. The code takes stellar evolution and chemical enrichment of the stellar populations and of the gas into account. A full description of the code used can be found in \citet{1999A&A...348..371B, SpurzemBMKLBMKB09}, to which we refer the reader for details on codification and the implementation.
We will use observational constraints in order to tune our models and simulations. We will proceed as follows:
\begin{enumerate}
\item We first realize multi-component self-gravitating models of a dwarf galaxy composed of a cuspy dark matter density profile, a stellar density distribution and a gas component.
\item We evolve these models in isolation to check their gravitational stability.
\item We activate the star formation (SF) processes to see the changes in the dark matter profiles (if any) and vary the physical parameters.
\item We compare the results to understand the interplay of SF processes and dynamics.
\item We turn off the stellar evolution processes to check again the dynamical equilibrium of the resulting system.
\end{enumerate}
\textit{Our goal is to explore how star formation influences the properties of the dark matter density profiles by changing the baryonic density profiles}.
The same approach will be tested on an initially cored dark matted density profile (see later sections.)
Our focus on \textit{isolated systems} is the \textit{first hypothesis} that makes our modeling not immediately comparable with observations since completely isolated galaxies probably do not exist in nature.
Our first hypothesis leads to the following corollaries:
\begin{enumerate}
\item We can assume a `$\gamma$-profile' (see below) that represents quite well both cored and cusped dark matter profiles in the inner part of the dwarf galaxies. The inner slope of a $\gamma$-profile has the advantage of depending on just one single parameter, i.e. $\gamma$. For the gas and stellar density profiles adopted in the starting models we can again assume gamma profiles as good representation of the projected luminosity profile of the dwarf galaxies in the de-Vaucouleurs form (e.g. \citet{1993MNRAS.265..250D,1994AJ....107..634T})
\item The outer part of the true dwarf galaxy is not expected to drop off as the $\gamma $-models. However, we are considering a synthetic spherical isolated model, so that for Newton's first theorem at any given radius $\hat r$ the dynamics within $\hat r$ are not affected by the mass distribution exterior to $\hat r$. As a consequence we expect \textit{a priori} no differences in the evolution of isolated models regardless of the slope by which our dark matter profile falls off, in particular in the outskirts of the galaxy whenever spherical symmetry is assumed. This permits us to work consistently with the $\gamma $-models.
\end{enumerate}
\textit{We point out that these two corollaries are not true if the first hypothesis is not assumed}. Nonetheless in the literature often different dark matter profiles are used, especially from the family of the King profiles and the Plummer spheres. But, \textit{even if the gravitational force at a given radius is not dependent on the matter distribution outside of that radius, this is not true for the tidal forces}. Therefore, we expect that in real systems, the tidal influence of external galaxies is relevant for the determination of both the shape of the DM density profile \textit{as well as its inner slope }(except in the unlikely case of a dwarf galaxy evolved at rest exactly at the center of a spherical cluster of galaxies).
The tidal effects are expected to act both by truncating the dark matter profile by tidal compression and by stretching the outer part of the dwarfs when evolved as satellite systems of a primary.
Thus, changes in the dark matter profiles in the directions of the leading and trailing tails are also expected, or shallower profiles in the outer part of the dwarfs have to be used if we want to introduce our isolated system in the cosmological framework, e.g., \citet{1997ApJ...490..493N} (hereafter NFW).
However, that is not our objective here (but see the comment in section \ref{evolution}).
\section{Isolated galaxy}\label{Settingup}
\subsection{The model}
Our goal is to provide a self-gravitating three-component model of a dwarf galaxy. The evolution proceeds initially without any astrophysical process other than gravity. This is done in order to check the consistency of the model and to permit us to isolate the different effects in the evolution of a true astrophysical system due to gravitation or stellar evolution.
In order to choose realistic parameters for our model of a dwarf galaxy, we look at a system that in the literature has been subject of extensive investigation, i.e. the Carina dwarf galaxy.
Our target is \textit{not }to realize a specific model of the Carina dwarf galaxy, but first to use observationally motivated input physics and then extend our analysis to the dwarf galaxies class of systems (but see \citet{Pasetto2009b}). The wealth of literature data available on this galaxy, starts already in 1983 with studies of the stellar populations of Carina (\citet{1983ApJ...273..530M}). Color magnitude diagrams were used to derive the star formation history \citep{ 1998AJ....115.1840H,1997AJ....114.1458M,1992PASAu..10...83M,1990A&AS...82..207M,1990A&AS...82....1M,2006MmSAI..77..270M,2006cams.book..272M,2004ASPC..310..133M,2004MSAIS...5...65M,2004MmSAI..75..114M,2003AJ....126..218M,2003MmSAI..74..909M,2004MmSAI..75..110R,2003ApJ...589L..85R}). Spectroscopic measurements yielded chemical abundances, e.g., \citet{2006AJ....131..895K,2008AJ....135.1580K}, the M/L ratio, e.g., \citet{2007ApJ...663..948G} and the dark matter distribution, e.g., \citet{2003ApJ...584..541H, 2007ApJ...663..948G}. Proper motion determinations were obtained by \citet{2003AJ....126.2346P} (see also \citet{2004AJ....128..951P}). Dynamical considerations are available for the orbit from timing arguments in \citet{2008ApJ...679..346M} and for Carina's tidal tails in \citet{2006ApJ...649..201M} which make this system well-suited for further orbit investigations (\citet{Pasetto2009b})
We are adopting the following density distribution profile:
\begin{equation}\label{DehnenProf}
\rho \left( r \right) = \frac{{3 - \gamma }}{{4\pi }}\frac{{Mr_s }}{{r^\gamma (r + r_s )^{4 - \gamma } }}.
\end{equation}
From the analysis of the CMD of Carina (e.g. \citet{1998AJ....115.1840H}) we infer that the oldest stellar population contributes no more than 15\% of the stellar mass. Moreover from the work of \citet{1993AJ....105..510M} we get a mass to light ratio in the V band of $\left( {\frac{M}{L}} \right)_V \sim 23\left( {\frac{{M_ \odot }}{{L_ \odot }}} \right)_V $. Thus from Table 3 of \citet{2003ApJ...584..541H} we get $L = 0.43 \cdot 10^6 L_ \odot $, which leads to a stellar mass estimate of $M \approx 23 \cdot 0.43 \cdot 10^6 M_ \odot = 9.89 \cdot 10^6 M_ \odot $. If we now assume that the mass lost by tidal interactions could reach an order of 95\%, following e.g., \citet{2003ApJ...584..541H} we get $M_{\text{Car}} = 2 \cdot 10^8 M_ \odot$. From this initial mass only $1/50$ is assumed to be observable. This baryonic material consists of 15\% stars, 85\% gas. We thus adopt a starting value of $M_{\text{gas}} = \frac{{M_{\text{dark}} }}{{50}}\frac{{85}}{{100}} = 3.3 \cdot 10^6 M_ \odot $
and $M_{\text{star}} = \frac{{M_{\text{dark}} }}{{50}}\frac{{15}}{{100}} = 5.9 \cdot 10^5 M_ \odot $ for our reference model. These numbers can be considered as starting values for an orbiting dwarf galaxy \citep{Pasetto2009b}, but here we will work only on an isolated system, thus for us the gas amount retained by the dwarf galaxy after its oldest stellar population is formed will be considered as a \textit{free parameter}. This approach is assumed because the present-day dSphs contain little or no gas, possibly because a large fraction of gas was expelled by the initial SNII component, by tidal interaction with a host galaxy or in general photoevaporation, rampressure, by star formation processes. Depending on its own individual evolution and on its orbital parameters, each dwarf ultimately evolves under different conditions depending on its orbital peri-center passages and a general study cannot be included in the present work.
The initial temperature for the gas can be inferred assuming a spherical collapse model for the initial dark matter model and a matter power spectrum $P(k)$ at redshift zero compatible with studies from the SDSS, e.g., \citet{2004ApJ...606..702T} or work on the analysis of the Lyman-$\alpha$ forest, e.g., \citet{2002MNRAS.334..107G}. In this case the typical RMS internal velocity of a halo within $M \le 10^9 M_ \odot $ is less than $20{\rm{km}} \cdot {\rm{s}}^{ - 1} $. This implies that e.g., hydrogen with a sound speed higher than $20{\rm{km}} \cdot {\rm{s}}^{ - 1} $ has a velocity $c_s $ greater than the escape velocity of the potential well in which it has to collapse. This gas is hence not in the condition to be retained by the dwarf. Therefore we will assume an initial sound speed in hydrogen of roughly around $c_s \cong 10{\rm{km}}\cdot{\rm{s}}^{ - 1} $ corresponding to a temperature of $T \cong 10000^\circ \text{K}$.
The evolution of the reference model presented here will be used as normalization case for the realization of a Carina-like dwarf galaxy model that is discussed in more detail in Paper II (\cite{Pasetto2009b}). Carina is one of the few dSph galaxies in the Local Group that shows long-lasting star formation and pronounced intermediate-age populations. This mean Carina resembles to much more isolated, low-mass dwarf irregulars and is one of the reasons why we chose it as a reference for the approximate masses of the different galactic components. We emphasize that it is not the intent of this paper to reproduce the star formation history of the actual orbiting Carina dSph. That is the subject of a separate paper. In fact, in our simulations we explore a larger range of stellar, gas and dark matter masses (Table \ref{Table01}) in order to bracket the classical dwarf spheroidal galaxies given in \cite{1998ARA&A..36..435M,1997RvMA...10...29G}.
Another parameter on which Eqn. \eqref{DehnenProf} depends, is the scale radius $r_c $. It is interesting to point out that in the analysis of \citet{2003ApJ...584..541H} of the possible extension of the dark matter halo, they deduce that the tidal radius $r_t $
of the assumed NFW shaped halos can greatly exceed the surface brightness cut-off $R_{co} $, by an order of $r_t \ge 20R_{co} $ and that the rate of mass loss can vary from 5\% to 95\% depending on the specific orbit chosen. We proceed by tuning the outer profiles for our dwarf model as follows. We can express the radius containing 90\% of the mass, $r_{90} $, as function of the scale radius for the profiles of Eqn. \eqref{DehnenProf} with $\gamma = 3/2$ by solving the simple Eqn. $4\pi \int_0^{r_{90} } {\rho \left( r \right)r^2 dr} = {\raise0.7ex\hbox{$9$} \!\mathord{\left/
{\vphantom {9 {10}}}\right.\kern-\nulldelimiterspace}
\!\lower0.7ex\hbox{${10}$}}M$, which yelds:
\[
r_c = \frac{3}{{19}}\left( {27 + 10 \cdot 30^{1/3} + 3 \cdot 30^{2/3} } \right)r_{90}
\]
For our model we adopt $r_t \cong r_{90} $ \citep{1995MNRAS.277.1354I} and we assume an initial tidal radius 20 times as large as the present one \citep{2003ApJ...584..541H}.
With these values the stellar and gas profiles have $r_{90} \cong \left( {0.5 \times 20} \right) \text{ kpc} = 10\text{ kpc}$, leading to a core radius of $r_{c,T_0 } \cong 0.72\text{ kpc}$ at the starting time $T_0 $ of our simulations\footnote{The definition of $T_0$ is provided in the following section.}.
Since we are not interested in the external radial dependence of the profile, as previously remarked, we could produce the same outer radial dependence for the dark matter profile by modifying the scale length in order to obtain the same $r_{90} $, radius containing the 90\% of the mass as used for the baryonic profile. However the approximation adopted for massive particles with Plummer spheres imposes a natural limit on the inner resolution radius we are going to analyze. The assumed smoothing length for our model (see below) is roughly $\varepsilon _{DM} \cong 0.120\text{pc}$
depending on the specific model. We will not attempt an analysis within the limiting radius $r_l = 2\varepsilon _{DM} $ in the following. Nonetheless we need to be able to discriminate between a dark matter profile with $\gamma = 2$ or $\gamma = 0$ for radii $r > r_l $. $\gamma$-models with an assumed total mass of $M_{DM} \cong 2 \cdot 10^9 M_ \odot $ are almost indiscernible for $r \in \left[ {r_l ,r_{90} } \right]$ for $r_{90} \cong 20\text{ kpc}$ at our particle resolution.
\section{Evolution}\label{evolution}
We will focus our attention here only on the properties of interest for dwarf galaxy systems, for which the following results hold. \textit{The role of the triaxiality and the extension to the more general class of elliptical galaxies is not investigated in the present paper}. A table with the range of masses considered here is presented in Table \ref{Table01}. Within this range we mostly sampled models with an $\gamma_{T_0} \equiv \left[M_{\text{dark}}/M_{\text{star}}\right]_{T_0} \in \left[16,688\right]$. This range of ratios partially excludes very massive `dwarfs', e.g., Large-Magellanic-Cloud-sized proto-systems (unrealistic for the scale radius adopted here but see \cite{Pasetto2009c}) and we completely excluded system with initial $M_{\text{star}}+M_{\text{gas}} \geq M_{\text{dark}}$.
The age of the Universe is 13.7 Gyr, or if expressed as lookback time, $T_{ini}=-13.7$ Gyr. We start our simulations much later, namely at a time when the dwarf galaxies we want to analyze have already formed their old populations.
Our final goal is to study the influence of star formation on the dark matter density profiles. In oder to isolate more clearly this effect we need to neglect the Hubble flow, and all the heating/cooling processes that influence the star formation at hight redshift. Hence, by starting at lookback time of $T_0=-9$Gyr we can follow the phase-space description of our models in a simpler way and implement the star formation processes and their feedback on the interstellar medium with higher numerical resolution.
Nevertheless, this approach forces us to introduce a preexisting stellar population in order to take into account the gas evolution from $T_{\text{ini}}$ to $T_0$. In this initial time the $\Lambda CDM$ paradigm predicts that the dwarf galaxies are the first structures to form as scale invariant perturbations having masses greater than the Jeans mass of isothermal perturbations at the time of recombination (masses of $ \cong 10^{5 \div 7} M_ \odot $). Hence, in this scenario we need to consider the time to violently relax the dark halo component within which the proto-cloud will congregate and produce the first stars.
It is clear that these complicated phases, treated in cosmological simulations, make it more difficult to clearly evidence and disentangle the single stellar evolution influence on the whole gravitational potential of a single system because two fundamental conditions we want to use are missing: \textit{equilibrium} and \textit{isolation}.
After this initial phase and after an equilibrium configuration has been reached, say at $T_0=-9$Gyr, we expect that an old stellar population has already formed.
\begin{table}%
\centerline{\begin{tabular}{lccl}
\hline
\hline
Total Mass $[M_{\odot}] $ & from & to & $\gamma$\\
\hline
$M_{\text{star}}$ & $5.1{\bf{ \times }}10^4 $ & $1.0{\bf{ \times }}10^8$ & $\gamma=1.5$ \\
$M_{\text{dark}}$ & $7.1{\bf{ \times }}10^7$ & $1.9{\bf{ \times }}10^9$ & $\gamma \in [0.5,2]$ \\
$M_{\text{gas}} $ & $5.5{\bf{ \times }}10^4$ & $6.9{\bf{ \times }}10^6$ & $\gamma=1.5$\\
\hline
\end{tabular}}
\caption{Range of masses within which the simulations are run. The scale lengths refer to $r_{90}$ as discussed in the text. The values listed here are our starting values at the initial time $T=T_0$}
\label{Table01}
\end{table}
\subsection{The evolution without the star formation processes.}\label{noSF}
Once the model has been realized we evolve the system in isolation in order to test the equilibrium of the system's density profiles. The critical parameter is of course the smoothing length of the particles, although in the treecodes the gravitational force estimation is largely performed through cell-cell interactions. Different studies in the literature can be found on the best-fit tuning of the smoothing length parameter (e.g. \citet{2001MNRAS.324..273D, 2000MNRAS.314..475A, 2006ApJ...639..617Z} and references therein). We mention here that different codes require to be tuned on different softening lengths to grant the stability of the starting model depending on the implementation of the code. The initial conditions evolved on a special purposed hardware (GRAPE) reach stability with $\varepsilon \cong 0.12\text{ kpc}$.
\begin{figure*}
\centering
\includegraphics[width=17cm]{DensityTimeAssieme.eps}
\caption{(left panel) Density profile temporal evolution for the model evolved in isolation under the action of gravity alone. The blue lines represent the dark matter profile at the beginning (solid) of the simulation and at the end (dashed). The red lines show the stellar profile. No apparent evolution of the dark density profile is seen confirming the equilibrium of the model simulated. The baryonic component remains exactly the same as the starting profile, hence the red dashed line is not visible here since it coincides with the solid red line. (right panel) Density temporal evolution for the same model of the right panel evolved in isolation \textit{and }considering the most relevant star formation processes. The blue solid line represents the dark matter profile at the beginning of the simulation $T=T_0$. The final profile is the blue dashed line. The same line types are used for the stellar profile (red lines). Intermediate thinner lines refer to intermediate evolutionary stages: $T \cong T_0+2.9$ Gyr and $T \cong T_0+6.1$ Gyr The graph scales are chosen to be directly comparable with similar results obtained in \cite{2006Natur.442..539M}.}
\label{DensityTimeAssieme}
\end{figure*}
As can be seen in Figure \ref{DensityTimeAssieme} (left panel), stability is achieved within the resolution used here. The figure represents a simulation with 35000 particles $\left( {n_{\text{gas}} ,n_{\text{star}} ,n_{\text{dark}} } \right) = \left( {1.0,0.5,2.0} \right) \times 10^4 $. The same stability is obtained with slightly different values for the smoothing length $\varepsilon $
of the dark matter, stars, and gas with a larger number of particles $\left( {n_{\text{gas}} ,n_{\text{star}} ,n_{\text{dark}} } \right) = \left( {2.0,1.0,4.0} \right) \times 10^4 $
and $\left( {n_{\text{gas}} ,n_{\text{star}} ,n_{\text{dark}} } \right) = \left( {1.0,0.5,2.0} \right) \times 10^5 $
for a total amount of 350000 particles \textit{in the beginning} of the highest resolution simulation. Even if the higher resolution would permit us to analyze regions closest to the core of the dwarf galaxy, we will see that our chosen distance of $0.5\text{ kpc}$
from the core to which we limit our consideration due to the smoothing length resolution of the dark matter particles, is enough to confirm our results. In Figure \ref{DensityTimeAssieme} (left panel), the overlapping lines represent the evolution of the system at $T = T_0$, i.e. 9 Gyr ago, and after $\Delta T = 9\text{ Gyr}$, i.e. now ($T=T_0+9=-9+9=0$ Gyr), respectively. The density profile of the model hardly changes with time as is desired, both for dark matter profile as for the stellar component, thus confirming that the initial conditions of our model are reasonable.
\subsection{The evolution with the star formation processes.}
Once the equilibrium of the density profile has been confirmed under the effect of gravity alone, we activate the star formation processes.
The understanding of the star formation processes was a primary task of the Astronomy during all the second half of the last century (e.g., \citet{1969MNRAS.145..405L, 1987IAUS..115..663S}) but, despite the great improvement of knowledge acquired, the micro-physics description still represents a challenge of the modern astrophysics. Great improvement has been achieved within the framework of the Smooth Particles Hydrodynamics star formation algorithms (\citet{1992ApJ...391..502K, 1993MNRAS.265..271N}) that we adopt in our description in its more updated recipe as explained in \citet{1999A&A...348..371B}.
We emphasized again that we start our simulation \textit{after }the old stellar populations have already formed. For instance, if it took the dwarf galaxy $\approx 3$ or 4 Gyr to accrete gas, to have it collapse and to experience its first measurable episode of star formation, then we assume this equilibrium stage as our initial time of evolution starting from which we analyze the system ($T=T_0$), 3 to 4 Gyrs after the Big Bang.
The assumed dynamical equilibrium of the model is as previously described in Section \ref{noSF} for the old stellar component, the gas and the dark matter, under the effect of gravity alone. The system can now evolve from an initial time $T_0 $ until the present time. Since we are interested in systems with cuspy dark matter density profiles, we only present their evolution here. Further considerations on flat dark matter profile systems can be found in Section \ref{conclusion}.
Depending on the initial amount of gas and on the total mass of the system, the general behavior of the system is to expel part of the gas due to the effect of SNII that gradually eject the gas to the outer part of the galaxy. Some of the gas is nevertheless conserved in the inner part of the system thanks to the steep potential well. The dwarf galaxy starts as an extremely dark matter dominated system especially in the inner part of the galaxy where the density profile is very cuspy ($M/L>200$ within 5 kpc from the center). In this sense, the steepness of the dark matter profile seems to be a necessity in the initial phase of the dwarf galaxy where just a few percent of the stellar population are formed after the primordial collapse. There are at least three reasons for this:
\begin{enumerate}
\item If the initial dark matter density profile is not cuspy but flat, the possibility for the dwarf galaxy to survive several pericenter passages as expected for a dwarf like Sagittarius (e.g. \citet{2001MNRAS.323..529H}) becomes more problematic.
\item If the initial dark matter profile is not cuspy but flat, the efficiency of the SNII to expel the gas is expected to be higher (e.g. \citet{1998A&A...337..338B, 2005MNRAS.356..107R}) thus increasing the amount of primordial gas necessary to reproduce tidally triggered bursts of star formation in galaxies as may be the case in, e.g., Carina (see Paper II). However, this result is strictly dependent on the initial environment in which the gas collapses and we start our simulations only after a time $T_0$ where all these effects have already taken place.
\item Cosmological N-body simulations predict cuspy profiles (e.g. \citet{1997ApJ...490..493N}) with an inner slope proportional to $r \propto r^{ - 1} $. The asymptotic behavior for the $\gamma $ models in the inner zones predicts, in the acceptable range studied here, $r \propto r^{ - \gamma } $, including, by the way, the NFW inner slope.
\end{enumerate}
It is difficult to constrain this residual fraction of gas. It seems to be a free parameter that can play a role in the periodic bursts of star formation recently suggested by \citet{2004A&A...422...55P} or may play a role in the dwarf evolving in an orbit where the SFR can be activated by strong tidal interactions.
As we can see from Figure \ref{DensityTimeAssieme} for the reference model of Section \ref{Settingup}, we experience a stellar density profile evolution with the same time scale, $\tau _s $, as the dark matter density profile time scale $\tau _d \cong \tau _s \cong 9 \text{ Gyr} $. The system evolves slowly over timescales of a few Gyr modifying its stellar profile thanks to stellar evolution: new stars are born and die interacting with the small fraction of gas to reach each time new configurations of dynamical equilibrium\footnote{This is obtained whenever the cell opening parameter of the treecode keeps a conservative value, say $\theta \approx 0.6^\circ $ in order to grant a better energy conservation, or when a more extensive use of the GRAPE hardware is adopted (e.g., \citet{2006sphe.workE..40B,2007sphe.work....5B})}. We point out here how not all the literature recipes for the SF are able to follow efficiently the gravitational changes of the baryonic component. The code really has to `make' the stars such that the number of particles increases regardless of the substantially increasing computational time required. Only these prescriptions, see e.g. Figure \ref{Npart} and \cite{1999A&A...348..371B,2000A&AT...18..829B,2001KFNT...17..213B}, together with the correct gravitational treatment (here provided by a devoted GRAPE hardware) permit one to follow correctly this slow change of the gravitational potential. In Figure \ref{Npart} we plot the range of particles used to test our results and their time evolution. The thick lines refer to the reference models and the thinner ones to the test models as defined in Section \ref{noSF}. The test model were developed in order to confirm the particle number independence of our results. The slight variation on the number of dark matter particles is due to a small evaporation in the outer part of the dwarf that occurs at distances out of 10 kpc from the center of the system. The number of particles for the stellar and gas components changes depending on the star formation processes (Section \ref{chemistrySF01}) as well as on the evaporation in the outer part of all the components (gas, stars and also dark matter). The similarity of the evolution any trend is remarkably good from the reference model up to the test model with the highest particle number.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{Npart.eps}}
\caption{Variation of the number of particles for the reference models and higher resolution test models. Black solid lines refer to the total number of particles, green dashed lines refer to the gas, red dotted lines refer the stars and blue dot-dashed lines refer to the dark matter. The reference model is marked with a thicker line.}
\label{Npart}
\end{figure}
Once new stars are added to the stellar component, they reshape the mass distribution (the density profile) and as a consequence the dark matter reacts with the same scale time $\tau _s \cong \tau _d $ passing thought several states of dynamical equilibrium.
This can be proven by simply stopping all the star formation processes and evolving the new equilibrium configuration under the influence of gravity alone. The phase-space configuration seems to be stable with respect to the simple dynamical evolution\footnote{This might seem a numerical test on the stability of the equilibrium configurations but it is \textit{not} because it is model dependent. The treecode approach prefers the cell-cell force computation for distant interactions by reducing the role played by the softening length (which is, on the other side, dependent on the growing number of particles used in the systems). The stability of the equilibrium configurations will not be subject of further investigation in this paper.}.
We observed that the inner flattening of the dark matter halo is accompanied by a simultaneous decrease of the density profile of the baryonic component (Fig \ref{DensityTimeAssieme}). That is simply an indication that the final density profile has to depend \textit{at least} on two parameters (other than the total mass or the central density). As an example, the family of $\gamma $ models or King models permit this kind of metamorphosis: we can obtain this result by simply fixing the radius containing 90\% of the mass $r_{90} = r_s \left[ {\left( {\frac{{10}}{9}} \right)^{\frac{1}{{3 - \gamma }}} - 1} \right]^{ - 1} $ and expressing in this way the scale radius as a function of $r_{90} $ and $\gamma $. At fixed $r_{90} $ and total mass $M$ in Eqn. \eqref{DehnenProf}, one can easily obtain central densities that differ by a few orders of magnitude.
More intuitive is the explanation of the evolution of the stellar density profile because it changes both central density and scale length due to the increasing number of particles and mass, while nonetheless roughly preserving the projected surface brightness.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{Vc.eps}}
\caption{Rotation curve for the dark matter profile (and the stellar component) of the reference model. The solid line represents the starting profile at $T=T_0$ and the dashed line the profile after $T=T_0+9\text{ Gyr}s$. The same colors and line types are used as in the previous figures.}
\label{Vcdark1}
\end{figure}
In Figure \ref{Vcdark1} we plotted the rotation curve $v_c = \sqrt {r\frac{{d\Phi }}{{dt}}} $ of the dark matter component at the starting time $T = T_0 $ and at the end of the simulation $T = T_0 + 9\text{ Gyr}$ (the intermediate-time snapshots have been omitted for the sake of clarity). The general trend expected as a consequence of the flattening of the profile is exactly the decrease of the central part of the rotation curve, as shown in Fig. \ref{Vcdark1}. We can however observe the small overlap of the rotation curve, which in a two parameter density-potential pair is an indication that the matter distribution at $T = T_0 + 9\text{ Gyr}$ is slightly more concentrated. A fit of the dark matter profile shows that the outer homoeoids for the dark matter profile have slightly contracted. The trend is generally present in our simulations even though the effect here is minimal. Nevertheless, this is a further indication of the fact that a good dark matter density profile has to be at least a 2-parameter model.
The cumulative mass density profile $M\left( r \right) = 4\pi \int_0^r {r'^2 \rho \left( {r'} \right)dr'}$ shows this trend as well (Fig. \ref{Masscum}).
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{masscumu3l.eps}}
\caption{(upper panel) The cumulative mass profile for the dark component (blue) and the stellar component (red). The starting profile is at $T=T_0$ (solid lines) and mostly lower than the profile for time $T=T_0+9\text{ Gyr}$ (dashed lines). (lower panel) A zoom on the dark matter component shows how the lines of the initial and final stage of the evolution cross. See text for details.}
\label{Masscum}
\end{figure}
The theoretical expectation for the curves in Fig. \ref{Masscum} indicates that the flatter density profile curve asymptotically converges under the cuspier density curve profile (once we assume the same total mass).
Considering the asymptotic behavior, the overlapping part of the final cumulative mass profile is an indication of a slightly reduced final scale radius: the same amount of matter is reached in inner homoeoids and starting from a central smaller value as shown in Fig. \ref{Masscum}.
Before discussing the kinematics we point out the limits of our simulations. Our finding that a two parameter dark matter density profile is favored, is based on
\begin{itemize}
\item the flattening of the density profile at a constant total amount of mass.
\item the analysis of the rotation curve and cumulative mass profile.
\end{itemize}
But only the first point is stringent evidence for the necessity of the two free parameters model in the dark matter component (in addition to the total mass), while the second point comes with an intrinsic theoretical bias and can \textit{not } be used to claim the necessity to add a further parameter to the NFW profile. The low reactivity of the outer homoeoids to the changes of the inner mass distribution is expected as direct consequence of Newton's second theorem. This theorem describes how the total force acting on the matter distribution out to, e.g., $r_{90}$, is the same as it would be if the shells' matter were concentrated into a point at the center of the dwarf galaxy, despite the different radial distribution it shows in spherical approximation. Thus the finding that the stable dark matter profile has to be at least a two-parameter density profile is just an indication that comes from the baryonic/non-baryonic interaction analysis and does not necessarily rule out the one-parameter density profiles as in \citet{1997ApJ...490..493N}. Moreover as already explained in Section \ref{Settingup}, a model evolved in isolation is expected to show a sharper decline of the external profile as compared to the one resulting from the cosmological collapse of a cube (as we effectively see by comparing a $\gamma$-model with a NFW model).
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{anisotropia.eps}}
\caption{Anisotropy parameter (light green) and radial velocity dispersion (red) for the stellar component. The left scale on the y axis is for the anisotropy parameter, the right scale for the radial velocity dispersion parameter. The thick solid lines show the starting stellar profile, here at $T_0=196$ Myr, and the dotted lines show the resulting profile at $T=T_0+9Gyr$.}
\label{sigmar}
\end{figure}
To investigate these processes further in detail we present in the following also some chemical considerations that will help us to improve the present theory on dwarf galaxy evolution. The idea to include some chemical considerations, emerges from the analysis of the anisotropy parameter, defined as $\beta \equiv 1 - \frac{1}{2}\frac{{\sigma _t^2 }}{{\sigma _r^2 }}$, where $\sigma _r$ is the radial velocity dispersion and $\sigma _t^2 \equiv \sigma _\theta ^2 + \sigma _\varphi ^2 $ is the tangential component in a suitable spherical coordinate reference system $\left( {O,r,\theta ,\varphi } \right)$. It is evident from this definition, and from assuming that, as for every reasonable Galactic potential $\sigma _r^2 > \sigma _t^2 $, that the anisotropy parameter $\beta \in \left[ { - \infty ,1} \right]$, where $\beta = - \infty $, corresponds to a system dominated by pure circular orbits $\sigma _r = 0$, in contrast to a system with $\beta = 1$ with pure radial orbits. As seen in Fig. \ref{sigmar} the system develops, within the timescale $\tau _d $ of the flattening of the dark matter cusp, a zone dominated by circular orbits, roughly out to a radius $r_{lim}=4$ kpc in radial distance. The kinematics in the outer part of the galaxy are left unchanged in this isolated model\footnote{Any further investigation of the link between the scale length of the starting model $r_c$ and this limiting distance is not the subject of this paper. Nevertheless when evolving this isolated system in an orbit around a host system, we expect that the external tidal force can perturb any relation that may exist between this limiting radius $r_{lim} $ and the scale length $r_c$.} and the radial dispersion velocity profile remains basically unchanged beyond $r>r_{lim}$. Similarly the anisotropy parameter approaches unity at larger radii, getting closer to unity after $r>r_{lim}\cong4$kpc. This indicates how circular orbits become more and more relevant when the dark matter cusp becomes larger, offering a plateau of uniform density and potential, where we expect an easier mixing of the stellar component. This will leave some observable traces in the radial gradient of the chemical composition that we now are going to explain.
\section{Chemo-dynamical evolution}\label{chemistry}
\subsection{Star formation recipes}\label{chemistrySF01}
The coding of the star formation processes is not the subject of our analysis. A more extended description can be found in the works by \citet{1999A&A...348..371B,2003Ap&SS.284..865B,2003CoKon.103..155B,2002Ap&SS.281..297B,2001KFNT...17..213B,2000KFNTS...3...91B}. Nevertheless the robustness of our evolution has been tested against the different star formation criteria that are discussed in the literature.
We found that there are several physical conditions that we \textit{have} to take into account in order to obtain the evolution of the dark matter profiles:
\begin{itemize}
\item A fluid element has to be instable, e.g., it has to satisfy the Jeans instability criteria $\tau _{sound} > \tau _{ff}$ between sound speed timescale and free fall timescale.
\item For the gas particles $\tau _{cool} \le \tau _{ff} $, the cooling timescale has to be lower than the free fall timescale (this is also often expressed in terms of the over-density criterion $\rho_i > \rho _{crit}$, i.e., the density of the gas particle $i$, $\rho_i$, has to be higher than a critical density $\rho_{crit}$).
\item The efficiency of the energy that heats the ISM in a dwarf galaxy can be reduced to a few percent $\epsilon=0.03$, say $\epsilon \eta_{SN I,II}$ where $\eta_{SN}$ is the standard energy injection per SN explosion. This is a result that was derived by the chemical modeling approach already presented in the literature, by, e.g., \citet{1998A&A...337..338B}.
\item A radius limit of the sphere of influence of a SNII event can be assumed when the sound speed of the front shock is cooled at the sound speed of the ISM. Simulations by \citet{1991ApJ...367...96C} have shown that a reasonable value is of the order of a few hundred parsec.
\end{itemize}
These are sufficient although not necessary conditions. There are other studies that correctly reproduce the observed dynamical effects, e.g., when using dynamical times, $\tau _{dyn} = \sqrt {3\pi /16G\rho } $ instead of the free fall time in the first criterion (e.g., \citet{1993MNRAS.265..271N}), or when using convergent fluxes with the criterion of the divergence of the velocity $\nabla \cdot {\bf{v}} < 0$ (e.g., \citet{1998MNRAS.297.1021C}), or even applying higher values of SN efficiency (e.g., \citet{2001MNRAS.322..800R}).
The first two criteria are universally accepted and do not require any particular fine tuning in our dwarf galaxy simulations to reproduce our results\footnote{ The only slight modification of the Jeans instability criterion is in the case of the presence of magnetic fields. In this situation also the first criterion can be slightly modified in favor of a critical mass depending on the Alfven and the sound speeds. Despite their relevance in the collapse and fragmentation of cold clouds, magnetic considerations have not been investigated in our simulations.} (see Fig. \ref{SFR}).
The last two criteria are \textit{not }tuning parameters. They are results that were already presented in the literature many years ago. They basically act as SN-`shock absorbers' in the dwarf galaxy environment. Thus the result we have presented for the dark matter evolution is \textit{not} dependent on any physical tuned mechanism acting on the baryonic component, nor does it require any `special' mechanism to explain the change of the baryonic density profile and in turn the resulting gravitational effect it produces (the conversion of a cuspy dark matter profile into a flat DM profile).
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{SFRcor.eps}}
\caption{Evolution of the star formation rate with time for the reference model. The total amount of mass (the sum under the area of the histograms), can change by an order of magnitude depending on the free parameter of the amount of gas that the dwarf galaxy can retain, e.g., in our case the initial amount of gas. Nevertheless it is \textit{always } less than a few percent of solar masses per year. The total absence of spurious initial bursts of star formation is also a confirmation of the good level of dynamical equilibrium obtained in Section \ref{noSF})}
\label{SFR}
\end{figure}
\subsection{Chemical Evolution}
In our simulations we \textit{assume} that the first episode of star formation and its associated SN explosions led to gas ejection to the outer regions of the dwarf galaxy as described in, e.g., \citet{1999ApJ...513..142M, 2004MNRAS.351.1338L, 2006MNRAS.371..643M}. This is not a direct result of our simulation because, as mentioned, we consider the evolution of a dwarf galaxy after its initial star formation episode which gives rise its old stellar population allowing us to better control the gas present at the initial stage $T=T_0$ of our simulations.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{FesuH_stars.eps}}
\caption{The evolution of the stellar $[Fe/H]$ over time. The contour plot shows the metallicity spread distribution with time. At the starting time $T_0$ the dwarf starts with a metal poor distribution associated with an old stellar population formed prior to the beginning of our simulation. The color scale is normalized to the highest number of stars (that is not constant in our simulations!) and to the peak of the metallicity per bin. In this way the plot better demostrates the increase of the metallicity with the time \textit{and }its distribution per temporal bin (due to the gradual filling of the distribution tails by SNIa events).}
\label{FesuHT}
\end{figure}
In Fig. \ref{FesuHT} we consider the metallicity evolution with time, without considering its spatial dependence (see Fig. \ref{GradientMet}), by plotting the trend of iron enrichment. Newly formed stars peak above the mean value of iron, $\left\langle {\left[ {Fe/H} \right]} \right\rangle $ which as consequence increases with time and is strongly dependent on the initial chemical composition assumed for the ISM and for the stellar component. We are actually not interested in reproducing any specific observed trend, but we want to understand the expected rate of enrichment. Starting with a rather metal-poor population ($\left\langle {\left[ {Fe/H} \right]} \right\rangle \cong - 3$ dex), we are able to see the general rate of enrichment independent from the specific initial conditions that differ from galaxy to galaxy in the true Universe. It is evident that a mean range of $\left\langle {\left[ {Fe/H} \right]} \right\rangle \in \left[ { - 1.5, - 0.5} \right]$ can be covered in $9 \text{ Gyr}$ of isolated evolution provided that the system has an initial range of rather low metallicity and then undergoes rapid early enrichment in the initial star formation episodes that need to have taken place during the first, say, 3 Gyr (e.g. \citet{1999ApJ...513..142M, 2004MNRAS.351.1338L}). We note that observationally, such a rapid initial increase does seem to take place, see, e.g., \citet{2007ApJ...657..241K, 2007AJ....133..270K} but we need to remark that a direct comparison with any real dwarf galaxy can not be carried out here. With the trend here evidenced we want to reproduce an 'upper-limit' to the observed chemical enrichment. Thus the results here presented are reference model that can be used to impose an upper limit constraint.
In Fig. \ref{SNs} we plot the number of SN events and PN as a function of time. This helps to explain how the spread in the previous figures originates. In line with the arguments of different authors (e.g. \citet{2006MNRAS.371..643M}) we see how the SNIa, mainly responsible for the iron enrichment of the ISM, operate to fill gradually the tail of the $(\left\langle {\left[ {Fe/H} \right]} \right\rangle-t)$ diagram, even though their number is only a few percent of the SNII. This can indeed be deduced from the Figure \ref{SNs} where the number evolution of SN events is presented as a function of time. Depending on the chemical enrichment we assumed for the oldest stellar population, at the end of the 8 to 9 Gyr of evolution we can obtain a mean present day metallicity of $\left\langle {\left[ {Fe/H} \right]} \right\rangle = - 0.5$ integrated along the radial direction (cfr. \citet{2004MNRAS.351.1338L, 2006MNRAS.371..643M}).
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{SNevents.eps}}
\caption{Cumulative number of SNII events in red, SNIa in magenta and planetary nebulae in gray. The number of SNII is almost constant due to the assumption of continuous star formation after the initial collapse which is assumed to to have occurred prior to time $T_0$, by creating the old stellar population of our dwarf galaxy.}
\label{SNs}
\end{figure}
The initial dominance of SNII is partially suppressed in this model thanks to the ad-hoc condition avoids the initial burst of star formation (see Fig \ref{SNsrate}) and the SNIa rate is decreasing with time as, e.g., predicted in \citet{2001ApJ...558..351M}). The reason is that in the collapse of the proto-cloud the star formation is enhanced by the increasing density ($SFR \propto \rho ^{1 \div 2} $) thus favoring the SN events, while in our simulation this phase is neglected and a starting Population II sample is always initially present at $T=T_0$.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{Rate.eps}}
\caption{Here the SNIa rate is presented. We find agreement with the literature prediction (e.g. \citet{1983A&A...118..217G}) of a decreasing SNIa rate with time (see also \citet{2001ApJ...558..351M})}
\label{SNsrate}
\end{figure}
Prior to the onset of contributions of SNIa, stellar metallicities ($[Fe/H]$) tend to increase while the $[\alpha/Fe]$ ratio remains rougly constant. This leads to a flat, plateau-like feature at low metallicities when plotting $[Fe/H]$ vs. $[\alpha/Fe]$ (\citet{2004MNRAS.351.1338L, 2006MNRAS.371..643M}). Since our simulations only start after the formation of the old stellar population, our Fig. \ref{FesuOmulti} does not show such a plateau. Instead we see the characteristic decline of $[\alpha/Fe]$ with increasing $[Fe/H]$ due to the combined action of Fe production in SNIa and $\alpha$ element production in SNII.
Moreover, differences in the initial gas fractions affect the metallicities reached during the first episodes of star formation. Lower gas fractions imply less enrichment. Fig. \ref{FesuOmulti} shows results for different initial gas fractions of $15\%$, $30\%$ and $45\%$ as an example. These result in differences in the amount of gas that the galaxy is able to retain at a fixed dark matter amount. Of course, the true behaviour of a dwarf galaxy depends on its intrinsic properties but also on the tidal interaction that may activate SF processes.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{chemoFesuOvFesuHs.eps}}
\caption{The evolution of the $\alpha$ elements (represented by oxygen) and iron for different initial gas fractions. We show the evolution of the mean $[Fe/H]$ and $[\alpha/Fe]$ in the gaseous and stellar components. The characteristic flat trend of $[\alpha/Fe]$ at low metallicities is not seen here since our simulations start after the first significant episode of star formation. The total mass of the gas is incremented from dark to lighter red: $M_{\text{gas}}=5.81 \cdot 10^5 M_ \odot $, $M_{\text{gas}}=6.55 \cdot 10^5 M_ \odot $, $M_{\text{gas}}=7.31 \cdot 10^5 M_ \odot $, $M_{\text{gas}}=3.23 \cdot 10^6 M_ \odot $ and $M_{\text{gas}}=6.46 \cdot 10^6 M_ \odot $. The supersolar evolution comes from a test model with $M_{\text{gas}}\cong 12.8 \cdot 10^8 M_ \odot $ inside the usual scale length of $r_{s,gas}=0.51$ kpc (see Section \ref{conclusion})}
\label{FesuOmulti}
\end{figure}
\subsection{Radial chemical gradients}
We now turn to the radial dependence of the chemical composition. The spatial distribution of stellar populations of different ages and metallicities is a matter of debate in different studies (see below). Here we want to show what is expected from the simple evolution of our isolated synthetic model.
Observationally one tends to find gradients such that the younger, and/or more metal-rich populations are more centrally concentrated (e.g. \citet{2001AJ....122.3092H,2006AJ....131..895K}).
The general expectation is that the SNe generate local enrichment at the beginning of the evolution, and that then diffusion tends to erase any radial gradient with time. This affects both iron and oxygen but on different time scales. While oxygen is produced by SNII and so distributed across the entire dwarf galaxy quite fast, iron is expected to be more concentrated in the central regions where the SNIa are more densely located (e.g. \citet{2006MNRAS.371..643M}). As result the gradient is expected to be stronger for iron and longer lasting, while smaller and shorter in duration for oxygen.
In this analysis we suggest a possible link between the timescale of the change of the dark matter density profiles and of the formation of chemical gradients which also involve the evolution of the stellar orbit families. If we consider Fig. \ref{GradientMet}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{Gradient.eps}}
\caption{Radial dependence of the mean $[Fe/H]$ value. $[Fe/H]$ is averaged along spherical shells. In red we have plotted the stellar metallicity radial gradient from $T=T_0$ (thick line) to $T=T_0+9$Gyr (the dashed line) with steps $\Delta T= 1$ Gyr . }
\label{GradientMet}
\end{figure}
together with the evolution of the anisotropy parameter, in Fig. \ref{sigmar}, we see how the progressive growth of the number of stars and the resulting changes of the stellar density profile cause a change in the overall gravitational potential, leading to a new equilibrium configuration for the dark matter, which changes its profile from cuspy to flat. For the stellar component we can study the evolution of the orbits, which helps us in interpreting the gradient and the link with the dark matter.
As can be seen in Fig. \ref{GradientMet}, in the first few Gyr, the gradient is present only in the central zones, where the star formation is more intense due to the higher density. The spatial extent of the gradient encompasses up to 1 kpc after 1 Gyr, up to 2 kpc after 3 Gyr etc. This effect is initially due to the higher fraction of radial orbits (Fig \ref{sigmar}) which mix the inner and outer regions of the system with high efficiency. As time passes the new density profile generated by the newly born stars gives rise to a growing fraction of circular orbits (preferred when a system is slowly evolving, because they are closer to the equilibrium and at lower energy). This can also be inferred from the slow decrease of the radial velocity dispersion which leads to a decrease of the anisotropy parameter. This is happening within the limiting radius $r_{lim}$, within which the chemical gradient and the anisotropy parameter develop for the 9 Gyr of our simulation. Once the fraction of circular orbits is increased (corresponding to the decrease of the anisotropy parameter seen in Fig. \ref{sigmar}), then the mixing efficiency decreases, because there are fewer orbits that extend to the galaxy outskirts. This, together with the gradually decrease of the gas fraction, acts to slow the radial expansion of the zones where the chemical gradient can be evidenced: as seen in Fig. \ref{GradientMet} the lines for T=7 Gyr, T=8 Gyr, T=9 Gyr lie closer and closer to the dotted line ($T=T_0+9$ Gyr) for $r<5$ kpc.
Evidence of such a radial gradient is already present in the literature on N-body simulations, see e.g., \citet{2002MNRAS.335..335C} where the massive spectrum of the elliptical galaxy and the role of the stellar winds are further investigated.
In our simulation we did not include galactic winds, thus excluding them as possible explanation of the chemical mixing and consequently a faster disappearance of the $[Fe/H]$ gradient. We point out that recently the same result has been found with a different mechanism, but still avoiding stellar winds by \citet{2008MNRAS.386.2173M}. We agree with the general result presented in their paper and extend it by showing that the reshaping of the dark matter profile seems to work on the same time scale of the metallicity gradient formation. The overlap of the shells of the SNII is also evidenced in Fig. \ref{shell}. In this figure an especially intense episode in the time evolution around $T_0+T=0.730$ Myr is plotted where two SN events are clearly visible (in red) and the remnants of at least other two other ones are still evident (in orange/yellow). Different spots of higher temperature/sound velocity are shown in the velocity surface cut arbitrary tilted by 45 degrees w.r.t. the vertical axis Z. This is a convenient representation of the velocity (or temperature) map that is partially comparable with Fig 3 of \citet{1991ApJ...367...96C} but here in a full 3D description. We clearly see how two green shells of irregular shape (sound speed contour of $25{\rm{km}} \cdot {\rm{s}}^{ - 1} $) from two different SN remnants are overlapping.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{shells.eps}}
\caption{Shells of isovelocity for the sound speed of the gas component. The sketch shows a cut tilted by 45 degrees w.r.t. the vertical axis Z in order to evidence the overlapping shells of isovelocity (see text for details).}
\label{shell}
\end{figure}
On the one hand, we are able to follow the behavior of the orbits and the evolution of the density profiles. On the other hand we can not reach the resolution of the chemical codes used by \citet{2006MNRAS.371..643M} thus preventing us from confirming the general efficiency of the mechanism invoked by these authors. Nevertheless we sustain their complementary picture because we also find the possibility of overlapping polluted regions of the SNII (see, e.g., Fig \ref{shell}). In this figure the isocontours of the sound speed velocity of the gas component are shown. A limiting radius for the sphere of influence of the SNII event feedback on the ISM is assumed when the sound speed of the front shock is cooled at the sound speed of the ISM, here ${\rm{11km}} \cdot {\rm{s}}^{ - 1} $ (see Section \ref{Settingup} and \citet{1991ApJ...367...96C}). This permits it to have overlapping phenomena of the shock fronts in the dwarf galaxies as proposed by \citet{2008MNRAS.386.2173M}.
\section{Conclusions, discussion and limits of the approach}\label{conclusion}
In this paper, we focused our attention on the problem of the dark matter density profiles by investigating the apparent incongruence between dynamical theoretical predictions of a cuspy dark matter profile and the observations that indicate a flatter profile. With our simple description based on the `classical' approach of smoothed particles hydrodynamics (e.g. \citet{1989ApJS...70..419H}) we have obtained the results that we briefly summarize here.
The classical description of the star formation processes seems to be sufficient to change a cuspy profile into a flatter one. This is shown with the classical smoothed particles approach to the hydrodynamic equations of motion completed with the star formation criteria based on the Jeans theory. The evolution of dark matter profiles could offer a natural explanation for the apparent flat density profiles found in observations (\citet{1995ApJ...447L..25B, 2002A&A...385..816D, 2005ApJ...634L.145G, 2005ApJ...634..227D}) once applied in a realistic context. If we introduce our theory in the context of the cosmological hierarchical merging picture, where the dwarf galaxies are the first stellar population systems that form, the theory developed here predicts that today the star formation processes should have flatted all the initially cuspy dark matter density profiles (for the range in masses explored). Any dwarf galaxy that contains a stellar population older than, say 8 Gyr, is expected to have a flat dark matter profile, because either the tidal gravitational effect has changed the dark matter profile, or because internal star formation processes acted on the dwarf in isolation as presented here. The combination of both effects is in fact expected to accelerate the profile flattening by tidally activating and enhancing the star formation processes.
We point out here that the dark matter density profile change happens as soon as the star formation processes act to modify the gravitational equilibrium of the stellar component and, in turn, the gravitational equilibrium of the dark matter. Once the star formation efficiency is reduced, i.e. the stellar over gas fraction exceeds the smaller value predicted in Table \ref{Table01}, then the effect is minor. An initially very gas deficient dwarf galaxy can not experience this profile change. On the other side, galaxies that reach dynamical equilibrium with a high amount of gas still bound to the system (say $M_{{\rm{gas}}} = 10^7 M_ \odot$ at $T>T_0$) do not have observational counterparts because they produce a dwarf system with super-solar $\left\langle {\left[ {Fe/H} \right]} \right\rangle $ (see, e.g., Fig \ref{FesuOmulti}).
Nevertheless dwarf galaxies with an initial old stellar population of less than $10^4 M_ \odot$ probably belong to a different dynamical regime that is not of interest for us here.
By analyzing the time evolution of the dark matter profiles, we extensively investigated their dynamical properties and evidenced the necessity for at least a two free parameter density model (in addition to the total mass) in order to correctly represent the shape resulting from the baryonic/non baryonic interaction. This is a result partially dependent on the isolation of the model and does necessary rule out NFW profiles, which are obtained only in a baryon-free simulations.
In our simulations, we used an extremely strongly peaked cuspy DM profile in order to study the effect expected in a case of a very deep potential well for the dark matter profile. As consequence these phenomena act even on smaller time scales for a central shallower dark matter density peak, or for small $M/L$ ratios. The same effect, in a different cosmological context, has been described by \citet{2006Natur.442..539M}, although they derived a very short timescale for the effect ($\tau _d < 800$ Myr). The most plausible explanation for their extremely short timescale $\tau _d$ is an `overheating' by SN of the interstellar medium. The role of the SN feedback (especially the SNII sphere of influence and the $\eta_{SMII}$) are more carefully taken in consideration in our present study as compared to earlier work by other authors. This leads us in a natural way to a longer $\tau _d$, which presents attractive advantages. In the context of dwarf galaxies tidally perturbed by the gravitational field of a primary host (i.e., a major galaxy, or a cluster of galaxies) the existence of deeply embedded systems (e.g. Sagittarius in our MW) could become more problematic whenever the orbital pericenter lies close to the center of the hosting system if $\tau _d$ is as short as predicted by \citet{2006Natur.442..539M}. Our models do not suffer this problem because the cuspy profile in our case is lasting several Gyr and is gradually flattened. As a result we are easily able to save the bulk of the orbiting dwarf galaxy around a primary system \citep{Pasetto2009b}. A second difference is that we treat completely isolated models while the \citet{2006Natur.442..539M} have to work with interacting objects. This offers us the advantage to eliminate spurious perturbations and to have a clean determination of the role of star formation alone. The peculiar orbital path of every different dwarf galaxy when evolved in interaction has a determining role on its SFH and on the density profile evolution. Our modeling of isolated system offers easier analysis and theoretical interpretation of the results.
We tried to link the dark matter profile change with an observable property. A suggested indication in the paper is the chemical radial gradient of metallicity. We performed an experiment that shows how a $\left[ {Fe/H} \right]$ radial gradient may originate, in absence of galactic winds, as simply a result of the baryonic reshaping of the density profiles on a timescale compatible with the dark matter profile reshaping. Thus, in this model, the presence of a radial gradient is always expected in the evolved dwarf as consequence of a smaller anisotropy parameter and a flat dark matter profile shape.
Finally, we want to point out a few general caveats:
\begin{itemize}
\item The fraction of gas has generally to be assumed as a free parameter (within the range indicated in Table \ref{Table01}) both in isolated and orbiting dwarf galaxies. We expect that different contents of gas left in the primordial dwarf galaxy can cause different star formation histories especially when the system evolves in orbit around a more massive host (e.g. \citet{2001Ap&SS.276..375M,2003A&A...405..931P}), leading to differences in the baryonic and dark matter phase space re-filling. We start with an already present stellar population formed in isolation. We will show in \citet{Pasetto2009b} that the gas consumption is mostly led by the orbit evolution.
\item Another type of observed SFH is the one presented by local systems as Dra and UMi. These systems show long-lasting early star formation episodes that led to considerable chemical enrichment. Obviously a substantial amount of the gas must thus have been retained in this early period (of, say, 13-10 Gyr). When this type of system lies within the well of an hosting system, the major role in the density profiles evolution is surely played by the external environment. Even in absence of gas, the tidal stripping is able to smooth system irregularities and to convert them in dSph (e.g., \citet{2001ApJ...559..754M}). This effect is shown to be efficient also in the MW dwarf companions (e.g., \citet{2003A&A...405..931P}) by morphologically evolving their density profiles. By extending these arguments to the more crowded cluster of galaxies (as Coma or Virgo), we enter the galaxy harassment regime \citep{1996Natur.379..613M} that is a pure gravitational effect on which the literature has still put too few constraints from the star formation and chemical point of view.
\item While the trend that leads to the chemical radial gradient is universally evident across the range of masses explored (and also generally present in the more extended sample of the elliptical galaxies, see. e.g. \citet{2002MNRAS.335..335C}) its gradient slope is not. $\frac{{d\left[ {Fe/H} \right]}}{{dr}}$ shows a strong dependence on the initial fraction of gas and stars that is not studied here.
\item The possible role of the triaxiality of the system is not investigated here. This is expected to play a minor role for the range of masses here investigated. Nevertheless the link between the $r_{lim}$, as defined in this paper, and the scale length of the model $r_c$ for the models here adopted has still to be investigated even for the more general class of elliptical galaxies and is the subject of a forthcoming study (\citet{Pasetto2009c}).
\end{itemize}
\begin{acknowledgements}
We thank C. Chiappini, F. Matteucci and C. Chiosi for useful discussions. We thank the referee G. Lanfranchi for useful comments. Simulations were performed on the GRACE supercomputer (grants I/80 041-043 of the Volkswagen Foundation and 823.219-439/30 and /36 of the Ministry of Science, Research and the Arts of Baden-W\"urttemberg). PB acknowledges the special support by the NAS Ukraine under the
Main Astronomical Observatory GRAPE/GRID computing cluster project. PB's studies are also
partially supported by the program Cosmomicrophysics of NAS Ukraine.
RS is supported by the Chinese Academy of Sciences Visiting Professorship for Senior International Scientists, Grant Number 2009S1-5 (The Silk Road Project).
\end{acknowledgements}
\bibliographystyle{apj}
|
\section{}
Formulating physics with more than one time leads to unsurmountable problems. However an elegant solution was found for the case of two times (2T) assuming the existence of a $Sp(2,R)$ gauge symmetry, involving a space with two timelike and an arbitrary number of spacelike directions, which reduces the theory effectively to one time after gauge fixing \cite{Bars:1998ph}. The simplest model we can formulate is for a particle moving in $D+2$ dimensions in a space with metric $(-,+,-,+,\dots,+)$ in $D$ spatial and two time dimensions having an $SO(D,2)$ isometry group. It was shown in a series of works that different gauge choices or different choices for the Hamiltonian among the generators of $SO(D,2)$ lead to different systems, like the massless relativistic particle in $D$ spacetime dimension \cite{Bars:1998pc}, a particle in $AdS_{D-n}\times S^n$ \cite{Bars:1998iy}, a massive non-relativistic particle in $D-1$ spatial dimensions \cite{Bars:1998iy}, the harmonic oscillator in $D-2$ spatial dimensions \cite{Bars:1998pc} or the hydrogen atom in $D-1$ spatial dimensions \cite{Bars:1998pc}. The common characteristic of these systems is that all of them share the original $SO(D,2)$ symmetry usually realized non-linearly in a hidden form. The formulation of 2T physics was also extended to the supersymmetric case \cite{Bars:1998ui}, to curved spaces \cite{Bars:2008sz}, and to field theories \cite{Bars:2000mz} and strings \cite{Bars:1997xb} with the same surprising results that different physical systems emerge out of it \cite{Bars:1998cs}. As stated above the emergence of different systems from the 2T theory is made in different settings.
Of course it would be desirable to have a formulation where all of its richness could be derived in just one single framework. This is precisely the aim of this paper. The transition amplitude for a particle in 2T physics is found by using the BRST quantization scheme in which only a single gauge choice is done. We will show that amplitudes for different lower dimensional systems are found when we take the initial and final points of the 2T particle in different hypersurfaces of the original $D+2$ space. Our results provide a more physical interpretation of 2T physics. The motion of the 2T particle restricted to different subspaces can be seen as a massless scalar field in Minkowski spacetime in $D$ dimensions or a massive scalar field in $AdS_D$ or any of the systems found in 2T physics. It provides an alternative way to understand 2T theories and strongly corroborates the results found earlier. Our results also have some of the flavor of M-theory in the sense that the restriction to different corners of the $D+2$ space reveals completely different theories in lower dimensions.
A particle moving in 2T theory has a phase space described by $(X^M, P_M)$, $M=0^\prime,1^\prime,0,1,\dots,D-1$. The gauge symmetries are generated by the three first class constraints $X^2 = X^M X_M, P^2 = P_M P^M$ and $PX = P_M X^M$ whose Poisson bracket algebra generates $Sp(2,R)$ \cite{Bars:1998cs}. Since the BRST-BFV formalism \cite{Fradkin:1975cq} is a powerful method to deal with gauge theories \cite{Rivelles:1995hm} we performed the BRST-BFV quantization of the 2T particle. The particle is described by its path $X^M(\tau)$ in the 2T space where $\tau$ is its evolution parameter. The action is composed by the usual kinetic term for a particle plus the constraints multiplied by their corresponding Lagrange multipliers $A^a(\tau)$, $a=1,2,3$, which will enforce them. The action is written as
\begin{equation} \label{1}
S = \int d\tau \left( P_M \dot{X}^M - A^1 X^2 - A^2 PX - A^3 P^2 \right),
\end{equation}
where dot means derivative with respect to $\tau$. This action is reparametrization invariant so its canonical Hamiltonian vanishes identically \cite{Bars:1998ph}. According to the BRST-BFV prescription we have to supplement the action (\ref{1}) with ghosts and momenta for the Lagrange multipliers so that the full action is invariant under a nilpotent BRST symmetry. Then a gauge fixing term, which is also BRST invariant, is further added to the action. The gauge choice that was made is $A^1 = A^2 = 0, \dot{A}^3 = 0$. The evaluation of the path integral representing the transition amplitude for this particle is rather involved and has many subtleties. It will be derived in full details in another publication \cite{new}. Here we will just present and discuss the resulting path integral. The transition amplitude for a particle in $D+2$ dimensions is giving by
\begin{equation} \label{2}
{\cal A}(X_i, X_f) = \delta(X^2_i) \delta(X^2_f) \int^{+\infty}_{-\infty} d\alpha \,\, \alpha^{D/2-2} \,\, e^{i\alpha X_i X_f},
\end{equation}
where $X_i = X(\tau_i)$ and $X_f = X(\tau_f)$ are the initial and final points of the particle respectively. Notice that it is manifestly invariant under $SO(D,2)$ but not under translations. Also, it does not depend on $\tau_i$ and $\tau_f$ To make sure that we have the right amplitude we can make a few checks. First of all, as just mentioned, it is invariant under the $SO(D,2)$ global symmetry of the classical theory as expected. It also has to obey Schr\H{o}edinger equation $H{\cal A} = -i \partial{\cal A}/\partial\tau$. Since the canonical Hamiltonian vanishes and the amplitude does not depend on $\tau$ this is also verified. Finally it has to satisfy all three constraints \footnote{The equations for a free scalar field in 2T physics \cite{Bars:2006dy} are the same as the ones being discussed here. The context however is different since we are working in a worldline formulation.}. The constraint $X^2=0$ is trivially satisfied for both arguments. The constraint $P^2 {\cal A} = - \partial^2 {\cal A}/\partial X^2 = 0$ requires some manipulations with delta functions but is also satisfied. Finally the last constraint has ordering problems but we find that the amplitude (\ref{2}) satisfies the symmetric ordered constraint $ (X^M \frac{\partial}{\partial X^M} + \frac{\partial}{\partial X^M} X^M) {\cal A} = 0$ \footnote{Notice that the solution to the constraints may not be unique. For instance, multiplying the integrand of (\ref{2}) by a factor of $\epsilon(\alpha)$ gives another solution of the constraint equations which, however, it is not derivable from the BRST-BFV formalism.}. In order to better handle the delta functions we will take the Fourier transform of the amplitude with respect to one of the extra coordinates. Using light cone coordinates $(X^{+^\prime},X^{-^\prime})$ for the sector $(X^{0^\prime},X^{1^\prime})$ we take a Fourier transform with respect to say $X^{+^\prime}$ obtaining
\begin{widetext}
\begin{equation}
\label{3}
\tilde{A}(p^{-^\prime}_i, X^{-^\prime}_i, X^\mu_i; p^{-^\prime}_f, X^{-^\prime}_f, X^\mu_f) = \frac{\exp \frac{1}{2} \left( \frac{p^{-^\prime}_i}{X^{-^\prime}_i} (X^\mu_i)^2 + \frac{p^{-^\prime}_f}{X^{-^\prime}_f} (X^\mu_f)^2 \right) }{|X^{-^\prime}_i X^{-^\prime}_f|} \int d\alpha \,\, \alpha^{D/2-2} \exp \left[ \frac{i}{2} \alpha \left( \sqrt{\frac{X^{-^\prime}_i}{X^{-^\prime}_f}} X^\mu_f - \sqrt{\frac{X^{-^\prime}_f}{X^{-^\prime}_i}} X^\mu_i \right)^2 \right],
\end{equation}
\end{widetext}
where $\mu=0,\dots,D-1$ and $p^{-^\prime}$ is the Fourier transformed variable.
We can now consider what happens when the initial and final points of the particle rest on some hyperplanes of the $D+2$ space. So let us choose $X^{-^\prime} = 1$ and $p^{-^\prime}=0$. We obtain immediately that
\begin{equation}
\label{4}
\tilde{A}( X^\mu_i; X^\mu_f) = \Delta_F(\Delta X^\mu),
\end{equation}
where $\Delta_F$ is the Feynman propagator for the massless Klein-Gordon field in Minkowski spacetime in $D$ dimensions. So if the initial and final points in the 2T particle amplitude reside in the surface defined by $X^{-^\prime} = 1$ and $p^{-^\prime}=0$ then the 2T particle behaves like the massless relativistic particle in two dimensions less, that is, in $D$ spacetime dimensions. It should be remarked that (\ref{4}) is not the transition amplitude for the relativistic particle computed, for instance, in the BRST formalism. The transition amplitude for the relativistic particle, which satisfies the relativistic particle constraint, is the Schwinger function $\Delta^{(1)}(\Delta X^\mu)$ satisfying the Klein-Gordon equation \cite{Henneaux:1992fc}. As we will see, this is a common feature for relativistic systems. We get the propagator of some field theory and not the transition amplitude for the associated particle. For the non-relativistic systems we will always get the quantum mechanical transition amplitude. Another remark is that instead of choosing $X^{-^\prime} = 1$ we could have chosen $X^{-^\prime} = \mbox{constant}$. Then the RHS of (\ref{4}) would be multiplied by $1/\mbox{constant}^2$ and the coefficient of the propagator would recall its higher dimensional origin.
The next corner we are going to look will show us the propagator of a scalar field in $AdS_D$. To this end we will make the change of variables from $(X^0, X^i, X^{D-1}, X^{-^\prime})$ to $(t, x^i, y, R)$, with $i=1,\dots,D-2$, given by
\begin{equation} \label{5}
X^0 = R t/y, \,\, X^i= R x^i/y, \,\, X^{D-1} = R, \,\, X^{-^\prime} = R/y.
\end{equation}
As we shall see $(t,x^i,y)$ will turn out to be Poincar\'e coordinates for $AdS_D$. With this choice we get
\begin{equation} \label{6}
\left( \sqrt{\frac{X^{-^\prime}_i}{X^{-^\prime}_f}} X^\mu_f - \sqrt{\frac{X^{-^\prime}_f}{X^{-^\prime}_i}} X^\mu_i \right)^2 = 2 R_i R_f u,
\end{equation}
where $u = [ (\Delta x^\mu)^2 + y^2_i + y^2_f ]/(2y_iy_f) -1$ is the chordal distance in $AdS_D$. After rescaling the integration variable by $\alpha \rightarrow \alpha/(R_i R_f)$ and choosing the hypersurface $p^{-^\prime} = 0$ and $ y - R^{D/2} = 0$, we can eliminate $R$ to get for (\ref{3}) (up to numerical factors)
\begin{equation} \label{7}
\tilde{A}(x^\mu_i, y_i; x^\mu_f, y_f) = \frac{\Gamma(D/2-1)}{(u + i\epsilon)^{D/2-1}},
\end{equation}
which is the propagator for a massive scalar field in $AdS_D$ coupled to the background in a Weyl invariant way \cite{D'Hoker:2002aw}. Its mass is given by $m^2 = - D(D-2)$ and this value is, of course, above the Breitenlohner-Freedman bound.
We can generalize the former result to a conformally flat spacetime in a slightly different way. Let us choose $p^{-^\prime} = 0$ and $X^{-^\prime} = \exp \sigma(x^\mu)$, and perform the change of variables $ X^\mu = \exp\sigma(x) x^\mu$. After the rescaling of the integration variable by $\alpha \rightarrow \alpha \exp( -\sigma(x_i) -\sigma(x_f) )$ we find that (\ref{3}) reduces to
\begin{equation} \label{8}
\tilde{\cal A}(x^\mu_i, x^\mu_f) = \exp\left[{-\frac{D}{2} (\sigma(x_i) + \sigma(x_f))}\right] \Delta_F(\Delta x^\mu),
\end{equation}
where, as before, $\Delta_F$ is the massless propagator of the Klein-Gordon field in flat Minkowski spacetime in $D$ dimensions. Clearly $\exp\sigma(x)$ is a conformal factor. The result (\ref{8}) is proportional to the propagator of the massless Klein-Gordon field conformally coupled to the background \cite{Birrell:1982ix}. There is a factor of $\exp((\sigma(x_i) + \sigma(x_f))$ missing in it so the result (\ref{8}) seems to recall its higher dimensional origin. Alternatively we could have started with $\exp((1-2/D)\sigma(x))$ everywhere instead of $\exp\sigma(x)$ and the right conformal factor would appear in (\ref{8}). Our results for the coupling to $AdS_D$ and to the conformally flat spacetime completely agree with the ones obtained in the context of 2T field theory \cite{Bars:2007kw,Bars:2000mz}.
Finally we will derive two nonrelativistic systems. The first is the nonrelativistic massive particle in $D-1$ spatial dimensions. To that end we will take a Fourier transform on $X^0_i$ and $X^0_f$ in (\ref{3}), with transformed variables $p^0_i$ and $p^0_f$ respectively. We choose the hypersurface $p^{-^\prime}=0$ and $X^{-^\prime}= 1/\sqrt{m}$ with $m$ constant. After rescaling the integral by $\alpha \rightarrow m\alpha$ and calling $p^0_f = \sqrt{mE}$, with $E$ another constant we get
\begin{widetext}
\begin{equation} \label{9}
\tilde{\cal A}(\vec{X}_i, \vec{X}_f, m, E) = m^{d/2} \int d\alpha \,\, \alpha^{d/2-2} \exp\left[ {\frac{i}{2} ( m\alpha \Delta\vec{X}^2 + \frac{E}{\alpha} )}\right] \delta(p^0_i + \sqrt{mE}) ,
\end{equation}
\end{widetext}
where $d=D-1$. Up to the delta function this is the amplitude for a nonrelativistic particle of mass $m$ and energy $E$ in $d$ spatial dimensions. This result is not so surprising since it could have been obtained directly from the propagator of the massless relativistic particle. The connection between the massless relativistic particle and the massive nonrelativistic particle is well know and it is due to the presence, again, of the $SO(D,2)$ symmetry \cite{Leiva:2003kd}.
For the last case we will use light-cone coordinates for the $X^\mu$ sector $(X^+, X^-, \vec{X})$. We start by taking the Fourier transform of $X^-_i,X^-_f$ with transformed variables $p^+_i, p^+_f$ respectively. This Fourier transform is used to get two delta functions and one of them can be used to perform the integral on $\alpha$. We now change variables from $(X^+, X^{-^\prime}, p^+, p^{-^\prime})$ to $(t, \tilde{x}^{-^\prime}, \tilde{p}^+, \tilde{p}^{-^\prime})$ according to
\begin{equation} \label{10}
X^+ = e^t, \,\, X^{-^\prime} = \tilde{x}^{-^\prime} e^{-t}, \,\, p^+ = \tilde{p}^+ e^t, \,\, p^{-^\prime} = \tilde{p}^{-^\prime} e^{-t},
\end{equation}
and choose the hypersurface defined by $\tilde{x}^{-^\prime} = 1$ and $\tilde{p}^+ + \tilde{p}^{-^\prime} = 0$. The final result is
\begin{widetext}
\begin{equation} \label{11}
\tilde{\cal A}(\tilde{p}^+_i, t_i, \vec{X}_i; \tilde{p}^+_f, t_f, \vec{X}_f) = (\tilde{p}^+_f)^{d/2-1} \,\,\, \sinh^{-d/2}\Delta t \,\,\, \exp\left[ \frac{i}{2} \tilde{p}^+_f \frac{\cosh t}{\sinh t} \left( \vec{X}_i^2 + \vec{X}_f^2 - \frac{2 \vec{X}_i \vec{X}_f}{\cosh t} \right) \right] \delta(\tilde{p}^+_i + \tilde{p}^+_f),
\end{equation}
\end{widetext}
where $d=D-2$. Notice the appearance of hyperbolic functions in (\ref{11}). It is proportional to the transition amplitude for the inverted harmonic oscillator in $d$ spatial dimensions with mass $\tilde{p}^+_f$. The inverted harmonic oscillator has the wrong sign for the potential but even being unbounded from below it can be quantized in order to study unstable systems \cite{Barton:1984ey}. Notice that to get the amplitude for the ordinary harmonic oscillator we should have to replace $t$ by $it$ in (\ref{10}) but this is not allowed by the change of variables if they are all real. In the usual formulation of 2T physics an ordinary harmonic oscillator is obtained \cite{Bars:1998pc}. Since the parametrization of our hypersurface does not seem to be related to the parametrization used in \cite{Bars:1998pc} it appears that we are looking at a different sector of the 2T theory. Again, (\ref{11}) is proportional to the inverted harmonic oscillator amplitude with the proportionality factors being $1/\tilde{p}^+_f$ and the delta function.
Finally some comments are in order. If we compare our hypersurfaces with the gauge choices made in the usual formulation of 2T physics \footnote{There are two nice tables in \cite{Bars:2007kw} where all of the cases studied here, except for the harmonic oscillator, are summarized.} we see some similarity. This is to be expected because in the amplitude (\ref{3}) we have explicitly to use the constrains $X^2=0$ which is also done when the gauge is fixed in the usual 2T formulation. So it is natural that some components of $X^M$ and $P^M$ do indeed coincide but the identification is not exact in all cases. It should also be noticed that the inverted harmonic oscillator was the only case with a non vanishing $p^{-^\prime}$ that we were able to analyze. We could not find the hypersurfaces which give rise to other systems like the massive relativistic particle or the hydrogen atom. It seems that the reason is related to the inclusion of the momenta in the definition of the hypersurface. In fact if we make a naive Fourier transformation to get the transition amplitude in phase space we find that it is problematic. Work in this direction is under way.
It is quite remarkable that a single transition amplitude in $D+2$ dimensions can give rise to several distinct amplitudes in two or three dimensions less. As fully explained in many papers on 2T physics they all share a common $SO(D,2)$ symmetry. This looks like a toy model realization of M-theory. Here we have the emergence of different systems when we fix the initial and final points of the 2T particle to rest on a given subspace of the $D+2$ original space. Even though a 2T formulation of supersymmetry and strings are known we can not leave aside the deep connection of $SO(D,2)$ and the AdS/CFT correspondence. We are planning to extend the formalism presented here to include supersymmetry and strings and try to shed some new light on these matters.
\begin{acknowledgments}
V.O.R. would like to thank Itzhak Bars for helpful discussions and comments.
The work of J.E.F. was supported in its initial stage by CNPq.
The work of V.O.R. is supported by CNPq grant No. 304495/2007-7, FAPESP grant No. 2008/05343-5 and PROSUL grant No. 490134/2006-8.
\end{acknowledgments}
\section{}
Formulating physics with more than one time leads to unsurmountable problems. However an elegant solution was found for the case of two times (2T) assuming the existence of a $Sp(2,R)$ gauge symmetry, involving a space with two timelike and an arbitrary number of spacelike directions, which reduces the theory effectively to one time after gauge fixing \cite{Bars:1998ph}. The simplest model we can formulate is for a particle moving in $D+2$ dimensions in a space with metric $(-,+,-,+,\dots,+)$ in $D$ spatial and two time dimensions having an $SO(D,2)$ isometry group. It was shown in a series of works that different gauge choices or different choices for the Hamiltonian among the generators of $SO(D,2)$ lead to different systems, like the massless relativistic particle in $D$ spacetime dimension \cite{Bars:1998pc}, a particle in $AdS_{D-n}\times S^n$ \cite{Bars:1998iy}, a massive non-relativistic particle in $D-1$ spatial dimensions \cite{Bars:1998iy}, the harmonic oscillator in $D-2$ spatial dimensions \cite{Bars:1998pc} or the hydrogen atom in $D-1$ spatial dimensions \cite{Bars:1998pc}. The common characteristic of these systems is that all of them share the original $SO(D,2)$ symmetry usually realized non-linearly in a hidden form. The formulation of 2T physics was also extended to the supersymmetric case \cite{Bars:1998ui}, to curved spaces \cite{Bars:2008sz}, and to field theories \cite{Bars:2000mz} and strings \cite{Bars:1997xb} with the same surprising results that different physical systems emerge out of it \cite{Bars:1998cs}. As stated above the emergence of different systems from the 2T theory is made in different settings.
Of course it would be desirable to have a formulation where all of its richness could be derived in just one single framework. This is precisely the aim of this paper. The transition amplitude for a particle in 2T physics is found by using the BRST quantization scheme in which only a single gauge choice is done. We will show that amplitudes for different lower dimensional systems are found when we take the initial and final points of the 2T particle in different hypersurfaces of the original $D+2$ space. Our results provide a more physical interpretation of 2T physics. The motion of the 2T particle restricted to different subspaces can be seen as a massless scalar field in Minkowski spacetime in $D$ dimensions or a massive scalar field in $AdS_D$ or any of the systems found in 2T physics. It provides an alternative way to understand 2T theories and strongly corroborates the results found earlier. Our results also have some of the flavor of M-theory in the sense that the restriction to different corners of the $D+2$ space reveals completely different theories in lower dimensions.
A particle moving in 2T theory has a phase space described by $(X^M, P_M)$, $M=0^\prime,1^\prime,0,1,\dots,D-1$. The gauge symmetries are generated by the three first class constraints $X^2 = X^M X_M, P^2 = P_M P^M$ and $PX = P_M X^M$ whose Poisson bracket algebra generates $Sp(2,R)$ \cite{Bars:1998cs}. Since the BRST-BFV formalism \cite{Fradkin:1975cq} is a powerful method to deal with gauge theories \cite{Rivelles:1995hm} we performed the BRST-BFV quantization of the 2T particle. The particle is described by its path $X^M(\tau)$ in the 2T space where $\tau$ is its evolution parameter. The action is composed by the usual kinetic term for a particle plus the constraints multiplied by their corresponding Lagrange multipliers $A^a(\tau)$, $a=1,2,3$, which will enforce them. The action is written as
\begin{equation} \label{1}
S = \int d\tau \left( P_M \dot{X}^M - A^1 X^2 - A^2 PX - A^3 P^2 \right),
\end{equation}
where dot means derivative with respect to $\tau$. This action is reparametrization invariant so its canonical Hamiltonian vanishes identically \cite{Bars:1998ph}. According to the BRST-BFV prescription we have to supplement the action (\ref{1}) with ghosts and momenta for the Lagrange multipliers so that the full action is invariant under a nilpotent BRST symmetry. Then a gauge fixing term, which is also BRST invariant, is further added to the action. The gauge choice that was made is $A^1 = A^2 = 0, \dot{A}^3 = 0$. The evaluation of the path integral representing the transition amplitude for this particle is rather involved and has many subtleties. It will be derived in full details in another publication \cite{new}. Here we will just present and discuss the resulting path integral. The transition amplitude for a particle in $D+2$ dimensions is giving by
\begin{equation} \label{2}
{\cal A}(X_i, X_f) = \delta(X^2_i) \delta(X^2_f) \int^{+\infty}_{-\infty} d\alpha \,\, \alpha^{D/2-2} \,\, e^{i\alpha X_i X_f},
\end{equation}
where $X_i = X(\tau_i)$ and $X_f = X(\tau_f)$ are the initial and final points of the particle respectively. Notice that it is manifestly invariant under $SO(D,2)$ but not under translations. Also, it does not depend on $\tau_i$ and $\tau_f$ To make sure that we have the right amplitude we can make a few checks. First of all, as just mentioned, it is invariant under the $SO(D,2)$ global symmetry of the classical theory as expected. It also has to obey Schr\H{o}edinger equation $H{\cal A} = -i \partial{\cal A}/\partial\tau$. Since the canonical Hamiltonian vanishes and the amplitude does not depend on $\tau$ this is also verified. Finally it has to satisfy all three constraints \footnote{The equations for a free scalar field in 2T physics \cite{Bars:2006dy} are the same as the ones being discussed here. The context however is different since we are working in a worldline formulation.}. The constraint $X^2=0$ is trivially satisfied for both arguments. The constraint $P^2 {\cal A} = - \partial^2 {\cal A}/\partial X^2 = 0$ requires some manipulations with delta functions but is also satisfied. Finally the last constraint has ordering problems but we find that the amplitude (\ref{2}) satisfies the symmetric ordered constraint $ (X^M \frac{\partial}{\partial X^M} + \frac{\partial}{\partial X^M} X^M) {\cal A} = 0$ \footnote{Notice that the solution to the constraints may not be unique. For instance, multiplying the integrand of (\ref{2}) by a factor of $\epsilon(\alpha)$ gives another solution of the constraint equations which, however, it is not derivable from the BRST-BFV formalism.}. In order to better handle the delta functions we will take the Fourier transform of the amplitude with respect to one of the extra coordinates. Using light cone coordinates $(X^{+^\prime},X^{-^\prime})$ for the sector $(X^{0^\prime},X^{1^\prime})$ we take a Fourier transform with respect to say $X^{+^\prime}$ obtaining
\begin{widetext}
\begin{equation}
\label{3}
\tilde{A}(p^{-^\prime}_i, X^{-^\prime}_i, X^\mu_i; p^{-^\prime}_f, X^{-^\prime}_f, X^\mu_f) = \frac{\exp \frac{1}{2} \left( \frac{p^{-^\prime}_i}{X^{-^\prime}_i} (X^\mu_i)^2 + \frac{p^{-^\prime}_f}{X^{-^\prime}_f} (X^\mu_f)^2 \right) }{|X^{-^\prime}_i X^{-^\prime}_f|} \int d\alpha \,\, \alpha^{D/2-2} \exp \left[ \frac{i}{2} \alpha \left( \sqrt{\frac{X^{-^\prime}_i}{X^{-^\prime}_f}} X^\mu_f - \sqrt{\frac{X^{-^\prime}_f}{X^{-^\prime}_i}} X^\mu_i \right)^2 \right],
\end{equation}
\end{widetext}
where $\mu=0,\dots,D-1$ and $p^{-^\prime}$ is the Fourier transformed variable.
We can now consider what happens when the initial and final points of the particle rest on some hyperplanes of the $D+2$ space. So let us choose $X^{-^\prime} = 1$ and $p^{-^\prime}=0$. We obtain immediately that
\begin{equation}
\label{4}
\tilde{A}( X^\mu_i; X^\mu_f) = \Delta_F(\Delta X^\mu),
\end{equation}
where $\Delta_F$ is the Feynman propagator for the massless Klein-Gordon field in Minkowski spacetime in $D$ dimensions. So if the initial and final points in the 2T particle amplitude reside in the surface defined by $X^{-^\prime} = 1$ and $p^{-^\prime}=0$ then the 2T particle behaves like the massless relativistic particle in two dimensions less, that is, in $D$ spacetime dimensions. It should be remarked that (\ref{4}) is not the transition amplitude for the relativistic particle computed, for instance, in the BRST formalism. The transition amplitude for the relativistic particle, which satisfies the relativistic particle constraint, is the Schwinger function $\Delta^{(1)}(\Delta X^\mu)$ satisfying the Klein-Gordon equation \cite{Henneaux:1992fc}. As we will see, this is a common feature for relativistic systems. We get the propagator of some field theory and not the transition amplitude for the associated particle. For the non-relativistic systems we will always get the quantum mechanical transition amplitude. Another remark is that instead of choosing $X^{-^\prime} = 1$ we could have chosen $X^{-^\prime} = \mbox{constant}$. Then the RHS of (\ref{4}) would be multiplied by $1/\mbox{constant}^2$ and the coefficient of the propagator would recall its higher dimensional origin.
The next corner we are going to look will show us the propagator of a scalar field in $AdS_D$. To this end we will make the change of variables from $(X^0, X^i, X^{D-1}, X^{-^\prime})$ to $(t, x^i, y, R)$, with $i=1,\dots,D-2$, given by
\begin{equation} \label{5}
X^0 = R t/y, \,\, X^i= R x^i/y, \,\, X^{D-1} = R, \,\, X^{-^\prime} = R/y.
\end{equation}
As we shall see $(t,x^i,y)$ will turn out to be Poincar\'e coordinates for $AdS_D$. With this choice we get
\begin{equation} \label{6}
\left( \sqrt{\frac{X^{-^\prime}_i}{X^{-^\prime}_f}} X^\mu_f - \sqrt{\frac{X^{-^\prime}_f}{X^{-^\prime}_i}} X^\mu_i \right)^2 = 2 R_i R_f u,
\end{equation}
where $u = [ (\Delta x^\mu)^2 + y^2_i + y^2_f ]/(2y_iy_f) -1$ is the chordal distance in $AdS_D$. After rescaling the integration variable by $\alpha \rightarrow \alpha/(R_i R_f)$ and choosing the hypersurface $p^{-^\prime} = 0$ and $ y - R^{D/2} = 0$, we can eliminate $R$ to get for (\ref{3}) (up to numerical factors)
\begin{equation} \label{7}
\tilde{A}(x^\mu_i, y_i; x^\mu_f, y_f) = \frac{\Gamma(D/2-1)}{(u + i\epsilon)^{D/2-1}},
\end{equation}
which is the propagator for a massive scalar field in $AdS_D$ coupled to the background in a Weyl invariant way \cite{D'Hoker:2002aw}. Its mass is given by $m^2 = - D(D-2)$ and this value is, of course, above the Breitenlohner-Freedman bound.
We can generalize the former result to a conformally flat spacetime in a slightly different way. Let us choose $p^{-^\prime} = 0$ and $X^{-^\prime} = \exp \sigma(x^\mu)$, and perform the change of variables $ X^\mu = \exp\sigma(x) x^\mu$. After the rescaling of the integration variable by $\alpha \rightarrow \alpha \exp( -\sigma(x_i) -\sigma(x_f) )$ we find that (\ref{3}) reduces to
\begin{equation} \label{8}
\tilde{\cal A}(x^\mu_i, x^\mu_f) = \exp\left[{-\frac{D}{2} (\sigma(x_i) + \sigma(x_f))}\right] \Delta_F(\Delta x^\mu),
\end{equation}
where, as before, $\Delta_F$ is the massless propagator of the Klein-Gordon field in flat Minkowski spacetime in $D$ dimensions. Clearly $\exp\sigma(x)$ is a conformal factor. The result (\ref{8}) is proportional to the propagator of the massless Klein-Gordon field conformally coupled to the background \cite{Birrell:1982ix}. There is a factor of $\exp((\sigma(x_i) + \sigma(x_f))$ missing in it so the result (\ref{8}) seems to recall its higher dimensional origin. Alternatively we could have started with $\exp((1-2/D)\sigma(x))$ everywhere instead of $\exp\sigma(x)$ and the right conformal factor would appear in (\ref{8}). Our results for the coupling to $AdS_D$ and to the conformally flat spacetime completely agree with the ones obtained in the context of 2T field theory \cite{Bars:2007kw,Bars:2000mz}.
Finally we will derive two nonrelativistic systems. The first is the nonrelativistic massive particle in $D-1$ spatial dimensions. To that end we will take a Fourier transform on $X^0_i$ and $X^0_f$ in (\ref{3}), with transformed variables $p^0_i$ and $p^0_f$ respectively. We choose the hypersurface $p^{-^\prime}=0$ and $X^{-^\prime}= 1/\sqrt{m}$ with $m$ constant. After rescaling the integral by $\alpha \rightarrow m\alpha$ and calling $p^0_f = \sqrt{mE}$, with $E$ another constant we get
\begin{widetext}
\begin{equation} \label{9}
\tilde{\cal A}(\vec{X}_i, \vec{X}_f, m, E) = m^{d/2} \int d\alpha \,\, \alpha^{d/2-2} \exp\left[ {\frac{i}{2} ( m\alpha \Delta\vec{X}^2 + \frac{E}{\alpha} )}\right] \delta(p^0_i + \sqrt{mE}) ,
\end{equation}
\end{widetext}
where $d=D-1$. Up to the delta function this is the amplitude for a nonrelativistic particle of mass $m$ and energy $E$ in $d$ spatial dimensions. This result is not so surprising since it could have been obtained directly from the propagator of the massless relativistic particle. The connection between the massless relativistic particle and the massive nonrelativistic particle is well know and it is due to the presence, again, of the $SO(D,2)$ symmetry \cite{Leiva:2003kd}.
For the last case we will use light-cone coordinates for the $X^\mu$ sector $(X^+, X^-, \vec{X})$. We start by taking the Fourier transform of $X^-_i,X^-_f$ with transformed variables $p^+_i, p^+_f$ respectively. This Fourier transform is used to get two delta functions and one of them can be used to perform the integral on $\alpha$. We now change variables from $(X^+, X^{-^\prime}, p^+, p^{-^\prime})$ to $(t, \tilde{x}^{-^\prime}, \tilde{p}^+, \tilde{p}^{-^\prime})$ according to
\begin{equation} \label{10}
X^+ = e^t, \,\, X^{-^\prime} = \tilde{x}^{-^\prime} e^{-t}, \,\, p^+ = \tilde{p}^+ e^t, \,\, p^{-^\prime} = \tilde{p}^{-^\prime} e^{-t},
\end{equation}
and choose the hypersurface defined by $\tilde{x}^{-^\prime} = 1$ and $\tilde{p}^+ + \tilde{p}^{-^\prime} = 0$. The final result is
\begin{widetext}
\begin{equation} \label{11}
\tilde{\cal A}(\tilde{p}^+_i, t_i, \vec{X}_i; \tilde{p}^+_f, t_f, \vec{X}_f) = (\tilde{p}^+_f)^{d/2-1} \,\,\, \sinh^{-d/2}\Delta t \,\,\, \exp\left[ \frac{i}{2} \tilde{p}^+_f \frac{\cosh t}{\sinh t} \left( \vec{X}_i^2 + \vec{X}_f^2 - \frac{2 \vec{X}_i \vec{X}_f}{\cosh t} \right) \right] \delta(\tilde{p}^+_i + \tilde{p}^+_f),
\end{equation}
\end{widetext}
where $d=D-2$. Notice the appearance of hyperbolic functions in (\ref{11}). It is proportional to the transition amplitude for the inverted harmonic oscillator in $d$ spatial dimensions with mass $\tilde{p}^+_f$. The inverted harmonic oscillator has the wrong sign for the potential but even being unbounded from below it can be quantized in order to study unstable systems \cite{Barton:1984ey}. Notice that to get the amplitude for the ordinary harmonic oscillator we should have to replace $t$ by $it$ in (\ref{10}) but this is not allowed by the change of variables if they are all real. In the usual formulation of 2T physics an ordinary harmonic oscillator is obtained \cite{Bars:1998pc}. Since the parametrization of our hypersurface does not seem to be related to the parametrization used in \cite{Bars:1998pc} it appears that we are looking at a different sector of the 2T theory.
Finally some comments are in order. If we compare our hypersurfaces with the gauge choices made in the usual formulation of 2T physics \footnote{There are two nice tables in \cite{Bars:2007kw} where all of the cases studied here, except for the harmonic oscillator, are summarized.} we see some similarity. This is to be expected because in the amplitude (\ref{3}) we have explicitly to use the constrains $X^2=0$ which is also done when the gauge is fixed in the usual 2T formulation. So it is natural that some components of $X^M$ and $P^M$ do indeed coincide but the identification is not exact in all cases. It should also be noticed that the inverted harmonic oscillator was the only case with a non vanishing $p^{-^\prime}$ that we were able to analyze. We could not find the hypersurfaces which give rise to other systems like the massive relativistic particle or the hydrogen atom. It seems that the reason is related to the inclusion of the momenta in the definition of the hypersurface. In fact if we make a naive Fourier transformation to get the transition amplitude in phase space we find that it is problematic. Work in this direction is under way.
It is quite remarkable that a single transition amplitude in $D+2$ dimensions can give rise to several distinct amplitudes in two or three dimensions less. As fully explained in many papers on 2T physics they all share a common $SO(D,2)$ symmetry. This looks like a toy model realization of M-theory. Here we have the emergence of different systems when we fix the initial and final points of the 2T particle to rest on a given subspace of the $D+2$ original space. Even though a 2T formulation of supersymmetry and strings are known we can not leave aside the deep connection of $SO(D,2)$ and the AdS/CFT correspondence. We are planning to extend the formalism presented here to include supersymmetry and strings and try to shed some new light on these matters.
\begin{acknowledgments}
V.O.R. would like to thank Itzhak Bars for helpful discussions and comments.
The work of J.E.F. was supported in its initial stage by CNPq.
The work of V.O.R. is supported by CNPq grant No. 304495/2007-7, FAPESP grant No. 2008/05343-5 and PROSUL grant No. 490134/2006-8.
\end{acknowledgments}
|
\section{Introduction}
In classical communication systems, information is usually encoded into orthogonal quantum or semi-classical states of light. An important example is binary phase shift keying with coherent states where the logical information, "0" and "1", is encoded as two coherent states with large amplitudes and a relative phase of $\pi$. Since the two coherent states each possess a large amplitude (with opposite signs), they are nearly orthogonal and thus they can be easily discriminated using an interferometric measurement strategy. On the other hand, if the amplitude becomes very small, which is the case for quantum communication as well as long distance amplification-free (e.g. free-space) classical communication, the two states are largely overlapping and thus non-orthognal. Due to this non-orthogonality, the states can no longer be perfectly discriminated. Although perfect discrimination is not possible, it is of high interest to find optimized strategies in order to minimize measurement errors, thus keeping the error rate as low as possible and increasing the mutual information between sender and receiver. Moreover, the search for such optimised strategies are of utmost importance for many applications in quantum communication with quantum key distribution (QKD) being the prime example~\cite{Bennett1984, Bennett1992, Grosshans2003}. Finally, we note that the problem of finding optimised measurement schemes associted with pre-defined alphabet is a fundamental problem in quantum mechnics~\cite{Helstrom1976, Ivanovic1987}.
There are basically two well-known discrimination strategies. In the first strategy, all measurement outcomes are used (that is, it is deterministic) and, therefore, the resulting conclusions will be infected by errors. The idea is to optimise the strategy such that the probability for making an error is minimized. This strategy is known as minimum error state discrimination and has been analyzed by Helstrom~\cite{Helstrom1976}.
The second discrimination strategy is probabilistic, and yields only a valid outcome when the conclusion drawn from the measurement is known to be error-free. Therefore, in this task the goal is to minimise the probability of inconclusive results (which are discarded). This strategy is known as unambiguous state discrimination (USD) and was originally proposed by Ivanovic, Dieks and Peres~\cite{Ivanovic1987,Dieks1988, Peres1988, Jaeger1995}.
A combination of the two discrimination schemes where one allows for both erroneous and inconclusive results has also been treated theoretically. More precisely, the minimal probability of errors for a fixed probability of inconclusive results has been derived for pure and mixed states in refs.~\cite{Chefles1998} and~\cite{fiurek_optimal_2003}, respectively.
For the discrimination of two coherent states with minimum error, several optimal and near-optimal receivers have been proposed~\cite{Kennedy1973, Dolinar1973, Sasaki1996, Geremia2004, Takeoka2005,Takeoka2008}.
Also, a device for implementing USD of coherent states was proposed by Huttner~\cite{Huttner1995} and later by Banaszek~\cite{Banaszek1999a}. Some of these schemes have been experimentally accessed, such as the Dolinar receiver~\cite{Cook2007}, the optimized displacement receiver~\cite{Wittmann2008c} and a programmable receiver implementing USD~\cite{bartkov_fiber-optics_2007}. However, the intermediate regime where errors as well as inconclusive results may occur has only recently been investigated experimentally~\cite{Wittmann2009d}.
In this paper we elaborate on the work of ref~\cite{Wittmann2009d}. We investigate two different receivers that belong to the intermediate regime. The first receiver is a standard homodyne detector, and the second receiver is a displacement controlled photon number resolving detector~\cite{Wittmann2009}. In both receivers the measurement outcomes are postselected to obtain a specific relation between errors and inconclusive results. The postselection based homodyne detector has been used in various protocols such as QKD~\cite{Silberhorn2002, Lorenz2004, Lance2005}, squeezed state and entangled state distillation~\cite{fiurek_experimentally_2007,Heersink2006,Dong2008,Franzen2006,Hage2008} and quantum state engineering\cite{Marek2009, Lance2006, Ourjoumtsev2007, Babichev2004}. Here we thoroughly characterize the detector in terms of the discrimination between two coherent states. In addition we conduct a thorough experimental analysis of the displacement based photon number resolving detector which was introduced in ref~\cite{Wittmann2009, Wittmann2009d}. We find that the displacement based PNR receiver outperforms the standard homodyne detection.
The paper is organized as follows. First we recapitulate the notion of intermediate measurement for coherent states in section~\ref{measurement}. In section~\ref{HD} and section~\ref{PNR}, we consider measurements in the intermediate regime with two strategies: a receiver using a homodyne detection and a receiver using an optimized displacement combined with a photon number resolving detector. In section~\ref{HD}, we also prove that the post-selection based homodyne scheme is
the optimal strategy for realizing the intermediate measurement
within all possible Gaussian operations and conditional dynamics. We demonstrate experimentally both receivers in section~\ref{exp}. Finally, we summarize the results in section~\ref{comp}.
\section{Intermediate measurement}
\label{measurement}
Let us assume a sender picks one signal state out of two pure and phase shifted coherent states and sends it through a communication channel that preserves the quantum property of the state. On the other end of the channel, a receiver has to tell which state was chosen by the sender. Let us also assume that the \textit{a priori} probabilities for the preparation of the states are $p_1=p_2=\frac{1}{2}$ and that the received states are $|-\alpha\rangle$ or $|+\alpha\rangle$.
The receiver measures the signal state and based on the measurement outcome guesses the state. Due to the non-orthogonality of the alphabet, however, the result will not be correct in all such attempts. In fact, the minimal error probability is given by the inner product of the states in the alphabet, $\sigma=|\langle-\alpha|\alpha\rangle|$. The maximally accessible information of the receiver is directly related to the minimal error rate.
The receiver can alternatively choose a measurement strategy which allows for inconclusive results. In this strategy, he will only accept states that are likely to be correctly identified, while he does not attempt to guess the results for signals associated with the inconclusive measurement results. This strategy is probabilistic as the outcomes are post selected. It can be shown, that for higher probability of inconclusive results $p_\mathrm{inc}$ (or lower acceptance probability $1-p_\mathrm{inc}$) a lower error probability $p_\mathrm{E}$ can be achieved.
This intermediate measurement strategy can be described by the three-component positive operator-valued measure (POVM) $\hat\Pi_i, i=1,2,?$ where $\hat\Pi_i > 0$ and $\hat\Pi_1+\hat\Pi_2+\hat\Pi_?=\hat I$. Consequently, an inconclusive result will occur with the probability
\begin{equation}
p_\mathrm{inc}=p_1\langle-\alpha|\hat\Pi_?|-\alpha\rangle+p_2\langle \alpha|\hat\Pi_?|\alpha\rangle.
\label{inconclusive}
\end{equation}
where $\langle-\alpha|\hat\Pi_?|-\alpha\rangle$ ($\langle \alpha|\hat\Pi_?|\alpha\rangle$) represents the probability of inconclusive results when $|-\alpha\rangle$ $(|\alpha\rangle)$ was prepared.
Furthermore, the average error probability is given by
\begin{equation}
p_\mathrm{E}=\frac{p_1\langle-\alpha|\hat\Pi_2|-\alpha\rangle+p_2\langle \alpha|\hat\Pi_1|\alpha\rangle}{1-p_\mathrm{inc}},
\label{errorrate}
\end{equation}
where $\langle-\alpha|\hat\Pi_2|-\alpha\rangle$ ($\langle\alpha|\hat\Pi_1|\alpha\rangle$) represents the error probability of mistakenly guessing $|\alpha\rangle$ $(|+\alpha\rangle)$.
Finally the measurement strategy is optimized, such that the receiver's error probability is minimized for a given probability of inconclusive results. The error probability according to~\cite{Chefles1998} is then given by
\begin{equation}
p_\mathrm{E}\ge \frac{1}{2} \left(1-\frac{\left(1-2p_\mathrm{inc}(1-\sigma)-\sigma^2\right)^{1/2}}{1-p_\mathrm{inc}}\right),
\label{opterror}
\end{equation}
where the error rate is lower bounded by the inner product and the tolerated rate of inconclusive results. A receiver scheme achieving this optimal bound is yet unknown. (To our knowledge this is also true for two non orthogonal qubits.) In the following two sections, we investigate two near-optimal receivers; the post-selection based homodyne receiver and displacement controlled photon number resolving detector.
\section{Homodyne receiver}
\label{HD}
\begin{figure}
\begin{tabular}{l}
\centerline{\includegraphics[width=8.4cm]{schemes.eps}} \\ [-5.7cm]
(a)\hspace{3.6cm}(c) \\[2.4cm]
(b)\hspace{3.6cm}(d) \\[2.4cm]
(e)\\[-0.2cm]
\centerline{\includegraphics[width=8cm]{ComparisonHDPNR.eps}} \\
\end{tabular}
\caption{\label{schemes} (a) Schematics of the homodyne receiver, showing that the signal (S) is interfered with a local oscillator (LO) on a 50/50 beam splitter. The photocurrents of two photo diodes are subtracted and results in a quadrature measurement along the encoding quadrature. (b) Marginal distribution of the two signal states. In the example, we assume a signal with $|\alpha|^2=0.24$. According to the measurement result the correct answer ($-, ?, +$) is guessed. (c) Schematics of the photon number resolving (PNR) receiver. The signal (S) is interfered with an auxiliary oscillator (AO) on a highly transmissive beam splitter. Finally, the signal is measured by a photon number resolving detector (PNRD) (d) Photon number distribution of two signal states. According to the measurement outcome of the PNRD the correct answer is guessed. In the example, we assume a signal with $|\alpha|^2=0.24$ and a displacement of $\beta=1$. Results of $n=1$ are considered inconclusive results. (e) Comparison of PNR receivers with $m=1$ to $3$ (solid lines) with homodyne receiver (dashed lines) at equal success rates. PNR receiver outperforms the homodyne receiver for all signal amplitudes. The dotdashed lines show the error rate of the optimal measurement.}
\end{figure}
A very simple receiver type, which is tunable in the probability of inconclusive results, is based on homodyne detection followed by postselection~\cite{Silberhorn2002, Lorenz2004, Lance2005}. The schematic of this receiver is shown in Fig.~\ref{schemes}(a). We now revisit homodyne detection with postselection in the context of state discrimination in the aforementioned intermediate regime.
In the homodyne measurement, the local oscillator is set along the excitation of the coherent states resulting in a distribution of quadrature values as shown in Fig.~\ref{schemes}(b). The conclusion of the receiver is deduced from the particular result, where positive measurement outcomes greater than a postselection threshold $B$ identify $|\alpha\rangle$ whereas negative outcomes less that the postselection threshold $-B$ identify $|-\alpha\rangle$. All measurement outcomes between the postselection thresholds are considered as inconclusive results. The corresponding POVMs are $\hat\Pi_1=\int_{-\infty}^{-B}|x\rangle\langle x|dx$, $\hat\Pi_2=\int_B^\infty|x\rangle\langle x|dx$ and $\hat\Pi_?=\hat I-\hat\Pi_1-\hat\Pi_2$, and result in the error probability
\begin{equation}
p_{\mathrm{E,HD}}=\frac{1-\mathrm{erf}(\sqrt{2}(B+|\alpha|))}{2(1-p_\mathrm{inc,HD})}.
\label{pH}
\end{equation}
The probability of inconclusive results is found to be
\begin{equation}
p_{\mathrm{inc,HD}}=\frac{1}{2}\left(\mathrm{erf}(\sqrt{2}(B+|\alpha|))+\mathrm{erf}(\sqrt{2}(B-|\alpha|))\right).
\label{pHinc}
\end{equation}
In the following, we prove that the post-selected homodyne scheme is
the optimal strategy for realizing the intermediate measurement
within all possible Gaussian operations and conditional dynamics
(classical feedback or feedforward).
For simplicity we assume $\alpha$ is real and positive.
Note that if the input alphabet as well as all operations
are Gaussian, conditional dynamics is useless
\cite{Eisert,Fiurasek,Giedke2002}.
In our case, however, the input alphabet consists of an ensemble
of two coherent states; $p_1 |{-}\alpha\rangle\langle{-}\alpha|
+ p_2 |\alpha\rangle\langle\alpha|$.
This is clearly non-Gaussian and thus we cannot discard
conditional dynamics as a tool to improve the discrimination task.
We first briefly introduce the characteristic functional formalism
of POVMs and then discuss the POVMs via
Gaussian operations with and without conditional dynamics.
Here we use a characteristic function formalism similar to the approach used to prove the optimality of the homodyne measurement
for discriminating binary coherent states with minimum error
under Gaussian operations \cite{Takeoka2008}.
In quantum optics, the characteristic function $\chi(\omega)$ is often
a useful tool to represent a continuous variable quantum state
\cite{Walls1994}.
In an $N$-mode infinite dimensional systems, the characteristic
function of a quantum state with the density matrix $\hat{\rho}$
is defined as
\begin{equation}
\label{eq:ch_func_state}
\chi_\rho(\omega) \equiv {\rm Tr}\left[ \hat{\rho}
\exp \left[ i \omega^T \hat{R} \right] \right],
\end{equation}
where $\omega \in \mathbb{R}^{2N}$ and
$\hat{R}= [\hat{x}_1, \cdots, \hat{x}_N, \hat{p}_1, \cdots, \hat{p}_N]$.
Here $\hat{x}_i$ and $\hat{p}_i$ are the quadrature operators of
mode $i$.
In particular, a Gaussian state is defined as the state whose
characteristic function is represented by a Gaussian function \cite{Ferraro2005};
\begin{equation}
\label{eq:ch_func_Gaussian_state}
\chi_\rho(\omega) = \exp \left[ -\frac{1}{4} \omega^T \Gamma \omega
+ i D^T \omega \right],
\end{equation}
where $\Gamma$ is a $2N \times 2N$ covariance matrix
and
$D$ is a $2N$ vector corresponding to the displacement.
A similar formalism is applicable for the representation of POVMs.
A single-mode POVM consisting of any Gaussian operation,
Gaussian auxiliary states, and homodyne measurements
can be described by a set of operators
$\{ \hat{\Pi}(d_\mathcal{M}) \}_{d_\mathcal{M}}$ whose
characteristic function is \cite{Takeoka2008}
\begin{eqnarray}
\label{eq:ch_func_POVM}
\chi_d (\omega) & = & {\rm Tr}\left[ \hat{\Pi}(D)
\exp \left[ i \omega^T \hat{R} \right] \right]
\nonumber\\
& = &
\frac{1}{\pi} \exp \left[ -\frac{1}{4}
\omega^T \Gamma_\mathcal{M} \omega
+ i d_\mathcal{M}^T \omega \right] ,
\end{eqnarray}
where $\Gamma_\mathcal{M}$ is a $2\times2$ covariance matrix and
$d_\mathcal{M}=[u,v]^T$ representing the measurement outcome.
A typical example of the measurement in this class is a heterodyne measurement
described by $\{ \hat{\Pi}(\beta) = |\beta\rangle\langle\beta|/\pi
\}_{\beta \in \mathbb{C}}$ whose covariance matrix is calculated
to be an identity matrix and
$d_\mathcal{M}=[\sqrt{2}{\rm Re}\beta, \sqrt{2}{\rm Im}\beta]^T$.
Another example may be a homodyne measurement:
A homodyne measurement with the phase $\varphi=0$ is
a projection measurement onto an X-quadrature and
its covariance matrix is given by
$\Gamma_\mathcal{M} = {\rm diag}[e^{-2r}, e^{2r}]$ with $r\to \infty$
and the element in the first row of $d_\mathcal{M}$ corresponds to
the measurement outcome.
Note that the POVM in Eq.~(\ref{eq:ch_func_POVM}) does not
include conditional dynamics.
In this formalism, the probability distribution
of detecting a state $\hat{\rho}$ by a POVM $\hat{\Pi}(d_\mathcal{M})$ is
calculated as
\begin{eqnarray}
\label{eq:prob_distribution}
P(d_\mathcal{M}) & = & {\rm Tr} \left[ \hat{\rho}
\hat{\Pi}(d_\mathcal{M}) \right]
\nonumber\\
& = &
\frac{1}{2\pi} \int d\omega \chi_\rho (\omega) \chi_d (-\omega).
\end{eqnarray}
More general characteristics of Gaussian state transformations
in the formalism of characteristic function are described, e.g.
in \cite{Giedke2002,Ferraro2005}
Let us construct the intermediate measurement
via a Gaussian measurement
described in Eq.~(\ref{eq:ch_func_POVM}),
i.e. without conditional dynamics,
and classical post-processing for a set of binary coherent states
$\{ |\alpha\rangle, |{-}\alpha\rangle \}$.
When the measurement is ``noise-free'' (i.e.
consisting of Gaussian unitary operations, pure Gaussian ancillary
states, and ideal homodyne measurements)
the covariance matrix is simply given by~\cite{Takeoka2008},
\begin{eqnarray}
\label{eq:cov_matrix_Gaussian}
\Gamma_\mathcal{M} = \left[
\begin{array}{cc}
\cosh 2r - \sinh2r \cos\varphi & \sinh2r \sin\varphi \\
\sinh2r \sin\varphi & \cosh 2r + \sinh 2r \cos\varphi
\end{array}
\right] , \nonumber\\
\end{eqnarray}
where $r$ and $\varphi$ are real parameters.
This noise-free restriction does not compromise generality
since one can always construct a general Gaussian measurement
by preparing a corresponding noise-free Gaussian measurement and
discarding a part of its measurement outcomes.
We also note that such a noise-free Gaussian measurement corresponds to
$\{ \frac{1}{\pi}|\psi_\zeta (u,v)\rangle\langle\psi_\zeta (u,v)| \}_{(u,v)}$,
where $|\psi_\zeta(u,v)\rangle = \hat{D}(u,v) \hat{S}(\zeta) |0\rangle$, and
$\hat{D}(u,v) = \exp[i(v\hat{x}-u\hat{p})]$ and
$\hat{S}(\zeta) = \exp[(\zeta^* \hat{a}^2 - \zeta \hat{a}^{\dagger\,2})]$
are a displacement and a squeezing operators, respectively,
and $\zeta=re^{i\varphi}$ is a complex squeezing parameter.
The characteristic functions of the coherent states $|\pm\alpha\rangle$
are given by $\chi_\pm (\omega) = \exp [-\frac{1}{4} \omega^T I
\omega + i d_\pm^T \omega]$, where $I$ is the identity matrix and
$d_\pm = [\pm\sqrt{2}\alpha, 0]^T$.
The probability distribution of detecting $|\pm\alpha\rangle$
with such a POVM is thus calculated to be
\begin{eqnarray}
\label{eq:prob_dist}
P(u,v|\pm) & = &
\frac{1}{2\pi} \int d\omega \chi_\pm(\omega) \chi_D (-\omega)
\nonumber\\ & = &
\frac{1}{\pi \sqrt{\det (\Gamma_\mathcal{M} + I)}}
\exp\left[ -\left\{ (\pm\sqrt{2}\alpha - u)^2 a \right. \right.
\nonumber\\ && \left. \left.
- 2 (\pm\sqrt{2}\alpha - u) v c + v^2 b \right\} \right],
\end{eqnarray}
where
\begin{equation}
\label{eq:gamma_inv}
\frac{1}{\Gamma_\mathcal{M} + I} =
\left[
\begin{array}{cc}
a & c \\
c & b
\end{array}
\right],
\end{equation}
and
\begin{eqnarray}
\label{eq:abc}
a & = & \frac{1+\cosh 2r+\sinh 2r \cos\varphi}{
2(\cosh 2r+1)}, \\
b & = & \frac{1+\cosh 2r-\sinh 2r \cos\varphi}{
2(\cosh 2r+1)}, \\
c & = & \frac{-\sin 2r \sin\varphi}{
2(\cosh 2r+1)}.
\end{eqnarray}
We first show that the optimal measurement in this class
is homodyne detection.
Let us denote the likelihood ratio of two signals as
\begin{equation}
\label{eq:likelihood}
\Lambda_1 = \frac{p_1 P(u,v|{-}\alpha)}{p_2 P(u,v|\alpha)}
= \frac{p_1}{p_2} \exp \left[-4\sqrt{2} \alpha (au+cv) \right] ,
\end{equation}
and $\Lambda_2{=}\Lambda_1^{-1}$.
According to the Bayesian strategy \cite{Helstrom1976},
an optimal signal decision for the fixed measurement is to guess
$|{-}\alpha\rangle$ for $\Lambda_1 {\ge} \Lambda_B$,
$|\alpha\rangle$ for $\Lambda_2 {\ge} \Lambda_B$,
and the inconclusive result otherwise,
where $\Lambda_B$ is the threshold.
The probabilities of having successive, erroneous, and
inconclusive results for each signal are then given by
\begin{eqnarray}
\label{eq:p_succ_E_inc}
p_s^{(\pm)} & = & \frac{1}{2} {\rm erfc}\left[ \sqrt{2a}\alpha
- \frac{\ln\Lambda_B \pm \ln(p_1/p_2)}{4\sqrt{2}\alpha} \right], \\
p_e^{(\pm)} & = & \frac{1}{2} {\rm erfc}\left[ \sqrt{2a}\alpha
+ \frac{\ln\Lambda_B \mp \ln(p_1/p_2)}{4\sqrt{2}\alpha} \right], \\
p_i^{(\pm)} & = & p_s^{(\pm)} - p_e^{(\pm)} .
\end{eqnarray}
The average error and inconclusive probabilities are given by
$p_E {=} (p_1 p_e^{(-)} {+} p_2 p_e^{(+)})/(1{-}p_{\rm inc})$
and $p_{\rm inc} {=} p_1 p_i^{(-)} {+} p_2 p_i^{(+)}$, respectively.
We find that, for a given $\Lambda_B$, these two probabilities are
simultaneously minimized for $\varphi=0$ and $r\to\infty$, i.e.
an ideal homodyne measurement with phase $\varphi {=} 0$.
This implies that the optimal measurement with only Gaussian operations
is the homodyne detector with a fixed phase of $\varphi {=} 0$.
Furthermore, in the follwing we prove that any conditional operation will not improve the discrimination task. To prove this, we consider two different Gaussian operators.
The first operation is a partial measurement of the signal which in general outputs
a measurement outcome (classical number) and a conditioned output state.
For an input of
$p_1 |{-}\alpha\rangle\langle{-}\alpha| + p_2 |\alpha\rangle\langle\alpha|$,
the conditioned output is given as
\begin{equation}
\label{eq:cond_output}
\hat{\rho}_{\rm out} = p'_1(d_\mathcal{M}) \hat{\rho}_-
+ p'_2(d_\mathcal{M}) \hat{\rho}_+ ,
\end{equation}
where $\hat{\rho}_\pm$ are general multi-mode states that preserve
Gaussian statistics with the joint covariance
matrix $\Gamma_{\rm out}$ and the displacement $\pm D$.
Here $d_\mathcal{M}$ denotes the measurement outcome and thus
only the posteriori probabilities in Eq.~(\ref{eq:cond_output})
depend on $d_\mathcal{M}$.
Moreover, it was shown that $\hat{\rho}_{\rm out}$ can be always
transformed into
another mixture of coherent states \cite{Takeoka2008},
\begin{equation}
\label{eq:cond_output2}
\hat{\rho}_{\rm out} \to \hat{\rho}_{\alpha'} \otimes \hat{\rho}_{\rm aux},
\end{equation}
where $\hat{\rho}_{\rm aux}$ is some Gaussian state and
\begin{equation}
\label{eq:cond_output3}
\hat{\rho}_{\alpha'} =
p'_1(d_\mathcal{M}) |{-}\alpha'\rangle\langle{-}\alpha'| +
p'_2(d_\mathcal{M}) |\alpha'\rangle\langle\alpha'| ,
\end{equation}
with real and positive $\alpha'$.
Such an additional Gaussian operation can be deterministic and
independent of the partial measurement outcome $d_{\mathcal{M}}$.
Since only the posteriori probabilities depend on $d_{\mathcal{M}}$,
the optimal second operation
is independent of $d_\mathcal{M}$ and given by
a fixed homodyne measurement ($\varphi{=}0$) as already shown.
We therefore conclude that any conditional dynamics is not useful
in the two-step measurement scenario.
An extension of the above conclusion to the multi-step measurement scenario
is straightforward, which proves the optimality of the homodyne detector
within all possible Gaussian operations and conditional dynamics.
\section{Photon number resolving receiver}
\label{PNR}
Quadrature measurements (measurements of the light's field amplitude) and photon counting measurements (measurements of the excitation of a light field) are fundamentally different. Therefore, it is of interest to investigate also a receiver based on the latter technique. In Ref.~\cite{Wittmann2009} we proposed to use a photon number resolving (PNR) receiver for the discrimination of two coherent states. It consists of two stages: a displacement operation $D(\beta)$ and a photon number resolving detector (PNRD), and it is sketched in Fig.~\ref{schemes}(c).
The post selection process of the PNR receiver is similar to the one of the homodyne detector: If the measurement outcome of the PNR detector is $n=0$, we guess $|-\alpha\rangle$, if it is $n>m$ (where $m$ is the threshold parameter), we guess $|\alpha\rangle$ and otherwise the measurment is inconclusive. This can be described by the projector $\hat\Pi_?=\sum_{n=1}^{m}{|n\rangle\langle n|}$ for $m>0$. Conclusive results are described by $\hat\Pi_1=|0\rangle\langle 0|$ and $\hat\Pi_2=\hat I -\hat\Pi_1-\hat\Pi_?$, where $\hat\Pi_1$ identifies $|-\alpha\rangle$ and $\hat\Pi_2$ identifies $|\alpha\rangle$. An example for the photon number distributions of two displaced coherent states is shown in Fig.~\ref{schemes}(d). The error rate is then given by
\begin{equation}
p_{\mathrm{E,PNR}}=\frac{\left(1-\frac{\Gamma \left(\text{m}+1,(\alpha -\beta )^2\right)}{\Gamma (\text{m}+1)}+ e^{-(\alpha +\beta )^2}\right)}{2(1-p_{\mathrm{D,inc}})},
\label{pPNR}
\end{equation}
where the Euler gamma funtion $\Gamma(z)$ and the incomplete gamma function $\Gamma(a,z)$ are defined as $\Gamma (z)=\int _0^{\infty }t^{z-1}e^{-t}dt$ and $\Gamma (a,z)=\int _z^{\infty }t^{a-1}e^{-t}dt$. The probability of inconclusive results is given by
\begin{eqnarray}
\label{pPNRinc}
p_{\mathrm{inc,PNR}}=\frac{ \Gamma \left(m{+}1,(\alpha {-}\beta )^2\right)+ \Gamma \left(m{+}1,(\alpha {+}\beta )^2\right)}{2\Gamma (m{+}1)} \\
- \frac{1}{2} e^{-(\alpha -\beta )^2}-\frac{1}{2} e^{-(\alpha +\beta )^2}.\nonumber
\end{eqnarray}
The displacement in the receiver can be chosen such that one of two input states is displaced to the vacuum state ($\beta=\alpha$) as suggested by Kennedy~\cite{Kennedy1973}. However, to minimize the error rate of the receiver the displacement must be optimized, i.e. $\mathrm{d}p_{\mathrm{E,PNR}}/\mathrm{d}\beta = 0$. A detailed discussion of this receiver can be found in Ref.~\cite{Wittmann2009}.
We compare the PNR receiver with homodyne receivers with different postselection thresholds. In this comparison, we choose the postselection parameter $B$ such that the rates of inconclusive results are equal for both strategies, i.e. $p_\mathrm{inc,HD}=p_\mathrm{inc,PNR}$. The error probability for the receivers with $m=0$ to $2$ and the corresponding homodyne receivers are plotted against the mean photon number of the signal in Fig.~\ref{schemes}(e). We find that the performance of the PNR receiver (solid lines) surpasses the performance of the homodyne receiver (dashed lines) for all signal amplitudes. The optimal discrimination strategy is shown by the grey) curve.
\section{Experimental Results}
\label{exp}
\begin{figure}
\begin{tabular}{l}
\centerline{\includegraphics[width=7.8cm]{Setup_long.eps}}\\[-0.1cm]
\end{tabular}
\caption{\label{SetupRaw} Simplified scheme of the experiment, where the abbreviated components are a fibre mode cleaner(FMC), beam splitters (BS, 50/50), a polarizing beam splitter (PBS), a half wave plate (HWP) and a photon number resolving detector (PNRD). The graphs show modulation and the corresponding recorded quadrature measurements and detection events of the PNRD.}
\end{figure}
\begin{figure*}
\begin{tabular}{ll}
\multicolumn{2}{c}{\includegraphics[width=0.7\textwidth,keepaspectratio]{Fig3.eps}}
\\ [-12.6cm]
(a) Acceptance probability $1-p_{\mathrm{inc,HD}}$ (th.)&\hspace{0.8cm}(c) Acceptance probability $1-p_{\mathrm{inc,HD}}$ (exp.)\\ [6.1cm]
(b) Error probability $\mathrm{log}(p_{\mathrm{E,HD}})$ (th.) &\hspace{0.8cm}(d) Error probability $\mathrm{log}(p_{\mathrm{E,HD}})$ (ex.)\\ [5.7cm]
\end{tabular}
\caption{\label{HDrates} (a) The acceptance probability for ideal homodyne detection is depending on the signals mean photon number $|\alpha|^2$ and the postselection threshold $B$. A dashed line shows for which postselection threshold $B$, the homodyne detection and USD have equal acceptance probability. (b) The error probabilities of ideal homodyne detection is shown in a logarithmic contour plot. The dashed line shows the error rate of homodyne detection for success rates equal to USD. (c) Experimentally measured acceptance probability. (d) Experimentally measured error probabilities. In (c) and (d), we corrected for the quantum efficiency of the receiver.}
\end{figure*}
In the following section, the receiver schemes are demonstrated with the experimental setup shown in Fig.~\ref{SetupRaw}. It consists of a preparation stage and two different receiver stages; the PNR receiver and a homodyne receiver. Our source is a grating stabilized CW diode laser at $810\:\mathrm{nm}$ (Toptica DL100). After passing a fiber mode cleaner (FMC), the linearly polarized beam is split asymmetrically in two parts to serve as a local oscillator of the homodyne receiver (LO) and an auxiliary oscillator (AO) for state preparation and displacement in the PNR receiver scheme. The signal state (S) is generated in a polarisation mode orthogonal to the auxiliary mode using an electro-optical modulator: The field amplitude of the auxiliary mode is coherently transfered into the signal polarisation and the excitation is controlled by the input voltage of the modulator. Note that the auxiliary oscillator remains in the polarisation mode orthogonal to the signal mode thus propagating along with the signal. After splitting the signal on a 50/50 beam splitter, two identical signal states (either $|\alpha\rangle^{\otimes 2}$ or $|-\alpha\rangle^{\otimes 2}$) are produced and subsequently directed to the two detection schemes.
The signal states are generated in time windows of $\tau=800\:\mathrm{ns}$ with a repetition rate of $100\:\mathrm{kHz}$. This was done by applying a constant voltage across the modulator during the measurement time. The birefringence induced by the EOM's input voltage causes a variable coupling between the S and AO mode similar to a variable beam splitter. We can therefore tune the signal amplitude continuously.
This modulation scheme is in contrast to the commonly used sideband modulation approach in experiments with homodyne detectors, where a RF modulation is applied to the modulator to create a pair of frequency sideband modes which defines the quantum state. Since the APD is not capable of selecting a specific pair of sideband modes, such a sideband approach cannot be used when the homodyne detector is used in conjunction with an APD. The quantum states are therefore defined as a pulse in the temporal frame of the local oscillator. The exact pulses measured by the two detectors are not identical as they have different frequency response. The effect of the detector response is described after the detailed description of the detection schemes.
We carefully characterise the prepared input signal and verify that the excess noise added to the quadrature by the signal preparation is only $5\cdot10^{-3}$ shot noise units (see Ref.~\cite{Wittmann2008c}). Such has purity is achieved by attenuation of the laser (the carrier) down to the single photon level thereby minimizing the thermal fluctuations at low frequencies prevailing the diode laser.
At the homodyne receiver the signal interferes with the local oscillator, the two resulting outputs are detected and the difference current is produced. This yields an integrated quadrature value for each signal pulse. The detected signal of the homodyne detector is filtered with a 7-pole Chebyshev low pass filter from DC to 10MHz, and subsequently, the signal is sampled with 20MS/s. For a single pulse, the number of produced samples was therefore 16. These data was then averaged and thus resulting in a single quadrature measurement for a 800ns pulse. The technical noise at low frequencies was removed by correcting for the base line shift occurring between consecutive signal states. The phase of the signal relative to the LO is estimated by sending a number of bright calibration pulses along with the signal pulses. Subsequently, the measurements are accepted or discarded according to the estimated phase, i.e. they are only accepted if the measurement was performed along the signal encoding quadrature. This substitutes a technically demanding phase locking method. A drawback is the increasing measurement time. The overall quantum efficiency of the homodyne receiver amounts to $\eta_{\mathrm{hom}}=85.8\%$; the interference contrast to the local oscillator is $96.6\pm0.1\%$ and the PIN-diode quantum efficiency is $92\pm3\%$. The electronic noise level is more than $23\:\mathrm{dB}$ below the shot noise level.
The PNR receiver is composed of a displacement operation and a fiber coupled avalanche photo diode (APD) operating in an actively gated mode to circumvent a dark count event and thus a dead time at the time the pulse is arriving. During the measurement time, the APD works as a primitive photon number resolving detector, as used in~\cite{Banaszek1999}. The quantum states are subsequently categorized according to the respective result of the photon number measurement, thus implementing the POVMs $\hat\Pi_1=|0\rangle\langle 0|$, $\hat\Pi_?=\sum_{n=1}^{m}{|n\rangle\langle n|}$ and $\hat\Pi_2=\hat I -\hat\Pi_1-\hat\Pi_?$. In contrast to the displacement operation depicted in Fig.~\ref{schemes}(b) where two spatially separated modes interfere on a beam splitter, in our setup the two modes (the auxiliary and the signal modes) are in the same spatial mode but have different polarisation modes (Fig.~\ref{schemes}(c)). The interference (and thus the displacement) is therefore controlled by a modulator and a polarizing beam splitter. This method facilitates the displacement operation and yields a very high interference contrast of 99.6\%. A high extinction ratio is of high importance as the mis-matched part of the auxiliary might impinge onto the APD and cause false detection events. Such false counts can be detrimental to the discrimination task for receivers with low error rate and especially if the signal amplitude is relatively large.
The detection efficiency of the scheme is estimated to $\eta_{\mathrm{on/off}}=55\%$, including the transmission coefficient of the modulator, the polarisation optics and the fiber of $89.1\%$ as well as the quantum efficiency of the APD of $63\pm3\%$. The latter efficiency was estimated by the APD click statistic for an input coherent state that was calibrated by the homodyne receiver. An optical isolator is used between the two detection schemes to prevent back scattering of the LO to the APD.
\begin{figure}%
\includegraphics[width=0.85\columnwidth]{modeoverlap.eps}%
\vspace{-0.3cm}
\caption{\label{modeoverlap}Comparison between the effective and ideal response function of the homodyne receiver}%
\end{figure}
As the two detectors have different response, they do not measure the exact same temporal mode. However, in the following we show that the two different temporal modes are nearly identical possessing an overlap of about 95\%. In the experiment, we probe the optical mode $\hat a(t_m) =1/T\int_{t_m-T/2}^{t_m+T/2} \hat a(t) \mathrm{d}t$. The APD is a broadband detector, and by neglecting the electronic jitter noise and the dead time, the APD detects the following mean photon number during the measurement time T,
\begin{equation}
\hat n(t_m) = \frac{1}{T}\int_{t_m-T/2}^{t_m+T/2} \hat a^\dagger(t) \hat a(t) \mathrm{d}t
\label{eq:APDmode}
\end{equation}
On the other hand, the homodyne detector has a finite detector bandwidth which means that the quadrature measurement, $\hat X_{sa}$, at time $t$ depends on the detector response $G(\tau)$. A single sample is thus described by~\cite{Ou1995}
\begin{equation}
\hat X_\mathrm{sa}(t) = \int \mathrm{d}\tau G(\tau) \hat X(t-\tau)
\label{eq:HDmode}
\end{equation}
where $G(\tau)$ is determined by the detectors frequency response $G(\omega)$ and accounts for the mean of $\hat X_{sa}$ over the measurement time T. The time-averaged measurement $\hat X_{av}$ at the time $t_m$ can then be written as
\begin{eqnarray}
\hat X_\mathrm{av}(t_m)&=& \frac{1}{T}\int_{t_m-T/2}^{t_m+T/2}\mathrm{d}t \int \mathrm{d}\tau G(\tau) \hat X(t-\tau)\\ \nonumber
&=& \frac{1}{T}\iint\mathrm{d}t \mathrm{d}\tau \mathrm{rect}(\frac{t-t_m}{T}) G(t-\tau) \hat X(\tau)\\ \nonumber
&=& \frac{1}{T}\int\mathrm{d}\tau\hat X(\tau)\int\mathrm{d}t \mathrm{rect}(\frac{t-t_m}{T}) G(t-\tau) \\ \nonumber
&\mathrel{\widehat{=}}&\int \mathrm{d}\tau G_\mathrm{eff}(t_m-\tau) \hat X(\tau)
\label{eq:effG}
\end{eqnarray}
where the measurement is described by the effective response function $G_\mathrm{eff}$. We compare this function to the ideal mode $\hat X_\mathrm{ideal}(t_m)= \frac{1}{T}\int_{t_m-T/2}^{t_m+T/2}\mathrm{d}t \hat X(t)$ as illustrated in Fig.~\ref{modeoverlap}. We estimate the mode overlap of the ideal and the effective mode with the normalized cross-correlation function $g_{12}=\langle \hat X_\mathrm{av}\hat X_\mathrm{ideal}\rangle/\sqrt{|\hat X_\mathrm{av}|^2|\hat X_\mathrm{ideal}|^2}$ (which is related to the visibility for the interference of two partially coherent waves~\cite{Saleh2007}). For our detector the cross-correlation is 95.5\,\%, and thus the similarity between the temporal mode seen by the APD and the one seen by the homodyne detector is about 95\%.
\begin{figure}
\begin{tabular}{l}
\centerline{\includegraphics[width=8cm]{Fig4_long.eps}} \\ [-18cm]
(a)\\[5.7cm]
(b)\\ [5.7cm]
(c)\\ [5.4cm]
\end{tabular}
\caption{\label{Drates} (a) Experimental data on the effect of the displacement $\beta$ on the error rate $p_\mathrm{E,PNR}$ for a given signal amplitude $|\alpha|^2=0.24$ (corrected for quantum efficiency) and different number of dropped results $m$. Theoretically predicted error probabilities for the receiver without imperfections are shown with dashed lines. Error rates at the optimal displacement $\beta_\mathrm{opt}$ for different receivers are marked. (b) Error rates for varying $m$ and optimized $\beta_\mathrm{opt}$. Error bars reflect the standard deviations of repeated measurements, which are larger than the statistical errors. Experimental data is compared to ideal receivers (dashed lines). (c) Experimental data for acceptance rates(points) and theoretical predictions (dashed lines). Below the curve for an optimal USD device (grey), states can be discriminated without error in principle. We corrected for the quantum efficiencies of the receiver.}
\end{figure}
We proceed by describing the procedure of the discrimination task. A PC controls the preparation of the states and the displacement in the PNR receiver by modulating two electro optical modulators. Simultaneously it acquires the homodyne and APD detection outcomes during the pulse sequence. An example of such a sequence is shown in Fig.~\ref{SetupRaw}, where we show the voltages applied to the amplifiers, the quadrature values and the recorded number of counts per measurement time. The outcomes of the homodyne receiver within the interval $[-B,B]$ are considered as inconclusive results. If the value is outside the interval and positive we guess $|\alpha\rangle$ and if the value is outside the interval and negative we guess $|-\alpha\rangle$. For the outcomes of the PNR receiver, we use the hypothesis that if the outcome is larger than $m$, we guess $|\alpha \rangle$, if it is zero, we guess $|-\alpha\rangle$, otherwise it is an inconclusive result. The error probability is therefore found by adding up all the false detections and relate it to the number of pulses that were accepted. The acceptance probability is the ratio of pulses that were accepted to the total number of pulses.
The theoretical prediction for the acceptance probability $1-p_\mathrm{inc,HD}$ and the error probability $p_\mathrm{E,HD}$ are shown in Fig.~\ref{HDrates}(a) and (b) respectively. It is shown, that for increasing signal amplitudes $|\alpha|^2$ the error probability $p_\mathrm{E,HD}$ drops and that an increase of the postselection threshold $B$ leads to a decreasing error probability on the expense of an increase in the probability for inconclusive results. An advantage of the homodyne receiver is the smooth dependence between postselection threshold and error rate. This allows one to chose exactly the error rate desired for a specific application.
For example in quantum key distribution the amount of mutual information between sender, adversary and receiver can be easily adjusted through the postselection threshold~\cite{Silberhorn2002}. The receivers performance is completely characterized by the error and acceptance rates. In the figures, we introduced a red dashed line, where the condition $p_{\mathrm{inc,HD}}=p_{\mathrm{inc,USD}}$ is met, with the probability of inconclusive results in a perfect USD measurement $p_{\mathrm{inc,USD}}$. This means, an error-free but probabilistic discrimination is in principle possible above this curve.
The experimental results for the acceptance and the error probability of the homodyne receiver are shown in Fig~\ref{HDrates}(b) and (d) respectively. The contour plots are generated from signals with 21 different amplitudes (with linearly increasing the mean photon number) and calculated for 41 postselection thresholds. We find very good agreement of theory and experiment with only minor deviations for very low error probabilities.
The PNR receiver is demonstrated for $m=0$, $1$ and $2$. In Fig.~\ref{Drates}(a), the dependence of the error probability on the displacement $\beta$ for a fixed signal amplitude is illustrated. We find that for any $m$ the displacement can be optimized such that the experimentally measured error rates reach a minimum. The optimal displacement is higher for higher $m$ and the minimum error rate after this optimization of the displacement is lowered for increasing $m$. When compared to the theoretical predictions, the experimental data fits well in the region of the minima, while the experimental imperfections dominate in the region of smaller displacement. We could also observe this for $m>2$.
We marked four data points in the plot. From left to right, they represent the error rates associated with the Kennedy receiver (black) (which is an early receiver for the minimum error discrimination~\cite{Kennedy1973} without optimized displacement, i.e. $\beta=\alpha$), the optimized displacement receiver with $m=0$ (red), and the PNR receivers with $m=1$ and $2$ (green and blue). The error rates for varying amplitudes are plotted in Fig.~\ref{Drates}(b). We find a maximal reduction of the error rate by a factor of $3.5$ going from $m=0$ (deterministic scheme) to $m=2$ (probabilistic scheme) at the signal amplitude $|\alpha|^2=0.47$. The corresponding penalty on the acceptance rates and the comparison with the theoretical predictions for the acceptance probability are shown in Fig.~\ref{Drates}(c).
Both detection schemes are compared to each other in Fig.~\ref{comprates}. We find that both receivers show the expected behavior. Especially for $m=1$, it is obvious that the PNR receiver outperforms the homodyne receiver for several data points.
\begin{figure}
\begin{tabular}{l}
\centerline{\includegraphics[width=8cm]{ComparisonHDPNRexp.eps}}\\
\centerline{\includegraphics[width=8cm]{improve_comparison.eps}}\\[-11.3cm]
(a)\\[5.9cm]
(b)\\ [4.5cm]
\end{tabular}
\caption{\label{comprates} (a) Experimental error rates and acceptance rates for the two receiver schemes are compared. For this comparison the success rate of both schemes is fixed to the one, that is theoretically reached by the PNR receiver. Experimental data is shown for PNR receivers with $m=1$ and $2$ (filled circles and triangles) and homodyne receiver (open circles and triangles). Additionally, theoretical predictions for the homodyne receiver (grey dashed line), the PNR receiver (solid line) and the optimal intermediate measurement (dotdashed lines) are shown. The mean photon number is varied for the different receivers. The PNR receiver again outperforms the homodyne receiver and we find a relatively good agreement of experimental data and theoretical predictions. (b) Error rate for various signal amplitudes. The PNR receiver surpasses the homodyne receiver. Statistical error bars show standard deviation of the random process. We corrected for the quantum efficiencies of the receivers.}
\end{figure}
In the following, we discuss the limitations of the different schemes. The quantum efficiency of homodyne detection is partly limited by the PIN-diode efficiency and partly by the mode-matching efficiency at the homodyne's beam splitter. For special made PIN diodes, the efficiency can reach nearly 100\%, and the beam splitter mode matching efficiencies beyond 99\% have been reported. The efficiency of the PNR is mainly limited by the quality of the displacement operation and the efficiency of the avalanche photo diode (APD). For higher values of m, we also find that also the probability of false detection events becomes important (see $m {=} 2$ in Fig.~\ref{comprates}(a)). We used a commercial available APD, but the development of photon number resolving detectors with very high quantum efficiency is rapidly progressing(see~\cite{Worsley2009} for a detailed list).
\section{Conclusion}
\label{comp}
In this paper we have thoroughly investigated two different probabilistic receivers of binary-encoded optical coherent states; the homodyne detector and the displacement controlled photon number resolving detector. These receivers yield inconclusive as well as error affected results and we have carefully conducted a detailed study of the relation between these two outcomes. Furthermore, we found, theoretically, that the homodyne detector is the optimal Gaussian receiver for minimizing the errors for a fixed probability of inconclusive results. Experimentally, we have implemented both receivers and through comparison we found that the performance of the new displacement controlled PNR is better than the homodyne receiver.
The new receiver is thus a promising alternative to the commonly used homodyne receiver. We find several advantages of the new scheme. For example, the power of the displacement beam is normally much lower than the power of the local oscillator beam required for homodyne detection, and thus less power is injected into the communication channel (e.g. an optical fibre). This has the obvious benefit of lowering the power consumption in the fibre but it also lowers the risk of scattering of auiliary photons into the signal state as such scattering mechanism is proportional to the power. We also note that instead of performing the displacement operation at receiving station, it can be already implemented at the sending station. This would completely remove the necessity of a phase reference. Finally, we note that the quadrature measurement can be also performed using a displacement operation followed by a single intensity detector similar to the setup of the PNR detector. However, in the former case the displacement must be macroscopic such that the quadratures are measured instead of the photon properties.
The new receiver is fundamentally different from the more commonly used and technically mature homodyne detector. Whether the PNR detector will be the future choice in real life implementations of binary detectors will depend on the future technical progress of the PNR detector technology. A future option is also to use both detector schemes in a receiving station, where the proper detection scheme is chosen according to the currently needed property, such as speed, low noise and the capability of performing a full tomography of the state~\cite{Qi2007d}.
The work has been supported by the EU project COMPAS and Lundbeckfonden(R13-A1274).
|
\section{Introduction}
A driven lattice gas, or driven diffusive system,
is a system of interacting particles that jump in a preferred direction on
a lattice.
The system cannot reach equilibrium but may attain a stationary
state with a steady current;
the model is a prototype for studies of nonequilibrium states
\cite{Schmittmann,Marro,Katz,Leung}. The simplest example of a driven
diffusive system, which has become one of the standard models of nonequilibrium statistical
mechanics, is the totally asymmetric simple exclusion process (TASEP)
\cite{MacDonald-Gibbs2,Andjel,Harris,Sptizer}.
In the TASEP with open boundaries
the edge sites are connected to particle reservoirs with fixed densities.
Introduced as a model of biopolymerization
\cite{MacDonald-Gibbs} and transport across membranes \cite{Heckmann}, over the years, this
model has been applied to other processes, e.g., traffic flow
\cite{Schadschneider}, and cellular transport \cite{Heijne,Kukla}.
From a mathematical point of view the model is of interest in the theory of interacting
particle systems since, despite its simplicity, it shows a nontrivial
behavior \cite{Harris,Sptizer,Ligget85,Spohn}. In the one-dimensional TASEP with open
boundaries, particles jump only to the right, along a
one-dimensional lattice whose sites can be empty or occupied by a single particle.
Particles are injected at the leftmost site
at rate $\alpha$ if this site is empty, and removed at the rightmost site
at rate $\beta$ if this site is occupied. The one-dimensional TASEP, which has been solved
exactly, exhibits three distinct
phases in the $\alpha$-$\beta$ plane
\cite{SchutzDomany,Derrida93i,Derrida96,Derrida98}.
The phase transition is discontinuous along the
line $\alpha=\beta<1/2$, where the density profile is linear, and continuous along the
lines $\alpha=1/2$, $\alpha>\beta$ and $\beta=1/2$, $\alpha<\beta$ (see Fig.~\ref{PhaDia}).
Although this model and and variants have been the subject of intensive theoretical
study, there is as yet no realization of a TASEP-like system in the laboratory.
The invention of optical tweezer arrays has permitted investigation of
the dynamics of colloidal particles in an external periodic potential
\cite{Sancho,Korda,cellsorting,Chiou,Lacasta,Roichman}. The motion at
long times and low friction consists of jumps between adjacent
potential minima. If the particle and potential well sizes are chosen properly, only
one particle can occupy a given well. The exclusion process is a caricature of this dynamics,
suggesting that a system of colloidal
particles in an optical tweezer array could be designed as a laboratory realization of the TASEP.
Motivated by this possibility, we propose a model in continuous space having the
same essential characteristics as the lattice TASEP.
Our model represents colloidal particles immersed in a fluid flowing at constant rate
through a one-dimensional optical tweezer array, restricting particle motion to the array axis.
Specifically, we study a one-dimensional system of interacting Brownian particles
subject to a periodic external potential,
and to a constant external force representing the drag due to
the fluid motion. A short-range (essentially hard-sphere) repulsion between particles
prevents more than one particle occupying the same potential well.
This continuous-space model is studied via numerical integration of Langevin equation
and dynamic Monte Carlo simulation.
We observe phase transitions similar to those
found in the lattice TASEP. Some differences in the detailed behavior nevertheless appear,
due to the lack of particle-hole symmetry in the continuous-space model.
Details on the model and simulation methods are given in Sec.~\ref{LAN} and \ref{MC}.
Simulation results are presented in Sec.~\ref{RES},
while our conclusions and prospects for future work are outlined in Sec.~\ref{CON}.
\section{Continuous-Space Model}\label{LAN}
Our aim is to study a continuous-space model sharing the same essential features
as the TASEP (defined on a lattice), as a first step toward
experimental realization of a TASEP-like system.
The model should possess the following characteristics: i) confinement of particles
to a one-dimensional structure; ii) localization of particles at potential minima (``wells") of a
linear periodic array with iii) multiple occupancy prohibited; iv) biased hopping between
adjacent wells; v) insertion (removal) of particles at the initial (final) well.
Criteria i)-iii) are realizable with a suitably tailored optical tweezer array.
The array consists of a series of spherically symmetric optical tweezers;
the particles flow along this line, which we take as the $x$ axis.
To avoid particles escaping the array,
there should be a substantial overlap between neighboring wells,
so that a potential maximum
at a point midway between two wells is in fact a saddle point in the full three-dimensional
space (see Fig.~\ref{two_gau_inv}).
For a TASEP-like
system, it is crucial that the probability of a particle escape from the array be
negligible on the time-scale of the experiment.
Effective confinement requires a potential barrier
to escape the array much larger than $k_BT$; the barrier
between adjacent minima should be smaller, to
allow transitions between
neighboring wells.
In what follows we shall assume this condition is satisfied, and consider,
for simplicity, a one-dimensional system. Fluctuations of the particle positions
in the directions perpendicular to the array will therefore be ignored, but should be
included in a more complete analysis.
The diameter of the optical tweezer well should be slightly greater than the
particle diameter, so that at most one particle can occupy the well at a given time.
Due to thermal fluctuations,
particles can occasionally overcome the potential barrier
separating neighboring wells. Since hopping must be
asymmetric, we impose a
steady fluid motion along the $+x$ axis, which effectively prohibits particle jumps in
the opposite direction.
In the lattice TASEP particles are inserted in the first site and removed from the last.
Experimental realization of this feature is subtler, but can in principle be achieved
with the help of
optical tweezers at the beginning and end of the array, which drag particles
into the first well and out of the last one at prescribed rates. We discuss
an alternative method of insertion and removal in Sec.~\ref{CON}.
The above sketch of an experimental setup motivates our study
of Brownian motion of interacting colloidal
particles in an optical tweezer array.
The Langevin equation for the $i$-$th$ particle is
\begin{equation}\label{ELUP}
m\ddot{x}_i(t) = -b \left[ \dot{x}_i(t) - v\right] - \frac{\partial V_{ext}(x_i)}{\partial x_i}
- \frac{\partial}{\partial x_i} \left[ V_{int}(x_{i,i-1}) + V_{int}(x_{i,i+1})\right] + m\Gamma_i(t),
\end{equation}
\noindent where $x_i$, $\dot{x}_i$ and $\ddot{x}_i$ are, respectively, the position, velocity and acceleration
of particle~$i$, $x_{i,j} \equiv x_j - x_i$, and $v$ is the terminal velocity of a particle
in the moving fluid, in the absence of the periodic external potential $V_{ext}$.
$V_{int}$ is the (strongly repulsive) potential between neighboring
particles.
(We assume that the range of $V_{int}$ to be short enough that only neighboring particles interact.)
The first term on the right side of Eq.~(\ref{ELUP}) represents damping of the particle
velocity relative to the fluid. For a sphere of radius $R$, Stokes' Law gives:
\begin{equation}\label{EscStokes}
b = 6\pi\mu R,
\end{equation}
where $\mu$ is the fluid viscosity.
Here it is important
to stress that the fluid is three dimensional although we treat the particle motion
as one dimensional. The final term is a random noise
with the following properties:
\begin{eqnarray}\label{PRA}
\langle\Gamma_i\rangle &=& 0, \\
\langle\Gamma_i(t)\Gamma_j(t')\rangle &=& \frac{2b k_B T}{m^2}\delta_{i,j}\delta(t-t'),
\end{eqnarray}
\noindent where $k_B = 1.3806504\times10^{-23}J/K$ is Boltzmann's constant, $T$ is temperature,
and $m$ is the particle mass.
It is convenient
to ignore the inertial term $m\ddot{x}_i(t)$ in Eq.~(\ref{ELUP}),
since the observational times of interest (microseconds or greater) are
much larger than the
relaxation time of the velocity, $m/b\sim10^{-8}\,$s, for our choice of parameters.
Then the velocity of particle $i$ follows,
\begin{eqnarray}\label{ELUP2}
\dot{x}_i(t) = v - \frac{1}{b}\frac{\partial V_{ext}(x_i)}{\partial x_i} -
\frac{1}{b}\frac{\partial}{\partial x_i} \left[ V_{int}(x_{i,i-1}) + V_{int}(x_{i,i+1})\right]
+ \frac{m}{b}\Gamma_i(t).
\end{eqnarray}
\subsection{Potentials}
The potential of an optical tweezer array can be represented by a sum of $n$ identical
Gaussian profiles of width $\sigma$ and spatial period $d$:
\begin{equation}
V_t (x) = - V_0 \sum_{n=0}^N e^{-(x-dn)^2/2\sigma^2}.
\end{equation}
We require neighboring wells to overlap, which can be accomplished
by setting $\sigma \stackrel < \sim d$; Figure~\ref{two_gau_inv} shows that
this condition is satisfied for
$\sigma=d/4$.
For this choice of parameters, the
potential is well approximated by a cosine, as can be seen from the
Fourier coefficients
\begin{equation}\label{CoeFou}
a_n = \frac{2}{d}\int\limits_0^d\cos\left(\frac{2n\pi x}{d}\right)
\left[ -\sum\limits_{n=0}^N e^{-(x-dn)^2/2\sigma^2}\right] dx.
\end{equation}
\begin{figure}[htb]
\begin{center}
\subfigure[\label{two_gau_inv}]{\includegraphics[height = 4.6cm]{two_gauss}}
\subfigure[\label{pot_efe}]{\includegraphics[height = 4.6cm]{pot_efet}}
\end{center}
\caption{(a) Two neighboring Gaussian profiles (red and black curves) and their sum (green);
(b) effective external potential.}
\end{figure}
\noindent Numerical evaluation yields
\begin{equation}\label{CoeFouNum}
\begin{array}{l}
a_0 = -0.627\\
a_1 = 0.365\\
a_2 = 9.010\times10^{-3}.
\end{array}
\end{equation}
Note that $a_2/a_1 \simeq 0.025$, allowing us to write,
to a good approximation, the array potential as a constant plus a
cosine term. For the one-dimensional model studied here, we define
\begin{equation}\label{EnePotExt}
V_{ext}(x) = -V_0 \cos\left(\pi \frac{x}{d_0}\right)
\end{equation}
as the periodic external potential.
The interaction between
colloidal particles is taken as purely repulsive; for convenience we use a truncated
$1/r^{12}$ potential:
\begin{equation}\label{EnePotInt}
V_{int}(r) = \left\lbrace \begin{array}{cc}
\displaystyle U_0\left[\left( \frac{a}{r}\right) ^{12} -1\right] , & r\leq a\\
0, & r\geq a,
\end{array}\right.
\end{equation}
where $a$ is the particle diameter and $r$ distance between neighboring particles.
\subsection{Parameter Values}
To specify the external and interaction potentials, we need to fix
$V_0$ and $U_0$. These values must be chosen so as to approximate the TASEP
dynamics given the length, time and energy scales characterizing the system.
We measure lengths in units of microns, time in seconds and energies in units of
$k_BT = 4.1419512\times10^{-21} $J, assuming a temperature of 300K.
We set $d_0=1$ (so that the period of the external potential is $2$ microns),
and take the particle diameter as $a=1.8$, so that a pair of particles occupying neighboring wells have
some freedom to fluctuate about the potential minimum. Taking the fluid as water
(with viscosity $\mu = 0.01$g/cm$\cdot$s) we fix the friction coefficient as
$b\cong4.096\;k_BT\cdot $s/($\mu$m)$^2$.
To determine the fluid velocity $v$ and external-potential intensity $V_0$ we examine
the effective external potential, defined as
\begin{equation}\label{PotExtEfe}
V_{eff} = -V_0\cos\left(\pi \frac{x}{d_0}\right)- b v x,
\end{equation}
\noindent i.e., the sum of external periodic potential and a fictitious potential representing
the constant friction force acting on particles (see Fig.~\ref{pot_efe}). If $b v d_0 \gg V_0$, particles
do not feel the periodic potential, while if $b v d_0 \ll V_0$ particle
hopping is essentially unbiased. Let $\Delta E_G$ ($\Delta E_P$)
denote the difference between a given maximum of $V_{eff}$ and the first minimum to
the right (left) of this maximum. In this way, $\Delta E_G$ ($\Delta E_P$) is the
potential barrier separating a given potential minimum from its left (right)
neighbor. A simple calculation shows that
\begin{equation}\label{DeltaE}
\begin{array}{l}
\Delta E_G - \Delta E_P = 2 b v d_0 \,,\\
\Delta E_G + \Delta E_P = 4V_0\cos(\pi x_0/d_0) + 4b v x_0 \,,
\end{array}
\end{equation}
with
\begin{equation}\label{x0}
x_0 = \frac{d_0}{\pi} \sin^{-1}\left( \frac{b v d_0}{\pi V_0}\right)\,,
\end{equation}
such that the potential minima occur at $x_j = 2jd_0+x_0$, and maxima at $y_j = 2jd_0+1-x_0$ for
$j$ an integer.
Given $\Delta E_G$ and $\Delta E_P$, we can determine $x_0$,
$v$ and $V_0$. We take $\Delta E_P \sim k_BT$, and $\Delta E_G \gg k_B T$, so that
a particle has a finite rate of jumping to the well on the right, and virtually
no chance of jumping in the opposite direction.
A good correspondence with lattice TASEP is obtained using
$\Delta E_G = 20k_BT$ and $\Delta E_P = 2k_BT$.
For these values we have,
\begin{equation}\label{V_0vx_0}
\begin{array}{l}
v = 2.1974\;\mu\mbox{m/s}\\
V_0 = 4.5680 \;k_BT\\
x_0 = 0.2158 \;\mu\mbox{m}.
\end{array}
\end{equation}
\noindent To maintain the interparticle repulsion in the presence of the
external potential, we must take $U_0$ substantially greater than $V_0$.
On the other hand, very large values of $U_0$ are inconvenient for
numerical integration of the Langevin equation, as a very small time increment
would be required to avoid spurious particle displacements.
We therefore use $U_0 = 40k_BT \simeq 9 V_0$.
\subsection{Single-particle dynamics}
To begin, we consider a single Brownian particle moving in the
system. Figure~\ref{position-time} shows a typical evolution of the particle position
over time. The graph exhibits plateaux whose size corresponds
to the time a particle stays in a given well. Studies of this kind allow us to determine
the mean transition time $\tau$ between neighboring wells as 6.5380(9)s. This quantity is needed in order to
define the insertion and removal rates. (Recall that in
the lattice TASEP these rates are defined in units of the hopping rate).
If the first well is empty
we insert a particle there (at the potential minimum position), at rate $\alpha/\tau$;
if the last well is occupied, we remove
the particle at rate $\beta/\tau$.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{traje}
\end{center}
\caption{Typical particle trajectory.}
\label{position-time}
\end{figure}
\section{Dynamic Monte Carlo Simulations}\label{MC}
The Langevin simulation outlined in the preceding section is valuable for fixing the
time scale of hopping between wells, and for confirming the basic phenomenology of the
model. It is, however, rather inefficient numerically, so that it is desirable to implement
a dynamic Monte Carlo (MC) simulation for large-scale studies.
We apply the Metropolis algorithm to evolve the particle positions
in time.
In this approach we use the following expression for the potential energy:
\begin{equation}\label{EneTotSis}
E = -\sum_{i=1}^N [V_0\cos(\pi x_i) + b v x_i ] + \sum_{i=1}^{N-1} V_{int}(x_{i+1}-x_i) ,
\end{equation}
where $V_{int}$ is given by Eq.~(\ref{EnePotInt}). The drag force due to the moving fluid
is represented by the effective potential $-b v x$.
In the MC dynamics,
a trial configuration is generated by selecting one of the $N$ particles at random
and subjecting it to a random displacement $\Delta x$, chosen from
a Gaussian distribution with mean zero and standard deviation $\sigma~=~0.2\mu$m.
This value is $10\%$ of the well size,
large enough to afford a substantial speedup, but small enough that the probability of a
particle displacement greater than $2\mu$m is negligible.
As is usual in Metropolis MC, trial moves such that the change in energy
$\Delta E \leq 0$ are always accepted,
while for $\Delta E > 0$ the trial move is accepted with probability $e^{-\Delta E/k_B T}$.
We determine the mean
number of Monte Carlo steps required for a particle move from one well to its
neighbor on the right as
$N_{MC} = 120.917(1)$.
Thus the time per Monte Carlo step is
\begin{equation}\label{tc}
\tau_0 = \frac{\tau}{N_{MC}} = 5.4070(8)\times10^{-2}\mbox{s}.
\end{equation}
\noindent Particle insertion and removal are done as in the Langevin simulations.
We verify below
that this method is equivalent to the latter approach; the MC algorithm is
$\sim \! 1000$ faster than numerical integration of the Langevin equation,
for L=100.
\vspace{1cm}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{PerDenLanb=04}
\end{center}
\caption{Density profile for $\beta = 0.4$ obtained via numerical integration of the
Langevin equation, for $\alpha$ values
as indicated, $L=100$.}
\label{profile}
\end{figure}
\section{Results}\label{RES}
During the simulations we monitor the mean occupation probability $\rho(i)$ at each well $i$,
and the current $J$,
measured by the mean number of particles leaving the system per unit time.
Examples of density profiles in the stationary regime
($\rho(i)$ versus $i$), are shown in
Fig.~\ref{profile} for several values
of $\alpha$, with $\beta = 0.4$.
(From here on $\alpha$ and $\beta$ are given in units of $1/\tau$, where $\tau$ is the
mean time required for hopping between wells.)
This figure shows that the continuous-space model exhibits the same basic phenomenology
as the lattice TASEP. For
$\alpha \leq 0.4$ the overall density grows with $\alpha$. On increasing
$\alpha$ from 0.4 to 0.5 there is a marked increase in density, but for further increases
the density changes very little.
While the Langevin simulation (LS) results already suggest that the model exhibits
phase transitions, we shall use the more precise results
of our MC simulations to perform a detailed analysis. Before proceeding, we
verify that the MC method yields results in agreement with the LS. In Fig.~\ref{MCLAComp}
we compare density profiles obtained via
LS and MC for the same values of $\alpha$ and $\beta$. The bulk densities
obtained using the two methods differ by $\le 1.6 \%$. Thus the MC method captures the behavior
found using the
Langevin equation to good precision.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{ComMC_LA_b=04_2}
\end{center}
\caption{Comparison between profiles obtained using LS (open symbols) and
MC (filled symbols), $\alpha=0.3$ (lower curves), $\alpha=0.8$ (upper curves), $\beta=0.4$ , $L=100$.}
\label{MCLAComp}
\end{figure}
We perform MC simulations of systems of $L=100$, 200 and 500 wells.
Far from the phase transition, density profiles depend only weakly
on system size, but near the transition there are significant
finite-size effects, as illustrated in Fig.~\ref{SizPro}. In this case, as $L$
increases, the profile tends to a near-constant value except for a
sharp increase near the exit.
\vspace{2em}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{ComPerSisTamDir3}
\end{center}
\caption{Density profiles for $\alpha = 0.7$ and $\beta = 0.6$, for systems of
$100$, $200$, and $500$ wells.} \label{SizPro}
\end{figure}
To minimize boundary effects we study the {\it bulk} density
$\rho$, defined as the mean density over the $10\%$ of sites nearest
the center:
\begin{equation}\label{density}
\rho = \frac{10}{L}\sum_{i=0.46L}^{0.55L}\rho(i).
\end{equation}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{DenComTam}
\end{center}
\caption{Bulk density versus $\alpha$ for $\beta = 0.4$, system
sizes $L=100$, $200$ and $500$. } \label{SysDenDifSiz}
\end{figure}
\noindent The bulk density as a function of $\alpha$, for $\beta = 0.4$, is
shown in Fig.~\ref{SysDenDifSiz} for the three system sizes studied. These
results strongly suggest the development of a discontinuity in
$\rho(\alpha)$ near $\alpha = 0.47$ as the system size is increased.
\begin{figure}[htb]
\center
\subfigure[\label{MosImpRes_a}]{\includegraphics[height = 4.6cm]{DenAlfFix2}}
\subfigure[\label{MosImpRes_b}]{\includegraphics[height = 4.6cm]{DenBetFix2}}
\\ \vspace{0.5cm}
\subfigure[\label{MosImpRes_c}]{\includegraphics[height = 4.6cm]{CorAlfFix2}}
\subfigure[\label{MosImpRes_d}]{\includegraphics[height = 4.6cm]{CorBetFix2}}
\caption
{Upper panels: bulk density for (a) fixed $\alpha$ and (b) fixed $\beta$. Lower panels:
current for (c) fixed $\alpha$ and (d) fixed $\beta$. System size $L = 200$.
}
\label{MosImpRes}
\end{figure}
\subsection{Phase diagram}
Of principal interest in determining the phase diagram
are the bulk density and the current as functions of the
rates $\alpha$ and $\beta$. These results are summarized in
Fig.~\ref{MosImpRes}, showing evidence of both continuous and discontinuous phase
transitions, depending on the rates.
We see that for low $\alpha$ ($\alpha<0.8$ or so) and $\beta<\alpha$
the system is in the high-density phase, in which density and current depend only on $\beta$,
whereas for $\beta > \alpha$, and $\beta < 0.6$ or so, the system is in the low-density phase in which
density and current depend only on $\alpha$. For larger values
($\alpha>0.8$ and $\beta>0.6$), the system is in the maximum-current phase, in which density and
current are independent of both $\alpha$ and $\beta$, and the current takes its maximum value.
Thus the continuous-space model exhibits the same
three phases observed in lattice model. As in the lattice model, the transition
between the low- and high-density phases is discontinuous, whereas transitions between
the maximum-current phase and the other phases are continuous.
(While the density is discontinuous in the former case, the current is always
continuous at the transition.)
We adopted the following procedure (``polynomial method") to determine the values of
$(\alpha,\beta)$ along the discontinuous
transition line. Consider the case of fixed $\alpha$.
For small $\beta$, the stationary current depends only on $\beta$ (see Fig.~\ref{MosImpRes_c}).
We therefore fit a polynomial $P(\beta)$ to the current in this regime, using data for
a large, fixed $\alpha$ (in practice, $\alpha \geq 1.4$).
For larger values of $\beta$, the current depends only on $\alpha$;
in Fig.~\ref{MosImpRes_c} this regime corresponds to one of the plateaux, $J = J^* (\alpha)$. The transition
point $\beta_c (\alpha)$ is taken as the value at which the plateau intersects the polynomial
fit to the small-$\beta$ data, i.e., $P(\beta_c) = J^* (\alpha)$.
Determination of $\alpha_c (\beta)$ follows an analogous procedure,
in which we fit a polynomial to the current data for small $\alpha$.
We find that a quadratic polynomial is sufficient to fit (to within uncertainty) the current
$j(\beta)$ at fixed $\alpha$, while a good fit of $j(\alpha)$ (at fixed $\beta$) requires a
quartic polynomial. In both cases the constant term in the polynomial is zero, since
the current vanishes for $\alpha$ and/or $\beta$ zero.
The above method is quite effective in locating points along the discontinuous transition line.
Although it can in principle be used to locate continuous transitions as well,
we found that in this case the estimates for $\alpha_c$ and $\beta_c$
are rather sensitive to one's choice of the range of values fit using the polynomial.
We found the following
approach (``derivative method") to be more useful for continuous transition points.
For fixed $\alpha$, we estimate the derivative $dJ/d\beta$ using a spline fit
(see Fig.~\ref{DC}). The derivative decreases in a linear fashion with increasing $\beta$,
except for a small roundoff region that we interpret as a finite-size effect. We then estimate
$\beta_c$ as the point where $d J/d \beta$ falls to zero, using the data in the linear region.
The procedure for fixed $\beta$ is analogous.
The continuous transition points obtained by this procedure are quite robust with
respect to changes in the region analyzed, as long as we exclude the roundoff region.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{der_J_beta_l200}
\end{center}
\caption{Current derivative for $\alpha=2.0$ and $L=200$.}
\label{DC}
\end{figure}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{diatasepcont}
\end{center}
\caption{Phase diagram of the TASEP on the lattice (dotted lines) and
in continuous-space (solid lines)
for L=500. The grey triangle denotes a region in which we are unable to
determine the phase boundaries precisely.}
\label{PhaDia}
\end{figure}
Using the methods described above we construct a phase diagram based on the data for each system size
studied; that for $L=500$ is shown in
Fig.~\ref{PhaDia}. The vertical boundary
at $\alpha=0.905(2)$, and the horizontal boundary at $\beta=0.714(1)$ are obtained
using the derivative method. (The polynomial method yields $\alpha = 0.8(1)$ and $\beta = 0.67(5)$,
respectively, for these boundaries.)
The grey triangle in Fig.~\ref{PhaDia} represents a region on which we could not determine precisely
the phase boundaries using either method, due to numerical uncertainty and finite-size effects.
While the simplest interpretation is that the three phase boundaries meet at the point of intersection between
the continuous transitions (i.e., horizontal and vertical) lines, the data in hand are not sufficient
to verify this.
The phase boundaries for $L=100$ and $L=200$ are nearly the same as for $L=500$,
suggesting that the latter
are already quite close to limiting (infinite-$L$) values.
Table~\ref{CTP} gives the $\alpha$ and $\beta$ values for the continuous transition
lines for the three sizes. There is little sign of systematic variation with system size,
other than a small ($\sim 2.5$\%) increase in $\beta_c$ on going from $L=200$ to $L=500$.
\begin{table}[htb]
\begin{center}
\begin{tabular}{ c c c }
\hline
$L$ & $\alpha_c$ & $\beta_c$ \\ \hline \hline
100 & 0.901(4) & 0.701(2) \\
200 & 0.901(2) & 0.697(1) \\
500 & 0.905(2) & 0.714(1) \\
\hline
\end{tabular}
\end{center}
\caption{Values of $\alpha$ and $\beta$ at the continuous transitions
obtained via the derivative method.}
\label{CTP}
\end{table}
While the continuous-space phase diagram is isomorphic to that of the lattice model,
there are some differences between the two cases that appear
likely to persist in the infinite-size limit.
Since the lattice model possesses particle-hole symmetry, the phase diagram is
invariant under the exchange of $\alpha$ and $\beta$. Thus the boundary
between the high- and low-density phases is a straight line extending from the
origin to the point $(1/2,1/2)$ in the $\alpha$-$\beta$ plane.
The phase diagram of the continuous-space model does not possess this symmetry;
the phase boundary between the high- and low-density
phases does not fall along the line $\alpha=\beta$, and appears to be somewhat curved.
One might inquire whether the differences in the phase boundaries of the
lattice and continuous-space models merely reflect finite-size effects
in the latter. We have verified that in the {\it lattice} TASEP with $L=200$, the phase boundaries
fall quite near their expected (infinite-size) positions. Comparison of the
phase boundaries (in continuous space) for $L=200$ and 500 suggests that
finite-size effects are somewhat stronger in the continuous space model than on the lattice.
Given the lack of particle-hole symmetry, however, it appears very unlikely that the
continuous-space phase boundaries will converge to those of the lattice model
in the infinite-size limit.
The differences between the lattice and continuous-space models reflect, in part, the absence
of particle-hole symmetry in the latter; particle positions fluctuate in continuous space,
but are fixed in the lattice model.
In continuous-space, moreover, particles occupying neighboring wells may influence one another
via the repulsive potential $V_{int}$. On the lattice model no such influence exists,
beyond simple exclusion. In continuous space,
repulsive interactions should tend to spread particles
more uniformly than on the lattice, promoting particle removal, and hindering insertion.
Thus the transition from high to low density occurs for $\beta < \alpha$.
Since repulsion is more significant for
higher densities (i.e., larger $\alpha$) the phase boundary should curve toward the $\alpha$
axis, as is observed. The smaller value of $\beta$ at the continuous transition (approximately
0.714(1) for $L=500$), compared to that of $\alpha$
(about 0.905(2) for $L=500$) may also be attributed
to repulsion between neighboring particles.
\subsection{Current: comparison with mean-field theory}
The absence of particle-hole symmetry is again evident in a plot of the current as a function
of density.
In the lattice model, mean field theory gives $J = \rho(1-\rho)$ \cite{Krug91,Derrida92}
which is in fact an exact expression.
Fig.~\ref{J_d} compares the current on
lattice with our results for continuous space. (Note that the latter
exhibit virtually no finite-size effects on the scale of the figure.)
Unlike in the lattice TASEP, here the current is not
symmetric about $\rho=1/2$; it takes
its maximum value at a density of about 0.57.
The fact that the maximum current occurs at a higher density than on the lattice
may again be attributed to interparticle repulsion.
\vspace{2em}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{J_r}
\end{center}
\caption{Current \textit{versus} density in the lattice model (exact) and the continuous-space model.}
\label{J_d}
\end{figure}
On the lattice, the current is equal to the
probability of having an occupied site with its neighbor on the right vacant, i.~e.,
$J~=~\langle\xi_i(1-\xi_{i+1})\rangle, \; \forall i$, where $\xi_i$ is an indicator variable
equal to one if site $i$ is occupied and $0$ if it is empty.
In mean-field theory the joint probability is factored so:
$\langle\xi_i (1-\xi_{i+1})\rangle = \langle\xi_i\rangle\langle(1-\xi_{i+1})\rangle$,
and setting $\langle\xi_i\rangle = \rho$ in the bulk, we obtain
$J = \rho(1-\rho)$.
In developing a mean-field theory for the continuous-space model,
one might argue that $J \simeq \rho(1-\rho)/\tau$, i.~e., that the current is simply
given by the transition rate $1/\tau$ for jumps between neighboring wells times the
probability of a given well being occupied and its right neighbor empty.
Due to repulsive interactions
between neighboring particles, however, a particle in the well to the left can also influence
the hopping rate. As a first approximation we write
\begin{equation}\label{MFTC}
J = P(0,1,0)j(0,1,0) + P(1,1,0)j(1,1,0),
\end{equation}
where $P(\xi_{i-1},\xi_i,\xi_{i+1})$ is the joint probability for
three adjacent wells, and $j(\xi_{i-1},\xi_i,\xi_{i+1})$
is the transition rate in this configuration. Factorizing the joint probability,
we have $P(0,1,0) = \rho (1-\rho)^2$ and $P(1,1,0) = \rho^2 (1-\rho)$. It remains to
evaluate the currents $j(0,1,0)$ and $j(1,1,0)$.
Consider first $j(0,1,0)$, the rate to overcome the barrier between wells $i$ and $i+1$,
given that both $i-1$ and $i+1$ are empty. The mean first-passage
time $\tau_{ca}$, for a particle to overcome the barrier,
is readily found via analysis of the one-dimensional
Fokker-Planck equation. From the standard result \cite{VanKampen} we have
\begin{equation}\label{FPT}
\tau_{ca} = \frac{1}{k_BT}\int_a^ce^{U(x')/k_BT}dx'\int_{-\infty}^{x'}e^{-U(x'')/k_BT}dx'',
\end{equation}
\noindent where $a=x_0$ and $c=x_0 + 2d_0$ are the positions of adjacent potential minima,
with $x_0$ given by Eq.~(\ref{x0}).
Using the effective external potential, Eq.~(\ref{PotExtEfe}), for $U(x)$
(since there are no interactions with other particles),
numerical evaluation
of Eq.~(\ref{FPT}) yields $\tau_{ca} = 6.538\,$s$\, = 1/j(0,1,0)$. This value agrees to within uncertainty
with our simulation result for the mean time $\tau$ for a particle to hop between adjacent wells
when there are no other particles in the system. Thus the mean-field curve in
Fig.~\ref{J_d} agrees with simulation in the low-density limit.
To estimate the transition rate $j(1,1,0)$,
we perform a Monte Carlo simulation to determine the mean time required for a particle
to hop to the next well, when the preceding
well is occupied, and there are no other particles in the system. This yields
$1/j(1,1,0) = 5.548\,$s. Although the presence of the trailing particle leads
to an increase of about 18\% in $j$, the effect is not sufficient to yield
quantitative agreement with the current observed at higher densities (Fig.~\ref{J_d}).
There are two possible sources for this
discrepancy. First, at moderate and high densities, strings of $n \geq 3$ occupied wells
occur with finite probability, and
the cumulative effect of repulsions along the chain should make the hopping rate of the
lead particle an increasing function of $n$.
Simulations of $n = 3$ - 9 occupied wells show that the transition rate of the first
particle grows with $n$, but not enough to account for the maximum value of the current observed.
A second point is that the mean-field factorizations
$P(0,1,0) = \rho(1-\rho)^2$ and $P(1,1,0) = \rho^2(1-\rho)$,
are not very accurate in the continuous-space model. For density $\rho \simeq 0.52$, for example,
we find
\begin{equation}
\frac{P(0,1,0)}{\rho(1-\rho)^2} \simeq 1.21
\end{equation}
and
\begin{equation}
\frac{P(0,1,1)}{\rho^2(1-\rho)} \simeq 1.04
\end{equation}
implying a significant correction to the mean-field theory predictions.
\subsection{Fluctuations}
To close this section we note an
interesting finding on fluctuations. The variance of the density,
as a function of $\alpha$, with
fixed $\beta$, exhibits a maximum at the discontinuous transition
(see Fig.~\ref{var_den}). For $\beta$ values such that the transition is continuous,
by contrast, no peak in var($\rho)$ is observed at the transition. (Similar
behavior is found, varying $\beta$ with $\alpha$ fixed.) The large density fluctuations
are associated with the presence of a shock separating high- and low-density regions, whose
position fluctuates over the entire system.
The position of maximum variance agrees to within uncertainty with the
lines of the (discontinuous) phase transitions reported in Fig.~\ref{PhaDia}.
Analysis of the variance, however, appears to furnish less precise
results than the method described above.
The total energy exhibits fluctuations similar to those observed in the density,
but we do not find any signal in var($J$) associated with the phase transitions.
\vspace{1cm}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6cm]{VarDenBetFix}
\end{center}
\caption{Density variance versus $\alpha$ for $\beta$ values as indicated. System size $L=200$.}
\label{var_den}
\end{figure}
\vspace{1em}
\section{Discussion}\label{CON}
We propose a continuous-space model of interacting Brownian particles in a periodic
potential as a possible realization of the TASEP.
The particles are
subject to a constant drive in a periodic external potential.
Using numerical integration of the Langevin equation and Monte Carlo simulation,
we study systems of L = 100, 200, and 500 wells.
Our results show that the continuous-space model exhibits continuous and discontinuous
phase transitions analogous to those observed in the lattice TASEP. The phase diagram of the
continuous-space model is similar to that of the lattice model, but exhibits some differences
due to the absence of particle-hole symmetry. This difference appears to be
associated with fluctuations of particle positions around potential minima.
Such fluctuations, together with the repulsive interactions
between neighboring particles, cause the current to attain its maximum value at a density
somewhat greater than 1/2, the density marking the maximum current in the lattice model.
We expect these changes (relative to the lattice model) to be generic for
continuous-space systems exhibiting TASEP-like phase transitions.
We believe that the present study demonstrates the possibility of observing TASEP-like
behavior in laboratory experiments on systems of interacting colloidal particles in a one-dimensional
optical tweezer array. The essential features of the TASEP - localization of particles in
potential wells, with multiple occupancy prohibited, and bias hopping along the line - are
readily accomplished with a appropriate choice of particle, fluid, and tweezer array parameters.
It is however less obvious how to implement random insertion and removal of particles at the
first and last wells of the array. Particle manipulation can be accomplished using
optical tweezers to transfer particles between wells and reservoirs.
To transfer particles in a random fashion, these tweezers would have to be intrinsically noisy
or chaotic,
controlled by a random number generator, or driven by a noise signal.
A simpler alternative may be {\it periodic} insertion and removal. In this case, one
inserts a particle into the first well (when empty) at intervals of $\tau/\alpha$, and
checks for occupancy of the final well at intervals of $\tau/\beta$, removing the particle
if the well is occupied.
A preliminary study of the continuous-space model using periodic insertion
and removal confirms that the three
TASEP phases are again found. The density profiles
under periodic and random particle transfer are very similar in the maximum current phase, where the
density profile is insensitive to small changes in the insertion and removal rates.
Small systematic differences do however appear in the other phases, as shown in Fig.~\ref{perfis_periodicos}.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=7.5cm]{perfis_periodico}
\end{center}
\caption{Comparison of density profiles obtained using periodic (black) and random (red)
insertion and removal, $L=100$.}
\label{perfis_periodicos}
\end{figure}
Although our study strongly suggests the feasibility of a laboratory realization of the TASEP,
a number of additional features would have to be included in the model, before a quantitative
comparison with experiment could be made. The principal modifications we expect to be necessary
are study of a three-dimensional model, allowing fluctuations in directions perpendicular to the
array axis, and inclusion of hydrodynamic interactions between the particles. We defer these
tasks to future work.
\begin{acknowledgments}
We thank O. N. Mesquita, U. Agero and J. G. Moreira for helpful comments.
This work was supported by CNPq and Fapemig, Brazil.
\end{acknowledgments}
|
\section{Introduction}
Recently an extension of the Standard Model (SM) based on the ideas of Lee and Wick~\cite{Lee:1969fy,Lee:1970iw} has been proposed\footnote{A Lee-Wick extension of the Higgs sector had been previously proposed in~\cite{Jansen:1992xv,Jansen:1992xx,Jansen:1993ji,Jansen:1993jj}.}~\cite{Grinstein:2007mp}. This features higher derivative kinetic terms for each SM field. As a consequence the field propagators fall off to zero more rapidly than the ordinary SM propagators, and the infinities associated with ultraviolet quantum fluctuations are either softened or even removed from the theory. In a scalar field theory all amplitudes turn out to be finite by power counting. In a gauge theory the higher derivative kinetic terms generate new momentum-dependent interactions which prevent the theory from being finite; however, a simple power counting argument shows that the only possible divergences are logarithmic. Thus the LW SM offers a potential solution to the hierarchy problem. This was the main motivation for studying the model in~\cite{Grinstein:2007mp}, and analyzing its phenomenological implications~\cite{Rizzo:2007ae,Rizzo:2007nf,Alvarez:2009af,Krauss:2007bz}.
If a higher-derivative kinetic term is added to the Lagrangian, the propagator of a LW SM field displays two poles, the lighter one corresponding to the ordinary SM particle, and the heavier one corresponding to a new degree of freedom, the LW partner. An equivalent formulation consists of separating the poles, in such a way that to each field there corresponds only one pole and one mass. The LW poles are then characterized by a negative residue, and thus act as Pauli-Villar regulators. However, unlike mere regulators, the LW fields nontrivially participate in gauge and Yukawa interactions.
In the LW SM, as in the ordinary SM, the largest one-loop contribution to the Higgs mass comes from an isospin violating sector of the theory: the top Yukawa coupling. There are two heavy partners of the top quark in the LW SM, one associated with the left-handed top-bottom doublet, with mass $M_q$, and the other with the right-handed top, with mass $M_t$. The contributions to the Higgs mass involving a single LW top are opposite in sign to those from a single SM top, so they cancel the quadratic divergence in $\delta m_h^2$. The net contribution is still logarithmically divergent, and for degenerate LW top quarks, $M_q=M_t$, is of the form
\begin{equation}
\delta m_h^2=\frac{3\lambda_t^2}{8\pi^2}M_q^2 \log\frac{\Lambda^2}{M_q^2} \ ,
\end{equation}
where $\Lambda$ is the cutoff. In the limit $M_q\to\infty$ the ordinary quadratic divergence reappears, with $M_q$ acting as a cutoff. Therefore, as already pointed out in Ref.~\cite{Alvarez:2008ks}, in order to avoid fine tuning the value of $M_q$ cannot be too large.
Because the dominant correction to the Higgs mass is associated with an isospin violating sector of the theory, it is important to check whether the LW tops cause a large contribution to the electroweak observables which are usually protected by custodial symmetry: $\Delta\rho$, and, for theories with heavy replicas of the top quark, the $Z b_L \bar{b}_L$ coupling~\cite{Agashe:2006at}~\cite{SekharChivukula:2009if}. Large contributions to these quantities would lead to a stringent lower bound on $M_q$, which would result in large corrections to $m_h^2$ and thus the necessity of fine-tuning the scalar sector of the theory.
In this paper we analyze the potential conflict between naturalness and isospin violation, by computing the contribution of the top-quark sector to the $\rho$ parameter and to the $Z b_L \bar{b}_L$ coupling. Furthermore, we compute the Barbieri {\em et. al.} ~\cite{Barbieri:2004qk,Chivukula:2004af} post-LEP electroweak parameters ($\hat{S}$, $\hat{T}$, $W$, and $Y$) to check for additional constraints. In terms of the post-LEP parameters,
we find a simple picture for the constraints on the Lee-Wick standard model. The dominant contributions to $\hat{T}$ come from the fermion sector at one loop, and limits on this parameter provide the strongest constraints on the top-quark sector\footnote{The dominant contributions to $\hat{S}$ likewise come from the fermion sector at one loop, but they are too small to provide strong constraints on the top quark sector.} . In contrast, the dominant contributions
to $Y$ and $W$ arise from the gauge sector at tree-level, and limits on these parameters therefore
provide the strongest constraints on the gauge sector. These results imply that the bounds on the LW fermions, coming almost entirely from $\hat{T}$, are essentially independent of the LW gauge masses.
Our results differ from those in Refs.~\cite{Alvarez:2008ks, Carone:2008bs} because their one-loop analysis of the effects of LW top quarks on electroweak observables rests on the incorrect assumption that the corrections are purely oblique \cite{Peskin:1991sw,Altarelli:1990zd,Altarelli:1991fk}. As discussed in Ref.~\cite{Underwood:2008cr} important non-oblique corrections arise at tree-level in the LW SM, in the form of non-zero values for $W$ and $Y$. Therefore, one must use the Barbieri {\em et. al.} parameters to compare the LW SM with experiment.
In section~\ref{sec:model}, we review the structure of the Lee-Wick standard model~\cite{Grinstein:2007mp}
and establish notation. In section~\ref{sec:effective} we present an effective field theory analysis of
the LW corrections to $\Delta \rho$ and to the $Zb_L\bar{b}_L$ coupling.
In section~\ref{sec:parameters} we present our analysis of the post-LEP electroweak parameters and the resulting constraints on the Lee-Wick standard model, while the constraints from the $Z b_L \bar{b}_L$ coupling are analyzed in section~\ref{sec:Zbb}. The leading logarithmic contributions to the electroweak observables in the full theory and the effective theory have to match, thus the results of section~\ref{sec:effective} provide an important check for those of sections~\ref{sec:parameters} and \ref{sec:Zbb}.
Questions concerning unitarity~\cite{Cutkosky:1969fq}, causality~\cite{Grinstein:2008bg}, and Lorentz invariance in LW theories, although potentially important, will not be considered in this analysis. A complete analysis of one-loop renormalization of the LW SM can be found in~\cite{Grinstein:2008qq}.
\section{The Lee-Wick Standard Model}\label{sec:model}
It is straighforward to write a higher derivative extension of the SM electroweak Lagrangian
\cite{Grinstein:2007mp}. Adopting a non-canonical normalization for the gauge fields, the gauge Lagrangian reads
\begin{eqnarray}
{\cal L}_{\rm gauge}^{\rm hd} = -\frac{1}{4g_1^2}\hat{B}_{\mu\nu}\hat{B}^{\mu\nu}
-\frac{1}{2 g_2^2}{\rm Tr}\left[\hat{W}_{\mu\nu}\hat{W}^{\mu\nu}\right]
+\frac{1}{2 g_1^2 M_1^2}\partial^\mu \hat{B}_{\mu\nu} \partial_\lambda \hat{B}^{\lambda\nu}
+\frac{1}{g_2^2 M_2^2}{\rm Tr}\left[\hat{D}^\mu \hat{W}_{\mu\nu} \hat{D}_\lambda \hat{W}^{\lambda\nu}\right] \ .
\label{eq:gaugelagrg}
\end{eqnarray}
The ``hat" notation indicates that the field's propagator contains not only the ordinary SM poles but also the LW poles. For example, in the limit of unbroken electroweak phase the $\hat{B}_\mu$ propagator has a massless pole, corresponding to the ordinary $B_\mu$ gauge field, and a mass-$M_1$ pole, corresponding to its LW counterpart. Notice also that additional dimension-six operators, quadratic in the field strength tensors could, in principle, be added to this Lagrangian. However these would lead to scattering amplitudes for the longitudinally polarized gauge bosons growing like $E^2$, and thus to a rather early violation of unitarity~\cite{Grinstein:2007iz}. We therefore do not include them in this analysis. Notice also that we only include one higher derivative term per SM field, which introduces a single corresponding LW pole. This is certainly fine for our purposes, since in this analysis we focus on the low momentum regime, where additional higher derivative terms are negligible. However at large momenta additional poles in the propagator can have important implications~\cite{Carone:2008iw,Carone:2009it}.
The higher derivative extension of the Higgs sector is
\begin{eqnarray}
{\cal L}_{\rm Higgs}^{\rm hd} = |\hat{D}_\mu \hat{\phi}|^2
-\lambda\left(\hat{\phi}^\dagger \hat{\phi} -\frac{v^2}{2}\right)^2
-\frac{1}{M_h^2} |\hat{D}^2 \hat{\phi}|^2 \ ,
\label{eq:higgslagrg}
\end{eqnarray}
where as usual the Higgs doublet may be written in component form as
\begin{eqnarray}
\hat{\phi} = \frac{1}{\sqrt{2}}
\left(\begin{array}{c}
i\sqrt{2}\hat{\phi}^+ \\ v+\hat{h}-i\hat{\phi}^0
\end{array}
\right)~.
\end{eqnarray}
Here and in Eq.~(\ref{eq:gaugelagrg}) the covariant derivative written with a hat is built with the hatted gauge fields. We will find it convenient to have a compact way of denoting $i \sigma^2 \hat{\phi}^\ast$ as we build operators that couple the Higgs to the right-handed top quark. Hence, we make the definition:
\begin{eqnarray}
\hat{\varphi} \equiv (i \sigma^2 \hat{\phi}^\ast) =
\frac{1}{\sqrt{2}}
\left(\begin{array}{c}
v+\hat{h}+i\hat{\phi}^0\\
i\sqrt{2}\hat{\phi}^-
\end{array}\right)~.
\end{eqnarray}
The field $\hat{\phi}$ contains both the ordinary Higgs doublet and a massive doublet\footnote{If $M_h$ is smaller than all other LW mass parameters, in a certain energy regime the model behaves like a two-Higgs doublet model, although one doublet is of LW type. This scenario was analyzed in~\cite{Carone:2009nu}.}
with mass $\sim M_h$.
In the fermion sector we only focus on the third quark generation, since this is the dominant source of isospin violation, and gives the largest correction to the Higgs mass.\footnote{Inclusion of the remaining flavors would introduce new mixing matrices, and, without the assumption of minimal flavor violation, potential sources of flavor changing neutral current. For a discussion see Ref.~\protect\cite{Dulaney:2007dx}.} The higher-derivative extension of the fermion Lagrangian is
\begin{eqnarray}
{\cal L}_{\rm quark}^{\rm hd} = \bar{\hat{q}}_L i \hat{\slashed{D}} \hat{q}_L
+ \bar{\hat{t}}^\prime_R i \hat{\slashed{D}} \hat{t}^\prime_R
+ \bar{\hat{b}}^\prime_R i \hat{\slashed{D}} \hat{b}^\prime_R
+\frac{1}{M_q^2} \bar{\hat{q}}_L i \hat{\slashed{D}}^3 \hat{q}_L
+\frac{1}{M_t^2} \bar{\hat{t}}^\prime_R i \hat{\slashed{D}}^3 \hat{t}^\prime_R
+\frac{1}{M_b^2} \bar{\hat{b}}^\prime_R i \hat{\slashed{D}}^3 \hat{b}^\prime_R \ ,
\label{eq:higher}
\end{eqnarray}
where $\hat{q}_L=(\hat{t}_L,\hat{b}_L)$. Notice that the right handed fields have been primed because, for example, $\hat{t}_L$ and $\hat{t}^\prime_R$ are not left and right component of the same Dirac spinor. In the unbroken electroweak phase $\hat{t}_L$ ($\hat{t}^\prime_R$) contains the ordinary massless SM left-handed (right-handed) top as well as a massive Dirac fermion of mass $M_q$ ($M_t$).
Finally we consider the Yukawa Lagrangian, which in the LW SM has no derivatives. Therefore, we write
\begin{eqnarray}
{\cal L}_{\rm Yukawa} = -y_t\ \bar{\hat{q}}_L\ \hat{\varphi}\ \hat{t}^\prime_R \ +\ {\rm h.c.} \ ,
\label{eq:yukhigher}
\end{eqnarray}
where the bottom Yukawa coupling has been ignored, since $y_b\ll y_t$.
As explained in the introduction this ``higher derivative" formulation of the theory, in which both the ordinary pole and the LW pole are contained in the same field, is equivalent to an ``ordinary formulation" in which: (i) the two poles belong to two different fields, and (ii) the kinetic and mass terms for the LW fields have the ``wrong'' sign. This alternative formulation is especially useful for calculating loop diagrams. In this paper we will compute loop diagrams with the top and bottom quarks in the loop. Thus we will find it helpful to replace the higher derivative fermion and Yukawa Lagrangians with the ordinary formulation Lagrangians
\begin{eqnarray}
{\cal L}_{\rm quark} = \bar{q}_L i \hat{\slashed{D}} q_L
+ \bar{t}^\prime_R i \hat{\slashed{D}} t^\prime_R
+ \bar{b}^\prime_R i \hat{\slashed{D}} b^\prime_R
-\bar{\tilde{q}}\left(i\hat{\slashed{D}}-M_q\right)\tilde{q}
-\bar{\tilde{t}}^\prime\left(i\hat{\slashed{D}}-M_t\right)\tilde{t}^\prime
-\bar{\tilde{b}}^\prime\left(i\hat{\slashed{D}}-M_b\right)\tilde{b}^\prime \ ,
\label{eq:lower}
\end{eqnarray}
and
\begin{eqnarray}
{\cal L}_{\rm Yukawa} = -y_t\ \left(\bar{q}_L-\bar{\tilde{q}}_L\right)\ \hat{\varphi}\
\left(t^\prime_R-\tilde{t}^\prime_R\right) \ + \ {\rm h.c.} \ ,
\label{eq:yuklower}
\end{eqnarray}
where
\begin{eqnarray}
\hat{q}_L \equiv q_L - \tilde{q}_L\ , \quad\quad \hat{t}^\prime_R\equiv t^\prime_R-\tilde{t}^\prime_R\ , \quad\quad
\hat{b}^\prime_R\equiv b^\prime_R-\tilde{b}^\prime_R \ ,
\end{eqnarray}
and where the fields with (without) a tilde are LW (SM) fields.
The equivalence between the higher derivative formulation, Eqs.~(\ref{eq:higher},\ref{eq:yukhigher}), and the ordinary formulation, Eqs.~(\ref{eq:lower},\ref{eq:yuklower}), can be easily proved; see for example~\cite{Grinstein:2007mp}. Notice that the ``wrong" sign in front of the kinetic and mass term makes the LW (tilde) fields act like Pauli-Villars regulators, with the difference that they also participate nontrivially in gauge and Yukawa interactions.
\section{Effective Field Theory for $\Delta \rho$ and $Zb\bar{b}$}\label{sec:effective}
The appearance of the LW fields in the Yukawa interactions, Eq.~(\ref{eq:yuklower}), suggest the presence of nonstandard sources of custodial isospin violation at energies below the LW scale. Dimension-six custodial violating operators can potentially arise from tree-level exchanges, and from loop diagrams with one or more LW fermions in the loop. The leading contribution to these operators, in inverse powers of the LW fermion masses, can be found by integrating out the LW fermions at tree-level, and computing loops in the resulting effective field theory. For LW fermion masses much larger than both the Higgs vacuum and the external momenta, the effective Lagrangian can be computed in powers of $\hat{\varphi}/M_{q,t}$ and $\hat{\slashed{D}}/M_{q,t}$. Including the leading nonstandard corrections, this leads to
\begin{eqnarray}
{\cal L}_{\rm eff} &=& \bar{q}_L i \hat{\slashed{D}} q_L
+ \bar{t}_R i \hat{\slashed{D}} t_R
+ \bar{b}_R i \hat{\slashed{D}} b_R
- y_t\left( \bar{q}_L \hat{\varphi}\ t_R + \bar{t}_R \hat{\varphi}^\dagger q_L \right) \nonumber \\
&-& \frac{y_t^2}{M_t^2}\ \bar{q}_L i \hat{\varphi}\hat{\slashed{D}}\left(\hat{\varphi}^\dagger q_L \right)
- \frac{y_t^2}{M_q^2}\ \bar{t}_R i \hat{\varphi}^\dagger \hat{\slashed{D}}\left(\hat{\varphi} t_R \right) \ .
\label{eq:efflagr0}
\end{eqnarray}
Notice that the primes have been removed from the right-handed fermion fields, because now left-handed and right-handed components are Dirac partners. Notice also that this Lagrangian assumes $M_q$ and $M_t$ to be of the same order, with no hierarchy between them. The leading logarithmic correction to observables will therefore be proportional to $\log M_q^2/v^2\sim \log M_t^2/v^2$. In what follows, we compute these leading-log corrections by constructing the operators which arise in
the effective theory appropriate for energy scales below $M_t\simeq M_q$, in which the LW partners have
been ``integrated out" but the top quark remains in the spectrum.
After electroweak symmetry breaking, the higher derivative operators lead to a renormalization of the fermion kinetic terms. An alternative approach consists of redefining $q_L$ and $t_R$ to make their kinetic terms canonically normalized both in the broken and the unbroken electroweak phase. This is achieved by the replacements
\begin{eqnarray}
q_L & \to & \left[1+\frac{y_t^2}{2 M_t^2} \hat{\varphi}\hat{\varphi}^\dagger
+ {\cal O}(1/M_t^3)\right] q_L \ , \nonumber \\
t_R & \to & \left[1+\frac{y_t^2}{2 M_q^2} \hat{\varphi}^\dagger\hat{\varphi}
+ {\cal O}(1/M_q^3)\right] t_R \ ,
\end{eqnarray}
which leads to a new Lagrangian, equivalent to ${\cal L}_{\rm eff}$:
\begin{eqnarray}
{\cal L}^\prime_{\rm eff} &=& \bar{q}_L i \hat{\slashed{D}} q_L
+ \bar{t}_R i \hat{\slashed{D}} t_R
+ \bar{b}_R i \hat{\slashed{D}} b_R \nonumber \\
&-& y_t\ \bar{q}_L \hat{\varphi}\left[1+\frac{y_t^2}{2}\left(\frac{1}{M_q^2}+\frac{1}{M_t^2}\right)
\hat{\varphi}^\dagger \hat{\varphi}\right]t_R + {\rm h.c.} \nonumber \\
&+& \frac{y_t^2}{2M_t^2}\ \bar{q}_L i \left[(\hat{D}_\mu\hat{\varphi})\hat{\varphi}^\dagger
-\hat{\varphi}(\hat{D}_\mu\hat{\varphi})^\dagger\right] \gamma^\mu q_L
+ \frac{y_t^2}{2M_q^2}\ \bar{t}_R \gamma^\mu t_R\ i \left[(\hat{D}_\mu\hat{\varphi})^\dagger \hat{\varphi}
-\hat{\varphi}^\dagger(\hat{D}_\mu\hat{\varphi})\right] \ .
\label{eq:efflagr}
\end{eqnarray}
As expected, there are custodial violating dimension-six operators. However at tree-level there is no non-standard contribution to $\Delta\rho$ or the $Z b_L \bar{b}_L$ coupling.
\begin{figure}[!t]
\includegraphics[width=5.5in,height=2.8in]{custodial.eps}
\caption{Diagrams contributing to the dimension-six four-$\hat{\phi}$ operators in the effective theory, with the LW fermions integrated out at tree-level.}
\label{fig:effectiverho}
\end{figure}
${\cal L}^\prime_{\rm eff}$ features terms coupling one, two, or three $\hat{\varphi}$ fields to a pair of fermions. Therefore, dimension-six four-$\hat{\varphi}$ operators arise both from vacuum polarization amplitudes and triangle diagrams, as shown in Fig.~\ref{fig:effectiverho}. The log-divergent parts of these diagrams (which yield the $\log(M^2_{t,q}/m^2_t)$ contributions)
can be computed in the unbroken electroweak phase, with fermions in the loop. The logarithmically divergent part of the amplitude is reproduced by the operators\footnote{There are also quadratic divergences which are completely absorbed by a counterterm of the form $|\hat{\varphi}|^4$, with no derivatives.}
\begin{eqnarray}
\frac{3y_t^4}{16\pi^2}\left[\frac{2}{M_t^2}+\frac{1}{M_q^2}\right]|\hat{D}\hat{\phi}|^2 |\hat{\phi}|^2 \cdot \frac{1}{\epsilon}
+\frac{3y_t^4}{16\pi^2}\left[\frac{1}{M_t^2}+\frac{2}{M_q^2}\right]|\hat{\phi}^\dagger\hat{D}\hat{\phi}|^2 \cdot \frac{1}{\epsilon} \ ,
\end{eqnarray}
where as usual $\epsilon=2-d/2$ in dimensional regularization. The first operator respects custodial symmetry, but the second operator does not, since it only contributes to the $Z$ boson mass. The second operator gives the dominant contribution to $\Delta\rho$, which is therefore of the order
\begin{eqnarray}
(\Delta\rho)_{\rm LW} \sim -\frac{3}{16\pi^2}\frac{2m_t^4}{v^2}\left[\frac{1}{M_t^2}+\frac{2}{M_q^2}\right]\log{\frac{M_q^2}{m_t^2}} \ ,
\label{eq:rholeading}
\end{eqnarray}
where the $1/\epsilon$ is replaced by the large $\log$ which arises in the effective theory scaling from the scale $M_q\sim M_t$ to the weak scale $m_t\sim v$. For LW fermions lighter than 1 TeV this is a large {\it negative} isospin violating effect. For example, taking $M_q=M_t=500$ GeV gives $\Delta\rho\sim -1.4\%$. Furthermore, since $\Delta\rho$ is always negative, a heavy Higgs is strongly disfavored in the LW SM.
The diagrams contributing to the left-handed fermionic gauge couplings up to one-loop order are shown in Fig.~\ref{fig:effectiveZbb}. The tree-level diagram (corrected by the field strength renormalizations) corresponds to the custodial violating operator proportional to $y_t^2/2M_t^2$, in Eq.~(\ref{eq:efflagr}). This operator contributes to the $Z t_L \bar{t}_L$ coupling, not to $Z b_L \bar{b}_L$\footnote{Including the bottom Yukawa coupling would lead to a tree-level operator contributing to $Z b_L \bar{b}_L$. However the top-loop contribution is dominant, since $16\pi^2 y_b^2\sim 0.1$.}. The remaining diagrams contain nonstandard logarithmic divergences which are reproduced by the operators
\begin{eqnarray}
\frac{y_t^4}{16\pi^2}\frac{1}{4M_q^2} i \left[\bar{q}_L \gamma^\mu \hat{D}_\mu q_L - \bar{q}_L \overleftarrow{\hat{D}^\dagger_\mu} \gamma^\mu q_L\right]\hat{\varphi}^\dagger \hat{\varphi} \cdot \frac{1}{\epsilon}
+\frac{y_t^4}{16\pi^2}\left[\frac{1}{M_t^2}+\frac{1}{4 M_q^2}\right]
\ \bar{q}_L \gamma^\mu q_L \ i \left[(\hat{D}_\mu\hat{\varphi})^\dagger \hat{\varphi}
-\hat{\varphi}^\dagger(\hat{D}_\mu\hat{\varphi})\right] \cdot \frac{1}{\epsilon} \ .
\label{eq:Zbbeffective}
\end{eqnarray}
In this expression the first (custodially symmetric) operator amounts to a renormalization of the standard gauge interactions, and does not contribute to nonstandard fermionic gauge couplings. The second operator violates custodial symmetry, and is only due to the triangle diagrams in Fig.~\ref{fig:effectiveZbb}. This contributes both to the $Z t_L \bar{t}_L$ coupling and the $Z b_L \bar{b}_L$ coupling. Expressing the latter in the form
\begin{equation}
\frac{e}{c_w s_w} g_L^{b\bar{b}}\ Z_\mu\ \bar{b}_L\gamma^\mu b_L \equiv \frac{e}{c_w s_w} \ \left[-\frac{1}{2} + \frac{1}{3}\sin ^2 \theta_W
+ (\delta g_L^{b\bar{b}})_{\rm SM} + (\delta g_L^{b\bar{b}})_{\rm LW} \right] \ Z_\mu\ \bar{b}_L\gamma^\mu b_L \ ,
\label{eq:totalZbb}
\end{equation}
where $(\delta g_L^{b\bar{b}})_{\rm SM}$ includes all higher order SM corrections, and replacing the $1/\epsilon$ poles with the large log arising from scaling in the
theory, we find that the second
operator of Eq.~(\ref{eq:Zbbeffective}) gives the dominant non universal LW contribution to $g_L^{b\bar{b}}$:
\begin{eqnarray}
(\delta g_L^{b\bar{b}})_{\rm LW} \sim -\frac{m_t^4}{32\pi^2 v^2}\left[\frac{4}{M_t^2}+\frac{1}{M_q^2}\right]\log\frac{M_q^2}{m_t^2} \ .
\label{eq:Zbbleading}
\end{eqnarray}
The SM prediction is already 1.96$\sigma$ below the observed central value. Hence the additional
negative correction in the LW theory goes in the direction opposite
to what is favored by experiment.
In the next two sections we compute perturbatively (in $v^2/M_q^2$ and $v^2/M_t^2$) and numerically
the values of $\Delta\rho$ and the $Z b_L \bar{b}_L$ coupling in the full LW theory. Our effective field
theory results, Eq.~(\ref{eq:rholeading}) and Eq.~(\ref{eq:Zbbleading}), provide a check of these full
calculations, since the leading logarithmic contributions have to match. More generally, below we compute the top-sector one loop contribution to all Barbieri {\em et. al.} \cite{Barbieri:2004qk,Chivukula:2004af} electroweak parameters, and provide lower bounds on $M_q$ and $M_t$ from comparison with experiment.
\begin{figure}
\includegraphics[width=5.5in,height=2.5in]{qq.eps}
\caption{Diagrams contributing to the dimension-six operators with two external $q_L$ and two $\hat{\varphi}$ fields. The triangle diagrams lead to the second operator of Eq.~(\ref{eq:Zbbeffective}), which contains non universal corrections to the $Z b_L \bar{b}_L$ coupling.}
\label{fig:effectiveZbb}
\end{figure}
\section{Constraints from Post-LEP Parameters}\label{sec:parameters}
In the language of Barbieri {\em et. al.} \cite{Barbieri:2004qk,Chivukula:2004af}, the observables $\hat{S}$, $\hat{T}$, $Y$, and $W$ parametrize the flavor-universal deviations from the SM at low energies. We now analyze the tree-level and the fermionic one-loop contibutions to these parameters, and use them to obtain constraints on the masses of the LW states.
\subsection{Tree-Level Contributions}
At tree-level it is straightforward to read from Eq.~(\ref{eq:gaugelagrg}) the vacuum polarization amplitudes:
\begin{eqnarray}
\Pi_{\hat{W}^+ \hat{W}^-}(q^2) &=& \frac{q^2}{g_2^2} -\frac{(q^2)^2}{g_2^2\ M_2^2}-\frac{v^2}{4} \ , \nonumber \\
\Pi_{\hat{W}^3 \hat{W}^3}(q^2) &=& \frac{q^2}{g_2^2} -\frac{(q^2)^2}{g_2^2\ M_2^2}-\frac{v^2}{4} \ , \nonumber \\
\Pi_{\hat{W}^3 \hat{B}}(q^2) &=& \frac{v^2}{4} \ , \nonumber \\
\Pi_{\hat{B} \hat{B}}(q^2) &=& \frac{q^2}{g_1^2} -\frac{(q^2)^2}{g_1^2\ M_1^2}-\frac{v^2}{4} \ .
\label{eq:twopoint}
\end{eqnarray}
Following \cite{Barbieri:2004qk}, we see that there is no tree-level correction to the Fermi constant
\begin{eqnarray}
\frac{1}{\sqrt{2}G_F} = - 4 \Pi_{\hat{W}^+ \hat{W}^-}(0) = v^2 \ .
\label{eq:GF}
\end{eqnarray}
Barbieri {\em et. al.} define the electroweak gauge couplings
\begin{eqnarray}
\frac{1}{g^2} \equiv \Pi^\prime_{\hat{W}^+ \hat{W}^-}(0) \ , \label{eq:g}\\
\frac{1}{g^{\prime 2}} \equiv \Pi^\prime_{\hat{B} \hat{B}}(0) \ , \label{eq:ggprime}
\end{eqnarray}
which in the LW SM gives $g'= g_1$ and $g = g_2$. We then compute the tree-level electroweak parameters~\cite{Underwood:2008cr},
\begin{eqnarray}
\hat{S} &\equiv & g^2\ \Pi^\prime_{\hat{W}^3 \hat{B}}(0) = 0 \ , \label{eq:treelevelShat}\\
\hat{T} &\equiv & g^2\left[\Pi_{\hat{W}^3 \hat{W}^3}(0)-\Pi_{\hat{W}^+ \hat{W}^-}(0)\right] = 0 \ , \\
Y &\equiv & \frac{1}{2}g^{\prime 2}m_W^2\ \Pi^{\prime\prime}_{\hat{B} \hat{B}}(0)=-\frac{m_W^2}{M^{2}_1} \ , \label{eq:treelevelY} \\
W &\equiv & \frac{1}{2}g^2m_W^2\ \Pi^{\prime\prime}_{\hat{W}^3 \hat{W}^3}(0)=-\frac{m_W^2}{M_2^2} \label{eq:treelevelW} \ ,
\end{eqnarray}
where in each equation the first equality is the definition of the corresponding post-LEP parameter \cite{Barbieri:2004qk}.
\subsection{Fermionic One-Loop Contributions}
The gauge current correlators receive important loop corrections from the top-bottom sector, through the diagrams shown in Fig.~{\ref{fig:loops}}. These vacuum polarization amplitudes contain two infinities, which are absorbed in the definitions of $g$ and $g^\prime$ given in Eqs.~(\ref{eq:g}) and (\ref{eq:ggprime}), respectively.
As a consequence the noncanonical normalization adopted in Eq.~(\ref{eq:gaugelagrg}) forces us to define renormalized LW gauge masses. A convenient scheme consists of defining $M$ and $M^\prime$ by
\begin{equation}
-\,\frac{2}{g^2 M^2} \equiv \Pi''_{\hat{W}^+ \hat{W}^-}(0) \qquad \qquad -\,\frac{2}{g^2 M'^2} \equiv \Pi''_{\hat{B}\hat{B}}(0)~,
\label{eq:newdef}
\end{equation}
which simplify the one-loop calculations below. At tree-level, from
Eq.~(\ref{eq:twopoint}), we see that $M=M_2$ and $M^\prime=M_1$, and both are related
to the masses of the LW partners of the gauge-bosons. Due to the power-counting
properties of LW theories, after the usual\footnote{Notice that the vacuum polarization diagrams involving only one LW fermion carry an overall negative sign. In fact this happens to make all zero-derivative functions, at $q^2=0$, finite. For this reason there is actually one less infinity compared to the ordinary SM, and the bare $v$ is finite \protect\cite{Grinstein:2007mp}.} coupling-constant and mass renormalizations,
all physical quantities remain finite \cite{Grinstein:2007mp}. Hence $M$ and $M'$ remain
finite at one-loop (and beyond). However, since they are defined by the zero-momentum properties
of the gauge-boson two-point functions their values only approximately equal the masses
of the LW partners of the gauge-bosons. This suffices for our purposes, since we are interested in low-energy observables; if we were studying quantities measured at higher energies, we would want to define $M$ and $M'$ based on propagators renormalized at high $q^2$ instead.
\begin{figure}[!t]
\includegraphics[width=0.6\textwidth]{VPA.eps}
\caption{Dominant vacuum polarization amplitudes for the LW SM gauge fields. These include the ordinary ($t_0$ and $b_0$) and the LW third generation quarks ($t_1$, $t_2$, $b_1$, and $b_2$) in the loop. These amplitudes contribute to the two-point functions of Eq.~(\ref{eq:twopoint}).}
\label{fig:loops}
\end{figure}
The propagators in the loops of Fig.~\ref{fig:loops} correspond to mass eigenstates, where the masses are obtained by diagonalizing the mass matrices by means of symplectic rotations: in this way the LW fields maintain the ``wrong-sign'' kinetic and mass terms. A perturbative diagonalization in $v^2/M_q^2$ and $v^2/M_t^2$~\cite{Alvarez:2008ks} requires considering two different scenarios: nondegenerate LW masses, $|M_q^2-M_t^2|\sim M_q^2$, and (near) degenerate LW masses, $|M_q^2-M_t^2|\ll M_q^2$. For nondegenerate LW top quarks the contributions to the electroweak parameters are quite lengthy. To leading order we obtain
\begin{eqnarray}
\hat{S} & = & -\frac{g^2 m_t^2}{48\pi^2 M_q^2}\Bigg[
\left(2+\frac{1}{r_t^2}\right)\log\frac{M_q^2}{m_t^2}
+\frac{1-3r_t^2+6r_t^4-r_t^6+3r_t^8}{r_t^2(1-r_t^2)^5}\log r_t^2
-\frac{5-17r_t^2+4r_t^4+12r_t^6-23r_t^8+7r_t^{10}}{2r_t^2(1-r_t^2)^4}\Bigg] \ , \nonumber \\
\hat{T} & = & -\frac{3 g^2 m_t^4}{32\pi^2 m_W^2 M_q^2}\Bigg[
\left(2+\frac{1}{r_t^2}\right)\log\frac{M_q^2}{m_t^2}
+\frac{1-3r_t^2+6r_t^6}{r_t^2(1-r_t^2)^5}\log r_t^2
-\frac{9-12r_t^2-21r_t^4+46r_t^6-68r_t^8+22r_t^{10}}{6r_t^2(1-r_t^2)^4}\Bigg] \ , \nonumber \\
Y & = & -\frac{m_W^2}{M^{\prime 2}} \ , \nonumber \\
W & = & -\frac{m_W^2}{M^2}+\frac{g^2 m_W^2}{640\pi^2 M_q^2}\Bigg[-7+\frac{3}{r_b^2}-\frac{9}{r_t^2}\Bigg] \ ,
\label{eq:nondeg}
\end{eqnarray}
where
\begin{eqnarray}
r_t\equiv M_t/M_q \ , \quad r_b \equiv M_b/M_q \ .
\end{eqnarray}
The electroweak parameters in the (near) degenerate case cannot simply be obtained by taking the $r_t, r_b \to 1$ limit in Eqs.~(\ref{eq:nondeg}), since the corresponding expressions diverge. Instead we must diagonalize the mass matrices perturbatively in $1/M_q^2$ (or $1/M_t^2$) and $|M_q^2-M_t^2|/M_q^2$, and then compute the electroweak parameters. For exact degeneracy, $M_q=M_t$, this gives
\begin{eqnarray}
\hat{S} & = & -\frac{g^2 m_t^2}{16\pi^2 M_q^2}\Bigg[\log\frac{M_q^2}{m_t^2}-\frac{12}{5}\Bigg] \nonumber \\
\hat{T} & = & -\frac{3 g^2 m_t^4}{32\pi^2 m_W^2 M_q^2}\Bigg[3\log\frac{M_q^2}{m_t^2}-\frac{141}{20}\Bigg] \nonumber \\
Y & = & -\frac{m_W^2}{M^{\prime 2}} \nonumber \\
W & = & -\frac{m_W^2}{M^2}-\frac{7 g^2 m_W^2}{640\pi^2 M_q^2} \ .
\label{eq:deg}
\end{eqnarray}
Note that the absence of fermionic one-loop corrections to the tree-level value of $Y$ is a direct consequence of the second definition in Eq.~(\ref{eq:newdef}): a different scheme choice would lead to an additional contribution. In the same way, changing the definition of $M$ would lead to a different fermionic one-loop expression for $W$; in any case, the second term\footnote{Note that the first definition in Eq.~(\protect\ref{eq:newdef}) pertains to
$\Pi''_{\hat{W}^+\hat{W}^-}$, whereas $W$ is defined in terms of $\Pi''_{\hat{W}_3 \hat{W}_3}$.}
in $W$ is numerically very small and can be neglected. We therefore conclude that the leading contributions to $Y$ and $W$ are those arising from the LW gauge-sector at tree-level, Eqs.~(\ref{eq:treelevelY}, \ref{eq:treelevelW}).
Since the tree-level values of $\hat{S}$ and $\hat{T}$ vanish, the leading LW contributions to both $\hat{S}$ and $\hat{T}$ arise from the top-quark sector at one loop. In the case of $\hat{T}$ this is not surprising since the dominant locus of isospin violation in the model is the splitting between the top and bottom quark masses. Because $\hat{T}$ is the same as $\Delta\rho$~\cite{Chivukula:2004af}, we may
compare the leading logarithmic correction in Eq.~(\ref{eq:nondeg}) with the result obtained in the effective theory, Eq.~(\ref{eq:rholeading}); we see that they agree. In the case of $\hat{S}$,
the situation is more subtle. The LW gauge-eigenstate fermion partners, being massive,
are not chiral and therefore, in the absence of electroweak contributions to the masses
which mix them with the light chiral gauge-eigenstates, their contribution to $\hat{S}$ vanishes.
Hence the dominant LW contributions to $\hat{S}$ also arise predominantly from the
top-sector of the theory.
Therefore, at tree-level plus one fermion loop we obtain a very simple conclusion: the fermion sector contributes to $\hat{S}$ and $\hat{T}$ only, while the gauge sector contributes to $Y$ and $W$ only. It is true that when gauge loops are included, there will be additional contributions. However, the gauge loop contributions will be sub-dominant compared to the quantities we have already calculated; the only potential exception is $\hat{S}$, for which all of the one-loop contributions are too small to be experimentally relevant. Thus, our existing results suffice for extracting constraints from the experimental data.
\subsection{Comparison with Data}
We begin with constraints on the masses of the LW partners of the gauge bosons. The previous subsection found that the only post-LEP parameters affected by the LW gauge boson masses are $W$ and $Y$, and also that the tree-level expressions for $W$ and $Y$, Eqs.~(\ref{eq:treelevelY}, \ref{eq:treelevelW}), suffice for comparison with data. The experimental constraints on $Y$ and $W$ are rather tight, and almost independent of the value of the Higgs mass~\cite{Barbieri:2004qk}. These translate into the 95\% CL lower bounds on $M_2$ and $M_1$ shown\footnote{These bounds are derived using the errors and correlation matrix given in Ref.~\cite{Barbieri:2004qk}} in Fig.~\ref{fig:gauge}. The left plot shows the bounds for arbitrary values of $M_1$ and $M_2$: for $m_h=115$ GeV the striped region is excluded, while for $m_h=800$ GeV the additional narrow yellow region is excluded as well. The right plot shows the 95\% C.L. ellipses in the $(Y,W)$ plane from the global fit to data~\cite{Barbieri:2004qk}, for $m_h=115$ and $m_h=800$, as well as the LW prediction for degenerate LW masses, $M_1=M_2$. All this is in agreement with the results of Ref.~\cite{Underwood:2008cr} and gives the constraints $M_1$, $M_2\ \gtrsim 2.4$ TeV.
\begin{figure}[!t]
\includegraphics[width=3in]{gauge.eps}
\includegraphics[width=3in]{gaugedegenerate.eps}
\caption{Left: Exclusion plot for the LW gauge-field masses $M_2$ and $M_1$. These bounds are due to the constraints on $Y$ and $W$, as shown by Eq.~(\ref{eq:treelevelY}) and Eq.~(\ref{eq:treelevelW}). For a light Higgs ($m_h=115$ GeV) the striped region to the left of both curves is excluded. For a heavy Higgs ($m_h=800$ GeV) the additional yellow strip between the curves is excluded as well. Right: 95\% C.L. ellipses in the ($Y$,$W$) plane, and the LW prediction for degenerate masses, $M_1=M_2$. The parametric plot is for $0.5\ {\rm TeV} < M_1 = M_2 < 10\ {\rm TeV}$ and the dots are equally spaced in mass. The lower bound on $M_1 = M_2$ is approximately 2.4 TeV for a light Higgs.}
\label{fig:gauge}
\end{figure}
Next, we seek constraints on the masses of the LW partners of the top quark. The previous subsection found that the post-LEP parameters sensitive to the LW fermion masses are $\hat{S}$ and $\hat{T}$, which do not depend on the LW gauge masses at the one-loop level. We should also note that, for a light Higgs, the LW prediction of $\hat{S}$ is very close to its central value, $\hat{S}\simeq0$. Furthermore from the global fit to the experimental data in Ref.~\cite{Barbieri:2004qk}, we conclude that $\hat{T}$ is only mildly correlated to $Y$ and $W$, the parameters that are most sensitive to the LW gauge boson masses in the LW SM. This confirms that the bounds on the LW fermions should be essentially independent of the LW gauge masses, and should come almost entirely from $\hat{T}$.
In Fig.~\ref{fig:That} we show the experimental mean value for $\hat{T}$ (red thick line), the $\pm 2 \sigma$ allowed region, the all-order (in $v^2/M_q^2$) LW prediction (solid blue curve), the leading order LW prediction from Eq.~(\ref{eq:deg}) (dashed blue curve), and the leading-log approximation (dotted blue curve), as functions of $M_q$, in the degenerate case. This figure reveals the bound $M_q=M_t\gtrsim$ 1.6 TeV on the LW fermion masses in the degenerate case. Note that although Eq.~(\ref{eq:deg}) appears to predict a positive $\hat{T}$ for small $M_q$ (dashed blue curve), the complete numerical evaluation (solid blue curve) shows that $\hat{T}$ is always negative, as Fig.~\ref{fig:That} shows explicitly; below $M_q=1$ TeV the perturbative diagonalization of the mass matrix is no longer valid, rendering the leading-order LW prediction unreliable in that mass regime.
\begin{figure}
\begin{center}
\includegraphics[width=4in]{That.eps}
\caption{$\hat{T}$ as a function of $M_q$ in the degenerate case, $M_q=M_t$.
The experimental mean value for $\hat{T}$ is shown by the thick red line, along with
the $\pm 2 \sigma$ allowed region. Also shown are the all-order (in $v^2/M_q^2$) LW prediction (solid blue curve), the leading order LW prediction, Eq.~(\protect\ref{eq:deg}) (dashed blue curve), and the leading-log curve, Eq.~(\protect\ref{eq:rholeading}) (dotted blue curve), as functions of $M_q$, in the degenerate case. Note that the leading-order prediction is not valid below $M_q \sim 1$ TeV. (See text for details.)}
\label{fig:That}
\end{center}
\end{figure}
If we relax the requirement of degenerate LW fermion masses, we obtain the 95\% C.L. bounds on $M_q$ and $M_t$ shown in Fig.~\ref{fig:fermionbounds} (left). For a light Higgs the striped region in Fig.~\ref{fig:fermionbounds} (left) is excluded, while for a heavy Higgs the whole (yellow) region is excluded. Note from Figs.~\ref{fig:gauge} and \ref{fig:fermionbounds} (left) that the mildest constraints on the LW masses are obtained in the fully degenerate case, $M=M^\prime$ and $M_q=M_t$.
Returning to the case of degenerate LW fermion masses, we show in Fig.~\ref{fig:fermionbounds} (right) the values of $\hat{S}$ and $\hat{T}$ as a function of $M_q=M_t$ for $0.5\ {\rm TeV} < M_q < 10\ {\rm TeV}$; the dots representing different values of $M_q$ are placed at regular intervals. The 95\% C.L. ellipses from the global fit to the data~\cite{Barbieri:2004qk} confirm the constraint $M_q\gtrsim 1.5$ TeV for a light Higgs, while a heavy Higgs scenario is disfavored for any LW fermion mass. In fact for a heavy Higgs the $\hat{T}$ parameter is expected to be positive, while the LW SM predicts a negative $\hat{T}$. This is a direct consequence of the negative sign in the LW fermion propagators, which results in an overall negative sign from the (dominant) diagrams involving a single LW fermion in the loop.
\begin{figure}[!t]
\includegraphics[width=3in]{fermion.eps}
\includegraphics[width=3in]{degenerate.eps}
\caption{Left: 95\% C.L. exclusion plots for the LW fermion masses $M_q$ and $M_t$. These bounds come almost entirely from the experimental constraints on $\hat{T}$. For a light Higgs the striped region to the left of the curve is excluded, while a heavy Higgs is completely excluded. Right: 95\% C.L. ellipses in the ($\hat{S}$,$\hat{T}$) plane, and the LW prediction for degenerate masses, $M_q=M_t$. The parametric plot is for $0.5\ {\rm TeV} < M_q < 10\ {\rm TeV}$ and the dots are equally spaced in mass. The lower bound on $M_q$ is approximately 1.5 TeV for a light Higgs.}
\label{fig:fermionbounds}
\end{figure}
Our results disagree with those of \cite{Alvarez:2008ks,Carone:2008bs} in two ways: their bounds on the LW fermion masses appear more stringent for a light Higgs and their limits appear to depend on the masses of the LW gauge boson partners. The disagreement arises because their study of one-loop electroweak corrections in the LW SM assumes the corrections to be purely oblique, and derives constraints by comparing the Peskin-Takeuchi $S$ and $T$~\cite{Peskin:1991sw} parameters to data. However, as clearly discussed in Ref.~\cite{Underwood:2008cr}, and confirmed above in Eqs.~(\ref{eq:treelevelY}) and (\ref{eq:treelevelW}), the LW SM features large non-oblique corrections, in the form of non-zero values for $Y$ and $W$ at tree-level. Hence, one must use the Barbieri {\em et. al.} parameters to compare the LW SM with experiment, as we have done.
\section{Constraints from the $Z b_L \bar{b}_L$ Coupling}\label{sec:Zbb}
The leading contribution to the $Z b_L \bar{b}_L$ coupling (in the gauge coupling expansion) can be obtained in the gaugeless limit from the $\phi^0 b_L \bar{b}_L$ coupling~\cite{Barbieri:1992nz,Barbieri:1992dq,Oliver:2002up,Abe:2009ni}, where $\phi^0$ is the Goldstone boson eaten by the $Z$. The loop diagram giving the largest correction involves the SM and LW top quarks\footnote{In the gaugeless limit of the LW SM, as in the SM itself, all external b-quark
wavefunction renormalization corrections are proportional to $y^2_b$ and
are therefore negligible. This should be contrasted with the situation
in the three-site Higgsless model [33].}
, and is shown in Fig.~\ref{fig:Zbb}. A detailed computation of the loop integral, valid for arbitrary models with heavy replicas of the top quark, is given in the appendix. At zero external momentum the amplitude corresponding to the diagram has the form
\begin{eqnarray}
i\ M = - A\ \slashed{p} P_L \ ,
\label{eq:MZbb}
\end{eqnarray}
where $P_L\equiv (1-\gamma^5)/2$ is the left-handed projector, $p$ is the incoming $\phi^0$ momentum, and the external fermion wavefunctions have been omitted. Then to leading order in $g$ the correction to the $Z b_L \bar{b}_L$ coupling is~\cite{Barbieri:1992nz,Barbieri:1992dq,Oliver:2002up,Abe:2009ni}
\begin{eqnarray}
\delta g_L^{b\bar{b}} = \frac{v}{2} A \ .
\label{eq:generalZBB}
\end{eqnarray}
Expanding the amplitude in powers of $m_t^2/M_q^2$ we obtain
\begin{eqnarray}
(\delta g_L^{b\bar{b}})_{\rm LW}=-\frac{m_t^4}{32\pi^2 v^2 M_q^2}\Bigg[\left(\frac{4}{r_t^2}+1\right)\log\frac{M_q^2}{m_t^2}
+\frac{4-11 r_t^2+9 r_t^4}{r_t^2(1-r_t^2)^3}\log r_t^2 -\frac{6-10 r_t^2 + 2r_t^4}{r_t^2(1-r_t^2)^2}\Bigg]
\label{eq:Zbbnondeg}
\end{eqnarray}
for nondegenerate LW fermion masses, and
\begin{eqnarray}
(\delta g_L^{b\bar{b}})_{\rm LW}=-\frac{m_t^4}{32\pi^2 v^2 M_q^2}\Bigg[5\log\frac{M_q^2}{m_t^2}-\frac{49}{6}\Bigg]
\label{eq:Zbbdeg}
\end{eqnarray}
for degenerate LW masses. Both of these expressions agree with the dominant contribution found in the effective theory, Eq.~(\ref{eq:Zbbleading}).
\begin{figure}[!t]
\includegraphics[width=3.5in,height=2.0in]{bb.eps}
\caption{Diagram giving the largest contribution to the $Z b_L \bar{b}_L$ coupling. The latter is related through the Ward identity to the $\phi^0 b_L \bar{b}_L$ coupling -- where $\phi^0$ is the Goldstone boson eaten by the $Z$ boson. The top quarks running in the loop are both ordinary and LW.}
\label{fig:Zbb}
\end{figure}
The experimental value of $g_L^{b\bar{b}}$ is derived from measurements of $R_b$, the ratio of the $Z\to b \bar{b}$ width to the width of the hadronic decays, and $A_b$, the forward-backward asymmetry for $Z$ decays into $b\bar{b}$~\cite{:2005ema}:
\begin{eqnarray}
(g_L^{b\bar{b}})_{\rm exp}=-0.4182\pm 0.0015 \ .
\end{eqnarray}
The SM value was computed using ZFITTER~\cite{Bardin:1999yd,Arbuzov:2005ma} in Ref.~\cite{SekharChivukula:2009if}, leading to
\begin{eqnarray}
(g_L^{b\bar{b}})_{\rm SM}=-\frac{1}{2} + \frac{1}{3}\sin ^2 \theta_W
+ (\delta g_L^{b\bar{b}})_{\rm SM}=-0.42114 \ ,
\end{eqnarray}
while the LW prediction is given by Eq.~(\ref{eq:Zbbnondeg}) and Eq.~(\ref{eq:Zbbdeg}). In Fig.~\ref{fig:Zbbexclusion} we show the experimental mean value (red thick horizontal line), the 2$\sigma$ allowed region below the mean value, the SM prediction (solid horizontal black line), the all-order (in $v^2/M_q^2$) LW prediction (solid blue curve), the leading order LW prediction from Eq.~(\ref{eq:Zbbdeg}) (dashed blue curve), and the leading-log approximation (dotted blue curve), as functions of $M_q$, in the degenerate case. Note that the dashed curve and Eq.~(\ref{eq:Zbbdeg}) are not reliable for $M_q\lesssim 1$ TeV, because the perturbative diagonalization of the mass matrix is not valid in that mass regime.
As anticipated by the effective field theory calculation, the LW correction is always negative: this is essentially due to the negative sign in front of the dominant nonstansdard triangle diagrams with one LW top and one SM top. It is large (for small values of $M_q$) because of the explicit breaking of custodial isospin symmetry. Since the SM value is already 1.96$\sigma$ below the experimental mean value, this correction goes in the direction opposite to what is needed. Agreement at the 2$\sigma$ level requires $M_q\gtrsim 4$ TeV; at 2.5$\sigma$ this bound relaxes to $M_q\gtrsim$ 700 GeV.
\begin{figure}[!t]
\includegraphics[width=4in]{Zbb.eps}
\caption{Constraints from the $Z b_L \bar{b}_L$ coupling. This graph features the experimental mean value (red thick horizontal line), the 2$\sigma$ allowed region below the mean value, the SM prediction (solid horizontal black line), the all-order (in $v^2/M_q^2$) LW prediction (solid blue curve), the leading order LW prediction from Eq.~(\ref{eq:Zbbdeg}) (dashed blue curve), and the leading-log approximation (dotted blue curve), as functions of $M_q$, in the degenerate case.}
\label{fig:Zbbexclusion}
\end{figure}
\section{Conclusions}\label{sec:conclusions}
There is significant tension between naturalness and isospin violation in the Lee-Wick Standard Model (LW SM). While corrections to the Higgs mass are smallest when the LW partners of the gauge bosons and fermions are light, isospin violation that must be present in the top sector to account for the large splitting between $m_t$ and $m_b$ tends to constrain the LW partners to have masses over a TeV. We have performed an effective field theory analysis of the corrections to $\hat{T}$ and the $Z b_L \bar{b}_L$ coupling in the LW SM and used it to confirm our full calculation of the LW effects on $\hat{S}$, $\hat{T}$, $W$, $Y$ and $(g_L^{b\bar{b}})$ including tree-level and fermionic one-loop contributions. The post-LEP parameters yield a simple picture in the LW SM: the gauge sector contributes to $Y$ and $W$ only, with leading contributions arising at tree-level, while the fermion sector contributes to $\hat{S}$ and $\hat{T}$ only, with leading corrections arising at one loop.
In agreement with \cite{Underwood:2008cr}, we find that experimental limits on $W$ and $Y$ jointly constrain the masses of the LW gauge bosons to satisfy $M_1, M_2 \gtrsim 2.3$ TeV at 95\% CL, with relatively little sensitivity to the Higgs mass.
We also conclude that the experimental limits on $\hat{T}$ require the masses of the LW fermions to satisfy $M_q, M_t \gtrsim 1.5$ TeV at 95\% CL if the Higgs mass is light and tend to exclude the LW SM for any LW fermion masses if the Higgs mass is heavy. This is because a model containing a heavy Higgs can be rendered consistent with the data only if some other sector of the model provides a large positive correction to $\hat{T}$. However, in the LW SM, the fermionic loops that provide the dominant contribution to $\hat{T}$ always make $\hat{T}$ more negative, due to the negative sign in the LW fermion propagators. The LW fermions simply cannot compensate for the presence of a heavy Higgs. Our results differ from those in Refs.~\cite{Alvarez:2008ks, Carone:2008bs} because their analysis incorrectly assumes that the electroweak corrections due to LW states are purely oblique. As explained in Ref.~\cite{Underwood:2008cr} the LW states actually induce important non-oblique corrections, and one must therefore use the Barbieri {\em et. al.} ~\cite{Barbieri:2004qk,Chivukula:2004af} post-LEP parameters to compare the LW SM with experiment, as we have done.
Weak isospin violation in the top-quark sector also manifests itself through corrections to the $Zb_L \bar{b}_L$ coupling. The SM prediction for $(g_L^{b\bar{b}})$ lies at the lower end of the range allowed by experiment at 95\% CL, so that new physics making negative contributions to the value of $(g_L^{b\bar{b}})$ would decrease the agreement with the data. As in the case of $\hat{T}$, however, we find that the negative sign in the LW fermion propagators translates into a negative contribution to $(g_L^{b\bar{b}})$; the lighter the LW fermions, the greater the disagreement between prediction and data. We find that the $Z b_L \bar{b}_L$ coupling places a lower bound of 4 TeV on the LW fermion masses at 95\% CL.
|
\section{Introduction}
\subsection{Notation}
Let $K$ be an algebraically closed field of characteristic $p>3$ and
let $W$ be the $p$--dimensional $\mathbbm{Z}$--graded restricted simple Witt
algebra with basis $e_{-1},e_0,\dots,e_{p-2}$. The
Lie bracket is given by
\[
[e_i,e_j]=(j-i)e_{i+j} \quad \text{for all} \quad -1\leq i,j\leq p-2,
\]
where by definition $e_i=0$ for $i\not\in\{-1,0,\dots,p-2\}$. The
$p$--mapping is given by
\[
e_i^{[p]}=\delta_{i0}e_i\quad \text{for all} \quad -1\leq i\leq p-2.
\]
For each $ -1\leq i\leq p-1$ we set \nomenclature[110]{$W_i$}{\nomrefpage}
\[
W_i=Ke_i\oplus\cdots\oplus Ke_{p-2}
\]
and we define \textit{the height} $r(\chi)$ of a functional $\chi\in W^*$ to be
\[
r(\chi)=\operatorname{min}\{i\mid -1\leq i\leq p-1 \text{ and
}\chi_{|W_i}=0\}.
\]
\subsection{Simple modules} The simple $W$--modules were determined by
Chang in \cite{Chang}. We describe here very briefly the simple $W$--modules
with a $p$--character $\chi$ of height $-1$, $0$, $1$ and $p-1$.
\subsubsection{} Suppose that $\chi\in W^*$ has height at most 1. For
each weight $\lambda\in\Lambda(\chi)=\{\lambda\in K\mid
\lambda^p-\lambda=\chi(e_0)^p\}$ consider the $\chi-$reduced Verma module
$V_\chi(\lambda)=U_\chi(W)\otimes_{U_\chi(W_0)} K_\lambda$ where
$K_\lambda$ denotes $K$ considered as a $U_\chi(W_0)$--module via
\[
e_0\cdot1=\lambda\cdot1\quad\text{and}\quad e_i\cdot1=0 \ \text{ for all
}\ i\neq0.
\]
\begin{Theorem}The following hold
\begin{enumerate}[(1)]
\item If $\chi$ has height $-1$, then there are $p$ isomorphism classes of simple $U_{\chi}(W)$--modules. These modules are represented by the one-dimensional trivial $W$--module $K$, the $(p-1)$-dimensional module $S=V_\chi(p-1)/\Soc_{U_\chi(W)}V_\chi(p-1)$ and the $p$-dimensional modules $V_\chi(\lambda)$ for $\lambda\in\{1,2,\dots,p-2\}$.
\item If $\chi$ has height $0$, then there are $p-1$ isomorphism classes of
simple $U_\chi(W)$--modules each of dimension $p$ and represented by
$V_\chi(\lambda)$ for $\lambda\in\{0,1,\dots,p-2\}$.
\item If $\chi$ has height $1$, then there are $p$ isomorphism classes of simple $U_\chi(W)$--modules each of dimension $p$ and represented by $V_\chi(\lambda)$ for $\lambda\in\Lambda(\chi)$.
\end{enumerate}
\end{Theorem}
For each $\mu\in\mathbbm{F}_p=
\mathbbm{Z}/p\mathbbm{Z}$ we let $[\mu]\in\{0,1,\dots,p-1\}$ denote the unique
representative of $\mu$. Note that $\lambda-\lambda'\in\mathbbm{F}_p$ for
all $\lambda,\lambda'\in\Lambda(\chi)$. Furthermore, if $\chi$ has height $1$ then $\lambda+\lambda'\not\in\mathbbm{F}_p$ since otherwise this would imply $\lambda,\lambda'\in\mathbbm{F}_p$.
\subsubsection{}If $\chi\in W^*$ has height $p-1$ then every simple
$U_\chi(W)$--module with one exception is projective. The remaining
simple module $L$ has a projective cover with two composition factors
both isomorphic to $L$, see \cite[Theorem 2.6]{FN}.
\begin{Remark}
If $1<r(\chi)<p-1$ then there exists a unique simple
$U_\chi(W)$--module up to isomorphism \cite[Hauptsatz 1]{Chang}. This module has a non-trivial self-extension.
\end{Remark}
\section{Extensions of the reduced Verma modules}\label{Verma}
\subsection{}Let $\chi\in W^*$ with $r(\chi)=-1,0,1$ and let $\lambda,\lambda'\in\Lambda(\chi)$. The isomorphism
\[
\Ext_{U_\chi(W)}(V_\chi(\lambda'),V_\chi(\lambda))\simeq \Ext_{U_{\chi} (W_0)}(K_{\lambda'},V_{\chi} (\lambda))
\]
reduces the problem of classifying the extensions of Verma modules to
that of classifying the extensions of $K_{\lambda'}$ by
$V_\chi(\lambda)$.
\subsection{}\label{Verma2}Consider a short exact sequence of $U_{\chi}(W_0)$--modules
\begin{equation}\label{seq1}
0\rightarrow V_\chi(\lambda)\xrightarrow{f}M\xrightarrow{g} K_{\lambda'}\rightarrow 0.
\end{equation}
The Verma module $V_\chi(\lambda)$ has a basis $v_0,v_1,\dots,v_{p-1}$
given by $v_i=e_{-1}^i\otimes1$. Set $w_i=f(v_i)$ for all $i$. Since $U_\chi(Ke_0)$ is a semisimple algebra, there exists $w'\in M$ such that
$e_0w'=\lambda' w'$ and $g(w')=1$. Thus we obtain a basis for $M$
\begin{equation}\label{basis}
w_0,w_1,\dots,w_{p-1},w',
\end{equation}
such that for all $0\leq j\leq p-2$
\begin{equation}\label{e_j1}
e_jw_i=
\begin{cases}
(-1)^{j}\frac{i!}{(i-j)!}((j+1)\lambda-i+j)w_{i-j}, & \text{if }j\leq i,\\
0, & \text{otherwise}.
\end{cases}
\end{equation}
Furthermore, if $1\leq j\leq p-2$ then
\[
e_0e_jw'=(e_je_0+je_j)w'=(\lambda'+j)e_jw',
\]
hence $e_jw'$ belongs to the 1-dimensional weight space $M_{\lambda'+j}$ spanned by $w_{[\lambda-\lambda'-j]}$. It follows that
\begin{equation}
\label{eq:3}
e_jw'=\mathfrak{a}_jw_{[\lambda-\lambda'-j]} \quad \text{for some}\quad \mathfrak{a}_j\in K.
\end{equation}
\subsection{}\label{welldef}The $w'$ from the previous section is not unique. Since the $\lambda'$ weight space in $M$ is
spanned by $w'$ and $w_{[\lambda-\lambda']}$ any different
choice for $w'$ has the form $w'+bw_{[\lambda-\lambda']}$ for some
$b\in K$. Obviously, this leads to the same $e_jw'$ if
$e_jw_{[\lambda-\lambda']}=0$. Thus, if $\lambda=\lambda'$ then
$(\a_1,\a_2,\dots,\a_{p-2})$ is determined by the extension \eqref{seq1}. Furthermore, we have
\begin{Lemma}[A]
Let $\lambda,\lambda'\in\Lambda(\chi)$. If
$(\lambda,\lambda')=(0,p-1)$ or $(\lambda',\lambda)=(p-1,0)$ then $e_jw_{[\lambda-\lambda']}=0$ for all $j\geq1$.
\end{Lemma}
Note that the assumptions of the lemma imply $r(\chi)<1$ as $\lambda,\lambda'\not\in\mathbbm{F}_p$ for $r(\chi)=1$.
We let $\Theta(\chi)\subset \Lambda(\chi)\times\Lambda(\chi)$ denote the subset given by \nomenclature[230]{$\Theta(\chi)$}{\nomrefpage}
\[
(\{(0,p-1),(p-1,0)\}\cap\Lambda(\chi)\times\Lambda(\chi))\cup \{(\mu,\mu)\mid \mu\in\Lambda(\chi)\}.
\]
Lemma A and the discussion at the beginning of this section show that all the $\mathfrak{a_j}$ are determined by $M$ if $(\lambda,\lambda')\in\Theta(\chi)$. This is not the case for $(\lambda,\lambda')\not\in\Theta(\chi)$. However, we have
\begin{Lemma}[B]
Let $\lambda,\lambda'\in\Lambda(\chi)$ such that $(\lambda,\lambda')\not\in\Theta(\chi)$. If furthermore $\lambda+\lambda'\neq p-1$
\textnormal{(}resp. $\lambda+\lambda'= p-1$\textnormal{)} then
there is a unique choice for $w'$ such that $e_1w'=0$
\textnormal{(}resp. $e_2w'=0$\textnormal{)}.
\end{Lemma}
The proof is straightforward, given the fact that $e_1w_{[\lambda-\lambda']}\neq0$
(resp. $e_2w_{[\lambda-\lambda']}\neq0$) if $\lambda+\lambda'\neq p-1$ (resp. $\lambda+\lambda'= p-1$).
\begin{Remark}For each pair $(\lambda,\lambda')\not\in\Theta(\chi)$,
we will always assume that $w'$ is the unique choice from the
lemma. It follows, then, that the tuple $(\a_1,\a_2,\dots,\a_{p-2})$ determines
the extension \eqref{seq1}. We call this tuple \textit{the $\a$--datum of \eqref{seq1}}.
\end{Remark}
\subsection{}\label{Phi}Consider now a second extension
\begin{equation}\label{seq2}
0\rightarrow V_\chi(\lambda)\xrightarrow{f'}M'\xrightarrow{g'}K_{\lambda'}\rightarrow 0
\end{equation}
of $U_{\chi}(W_0)$--modules. If \eqref{seq1} and \eqref{seq2} are equivalent
extensions then there is an isomorphism $h:M\xrightarrow{\sim}M'$
compatible with the identities in both $V_\chi(\lambda)$ and $K_{\lambda'}$
\begin{equation*}
\begin{CD}
0 @>>> V_\chi(\lambda) @>f>> M @>g>> K_{\lambda'} @>>> 0\\
@. @| @VVhV @| \\
0 @>>> V_\chi(\lambda) @>f'>> M' @>g'>> K_{\lambda'} @>>> 0.
\end{CD}
\end{equation*}
We can choose $w''=h(w')$ to be the analogue to $w'$. Then $e_iw''=h(e_iw')=h(f(\mathfrak{a}_iv_{[\lambda-\lambda'-i]}))=f'(\mathfrak{a}_iv_{[\lambda-\lambda'-i]})$ for all $1\leq i\leq p-2$. Hence $f^{-1}(e_iw')$ depends only on the class of the extension \eqref{seq1} and we obtain a well-defined map
\begin{equation}\label{hom}
\Phi^i_{\lambda,\lambda'}:\Ext_{U_{\chi}(W_0)}(K_{\lambda'},V_\chi(\lambda))\rightarrow K,
\end{equation}
that sends the class of \eqref{seq1} to $\a_i$. One checks that
$\Phi^i_{\lambda,\lambda'}$ is linear. In Section \ref{a_i} we show
that every $\mathfrak{a}_i$ with $i\geq 3$ can be expressed as a
linear combination of $\mathfrak{a}_1$ and $\mathfrak{a}_2$. It
follows that $\Phi_{\lambda,\lambda'} =(\Phi^1_{\lambda,\lambda'},\Phi^2_{\lambda,\lambda'})$ maps $\Ext_{U_\chi(W_0)}(K_{\lambda'},V_\chi(\lambda))$ injectively into $K^2$.
\begin{Remark}If $(\lambda,\lambda')\not\in\Theta(\chi)$ then $\dim \Ext_{U_\chi(W_0)}(K_{\lambda'},V_\chi(\lambda))\leq1$. This is a consequence of Lemma \ref{welldef} B.
\end{Remark}
\subsection{}\label{a_i}Since $M$ is a $U_{\chi}(W_0)$--module, we have
\begin{equation*}
[e_i,e_j]w'=(e_ie_j-e_je_i)w' \ \text{ for all }i \text{ and }j.
\end{equation*}
We will need the full strength of this formula later, but for now we
are content to remark that in the case where $i=1$ and $2\leq j\leq
p-3$, it yields
\begin{align*}
(j-1)\mathfrak{a}_{j+1}={}&-(-1)^j\biggl(\prod_{k=1}^{j}(\lambda-\lambda'-k)\biggr)(j(\lambda+1)+\lambda'+1)\mathfrak{a}_1\\
&-(\lambda-\lambda'-j)(\lambda+\lambda'+j+1)\mathfrak{a}_j,
\end{align*}
or, equivalently, by induction
\begin{equation}\label{A1}
\mathfrak{a}_j=A_j\mathfrak{a}_1+B_j\mathfrak{a}_2 \quad \text{for all}\quad 3\leq j\leq p-2,
\end{equation}
where
\begin{align*}
A_j={}&\frac{(-1)^j}{j-2}\biggl(\prod_{k=1}^{j-1}(\lambda-\lambda'-k)\biggr)\biggl((j-1)\lambda+\lambda'+j+\sum_{k=4}^{j}\frac{(j-k)!}{(j-3)!}\biggr.\\
&\cdot((j+2-k)\lambda+\lambda'+(j+3-k))\prod_{l=0}^{k-4}(\lambda+\lambda'+j-l)\bigr),
\end{align*}
(the summation $\sum_{k=4}^j$ is understood to be 0 when $j=3$) and
\[
B_j=\frac{(-1)^j}{(j-2)!}\prod_{k=2}^{j-1}(\lambda-\lambda'-k)(\lambda+\lambda'+k+1).
\]
Consequently, all the $\mathfrak{a}_i$ with $i\geq 3$ are determined by $\mathfrak{a}_1$ and $\mathfrak{a}_2$.
\subsection{}\label{M_a}Let $\lambda,\lambda'\in\Lambda(\chi)$. We
want to describe all possible $\a$--data of extensions as in
\eqref{seq1}. To this end, consider an arbitrary pair $(\a_1,\a_2)\in
K^2$ and use \eqref{A1} to extend it to a tuple
$\a=(\a_1,\a_2,\dots,\a_{p-2})$ in $K^{p-2}$. In order to simplify
notation set $\a_i=0$ for all $i>p-2$. Consider a vector space
$M_\a$ with a basis $w_0,w_1,\dots,w_{p-1},w'$ and define
endomorphisms $E_j\in\End_K(M_\mathfrak{a})$ for $0\leq j\leq p-2$ such that
\begin{equation*}
E_jw_i=
\begin{cases}
(-1)^{j}\frac{i!}{(i-j)!}((j+1)\lambda-i+j)w_{i-j}, & \text{if }j\leq i,\\
0, & \text{otherwise},
\end{cases}
\end{equation*}
and
\begin{equation*}
E_jw'=
\begin{cases}
\lambda'w', & \text{if }j=0,\\
\mathfrak{a}_jw_{[\lambda-\lambda'-j]}, & \text{otherwise}.
\end{cases}
\end{equation*}
Furthermore we set $E_j=0$ for $j\not\in\{0,1,\dots,p-2\}$. A comparison with the formulas in Section \ref{Verma2} shows that $\a$
is the $\a$-datum of an extension as in \eqref{seq1} if and only if
$M_\a$ has a $W_0$--module structure such that $e_jw=E_jw$ for all
$w\in M_\a$ and all $j\geq0$ and if this $W_0$--module then has
$p$--character $\chi$. We have on $M_\a$ a structure as $W_0$--module
if and only if
\begin{equation}\label{act-1}
[E_i,E_j]w=(j-i)E_{i+j}w\quad\text{for all } i,j\geq0\text{ and } w\in M_\a.
\end{equation}
It suffices to check this for all $w$ in our basis. The linear map
$f\:V_\chi(\lambda)\rightarrow M_\a$ with $f(v_i)=w_i$ for all $i$ satisfies
$E_jf(v)=f(E_jv)$ for all $v\inV_\chi(\lambda)$ and all $j$. Consequently, \eqref{act-1}
holds for all $w_i$ and $M_\a$ has the desired structure if and only
if
\begin{equation}\label{act0}
[E_i,E_j]w'=(j-i)E_{i+j}w' \ \text{ for all }i,j\geq0.
\end{equation}
\subsection{}\label{W--module}
One verifies easily that \eqref{act0} holds for $i=0$. It also holds for
$i=j$. For $i=1$ and $2\leq j\leq p-3$ it is equivalent to \eqref{A1} which holds by the definition of the $\a_j$ ($j\geq3$). Furthermore, for every $(i,j)$ in
\[
\{(1,p-2),(p-2,1)\}\cup\{(a,b)\in\mathbbm{Z}^2\mid a\neq b \text{ and }2\leq a,b\leq p-2\},
\]
formula \eqref{act0} is equivalent to the following conditions
\medskip
\noindent If $i>[\lambda-\lambda'-j]$ and $j>[\lambda-\lambda'-i]$, then
\begin{equation}\label{A2}
\a_{i+j}=0.
\end{equation}
If $i>[\lambda-\lambda'-j]$ and $j\leq[\lambda-\lambda'-i]$, then
\begin{equation}\label{A3} (j-i)\a_{i+j}=(-1)^{j+1}\frac{[\lambda-\lambda'-i]!}{([\lambda-\lambda'-i]-j)!}(j(\lambda+1)+\lambda'+i)\a_i.
\end{equation}
If $i\leq[\lambda-\lambda'-j]$ and $j\leq[\lambda-\lambda'-i]$, then
\begin{align}\label{A4}
(j-i)\a_{i+j}={}& (-1)^{i}\frac{[\lambda-\lambda'-j]!}{([\lambda-\lambda'-j]-i)!}(i(\lambda+1)+\lambda'+j)\a_j\\
&-(-1)^{j}\frac{[\lambda-\lambda'-i]!}{([\lambda-\lambda'-i]-j)!}(j(\lambda+1)+\lambda'+i)\a_i\notag.
\end{align}
\subsection{}\label{char}Suppose now that $\a$ satisfies \eqref{A2}--\eqref{A4} and hence
that $M_\a$ is a $W_0$--module with each $e_i$ acting as $E_i$. We
have then a short exact sequence
\begin{equation}\label{seq11}
0\rightarrowV_\chi(\lambda)\xrightarrow{f}M_\a\xrightarrow{g}K_{\lambda'}\rightarrow0.
\end{equation}
Evidently, $M_\a$ has $p$-character $\chi$ if and only if $e_i^p
w'=0$ for all $i\geq1$, or equivalently, if
$\a_ie_i^{p-1}w_{[\lambda-\lambda'-i]}=0$ for all $i\geq1$. We have
\[
\a_1e_1^{p-1}w_{p-1}=-\a_1\biggl(\prod_{j=0}^{p-2}(2\lambda-j)\biggr)w_0.
\]
Observe that the term inside the bracket does not vanish if the height of $\chi$ is
1 because $\lambda\not\in\mathbbm{F}_p$ in this case. Since furthermore
$e_i^{p-1}w_{[\lambda-\lambda'-i]}=0$ if $i>1$ or
$\lambda\neq\lambda'$ we obtain the following lemma
\begin{Lemma}$M_\a$ has $p$-character $\chi$ if and only if one of the
following conditions holds
\begin{enumerate}
\item If $r(\chi)\in\{-1,0\}$ and $\lambda=\lambda'$ and $2\lambda=p-1$ then $\a_1=0$.
\item If $r(\chi)=1$ and $\lambda=\lambda'$ then $\a_1=0$.
\end{enumerate}
\end{Lemma}
\subsection{}We first look at the case where $\lambda=\lambda'$.
\begin{Proposition}\label{Self-Ext1}We have
\begin{equation*}
\Ext_{U_{\chi}(W_0)}(K_\lambda,V_{\chi}(\lambda))\simeq
\begin{cases}
K, &\text{if }\lambda\in\{0,p-1\} ,\\
0, &\text{otherwise}.
\end{cases}
\end{equation*}
\end{Proposition}
It is convenient to break up the proof of the proposition into two lemmas.
\begin{Lemma}[A]If $r(\chi)=1$ then
\[
\Ext_{U_\chi(W_0)}(K_\lambda,V_{\chi}(\lambda))=0.
\]
\end{Lemma}
\begin{proof}
Every extension of $K_\lambda$ by $V_{\chi}(\lambda)$ can be represented by a short exact sequence of $U_\chi(W_0)$--modules
\[
0\rightarrow V_{\chi}(\lambda)\rightarrow M_\a\rightarrow K_\lambda\rightarrow0.
\]
The claim to be proved amounts to saying that the above sequence splits, or equivalently, that $\a=0$. Since all the $\a_j$ can be expressed as a linear combination of $\a_1$ and $\a_2$, it suffices to prove that both $\a_1$ and $\a_2$ are 0. Note that $\a_1=0$ follows from Lemma \ref{char}. Thus, by definition, we have
\[
\a_j=(j-1)\biggl(\prod_{k=3}^{j}(2\lambda+k)\biggr)\a_2 \quad \text{for}\quad 3\leq j\leq p-2.
\]
Since the height of $\chi$ is 1 it follows that $2\lambda+k\neq0$ for all $k$. Therefore to prove $\a_2=0$ it suffices to show $\a_j=0$ for any $2\leq j\leq p-2$. Now, if we insert $(i,j)=(1,p-2)$ into \eqref{A4} and use the fact that $(p-1)!=-1$ we get
\begin{equation*}
-2(2\lambda-1)\a_{p-2}=0,
\end{equation*}
proving the claim.
\end{proof}
\begin{Lemma}[B]If $r(\chi)\in\{-1,0\}$ then
\[
\Ext_{U_\chi(W_0)}(K_\lambda,V_\chi(\lambda))\simeq
\begin{cases}
K, &\text{if }\lambda\in\{0,p-1\} ,\\
0, &\text{otherwise}.
\end{cases}
\]
\end{Lemma}
\begin{proof}
Let $(\a_1,\a_2)$ be an arbitrary pair in $K^2$ and consider the vector space $M_\a$ constructed in Section \ref{M_a}. By definition, we have
\begin{equation}\label{C1}
\a_j=A_j\a_1+B_j\a_2 \quad \text{for all} \quad 3\leq j\leq p-2,
\end{equation}
where
\begin{align*} A_j=-\frac{\lambda+1}{j-2}\biggl(j!+\sum_{k=4}^{j}(k-1)!\prod_{l=k}^{j}\frac{l-1}{l-3}(2\lambda+l)\biggr),
\end{align*}
(the summation $\sum_{k=4}^j$ is understood to be 0 when $j=3$) and
\begin{equation*}
B_j=(j-1)\prod_{k=3}^{j}(2\lambda+k).
\end{equation*}
We rewrite \eqref{A2}--\eqref{A4} for the present case. First, note that \eqref{A2} and \eqref{A3} hold trivially since $[\lambda-\lambda-i]=p-i$ and $[\lambda-\lambda-j]=p-j$ for all $1\leq i,j\leq p-2$. If we insert $(i,j)=(1,p-2)$ into \eqref{A4} and use $(p-1)!=-1$, we
see that
\begin{equation}\label{C2}
2(2\lambda-1)\a_{p-2}=(\lambda+1)\a_1.
\end{equation}
If $2\leq i,j\leq p-2$ and $i+j\leq p$, the same formula yields
\begin{align}\label{C3}
(j-i)\a_{i+j}={}&\frac{(i+j-1)!}{(j-1)!}((i+1)\lambda+i+j)\a_j\\
&-\frac{(i+j-1)!}{(i-1)!}((j+1)\lambda+i+j)\a_i\notag.
\end{align}
We now proceed by showing how these formulas are related. Let's
begin by assuming that $\lambda\not\in\{0,p-1\}$. If $2\lambda=p-1$ then Lemma \ref{char} implies $\a_1=0$, which together with \eqref{C1} implies
\begin{equation}\label{A1a}
\a_j=(j-1)(j-1)!\a_2 \quad \text{for all}\quad 3\leq j\leq p-2.
\end{equation}
It follows that $\a_2=0$ because otherwise this would mean that $\a_{p-2}\neq0$ and then by \eqref{C2} that $2\lambda=1$ in contradiction with our assumption.
\medskip
Suppose next that $2\lambda\neq p-1$. If we insert $(i,j)=(2,p-2)$ into \eqref{C3}, we obtain
\begin{equation}\label{p-2}
6\a_{p-2}=\a_2.
\end{equation}
\medskip
\noindent Assume first that $2\lambda=1$. Since $\lambda\neq p-1$ eq. \eqref{C2} implies $\a_1=0$ hence \eqref{C1} reduces to
\begin{equation*}
\a_j=\frac{1}{6}(j-1)(j+1)!\a_2 \quad \text{for all} \quad 3\leq j\leq p-2.
\end{equation*}
In particular, we have $2\a_{p-2}=\a_2$, which together with \eqref{p-2}, implies $\a_2=0$, as desired.
\medskip
\noindent Assume next that $2\lambda\neq1$. Eq. \eqref{C2} and \eqref{p-2}
give the following relation
\begin{equation}\label{rel:a_1,a_2}
3(\lambda+1)\a_1=(2\lambda-1)\a_2,
\end{equation}
which means that $\a_1$ and $\a_2$ are either both 0 or both nonzero. We assume the latter and derive a contradiction. By definition, we have
$\a_3=-6(\lambda+1)\a_1+2(2\lambda+3)\a_2$ hence it follows from \eqref{rel:a_1,a_2} that $\a_3=8\a_2$.
Thus, if we insert $(i,j)=(3,p-3)$ into \eqref{C3}, we obtain
\begin{equation*}
3\a_{p-3}=-\a_2.
\end{equation*}
Likewise, if we insert $(i,j)=(2,p-3)$ into \eqref{C3}, we obtain
\[
6(3\lambda-1)\a_{p-3}=-(2\lambda+1)\a_2.
\]
Since $2\lambda+1\neq0$ we can determine $\lambda$ by eliminating $\a_2$ and $\a_{p-3}$
\[
4\lambda=3.
\]
Note $3\a_{p-3}=-\a_2$ implies $p>5$. If we insert $(i,j)=(2,3)$ into \eqref{C3}, we see
\[
\a_5=(192\lambda+360)\a_2=504\a_2.
\]
However, by definition we have
\[
\a_5=\frac{1}{3}(128\lambda^2+656\lambda+1080)\a_2=548\a_2,
\]
which implies first that $p=11$ and then that $\lambda=9$. If we put
all this together and use \eqref{C1} we see $\a_1=\a_4$. But inserting
$(i,j)=(4,6)$ into \eqref{C3} implies $\a_4=0$ and thus $\a_1=0$. Contradiction!
\medskip
We now move to the case where $\lambda=0$. A necessary condition for $M_\a$ to be a $U_\chi(W_0)$--module is that $2\a_{p-2}=-\a_1$, cf. \eqref{C2}. Since by definition
\[
\a_{p-2}=4\a_1-\frac{3}{2}\a_2,
\]
this implies $\a_2=3\a_1$. Now, the claim to be proved amounts to saying that each pair $(\a_1,\a_2)$ with $\a_2=3\a_1$ induces a $U_\chi(W_0)$--module $M_\a$ in the way described in Section \ref{M_a}. Using our previous notation, we have
\[
A_j=-(j-2)j! \quad \text{and} \quad B_j=\frac{1}{2}(j-1)j!,
\]
hence $\a_2=3\a_1$ implies
\begin{equation*}
\a_j=\frac{1}{2}(j+1)!\a_1 \quad \text{for all} \quad 1\leq j\leq p-2.
\end{equation*}
In particular, we have $\a_{p-2}=-1/2\a_1$ in consistence with \eqref{C2}. A straightforward computation shows that
\begin{align*}
(j-i)A_{i+j}&=(i+j)!\left(\frac{A_j}{(j-1)!}-\frac{A_i}{(i-1)!}\right),\\
(j-i)B_{i+j}&=(i+j)!\left(\frac{B_j}{(j-1)!}-\frac{B_i}{(i-1)!}\right).
\end{align*}
proving \eqref{C3} for $i+j\leq p-2$. Likewise, a simple computation shows
\[
\a_j(p-j-2)!=\a_{p-j-1}(j-1)!,
\]
proving \eqref{C3} for $i+j=p-1$. Since $i+j=p$ gives 0 on both sides of \eqref{C3} the claim follows from Lemma \ref{char}.
\medskip
Finally, we consider the case where $\lambda=p-1$. Here we have $A_j=0$ and it
follows from \eqref{C2} that $\a_{p-2}=0$. Therefore since
\[
\a_j=(j-1)!\a_2 \quad \text{for all} \quad 2\leq j\leq p-2,
\]
this implies first $\a_2=0$ and then that $\a_j=0$ for all $j\geq2$. The rest of the proof is straightforward.
\end{proof}
\subsection{}\label{(0,p-1)} Next, we address the cases where $(\lambda,\lambda')=(0,p-1)$ and $(\lambda,\lambda')=(p-1,0)$. This can only occur if the height of $\chi$ is less than 1 because $\lambda,\lambda'\not\in\mathbbm{F}_p$ for $r(\chi)=1$. The arguments presented previously would apply equally well to these cases, but we will give another, simpler, proof which relies on the fact that $\Ext_{U_\chi(W_0)}(K_0,V_\chi(0))$ is 1-dimensional, cf. Proposition \ref{Self-Ext1}.
\begin{Proposition} If $(\lambda,\lambda')=(0,p-1)$ or $(\lambda,\lambda')=(p-1,0)$, then
\begin{equation*}
\Ext_{U_\chi(W_0)}(K_{\lambda'},V_\chi(\lambda))\simeq K.
\end{equation*}
\end{Proposition}
\begin{proof}
Let $\eta\in W^*$ be a functional of height $0$. We have
\[
\Ext_{U_\eta(W)}(V_\eta(0),V_\eta(0))\simeq\Ext_{U_\chi(W_0)}(K_0,V_\chi(0))\simeq K.
\]
Since the reduced Verma modules $V_\eta(0)$ and $V_\eta(p-1)$ are isomorphic \cite [Hilfssatz 7]{Chang}, we obtain
\begin{equation*}
\Ext_{U_\chi(W_0)}(K_0,V_\chi(p-1))\simeq\Ext_{U_{\eta}(W)}(V_\eta(0),V_\eta(p-1))\simeq K,
\end{equation*}
as claimed. The other case can be handled similarly.
\end{proof}
\subsection{}\label{neq p-1}It is convenient to divide the case $(\lambda,\lambda')\not\in \Theta(\chi)$ into two subcases depending on whether or not $\lambda+\lambda'=p-1$. Recall that $\lambda+\lambda'=p-1$ can only occur if $r(\chi)<1$. The following lemma will be needed in the sequel.
\begin{Lemma}
Let $i$ and $j$ be two integers with $2\leq i,j\leq p-2$.
\begin{enumerate}
\item If $i>[\lambda-\lambda'-j]$ then $i+j>[\lambda-\lambda']$.
\item If $j\leq[\lambda-\lambda'-i]$ and $i+j>[\lambda-\lambda']$ then $i>[\lambda-\lambda']$.
\end{enumerate}
\end{Lemma}
\subsection{}We denote the roots of the polynomial $x^2-a\in K[x]$ by $\pm\sqrt{a}$. \nomenclature[250]{$\pm\sqrt{a}$}{\nomrefpage}
\begin{Proposition}\label{neq}If $(\lambda,\lambda')\not\in \Theta(\chi)$ and $\lambda+\lambda'\neq p-1$, then
\begin{equation*}
\begin{array}{c}
\Ext_{U_{\chi}(W_0)}(K_{\lambda'},V_{\chi}(\lambda))\\ \ \\
\simeq\begin{cases}
K,&\text{if }[\lambda-\lambda']=1 \text{ and }\lambda\in\{1,p-1\},\\
K,&\text{if }[\lambda-\lambda']\in\{2,3\},\\
K,&\text{if }[\lambda-\lambda']=4 \text{ and }p>5,\\
K,&\text{if }[\lambda-\lambda']=4 \text{ and }\lambda\in\{0,3\},\\
K,&\text{if }[\lambda-\lambda']= 5 \text{ and }\lambda\in\{0,4\},\\
K,&\text{if }[\lambda-\lambda']= 6 \text{ and
}2\lambda=5\pm\sqrt{19}\text{ and }p>7,\\
0,&\text{otherwise}.
\end{cases}
\end{array}
\end{equation*}
\end{Proposition}
\begin{proof}
Every extension of $K_{\lambda'}$ by $V_{\chi}(\lambda)$ can be represented by a short exact sequence of $U_{\chi}(W_0)$--modules
\begin{equation}\label{seqv}
0\rightarrow V_{\chi}(\lambda)\rightarrow M_\a\rightarrow K_{\lambda'}\rightarrow 0,
\end{equation}
where $\a_1=0$ (see Lemma \ref{welldef} B and Remark \ref{welldef}) and
\begin{equation}\label{D1}
\a_j=\frac{(-1)^j}{(j-2)!}\biggl(\prod_{k=2}^{j-1}(\lambda-\lambda'-k)(\lambda+\lambda'+k+1)\biggr)\a_2, \quad 3\leq j\leq p-2.
\end{equation}
Note that $\a_j=0$ for all $j>[\lambda-\lambda']$ if $[\lambda-\lambda']\geq2$. This is a very useful observation which we shall use several times, often without any reference. We divide the proof into several steps depending on $[\lambda-\lambda']$.
\medskip
\noindent\textbf{Case 1.} Suppose that $[\lambda-\lambda']=1$.
\medskip
\noindent For the sake of simplicity, we assume that $p>5$. Eq. \eqref{D1} becomes
\[
\a_j=\biggl(\prod_{k=2}^{j-1}(2\lambda+k)\biggr)\a_2 \quad \text{for all} \quad 3\leq j\leq p-2.
\]
If we insert $(i,j)=(2,3)$ into \eqref{A4} and then use the above
formula to express $\a_3$ as
a scalar multiple $\a_2$ we obtain
\[
\a_5=(2\lambda+2)(18\lambda+12)\a_2.
\]
A necessary condition for \eqref{seqv} to be non-split is that $\a\neq0$, or equivalently, that $\a_2\neq0$. If this is the case, then
\[
\prod_{k=2}^{4}(2\lambda+k)=(2\lambda+2)(18\lambda+12),
\]
which implies $\lambda\in\{1,p-1\}$. (It should be noted that $\lambda=0$ is a solution to the above equation, but it has been deliberately omitted because otherwise we would have $\lambda'=p-1$ and thus $\lambda+\lambda'=p-1$.)
\medskip
Conversely, suppose that $\lambda=1$ or $\lambda=p-1$. Since
$\lambda-\lambda'=1$, this implies $\lambda'=0$ or $\lambda'=p-2$,
respectively. Let $(\a_1,\a_2)$ be a pair in $K^2$ such that $\a_1=0$
and consider the corresponding vector space $M_\a$. If $\lambda=p-1$,
then $\lambda+\lambda'=p-3$ hence by definition $\a_j=0$ for all
$j\neq2$ and a very simple computation shows that \eqref{A2}--\eqref{A4} hold. Suppose next that $\lambda=1$. We have then
\[
p-2>[\lambda-\lambda'-1]\quad\text{and}\quad 1<[\lambda-\lambda'-(p-2)].
\]
If we insert $(i,j)=(p-2,1)$ into \eqref{A3}, we obtain zero on both sides because $(\lambda+1)+\lambda'-2=0$ and $\a_{p-1}=0$. Furthermore, if $2\leq i,j\leq p-2$ then
\[
[\lambda-\lambda'-i]=p+1-i \quad \text{and}\quad [\lambda-\lambda'-j]=p+1-j.
\]
Thus, we may assume that
\[
i\leq[\lambda-\lambda'-j] \quad \text{and}\quad j\leq[\lambda-\lambda'-i],
\]
or equivalently that $i+j\leq p+1$. We have
\[
\a_k=\frac{1}{6}(k+1)!\a_2 \quad \text{for all}\quad 2\leq k\leq p-2,
\]
which, when inserted into \eqref{A4}, yields
\begin{align*}
(j-i)\a_{i+j}={}& \frac{1}{6}(i+j-2)!(2i+j)(j-1)j(j+1)\a_2\\
&-\frac{1}{6}(i+j-2)!(2j+i)(i-1)i(i+1)\a_2.
\end{align*}
Since
\[
(j-i)\a_{i+j}=
\begin{cases}
\frac{1}{6}(j-i)(i+j+1)!\a_2, &\text{if }i+j\leq p-2,\\
0, &\text{if }i+j\in\{p,p\pm1\},
\end{cases}
\]
the claim follows from a straightforward computation.
\medskip
\noindent\textbf{Case 2.} Suppose that $[\lambda-\lambda']\in\{2,3\}$.
\medskip
\noindent Let $(\a_1,\a_2)$ be a pair in $K^2$ such that $\a_1=0$ and
consider the corresponding vector space $M_\a$. Lemma \ref{neq p-1}
together with the remark following eq. \eqref{D1} yields \eqref{A2}--\eqref{A4}
for $2\leq i,j\leq p-2$. Since furthermore $\a_{p-2}=0$ for $p>5$ this
proves \eqref{A2}--\eqref{A4} for all $i,j$ in the case $p>5$. (Note that the above reasoning works just as well for $\lambda-\lambda'=4$.) The case $p=5$ is straightforward and is omitted here.
\medskip
\noindent\textbf{Case 3.} Suppose that $[\lambda-\lambda']=4$.
\medskip
\noindent As noted above, we may assume $p=5$ and $(i,j)\in\{(1,3),(3,1)\}$. We have
\[
3=[\lambda-\lambda'-1] \quad \text{and}\quad 1=[\lambda-\lambda'-3].
\]
Thus, if we insert $(i,j)=(1,3)$ into \eqref{A4} and then use the definition of $\a_3$ to express it as a scalar multiple $\a_2$, we obtain
\[
2(\lambda+\lambda'+3)(\lambda+\lambda'+4)\a_2=0.
\]
Now, every extension of $K_{\lambda'}$ by $V_{\chi}(\lambda)$ can be represented by a short exact sequence as in \eqref{seqv} and such that $\a_1=0$. The above computation shows that a necessary (and sufficient) condition for \eqref{seqv} to be non-split is that $\lambda+\lambda'\in\{1,2\}$, or equivalently, that $\lambda\in\{0,3\}$.
\medskip
\noindent\textbf{Case 4.} Suppose that $[\lambda-\lambda']=5$.
\medskip
\noindent First, note that the assumption implies $p>5$ since $[\lambda-\lambda']<p$. We consider a short exact sequence of $U_{\chi}(W_0)$--modules as in \eqref{seqv}. If we insert
$(i,j)=(2,3)$ into \eqref{A4} and then use the definition of
$\a_5$ to express it as a scalar multiple $\a_2$, we get
\begin{align}\label{23} \frac{1}{6}\biggl(\prod_{k=3}^5(\lambda+\lambda'+k)\biggr)\a_2=&\ (2\lambda+\lambda'+5)(\lambda+\lambda'+3)\a_2\\ \notag
&-(3\lambda+\lambda'+5)\a_2,
\end{align}
which since $\lambda+\lambda'\neq p-1$ can be rewritten as
\begin{equation*}
((\lambda+\lambda')(\lambda+\lambda'+5)-6\lambda)\a_2=0.
\end{equation*}
Thus, the term inside the parentheses must equal 0 in order for \eqref{seqv} to be non-split. Note that the above reasoning works just as well in the case $[\lambda-\lambda']\geq5$. We shall make use of this later, but for now we are
content to remark that in our present case this implies $\lambda\in\{0,4\}$.
\medskip
Conversely, suppose that $\lambda\in\{0,4\}$ and let $(\a_1,\a_2)$ be
a pair in $K^2$ such that $\a_1=0$. We extend $(\a_1,\a_2)$ as usual
to a tuple $\a\in K^{p-2}$. The choice of $\lambda$ together with
Lemma \ref{neq p-1} implies \eqref{A2}--\eqref{A4} for $2\leq i,j\leq
p-2$ hence we may assume $(i,j)\in\{(1,p-2),(p-2,1)\}$.
If $p=7$ then
\[
p-2>[\lambda-\lambda'-1]\quad \text{and}\quad 1>[\lambda-\lambda'-(p-2)].
\]
Since by definition $\a_{p-1}=0$, this proves \eqref{A2}. If $p>7$, then
\[
p-2>[\lambda-\lambda'-1]\quad \text{and}\quad 1<[\lambda-\lambda'-(p-2)].
\]
The remark following eq. \eqref{D1} implies $\a_{p-2}=0$ and thus \eqref{A3}, thereby proving the claim.
\medskip
\noindent\textbf{Case 5.} Suppose that $[\lambda-\lambda']=6$.
\medskip
\noindent As previously noted, every nontrivial extension of
$K_{\lambda'}$ by $V_{\chi}(\lambda)$ can be represented by a sequence
of $U_{\chi}(W_0)$--modules as in \eqref{seqv} and such that $\a_1=0$
and $(\lambda+\lambda')(\lambda+\lambda'+5)=6\lambda$. In our present case, this amounts to
\[
\lambda=\frac{1}{2}(5\pm\sqrt{19}).
\]
For $r(\chi)<1$ we have $\lambda\in\mathbbm{F}_p$ and since $x^2-19$ does not split in $\mathbbm{F}_7[x]$ this shows that $p>7$. For $r(\chi)=1$ we have $\lambda+\lambda'\not\in\mathbbm{F}_p$. If $p=7$ then
\[
p-2=[\lambda-\lambda'-1] \quad \text{and}\quad 1=[\lambda-\lambda'-(p-2)].
\]
But if we insert $(i,j)=(1,p-2)$ into \eqref{A4} and then express $\a_{p-2}$ as a multiple of $\a_2$ we obtain
\[
\frac{1}{6}(\lambda+\lambda'-1)\biggl(\prod_{k=2}^{4}(\lambda-\lambda'-k)(\lambda+\lambda'+k+1)\biggr)\a_2=0.
\]
This is only possible if $\a_2=0$, or equivalently, if $\a=0$ contradicting the fact that the sequence \eqref{seqv} does not split.
\medskip
Conversely, assume that $p>7$ and let $\lambda$ be as above. Let
$(\a_1,\a_2)\in K^2$ be a pair such that $\a_1=0$ and consider the
corresponding vector space $M_\a$. The remark following eq. \eqref{D1}
gives \eqref{A2}--\eqref{A4} for $(i,j)\in\{(1,p-2),(p-2,1)\}$. By
virtue of Lemma \ref{neq p-1} we may assume that
$i+j\leq[\lambda-\lambda']$. Since furthermore we have chosen $\lambda$ in such a way that \eqref{A4} holds for $(i,j)=(2,3)$, it suffices to consider the case $(i,j)\in\{(2,4),(4,2)\}$. We have
\[
4=[\lambda-\lambda'-2] \quad \text{and}\quad 2=[\lambda-\lambda'-4].
\]
If we insert $(i,j)=(2,4)$ into \eqref{A4} and then use the definition of $\a_4$ to express it as a scalar multiple $\a_2$, we obtain
\begin{align*}
\frac{1}{12}\biggl(\prod_{k=3}^6(\lambda+\lambda'+k)\biggr)\a_2={}& \frac{1}{2}(2\lambda+\lambda'+6)\biggl(\prod_{k=3}^4(\lambda+\lambda'+k)\biggr)\a_2\\ \notag
&-(4\lambda+\lambda'+6)\a_2,
\end{align*}
which gives the equation
\[
(\lambda+\lambda'+1)(\lambda+\lambda'+6)((\lambda+\lambda')(\lambda+\lambda'+5)-6\lambda)\a_2=0,
\]
that holds by the definition of $\lambda$.
\medskip
\noindent\textbf{Case 6.} Suppose that $[\lambda-\lambda']\geq7$.
\medskip
\noindent Every extension of $K_{\lambda'}$ by $V_{\chi}(\lambda)$ can
be represented by a short exact sequence of $U_{\chi}(W_0)$--modules
as in \eqref{seqv} and such that $\a_1=0$. A necessary condition for
\eqref{seqv} to be non-split is that
$(\lambda+\lambda')(\lambda+\lambda'+5)=6\lambda$ which, in our present case, implies
\begin{equation}\label{relat}
\lambda'=\frac{1}{2}( -5\pm\sqrt{24\lambda+25})-\lambda.
\end{equation}
Assume towards contradiction that $\a_2\neq0$. We have
\[
5\leq[\lambda-\lambda'-2] \quad \text{and}\quad 2\leq[\lambda-\lambda'-5].
\]
Thus, if we insert $(i,j)=(2,5)$ into \eqref{A4} and then use the definition of $\a_5$ to write it as a scalar multiple $\a_2$, we get
\begin{align*}
\frac{1}{40}\biggl(\prod_{k=3}^7(\lambda+\lambda'+k)\biggr)\a_2={}& \frac{1}{6}(2\lambda+\lambda'+7)\biggl(\prod_{k=3}^5(\lambda+\lambda'+k)\biggr)\a_2\\ \notag
&-(5\lambda+\lambda'+7)\a_2.
\end{align*}
which, together with \eqref{relat}, yields
\[
\pm(\lambda^2+5\lambda+4)\sqrt{24\lambda+25}=15\lambda^2+35\lambda+20.
\]
Squaring both sides and subtracting one from the other yields the equation
\[
\frac{9}{50}\lambda(\lambda-1)(\lambda+1)^2(3\lambda+2)=0,
\]
which in turn yields $\lambda\in\{0,\pm1,-2/3\}$. Now we can insert this back into \eqref{relat} to determine $\lambda'$. But first note that if
\begin{equation*}
\lambda'=\frac{1}{2}( -5-\sqrt{24\lambda+25})-\lambda,
\end{equation*}
then $\lambda=p-1$ because $0,1,-2/3$ are not solutions to
\begin{equation*}
-(\lambda^2+5\lambda+4)\sqrt{24\lambda+25}=15\lambda^2+35\lambda+20.
\end{equation*}
Putting all of this together, we obtain
\[
(\lambda,\lambda')\in\{(p-1,p-1),(p-1,p-2),(1,0),(0,0),(-2/3,-1/3)\}.
\]
But none of these cases is possible since by assumption
$[\lambda-\lambda']\geq7$ and $\lambda+\lambda'\not=p-1$; a contradiction which can only be avoided if $\a_2=0$. This completes the proof of the proposition.
\end{proof}
\subsection{}We are left with the case where
$(\lambda,\lambda')\not\in \Theta(\chi)$ and
$\lambda+\lambda'=p-1$. Note that this implies $r(\chi)<1$.
\begin{Proposition} If $(\lambda,\lambda')\not\in \Theta(\chi)$ and $\lambda+\lambda'=p-1$, then
\begin{equation*}
\Ext_{U_\chi(W_0)}(K_{\lambda'},V_\chi(\lambda))\simeq
\begin{cases}
K,&\text{if } [\lambda-\lambda']\in\{2,3,4\}, \\
K,&\text{if } [\lambda-\lambda']=6 \text{ and }p=19,\\
0,&otherwise.
\end{cases}
\end{equation*}
\begin{proof}
Let $(\lambda,\lambda')\not\in \Theta(\chi)$ such that
$\lambda+\lambda'= p-1$. Every extension of $K_{\lambda'}$ by
$V_\chi(\lambda)$ can be represented by a short exact sequence of
$U_\chi(W_0)$--modules as \eqref{seqv}. The assumption
$\lambda+\lambda'= p-1$ implies $\a_2=0$, see Lemma \ref{welldef} B
and Remark \ref{welldef}. Using the foregoing notation, we have
\begin{align*} A_j={}&\frac{(-1)^j}{j-2}\biggl(\prod_{k=0}^{j-2}(2\lambda-k)\biggr)\biggl((j-2)\lambda-1+j+\sum_{k=4}^{j}\frac{(j-k)!}{(j-3)!}\biggr.\\ \notag
&\cdot((j+1-k)\lambda+(j+2-k))\prod_{l=1}^{k-3}(j-l)\biggr).
\end{align*}
(The summation $\sum_{k=4}^j$ is understood to be 0 when $j=3$.) Note that $\a_j=0$ for all $j>[\lambda-\lambda']$. Furthermore, it should also be noted that $\lambda-\lambda'\neq1$ since otherwise we would have $(\lambda,\lambda')=(0,p-1)$ in contradiction to the assumption that $(\lambda,\lambda')\not\in\Theta(\chi)$.
\medskip
\noindent\textbf{Case 1.} Suppose that $[\lambda-\lambda']\in\{2,3,4\}$.
\medskip
\noindent Conversely, every pair $(\a_1,\a_2)\in K^2$ with $\a_2=0$
gives rise to a module $M_\a$ as described in Section \ref{M_a}: Lemma
\ref{neq p-1} yields \eqref{A2}--\eqref{A4} for $2\leq i,j\leq
p-2$. For $(i,j)\in\{(1,p-2),(p-2,1)\}$ the computations are simple
and are omitted for brevity.
\medskip
\noindent\textbf{Case 2.} Suppose that $[\lambda-\lambda']=5$.
\medskip
\noindent Consider a short exact sequence of $U_\chi(W_0)$--modules as in \eqref{seqv}. If we insert $(i,j)=(2,3)$ into \eqref{A4} and then insert the expressions $\a_3=A_3\a_1$ and $\a_5=A_5\a_1$ into the result we get
\begin{align*}\lambda(\lambda+2)(\lambda+4)\biggl(\prod_{k=1}^3(2\lambda-k)\biggr)\a_1=\frac{1}{3}\lambda(13\lambda+22)\biggl(\prod_{k=1}^3(2\lambda-k)\biggr)\a_1.
\end{align*}
Now, the assumptions $\lambda+\lambda'=p-1$ and $(\lambda,\lambda')\not\in\Theta(\chi)$ imply $\lambda\neq0$. Since, in addition, $[\lambda-\lambda']=5$, the above equation reduces to
\[
3(\lambda+2)(\lambda+4)\a_1=(13\lambda+22)\a_1,
\]
which, by subtracting the right side from the left side, gives
\[
(3\lambda^2+5\lambda+2)\a_1=0.
\]
We get a quadratic equation which can be solved easily; the roots are $p-1$ and $-2/3$. But, we can immediately exclude $\lambda=p-1$ since this would imply $(\lambda,\lambda')=(p-1,0)$. All the above reasoning works just as well in the case where $[\lambda-\lambda']\geq5$. We shall make use of this later, but for now we are content to remark that in our present case $\lambda+\lambda'=p-1$ implies $(\lambda,\lambda')=(-2/3,-1/3)$ contradicting the fact that $\lambda-\lambda'=5$.
\medskip
\noindent\textbf{Case 3.} Suppose that $[\lambda-\lambda']=6$.
\medskip
\noindent As noted previously, a necessary condition for \eqref{seqv}
to be non-split is that $\lambda=-2/3$. The assumption
$\lambda+\lambda'=p-1$ implies $\lambda-\lambda'=2\lambda+1$ which
since $\lambda-\lambda'=6$ means $\lambda=5/2$ and hence
$p=19$. Conversely, let $p=19$ and $(\lambda,\lambda')=(12,6)$. (Note
$12=5/2$ in characteristic $p=19$.) Let $(\a_1,\a_2)$ be a pair in
$K^2$ such that $\a_2=0$ and consider the corresponding vector space
$M_\a$. Eq. \eqref{A2}--\eqref{A4} follow immediately for
$(i,j)\in\{(1,p-2),(p-2,1)\}$ since $\a_{p-2}=0$. By virtue of Lemma \ref{neq p-1} we may assume that $(i,j)\in\{(2,4),(4,2)\}$. We have
\[
2=[\lambda-\lambda'-4]\quad \text{and}\quad 4=[\lambda-\lambda'-2].
\]
If we insert $(i,j)=(2,4)$ into \eqref{A4} and then insert the expressions $\a_4=A_4\a_1$ and $\a_6=A_6\a_1$ into the result, we get
\[ \frac{1}{3}\lambda(77\lambda+125)\biggl(\prod_{k=1}^{4}(2\lambda-k)\biggr)\a_1=\lambda(\lambda+5)(5\lambda+9)\biggl(\prod_{k=1}^{4}(2\lambda-k)\biggr)\a_1.
\]
Keeping in mind that the characteristic is 19, a very simple
computation shows that the above equation holds for all $\a_1$.
\medskip
\noindent\textbf{Case 4.} Suppose that $[\lambda-\lambda']\geq7$.
\medskip
\noindent Suppose we are given a short exact sequence of
$U_\chi(W_0)$--modules as in \eqref{seqv} and such that $\a_2=0$ and
$\lambda=-2/3$. We have
\[
2\leq[\lambda-\lambda'-5]\quad\text{and}\quad 5\leq[\lambda-\lambda'-2].
\]
If we insert $(i,j)=(2,5)$ into \eqref{A4}, we obtain
\begin{align*} \frac{6}{10}\lambda(87\lambda+137)\biggl(\prod_{k=1}^{5}(2\lambda-k)\biggr)\a_1={}&\frac{2}{3}\lambda(\lambda+6)(13\lambda+22)\\
&\cdot\biggl(\prod_{k=1}^{5}(2\lambda-k)\biggr)\a_1.
\end{align*}
Since $\lambda\neq0$ and $[\lambda-\lambda']\geq7$, the above equation reduces to
\[
9\left(87\lambda+137\right)\a_1=10(\lambda+6)(13\lambda+22)\a_1,
\]
which, by subtracting the left side from the right side, gives
\[
(130\lambda^2+217\lambda+87)\a_1=0.
\]
The claim follows since $-2/3$ is not a root of the polynomial inside the brackets.
\end{proof}
\end{Proposition}
\subsection{}\label{Theorem1}We summarize the preceding results as a theorem as follows.
\begin{Theorem}\label{sum1} We have the following three cases
\begin{enumerate}
\item If $\lambda'\in\{0,p-1\}$, then
\begin{equation*}
\Ext_{U_\chi(W)}(V_\chi(\lambda'),V_\chi (\lambda))\simeq
\begin{cases}
K,& \text{if }\lambda\in\{0,1,2,3,4,p-1\},\\
0,&\text{otherwise}.
\end{cases}
\end{equation*}
\item If $\lambda\in\{0,p-1\}$, then
\begin{equation*}
\Ext_{U_\chi(W)}(V_\chi(\lambda'),V_\chi(\lambda))\simeq
\begin{cases}
K,& \text{if }\lambda'\in\{0\}\cup\{p-i\mid 1\leq i \leq 5\},\\
0,&\text{otherwise}.
\end{cases}
\end{equation*}
\item If $\lambda,\lambda'\not\in\{0,p-1\}$, then
\begin{equation*}
\begin{array}{c}
\Ext_{U_\chi(W)}(V_\chi(\lambda'),V_\chi(\lambda))\\ \ \\
\simeq \begin{cases}
K,& \text{if }[\lambda-\lambda']\in\{2,3\},\\
K,& \text{if }[\lambda-\lambda']=4 \text{ and }p\neq5,\\
K,& \text{if }[\lambda-\lambda']=6 \text{ and }2\lambda=5\pm\sqrt{19}\text{ and }p>7,\\
0,&\text{otherwise}.
\end{cases}
\end{array}
\end{equation*}
\end{enumerate}
\end{Theorem}
\section{Height $-1$}\label{ch3}
Throughout this section we will assume that $r(\chi)=-1$, or equivalently, that $\chi=0$. Furthermore, we let as usual $\lambda$ and $\lambda'$ be elements in $\Lambda(0)\simeq\mathbbm{F}_p$.
\subsection{$V_0(\lambda)$ and $V_0(\lambda')$}
This case is fully described in Section \ref{Verma}. The reader is referred to Theorem \ref{sum1}.
\subsection{$V_0(\lambda)$ and $K$}
Our approach will follow the one taken earlier in Section
\ref{Verma}. We have an isomorphism of vector spaces
\[
\Ext_{U_0(W)}(V_0(\lambda),K)\simeq \Ext_{U_0(W_0)}(K_{\lambda},K),
\]
where $K$ denotes the trivial $W$-module. Suppose that we have a short exact sequence of $U_0(W_0)$-modules
\begin{equation}\label{seq3}
0\rightarrow K\xrightarrow{f} M\xrightarrow{g} K_\lambda\rightarrow 0,
\end{equation}
and fix bases $v$ and $v'$ for $K$ and $K_\lambda$, respectively. Let $\{w,w'\}$ be a basis for $M$ such that $f(v)=w$ and $g(w')=v'$. Since $U_0(Ke_0)$ is semisimple we may choose $w'$ such that $e_0w'=\lambda w'$. Note that the $\lambda$ weight space in $M$ is one dimensional so $w'$ is unique with these properties. Now, we have clearly $e_iw=0$ for all $i$. Furthermore, if $e_1w'$ and $e_2w'$ are both 0, then $e_iw'=0$ for all $i>0$. This can be seen by a simple induction argument since
\[
(i-1)e_{i+1}w'=(e_1e_i-e_ie_1)w'.
\]
The above observations give a necessary condition for \eqref{seq3} to be non-split, namely, that $e_1w'\neq0$ or $e_2w'\neq0$. We can interpret this condition in terms of $\lambda$. Indeed, since $e_iw'$ belongs to the weight space $M_{\lambda+i}$, we have $e_iw'=0$ if $\lambda+i\neq0$ and $i>0$. Thus $e_1w'=0$ and $e_2w'=0$ for $\lambda\not\in\{p-1,p-2\}$. This leads to the following lemma
\begin{Lemma} If $\lambda\not\in\{p-1,p-2\}$ then
\begin{equation*}
\Ext_{U_0(W_0)}(K_\lambda,K)=0.
\end{equation*}
\end{Lemma}
We consider the cases $\lambda\in\{p-1,p-2\}$ separately. Note that $\lambda=p-1$ is not interesting for our purpose because $V_0(p-1)$ is not simple; we shall, nevertheless, include it in our study for the sake of completeness. Fix $\lambda=p-k$ for $k\in\{1,2\}$. There exists $\a \in K$ such that for every $i>0$
\[
e_iw'=
\begin{cases}
\a w, &\text{if } i=k,\\
0, &\text{otherwise}.
\end{cases}
\]
As in Section \ref{Verma}, we construct a well-defined injective homomorphism
\[
\Ext_{U_0(W_0)}(K_\lambda,K)\rightarrow K,
\]
which sends the class of \eqref{seq3} to $\a$. Conversely, for each $\a\in K$ consider a short exact sequence of vector spaces
\begin{equation*}
0\rightarrow K\xrightarrow{f} M_{\a}\xrightarrow{g} K_\lambda\rightarrow 0,
\end{equation*}
and choose a basis $\{w,w'\}$ for $M_{\a}$ such that $f(v)=w$ and $g(w')=v'$ where, as before, $v$ and $v'$ are bases for $K$ and $K_\lambda$, respectively. For each $0\leq i\leq p-2$ define an endomorphism $E_i\in \End_K(M_{\a})$ of $M_{\a}$ such that $E_iw=0$ for all $i$ and
\[
E_iw'=
\begin{cases}
\lambda w',&\text{if }i=0,\\
\a w,&\text{if }i=k,\\
0,&\text{otherwise}.
\end{cases}
\]
Furthermore, set $E_i^{[p]}=\delta_{i0}E_i$ for all $i=0,1,\dots,p-2$
and $E_i=0$ for $i\neq0,1,\dots,p-2$. We have
\[
(E_i^p-E_i^{[p]})w'=0\quad \text{for all} \quad 0\leq i\leq p-2,
\]
and
\[
[E_i,E_j]w'=(j-i)E_{i+j}w' \ \text{ for all}\ i \text{ and }j.
\]
To prove the above formula one can assume $i\in\{0,k\}$ and that
$i\neq j$ since otherwise one would obtain 0 on both sides of the equation.
We summarize the results in this section into the following Proposition
\begin{Proposition}We have
\[
\Ext_{U_0(W)}(V_0(\lambda),K)\simeq
\begin{cases}
K,&\text{if }\lambda\in\{p-1,p-2\},\\
0,&\text{otherwise}.
\end{cases}
\]
\end{Proposition}
Evidently, $K$ is a self-dual module in the sense that it is
isomorphic to its dual. Thanks to the isomorphism
\[
\Ext_{U_0(W)}(K,V_0(\lambda))\simeq \Ext_{U_0(W)}(V_0(\lambda)^*,K^*),
\]
and the fact that $V_0(\lambda)^*$ and $V_0(p-1-\lambda)$ are isomorphic we obtain
\begin{Corollary} We have
\[
\Ext_{U_0(W)}(K,V_0(\lambda))\simeq
\begin{cases}
K,&\text{if }\lambda=\{0,1\},\\
0,&\text{otherwise}.
\end{cases}
\]
\end{Corollary}
\subsection{$V_0(\lambda)$ and $S$}\label{3.1}We proceed to describe
$\Ext_{U_0(W)}(V_0(\lambda),S)$ where $S$ denotes the
$(p-1)$-dimensional simple $U_0(W)$-module. We have an isomorphism
\[
\Ext_{U_0(W)}(V_0(\lambda),S)\simeq \Ext_{U_0(W_0)}(K_{\lambda},S).
\]
Thus, classifying the extensions of $V_0(\lambda)$ by $S$ reduces to classifying the extensions of $K_\lambda$ by $S$. Suppose that we have a short exact sequence of $U_0(W_0)$-modules
\begin{equation}\label{seq4}
0\rightarrow S\xrightarrow{f} M\xrightarrow{g} K_\lambda\rightarrow 0,
\end{equation}
and fix bases $v_0,v_1,\dots,v_{p-2}$ and $v'$ of $S$ and $K_\lambda$, respectively. Let
\begin{equation}\label{b1}
w_0,\dots,w_{p-2},w'
\end{equation}
be a basis of $M$ such that $f(v_i)=w_i$ for all $i$ and $g(w')=v'$; we may choose $w'$ such that $e_0w'=\lambda w'$. We have
\begin{equation*}
e_jw_i=
\begin{cases}
(-1)^{j+1}\frac{(i+1)!}{(i-j)!}w_{i-j}, & \text{if }j\leq i,\\
0, & \text{otherwise}.
\end{cases}
\end{equation*}
By weight considerations there exists $\a_j\in K$ ($j>0$) such that
\[
e_jw'=\begin{cases}
\a_jw_{[-\lambda-j-1]},&\text{if }\lambda+j\neq0,\\
0,&\text{otherwise}.
\end{cases}
\]
Furthermore, we set
\begin{equation}\label{addcond}
\a_j=0\quad\text{if}\quad\lambda+j=0 \ \text{ or }\ j\not\in\{1,2,\dots,p-2\},
\end{equation}
For $\lambda\neq0$ the $\lambda$ weight space in $M$ is two-dimensional; it is generated by $w'$ and $w_{[-\lambda-1]}$ so any different choice for $w'$ has the form $w'+bw_{[-\lambda-1]}$ for some
$b\in K$. Obviously, this leads to the same $e_jw'$ if
$e_jw_{[-\lambda-1]}=0$. In particular, we see that all the $\mathfrak{a}_j$ are determined by $M$ if $\lambda=p-1$. The same holds for $\lambda=0$ since $w'$ is unique in this case; we have $M_0=Kw'$.
\begin{Lemma}[A]If $\lambda\in\{0,p-1\}$ then all the $\a_j$ are determined by $M$.
\end{Lemma}
For $\lambda\not\in\{0,p-1\}$ we have $e_1w_{[-\lambda-1]}\neq0$. Thus, the discussion preceding Lemma A shows that we can choose $w'$ uniquely such that $e_1w'=0$.
\begin{Lemma}[B]
If $\lambda\not\in\{0,p-1\}$ then there is a unique choice for $w'$ such that $e_1w'=0$.
\end{Lemma}
We will henceforth always assume that $w'$ is the unique choice from
Lemma B if $\lambda\not\in\{0,p-1\}$. As in Section \ref{Verma} we
obtain an injective homomorphism
$\Ext_{U_0(W_0)}(K_{\lambda},S)\rightarrow K^2$ which sends the class
of \eqref{seq4} to $(\a_1,\a_2)\in K^2$ and such that $\a_1=0$ for
$\lambda\not\in\{0,p-1\}$. We have
\begin{equation}\label{extend}
\a_j=A_j\a_1+B_j\a_2 \quad\text{for all}\quad 3\leq j\leq p-2,
\end{equation}
where
\begin{align*}
A_j&=-\frac{1}{j-2}\biggl(\prod_{k=1}^{j}(\lambda+k)\biggr)\biggl(1+\sum_{k=4}^{j}\frac{(j-k)!}{(j-3)!}\prod_{l=1}^{k-3}(\lambda+j-l)\biggr),
\end{align*}
(the summation $\sum_{k=4}^j$ is understood to be 0 when $j=3$) and
\[
B_j=\frac{(-1)^j}{(j-2)!}\prod_{k=2}^{j-1}(-\lambda-1-k)(\lambda+k).
\]
Conversely, let $(\a_1,\a_2)$ be a pair in $K^2$ such that, cf. \eqref{addcond}
\[
\a_1=0\quad\text{if}\quad\lambda=p-1,
\]
\[
\a_2=0\quad\text{if}\quad\lambda=p-2.
\]
We extend $(\a_1,\a_2)$ to a tuple $\a=(\a_1,\a_2,\dots,\a_{p-2}) \in K^{p-2}$ by using \eqref{extend} and consider a short exact sequence
\begin{equation}\label{seq6}
0\rightarrow S\xrightarrow{f} M_\a\xrightarrow{g} K_\lambda\rightarrow 0,
\end{equation}
where $M_\a$ is a vector space with a basis as in \eqref{b1}. For each $0\leq j\leq p-2$ we define an endomorphism $E_j\in \End_K(M_\a)$ such that
\begin{equation*}
E_jw_i=
\begin{cases}
(-1)^{j+1}\frac{(i+1)!}{(i-j)!}w_{i-j}, & \text{if }j\leq i,\\
0, & \text{otherwise},
\end{cases}
\end{equation*}
and $E_0w'=\lambda w'$ and for every $1\leq j\leq p-2$
\[
E_jw'=\begin{cases}
\a_jw_{[-\lambda-j-1]},&\text{if }\lambda+j\neq0,\\
0,&\text{otherwise}.
\end{cases}
\]
Furthermore, we set $E_j^{[p]}=\delta_{j0}E_j$ for all $j$ and $E_j=0$ for $j\not\in\{0,1,\dots,p-2\}$. We have
\[
(E_j^p-E_j^{[p]})w'=0 \quad \text{for all}\quad j.
\]
Except for the change in notation, the formula
$[E_i,E_j]w'=(j-i)E_{i+j}w' $ leads to the same equations as in
Section \ref{W--module}; there is no need to include them here.
\begin{Proposition} We have
\[
\Ext_{U_0(W)}(V_0(\lambda),S)\simeq
\begin{cases}
K,&\text{if }\lambda\in\{0\}\cup\{p-i\mid 3\leq i\leq5\},\\
0,&\text{otherwise}.
\end{cases}
\]
\end{Proposition}
\begin{proof} The short exact sequence
\begin{equation}\label{eq:5}
0\rightarrow S\rightarrow V_0(0)\xrightarrow{\pi} K\rightarrow 0
\end{equation}
induces the long exact sequence of vector spaces
\begin{align}\label{seq7}
0&\rightarrow\Hom_{U_0(W)}(V_0(\lambda),S)\rightarrow\Hom_{U_0(W)}(V_0(\lambda),V_0(0)) \\ \notag
&\rightarrow\Hom_{U_0(W)}(V_0(\lambda),K)\rightarrow\Ext_{U_0(W)}(V_0(\lambda),S)\\ \notag
&\rightarrow\Ext_{U_0(W)}(V_0(\lambda),V_0(0))\rightarrow\Ext_{U_0(W)}(V_0(\lambda),K)\rightarrow\cdots.
\end{align}
Due to Schur's Lemma, we have $\Hom_{U_0(W)}(V_0(\lambda),K)=0$ for $\lambda\not\in\{0,p-1\}$. We obtain, in these cases, the exact sequence
\begin{align}\label{seq8}
0&\rightarrow\Ext_{U_0(W)}(V_0(\lambda),S)\rightarrow\Ext_{U_0(W)}(V_0(\lambda),V_0(0))&\\ \notag
&\rightarrow\Ext_{U_0(W)}(V_0(\lambda),K)\rightarrow\cdots.
\end{align}
Now $\Hom_{U_0(W)}(V_0(0),K)$ is 1-dimensional and generated by the surjection $\pi$ in \eqref{eq:5}. The map $\Hom_{U_0(W)}(V_0(0),V_0(0))\rightarrow\Hom_{U_0(W)}(V_0(0),K)$ appearing in \eqref{seq7} is surjective because it maps the identity $\id_{U_0(W)}$ on $U_0(W)$ to $\pi$. Thus, we obtain an exact sequence as in \eqref{seq8}
\begin{align*}
0&\rightarrow\Ext_{U_0(W)}(V_0(0),S)\rightarrow\Ext_{U_0(W)}(V_0(0),V_0(0))&\\
&\rightarrow\Ext_{U_0(W)}(V_0(0),K)\rightarrow\cdots.
\end{align*}
Consequently, we have for every $\lambda\neq p-1,p-2$
\[
\Ext_{U_0(W)}(V_0(\lambda),S)\simeq\Ext_{U_0(W)}(V_0(\lambda),V_0(0)),
\]
which, by Proposition \ref{sum1}, implies
\[
\Ext_{U_0(W)}(V_0(\lambda),S)\simeq
\begin{cases}
K,&\text{if }\lambda\in\{0\}\cup\{p-i\mid 3\leq i\leq5\},\\
0,&\text{if }\lambda\not\in\{0\}\cup\{p-i\mid 1\leq i\leq5\}.
\end{cases}
\]
Next, suppose that $\lambda\in\{p-1,p-2\}$. We prove that every short
exact sequence of $U_0(W)$-modules as in \eqref{seq6} splits, or
equivalently, that $\a_1=0$ and $\a_2=0$. The case $\lambda=p-2$ comes
almost at once from the observations preceding the proposition; indeed
$\a_1=0$ because of Lemma B and $\a_2=0$ by definition, see
\eqref{addcond}. For $\lambda=p-1$ it follows from \eqref{addcond}
that $\a_1=0$. To prove $\a_2=0$ note first that \eqref{extend}
implies $\a_2=-2 \a_{p-2}$. However, since
\[
1\leq[-1-\lambda-(p-2)]\quad\text{and}\quad p-2\leq[-1-\lambda-1],
\]
we see by inserting $(i,j)=(1,p-2)$ into \eqref{A4} that $\a_{p-2}=0$. This is only possible if $\a_2=0$ so the proposition is proved.
\end{proof}
The module $S$ is a self-dual module. Thanks to the isomorphism
\[
\Ext_{U_0(W)}(S,V_0(\lambda))\simeq \Ext_{U_0(W)}(V_0(\lambda)^*,S^*)
\]
and the fact that $V_0(\lambda)^*$ and $V_0(p-1-\lambda)$ are isomorphic we obtain
\begin{Corollary} We have
\begin{equation*}
\Ext_{U_0(W)}(S,V_0(\lambda))\simeq
\begin{cases}
K,&\text{if }\lambda\in\{2,3,4,p-1\},\\
0,&\text{otherwise}.
\end{cases}
\end{equation*}
\end{Corollary}
\subsection{Self-extensions of $K$}
There are no nontrivial self-extensions of the trivial module $K$.
\begin{Proposition} We have
\begin{equation*}
\Ext_{U_0(W)}(K,K)=0
\end{equation*}
\end{Proposition}
\begin{proof}
Suppose we have a short exact sequence of $U_0(W)$-modules
\begin{equation}\label{seq5}
0\rightarrow K\xrightarrow{f} M\xrightarrow{g} K\rightarrow 0,
\end{equation}
and let $\{w,w'\}$ be a basis of $M$ such that $Kw=\Img f$. We have clearly $Ww=0$. Furthermore, $W w'\subset Kw$ since $Ww'\subset\ker g$. Thus, every $x\in W$ acting on $M$ can be represented by a matrix
\begin{equation*}
\begin{bmatrix}
0&\phi(x)\\
0&0
\end{bmatrix},
\end{equation*}
where $\phi:W\rightarrow K$ is a homomorphism of Lie algebras. That $\phi$ preserves the Lie algebra structure follows from the fact that $M$ is a $W$-module and hence that $W\rightarrow \mathfrak{gl}(M)\simeq\mathfrak{gl}(2,K)$ is a homomorphism of Lie algebras. Now, if \eqref{seq5} is non-split, then $\phi(x)\neq0$ for some $x\in W$. The kernel of $\phi$ is then an ideal of codimension 1, in apparent contradiction with the fact that $W$ is simple.
\end{proof}
\subsection{Self-extensions of $S$}
There are no nontrivial self-extensions of the module $S$.
\begin{Proposition}We have
\[
\Ext_{U_0(W)}(S,S)=0.
\]
\end{Proposition}
\begin{proof}
The short exact sequence
\[
0\rightarrow S\rightarrow V_0(0)\rightarrow K\rightarrow 0
\]
induces the long exact sequence
\begin{align*}
0&\rightarrow\Hom_{U_0(W)}(S,S)\rightarrow\Hom_{U_0(W)}(S,V_0(0))\\
&\rightarrow\Hom_{U_0(W)}(S,K)\rightarrow\Ext_{U_0(W)}(S,S)\\
&\rightarrow\Ext_{U_0(W)}(S,V_0(0))\rightarrow\cdots,
\end{align*}
which by Corollary \ref{3.1} and Schur's lemma implies the claim.
\end{proof}
\subsection{$S$ and $K$}The proof of the next proposition makes
use of the fact that $\Hom_{U_0(W)}(V_0(0),S)=0$.
\begin{Proposition}We have
\[
\Ext_{U_0(W)}(K,S)\simeq K^2.
\]
\end{Proposition}
\begin{proof}
The short exact sequence
\[
0\rightarrow S\rightarrow V_0(0)\rightarrow K\rightarrow 0
\]
induces the long exact sequence
\begin{align*}
0&\rightarrow\Hom_{U_0(W)}(K,S)\rightarrow\Hom_{U_0(W)}(V_0(0),S)\\
&\rightarrow\Hom_{U_0(W)}(S,S)\rightarrow\Ext_{U_0(W)}(K,S)\\
&\rightarrow\Ext_{U_0(W)}(V_0(0),S)\rightarrow\Ext_{U_0(W)}(S,S)\rightarrow\cdots,
\end{align*}
which in turn induces the exact sequence
\begin{align*}
0&\rightarrow\Hom_{U_0(W)}(S,S)\rightarrow\Ext_{U_0(W)}(K,S)\rightarrow\Ext_{U_0(W)}(V_0(0),S)\rightarrow0.
\end{align*}
The claim follows from Schur's lemma and Proposition \ref{3.1}.
\end{proof}
Since $K$ and $S$ are self-dual modules, the isomorphism
\[
\Ext_{U_0(W)}(K,S)\simeq \Ext_{U_0(W)}(S^*,K^*)
\]
implies the following corollary
\begin{Corollary} We have
\[
\Ext_{U_0(W)}(S,K)\simeq K^2.
\]
\end{Corollary}
\section{Height 0 and 1}
All the work has been done in Section \ref{Verma}; Theorem \ref{sum1}
gives a complete classification of the $\chi$-reduced Verma modules
having character $\chi$ of height 0 and 1.
\section{Height $p-1$}
Recall that every simple $U_\chi(W)$--module with one exception is
projective. The remaining simple module $L$ has a projective cover
with two composition factors both isomorphic to $L$. We have a
short exact sequence
\[
0\rightarrow L\rightarrow P\rightarrow L\rightarrow 0,
\]
where $P$ is a projective module. This induces the long exact sequence
\begin{align*}
0&\rightarrow\Hom_{U_\chi(W)}(L,L)\rightarrow\Hom_{U_\chi (W)}(P,L)\\
&\rightarrow\Hom_{U_\chi (W)}(L,L)\rightarrow\Ext_{U_\chi (W)}(L,L)\\
&\rightarrow\Ext_{U_\chi (W)}(P,L)\rightarrow\cdots,
\end{align*}
which in turn implies $\Ext_{U_\chi (W)}(L,L)\simeq K$. We have therefore
\begin{Proposition} If $M$ and $N$ are two simple $U_\chi(W)$--modules then
\[
\Ext_{U_\chi(W)}(M,N)\simeq
\begin{cases}
K,&\text{if }N=M=L,\\
0,&\text{otherwise}.
\end{cases}
\]
\end{Proposition}
|
\section{Introduction}
The number $F(G)$ of stable sets (or independent sets) of a graph $G$
was first considered by Prodinger and Tichy~\cite{Prodinger82}. They called it the {\em Fibonacci number}
of a graph, based on the observation that the number of stable sets in a path on $n$
vertices is exactly the $n+2$-th Fibonacci number.
The invariant $F(G)$ is also known as the {\em Merrifield-Simmons index} or {\em $\sigma$-index}
of the graph $G$~\cite{MS89}.
There is a rich literature dealing with extremal questions regarding $F(G)$, and the
case where $G$ is a tree received much attention.
For instance, Heuberger and Wagner~\cite{Heuberger08} characterized
the $n$-vertex trees $T$ with maximum degree $\Delta$ maximizing $F(T)$, for given $n$ and $\Delta$.
Li, Zhao, and Gutman~\cite{Li05}, and independently Pedersen and Vestergaard~\cite{Pedersen07}
determined the $n$-vertex trees $T$ maximizing $F(T)$ when the diameter is fixed.
Yu and Lv~\cite{YuLv07} considered similarly the case of trees with a fixed number of leaves.
The reader is referred to \cite{Knopfmacher07, LinLin95, Wang08}
and the references therein for other results of this kind on trees.
Other classes of graphs have been considered as well; this includes
unicyclic graphs~\cite{LiZhu09, Pedersen05, Pedersen06, Wang08a},
bicyclic graphs~\cite{DengChenZhang08},
tricyclic graphs~\cite{ZhuLiTan10},
quasi-trees~\cite{LiLiJing09},
maximal outerplanar graphs~\cite{Alameddine98},
connected graphs~\cite{Bruyere09},
and bipartite $d$-regular graphs~\cite{Kahn01}.
In this paper, we propose to revisit an old result of Prodinger and Tichy~\cite{Prodinger82}
which is well-known in this area: among all trees $T$ on $n\geq 2$ vertices,
the path $P_{n}$ and the star $K_{1,n-1}$ respectively minimizes and maximizes $F(T)$.
Observe that the stability number (the largest cardinality of a stable set)
of $P_{n}$ is $\lceil n/2 \rceil$, while that of $K_{1,n-1}$ is $n-1$.
Since the stability number $\alpha$ of every tree on $n$ vertices is between these two
extreme values, one may wonder which are the trees $T$
minimizing or maximizing $F(T)$ for fixed $\alpha$.
In a previous contribution~\cite{Bruyere09}, a subset of the authors showed
that the trees maximizing $F(T)$ have a rather simple structure: they are exactly the
trees that can be obtained from the Tur\'an graph\footnote{
The {\DEF Tur\'an graph} $T(n,\alpha)$ is the $n$-vertex graph consisting of
$\alpha$ vertex-disjoint cliques which are as balanced as possible.}
$T(n,\alpha)$ by adding $\alpha -1$ edges.
Here, we focus on the trees minimizing $F(T)$ for fixed $\alpha$.
As will become apparent in the next sections, the structure of the
corresponding extremal trees is less straightforward than when maximizing $F(T)$.
Our main result is that the edges of a (non-trivial) extremal tree can uniquely
be partitioned into $n-\alpha$ stars, each of size $\left \lceil \frac{n-1}{n-\alpha} \right \rceil$
or $\left \lfloor \frac{n-1}{n-\alpha} \right \rfloor$, so that
every vertex is included in at most two distinct stars,
and the centers of these stars form a stable set of the tree.
(See Figure~\ref{fig-intro} for an illustration.)
\begin{figure}
\centering
\includegraphics[scale=0.9]{fig-intro}
\caption{\label{fig-intro}An extremal tree for $n= 18$ and $\alpha = 13$.
White vertices are the centers of the stars.}
\end{figure}
It follows that an extremal tree can be completely described by specifying which
stars have a vertex in common. This is captured by the following labeled tree,
which we call {\em center-tree}:
the tree has $n-\alpha$ vertices (representing the stars);
there is an edge between two vertices if the corresponding stars have a vertex in common,
and each vertex is labeled by the size of the corresponding star.
However, characterizing the {center-tree s} of extremal trees appears to be challenging.
We obtained such a characterization in a very restricted case, namely when
$(n - 1) \bmod (n-\alpha) \in \{0, n-\alpha - 2, n-\alpha - 1\}$, but
the general case is far from understood and is left as an open problem.
The paper is organized as follows. In Section~\ref{sec_prelim}, we provide the necessary
definitions and a few basic lemmas. Sections~\ref{sec-task1} and~\ref{sec-task2},
which constitute the bulk of this paper,
are devoted to the proof that extremal trees have a partition into stars as described above
(cf.\ Theorem~\ref{th-task2}).
Finally, in Section~\ref{sec-task3}, we conclude with some remarks on
the problem of characterizing the {center-tree s} of extremal trees, and provide a solution for the
particular case mentioned above.
\section{Definitions and Preliminaries} \label{sec_prelim}
This section is devoted to basic definitions and notations used throughout the text.
Also, we introduce the notion of {\ts}, and prove some useful facts about these trees.
All graphs are assumed to be finite, simple, and undirected.
We generally follow the terminology and notations of Diestel~\cite{Diestel05}.
The neighborhood of a vertex $v$ of a graph $G$ is
denoted $N_{G}(v)$; also, we write $N_{G}[v]$ for the set $N_{G}(v) \cup \{v\}$.
The degree of $v$ is denoted by $\deg_{G}(v)$.
(We often drop the subscript $G$ when the graph is clear from the context.)
We simply write $|G|$ for the order $|V(G)|$ of $G$.
A subset $S\subseteq V(G)$ of vertices of a graph $G$ is a
{\DEF stable set} if the vertices in $S$ are pairwise non adjacent in $G$.
The maximum cardinality of a stable set in $G$ is the {\DEF stability
number} of $G$, and is denoted $\alpha(G)$.
Recall that $F(G)$ denotes the number
of stable sets of $G$ (including the empty set).
In this paper, we call this invariant the {\DEF Fibonacci number} of $G$, as in~\cite{Prodinger82}.
The following easy properties (see \cite{Gutman86, Li05, Prodinger82})
of the {Fibonacci number} are used throughout the paper:
\begin{lem} \label{lem:basic}
Let $G$ be a graph. If $G$ is not empty, then,
\begin{itemize}
\item $F(G) = F(G -v) + F(G - N[v])$ for every vertex $v$ of $G$, and
\item $F(G) = \prod_{i=1}^k F(G_i)$
if $G$ is the disjoint union of $k$ graphs $G_1, G_{2}, \dots, G_{k}$.
\end{itemize}
In particular,
\begin{itemize}
\item $F(G) > F(G -v)$ for every vertex $v$ of $G$, and
\item $F(G) \le 2 F(G-v)$,
with strict inequality if $\deg_{G}(v) \geq 1$.
\end{itemize}
\end{lem}
A tree $T$ is {\DEF extremal} if $T$ has minimum {Fibonacci number} among
all trees on $|T|$ vertices with {stability number} $\alpha(T)$.
As mentioned in the introduction,
the structure of these extremal trees is the main topic of this paper;
in particular, we will see that non-trivial extremal trees are {\tss} that are balanced
(see below for the definitions).
In this paper, a {\DEF leaf} of a tree is defined as a vertex of degree at most $1$. Thus in particular if the tree has a unique vertex, then this vertex is considered to be a leaf. (Usually leaves are required to have degree exactly $1$, but this definition will be more convenient for our purposes.).
A {\DEF \ts} is defined inductively as follows:
\begin{itemize}
\item a single vertex is a {\ts};
\item if $T_1, T_2, \ldots, T_k$ ($k \ge 2$) are disjoint {\tss} and $v_i$ is a leaf of $T_i$ for $i=1, 2, \ldots, k$,
then the tree obtained from $T_{1} \cup \cdots \cup T_{k}$ by adding a new vertex $w$ adjacent to every vertex $v_{i}$ is also a {\ts}.
\end{itemize}
From the inductive definition given above, it is easy to check that, in a {\ts} $T$,
the distances between a fixed vertex $v\in V(T)$ and all the leaves of $T$ have the same
parity. The vertex $v$ is a {\DEF center} of $T$ is these distances are odd.
An example of a {\ts} is given in Figure~\ref{fig-exa-tos}.
\begin{figure}[!htb]
\begin{center}
\includegraphics[scale=0.9]{fig-exa-tos}
\caption{\label{fig-exa-tos}
A {\tos}. The white vertices are the centers of the tree.
}
\end{center}
\end{figure}
We note that {\tss} can equivalently be defined as follows. First observe that the vertex set of a tree $T$ can uniquely be partitioned into two stable sets $A$ and $B$. Then $T$ is a {\ts} if and only if one of these two sets, say $A$, contains all leaves of $T$ and no vertex with degree at least $3$ (thus all vertices in $A$ have degree at most $2$). In that case, the centers of $T$ are exactly the vertices in $B$. (We leave it to the reader to check that this definition of {\tss} is indeed equivalent to the original one.)
A {\ts} is {\DEF balanced} if the degrees of any two of its centers differ by at most~$1$
(for instance, the {\ts} in Figure~\ref{fig-intro}
is balanced while the one in Figure~\ref{fig-exa-tos} is not).
Let $T$ be a {\ts}. The set of {\ct s} of $T$ is denoted by $\C(T)$.
Observe that a vertex of $T$ with degree at least 3 is always a {\ct}; thus, the set
$V(T) - \C(T)$ includes only vertices with degree at most 2. Also,
notice that $(\C(T), V(T) - \C(T))$ is a partition of $V(T)$ into two stable sets.
Finally, note that a path $P$ is a {\ts} if and only if $P$ is even\footnote{
$P$ is odd (even) if $P$ has odd (respectively even) length, where the length
is defined as the number of edges.}.
\begin{lem}
\label{lem-mss}
Let $T$ be a {\ts}. Then $V(T) - \C(T)$ is the unique maximum stable set of $T$.
\end{lem}
\begin{proof}
We use induction on $|\C(T)|$. If $|\C(T)| = 0$, then $T$ is a single vertex
and the claim trivially holds. Suppose $|\C(T)| \ge 1$, and
let $S$ be any {maximum stable set} of $T$. Also, let $v$ be a leaf of $T$
and $w$ be its unique neighbor.
Let $T_{1}, \dots, T_{k}$ be the components of $T - \{v, w\}$. Notice that each
$T_{i}$ is a {\ts} and $\C(T) = \{w\} \cup \C(T_{1}) \cup \cdots \cup \C(T_{k})$.
First, suppose that $S\cap V(T_{i}) \neq V(T_{i}) - \C(T_{i})$
for some $i\in \{1,\dots,k\}$. Then, by induction, $|S\cap V(T_{i})| < |V(T_{i}) - \C(T_{i})|$,
and the set
$$
S' := \big(S - \left(S\cap V(T_{i})\right)\big) \cup \big(V(T_{i}) - \C(T_{i})\big)
$$
has cardinality larger than $S$. Since $S'$ cannot be a stable set of $T$, it follows $w \in S'$,
and thus $v \notin S'$.
But then $(S' - \{w\}) \cup \{v\}$ is a stable set of $T$, with cardinality equal to
that of $S'$, a contradiction.
Therefore, $S\cap V(T_{i})=V(T_{i}) - \C(T_{i})$ for
every $i\in \{1,\dots,k\}$, which implies $w \notin S$, and hence $v\in S$. It follows that $S=V(T) - \C(T)$.
\end{proof}
Suppose that $T$ is a tree not isomorphic to a path and that $T$ has
a leaf $w$ such that the tree $T'$ obtained by adding a new vertex $v$ adjacent to $w$ is a {\tos}.
Then $T$ is said to be {\DEF almost a {\tos}}, and
a vertex of $T$ is considered to be a center of $T$ if it is a center of $T'$.
Note that the requirement that $T$ is not a path ensures that the leaf $w$ is uniquely determined.
Hence, the set of centers of $T$ is well-defined.
The vertex $w$ is the unique center of $T$ which is also a leaf, and is said to be the {\DEF exposed center} of $T$.
See Figure~\ref{fig-exa-ats} for an illustration.
\begin{figure}[!htb]
\begin{center}
\includegraphics[scale=0.9]{fig-exa-ats}
\caption{\label{fig-exa-ats}
A tree which is almost a {\tos}. Its {exposed center} is drawn in grey.}
\end{center}
\end{figure}
Let us observe that, if $v$ is a leaf of a {\tos} $T$ with $|T| \geq 3$, then $T-v$ is exactly one of the following:
\begin{itemize}
\item a {\tos};
\item almost a {\tos};
\item an odd path.
\end{itemize}
This fact will be repeatedly used in our proofs.
\section{Extremal Trees are Trees of Stars}
\label{sec-task1}
The following theorem is a first step towards understanding the structure of extremal trees.
\begin{thm}
\label{th-task1}
Every extremal tree is either a {\ts} or an odd path.
\end{thm}
The proof of Theorem~\ref{th-task1} is based on the concept of rotating an edge in a tree $T$, which consists in first removing some edge $uv$ of $T$, and then adding an edge $uv'$, where $v'$ is some vertex of the component of $T-uv$ containing $v$.
We will show that if $T$ is neither a {\ts} nor an odd path, then
there exists an edge of $T$ that can be rotated in such a way that the resulting tree $T'$
satisfies both $\alpha(T')=\alpha(T)$ and $F(T') < F(T)$.
To this aim, we introduce a few technical lemmas.
\subsection{Trees of Stars and the Golden Ratio}
The golden ratio $\phi := \frac{1 + \sqrt 5}{2}\simeq 1.618$
appears naturally when considering {\tss},
as illustrated by the following two lemmas.
\begin{lem}
\label{lem-ToS-leaf}
Let $T$ be a {\ts} and let $v$ be a leaf of $T$. Then
$F(T) > \phi \cdot F(T-v)$. Moreover, this remains true if $T$ is almost a {\ts}, provided $v$ is not the {exposed center}.
\end{lem}
\begin{proof}
The proof is by induction on $|T|$. The claim is true when $|T|=1$, since then
$F(T)=2>\phi \cdot F(T-v)=\phi$.
For the inductive step, assume $|T| > 1$. Let $w$ be the center adjacent to $v$
and $T_{1}, \dots, T_{k}$ be the components of $T-\{v,w\}$ (thus $k\geq 1$).
Let also $v_{i}$ be the leaf of $T_{i}$ that is adjacent to $w$ in $T$. Letting
$$
\gamma := \prod_{i=1}^{k} \frac{F(T_{i})}{F(T_{i} - v_{i})},
$$
we deduce
$$
\frac{F(T)}{F(T-v)} = \frac{2\prod_{i=1}^{k} F(T_{i})
+ \prod_{i=1}^{k} F(T_{i}-v_{i})}{\prod_{i=1}^{k} F(T_{i})
+ \prod_{i=1}^{k} F(T_{i}-v_{i})} = 2 - \frac{1}{1 + \gamma}.
$$
If $T$ is a {\ts}, then $T_{i}$ is also a {\ts} and $F(T_i) > \phi \cdot F(T_i-v_i)$ by the induction hypothesis,
for every $i\in\{1, \dots, k\}$. Hence, $\gamma >\phi^{k} \geq \phi$.
If, on the other hand,
$T$ is almost a {\ts}, then all trees $T_i$ are {\tss} except for exactly one, say $T_{1}$.
Thus, by induction, $F(T_i) > \phi \cdot F(T_i-v_i)$ for every $i\in\{2, \dots, k\}$.
Moreover, $T_{1}$ is either almost a {\ts} or an odd path. In the first case, $v_{1}$ is not the {exposed center} of $T_{1}$, and
thus $F(T_1) > \phi \cdot F(T_1-v_1)$ by induction, which again implies $\gamma >\phi^{k} \geq \phi$.
In the second case, we have $k\geq 2$ since $T$ is not a path, and
$$
\gamma \geq \prod_{i=2}^{k} \frac{F(T_{i})}{F(T_{i} - v_{i})} > \phi^{k-1} \geq \phi,
$$
by induction.
Therefore, $\gamma > \phi$ holds in all possible cases, and it follows
$$
\frac{F(T)}{F(T-v)} = 2 - \frac{1}{1 + \gamma}>
2 - \frac{1}{1 + \phi} = \phi.
$$
(The last equality is derived using the fact $\phi^{2} = \phi + 1$.)
\end{proof}
A similar but opposite property holds for centers:
\begin{lem}
\label{lem-ToS-center}
Let $T$ be a {\ts} and let $w$ be a center of $T$. Then
$F(T) < \phi \cdot F(T-w)$. This moreover remains true if $T$ is almost a {\tos}.
\end{lem}
Note that, contrary to Lemma~\ref{lem-ToS-leaf}, here the vertex $w$ is allowed to be the {exposed center} of $T$ if $T$ is almost a {\tos}.
\begin{proof}[Proof of Lemma~\ref{lem-ToS-center}]
Let $T_{1}, \dots, T_{k}$ be the components of $T-w$. (Thus $k\geq 1$ and $k=1$ if and only if $w$ is the {exposed center} of $T$.)
We denote by $v_i$ the leaf of $T_i$ that is adjacent to $w$. Let
$$
\gamma := \prod_{i=1}^{k} \frac{F(T_{i})}{F(T_{i} - v_{i})}.
$$
We have
$$
\frac{F(T)}{F(T-w)} = \frac{\prod_{i=1}^{k} F(T_{i})
+ \prod_{i=1}^{k} F(T_{i}-v_{i})}{\prod_{i=1}^{k} F(T_{i})} = 1 + \frac{1}{\gamma}.
$$
If $T$ is a {\ts}, then $T_i$ is a {\ts} for every $i\in \{1, \dots, k\}$. The same is true if $T$ is almost a {\ts}
and $w$ is the {exposed center} of $T$. Thus, in these two cases,
we have $F(T_i) > \phi \cdot F(T_i-v_i)$ for every $i\in \{1, \dots, k\}$ by Lemma~\ref{lem-ToS-leaf},
which implies $\gamma > \phi^{k} \geq \phi$.
Now, if $T$ is almost a {\ts} and $w$ is {\em not} the {exposed center} of $T$, then $k\geq 2$ and all trees $T_i$ are {\tss} except for exactly one,
say $T_{1}$. Thus, $F(T_i) > \phi \cdot F(T_i-v_i)$ for every $i\in\{2, \dots, k\}$ by Lemma~\ref{lem-ToS-leaf}.
Since $F(T_{1}) \geq F(T_1-v_1)$, this implies
$$
\gamma \geq \prod_{i=2}^{k} \frac{F(T_{i})}{F(T_{i} - v_{i})} > \phi^{k-1} \geq \phi.
$$
It follows that $\gamma > \phi$ always holds, implying
$$
\frac{F(T)}{F(T-w)} = 1 + \frac{1}{\gamma} <
1 + \frac{1}{\phi} = \phi.
$$
\end{proof}
\subsection{Edge Rotations}
\label{sec-rotations}
Let $T$ be a tree, $xy$ be one of its edges, and
$x'$ be a vertex distinct from $x$ in the component of $T -xy$ containing $x$.
Then the pair $\rho:=(yx, yx')$ is called a
{\DEF rotation}, and we let $\rho(T):= (T - yx) +yx'$ denote the
tree resulting from the rotation.
The rotation $\rho$ is {\DEF good} if $\alpha(\rho(T)) = \alpha(T)$ and $F(\rho(T)) < F(T)$.
(Thus, in order to prove that a tree is {\em not} extremal, it is enough to show
that it admits a good rotation.)
When the rotation $\rho=(yx, yx')$ is clear from the context, we use the following notations.
\begin{nota}
\label{nota-rot}
Denote by $z$ ($z'$) the neighbor of $x$ (respectively $x'$) that
is included in the unique $xx'$-path in $T$.
(Observe that $z=x'$ and $z'=x$ if $xx' \in E(T)$; similarly, $z=z'$
if $x$ and $x'$ are at distance $2$.)
Let $x_{1}, \dots, x_{\l}$ be the neighbors of $x$ distinct from $z$ and $y$,
and $x'_{1}, \dots, x'_{\ell'}$ be the neighbors of $x'$ distinct from $z'$.
These notations are illustrated in Figure~\ref{fig-rotating-edge}.
Notice that $x$ and $x'$ have degree $\l+2$ and $\ell' +1$, respectively.
\begin{figure}[!ht]
\begin{center}
\includegraphics{fig-rotating-edge}
\caption{\label{fig-rotating-edge}
A rotation $\rho=(yx, yx')$.
}
\end{center}
\end{figure}
Furthermore, we refer to the components of $T - \{ x,x'\}$ as follows:
If $v \in V(T) - \{x, x'\}$, then $T_{v}$ denotes the component of $T - \{ x,x'\}$
that includes $v$. Also, it will be convenient to define $T_{v}$ as the empty tree if $v\in\{x, x'\}$.
(In particular, $T_{z}=T_{z'}$, and the latter tree is empty if and only $xx' \in E(T)$.)
Finally, we define ten numbers corresponding to these components:
$$
\begin{array}{l l}
X := \prod_{i=1}^{\l} F(T_{x_i}), &
\overline X := \prod_{i=1}^{\l} F(T_{x_i} - x_{i}),\\[7pt]
X' := \prod_{i=1}^{\ell'} F(T_{x'_i}), &
\overline X' := \prod_{i=1}^{\ell'} F(T_{x'_i} - x'_{i}),\\[7pt]
Y := F\left(T_{y} \right), &
\overline Y := F \left(T_{y} - y \right),\\[7pt]
Z := F \left(T_{z} \right), &
\overline Z_{z} := F \left(T_{z} - z \right),\\[7pt]
\overline Z_{z'} := F \left(T_{z} - z' \right) , \textrm{ and} & \overline Z_{\{z,z'\}} := F \left(T_{z} - {\{z,z'\}} \right).
\end{array}
$$
Since the {Fibonacci number} of an empty graph is $1$, by convention we let $X = \overline X := 1$ if $\l = 0$. Similarly,
$X' = \overline X' := 1$ if $\ell' = 0$, and $Z = \overline Z_{z} = \overline Z_{z'} = \overline Z_{\{z,z'\}} := 1$
if $x$ is adjacent to $x'$.
\end{nota}
\begin{lem}
\label{lem-rotating-edge}
With Notation~\ref{nota-rot}, we have
$F(\rho(T)) < F(T)$ if and only if $X \overline X' \overline Z_{z'} > \overline X X' \overline Z_{z}$.
\end{lem}
\begin{proof}
Let us compute $F(T)$ by applying twice Lemma~\ref{lem:basic} (first item) with vertices $x$ and $x'$.
We obtain
$$
F(T) = X Y \left( X' Z + \overline X' \ \overline Z_{z'} \right) + {\overline X} \ {\overline Y} \left( X' \overline Z_{z} + \overline X' \ \overline Z_{\{z,z'\}} \right).
$$
Similarly,
$$
F(\rho(T)) = X' Y \left( X Z + \overline X \ \overline Z_{z} \right) + {\overline X}' \ {\overline Y} \left( X \overline Z_{z'} + \overline X \ \overline Z_{\{z,z'\}} \right).
$$
Hence, we get
\begin{align*}
F(T) - F(\rho(T)) &= X \overline X' \overline Z_{z'} \left( Y - \overline Y \right) + \overline X X' \overline Z_{z} \left(\overline Y - Y \right),\\
&= \left(X \overline X' \overline Z_{z'} - \overline X X' \overline Z_{z} \right) \cdot \left(Y - \overline Y\right).
\end{align*}
Since $Y - \overline Y > 0$, we deduce that
$F(\rho(T)) < F(T)$ if and only if $X \overline X' \overline Z_{z'} > \overline X X' \overline Z_{z}$.
\end{proof}
For the remainder of this section, let $T$ be a tree with a vertex $v$ of degree $k + 1 \geq 3$ such that the components of $T - v$ can be denoted as $T', T_{1}, T_{2}, \dots, T_{k}$ in such a way that,
letting $v_{i}$ ($i\in \{1, \dots, k\}$) be the neighbor of $v$ in $V(T_{i})$
and letting $T^{+}_{i}$ be the tree obtained from $T_i$ by adding the vertex $v$ and the edge $vv_i$, the following holds:
\begin{itemize}
\item $T^{+}_{1}$ is a {\tos}, and
\item for each $i\in \{2, \dots, k\}$, at least one of the following two conditions holds:
\begin{enumerate}[(C1)]
\item \label{type1} $T_{i}$ is a {\tos} and $v_{i}$ is a leaf of $T_{i}$,
\item \label{type2} $T^{+}_{i}$ is a {\tos}.
\end{enumerate}
\end{itemize}
Also, let $v'$ be the neighbor of $v$ in $V(T')$, and let
$w$ be a leaf of $T_{1}$ distinct from the vertex $v_{1}$ (such a leaf exists
since $|T^{+}_{1}| \geq 3$).
Finally, let us emphasize that no assumption is made on the component $T'$, that is, $T'$
is an arbitrary non-empty tree.
The following lemma is a crucial tool in our proof of Theorem~\ref{th-task1}:
\begin{lem} \label{lem-gr}
Let $T$ be as above. Then there exists a {good rotation} in $T$.
\end{lem}
In order to prove Lemma~\ref{lem-gr}, we distinguish three cases:
\begin{itemize}
\item $T_{1}$ is a path and (C\ref{type1}) holds for some $i\in\{2,\dots, k\}$;
\item $T_{1}$ is not a path and (C\ref{type1}) holds for some $i\in\{2,\dots, k\}$, and
\item (C\ref{type2}) holds for every $i\in\{2,\dots, k\}$.
\end{itemize}
(Observe that $T$ falls in at least one of these cases.)
We consider each of these cases separately;
Lemma~\ref{lem-gr} is obtained by combining
Lemmas~\ref{lem-case1-gr}, \ref{lem-case2-gr} and \ref{lem-case3-gr}, which
respectively address the first, second, and third case.
The latter lemmas rely in turn on Lemmas~\ref{lem-case1_2-mss} and~\ref{lem-case3-mss}
below, showing that some specific rotations do not change the stability number of $T$.
Let $\rho_1$ and $\rho_2$ denote the two rotations $\rho_1 := (v'v, v'w)$ and $\rho_2 := (vv_1, vw)$
(see Figure~\ref{fig-two-rot} for an illustration).
\begin{figure}
\begin{center}
\includegraphics{fig-two-rot}
\caption{\label{fig-two-rot}
Rotations $\rho_1 = (v'v, v'w)$ and $\rho_2 = (vv_1, vw)$.
}
\end{center}
\end{figure}
\begin{lem} \label{lem-case1_2-mss}
Suppose that (C\ref{type1}) holds for some $i\in\{2,\dots, k\}$.
Then $\alpha(T)=\alpha(\rho_2(T))$, and $\alpha(T)=\alpha(\rho_1(T))$ if $T_{1}$ is a path.
\end{lem}
\begin{proof}
Assume without loss of generality that (C\ref{type1}) holds for $i=2$.
We first show:
\begin{equation}
\label{eq:avoid-v}
\textrm{Each of $T$, $\rho_1(T)$, and $\rho_2(T)$ contains a {maximum stable set} not including $v$.}
\end{equation}
First, consider the tree $T$. Let $S$ be a {maximum stable set} of $T$. Suppose that $v \in S$ (otherwise, we are trivially done). Then $v_{2} \notin S$.
Since $T_2$ is a {\tos}, $T_{2}$ has a unique {maximum stable set}, namely $S_2:=V(T_{2}) - \C(T_{2})$
(see Lemma~\ref{lem-mss}).
In particular, $v_{2} \in S_{2}$, implying $| S \cap V(T_2) | < | S_2 |$. Let
$$
S' := \big(S - (V(T_2)\cup \{v\})\big) \cup S_2.
$$
Observe that $S'$ is also a stable set of $T$, and $|S'| \geq |S|$. It follows that $S'$ is a {maximum stable set} of $T$ with $v\notin S'$, as claimed.
Finally, note that the above argument still holds if we replace the tree $T$ by $\rho_1(T)$ or $\rho_2(T)$,
which completes the proof of \eqref{eq:avoid-v}.
By~\eqref{eq:avoid-v}, there is a {maximum stable set} $S$ of $T$ such that $v\notin S$.
The set $S$ is also a stable set of $\rho_2(T)$, implying $\alpha(\rho_2(T)) \ge \alpha(T)$.
Conversely, there is a {maximum stable set} $S'$ of $\rho_2(T)$ with $v\notin S'$.
Since $S'$ is a stable set of $T$, this shows $\alpha(\rho_2(T)) \le \alpha(T)$, and hence
$\alpha(T) = \alpha(\rho_2(T))$.
Now, suppose that $T_{1}$ is a path. Observe that $T_{1}$ has odd length, since $T^+_1$ is a {\tos}.
By~\eqref{eq:avoid-v}, there is a {maximum stable set} $S'$ of $\rho_1(T)$ such that $v\notin S'$.
This set is also a stable set of $T$, implying $\alpha(T) \ge \alpha(\rho_1(T))$.
Let $S$ be a {maximum stable set} of $T$. If $v'\notin S$ or $w \notin S$ then $S$ is also a stable set of $\rho_1(T)$, showing
$\alpha(T) \le \alpha(\rho_1(T))$, and hence $\alpha(T) = \alpha(\rho_1(T))$. Assume thus $v',w \in S$.
Then $v \notin S$. Since $T_{1}$ is a path with an even number of vertices, $V(T_{1})$ can be uniquely partitioned into two maximum stable sets $S_{1}$ and $S_{2}$ of $T_{1}$. Assuming without loss
of generality $w\in S_{1}$, we then have $S \cap V(T_1) = S_{1}$, because otherwise we could modify $S$
and find a larger stable set of $T$. It follows that
$$
\tilde S := (S - S_{1}) \cup S_{2}
$$
is another stable set of $T$ with $|\tilde S|=|S|$ and $w\notin \tilde S$, and we deduce
that $\alpha(T) = \alpha(\rho_1(T))$, as previously.
\end{proof}
\begin{lem} \label{lem-case3-mss}
Suppose that (C\ref{type2}) holds for every $i\in\{2,\dots, k\}$. Then $\alpha(T) = \alpha(\rho_1(T))$.
\end{lem}
\begin{proof}
First, we show that both $T$ and $\rho_1(T)$ contain a {maximum stable set} that does not include $v'$.
Let $S$ be a {maximum stable set} of $T$ (respectively, $\rho_1(T)$), and suppose that $v' \in S$. Then, $v \notin S$ (respectively, $w \notin S$).
The tree $T^+_i$ ($i\in\{1,\dots, k\}$) is a {\tos}, and hence
contains a unique {maximum stable set} $S^+_i$ by Lemma~\ref{lem-mss}. (Note that $v, w \in S^+_1$, and
$v \in S^+_i$ for every $i\in\{2,\dots, k\}$.)
Since $v \notin S$ or $w \notin S$, it follows that $| S \cap V(T^+_1) | < | S^+_1 |$. Also,
$| S \cap V(T^+_i) | \le | S^+_i |$ for every $i\in\{2,\dots, k\}$.
Let
$$
S' := \big(S \cap (V(T') - \{v'\})\big) \cup S^+_1 \cup S^+_2 \cdots \cup S^+_k.
$$
The set $S'$ is a stable set of $T$ (respectively, $\rho_1(T)$), and $|S'| \geq |S|$ by the previous observations, implying that
$S'$ is a {maximum stable set} with $v' \notin S'$.
Now, observe that every stable set of $T$ (respectively $\rho_1(T)$) that does not include $v'$ is also a stable set of $\rho_1(T)$
(respectively $T$). Since there exists a {maximum stable set} not including $v'$ in each of these two trees, it follows $\alpha(T) = \alpha(\rho_1(T))$.
\end{proof}
\begin{lem} \label{lem-case1-gr}
Suppose that $T_{1}$ is a path and that (C\ref{type1}) holds for some $i\in\{2,\dots, k\}$.
Then the rotation $\rho_1$ is good for $T$.
\end{lem}
\begin{proof}
We have that $\alpha(T) = \alpha(\rho_1(T))$ by Lemma~\ref{lem-case1_2-mss}, thus
it remains to prove that $F(\rho_1(T)) < F(T)$.
Let $x:=v, x':=w, y:=v'$, and consider Notation~\ref{nota-rot}
with respect to the rotation $\rho_1=(v'v, v'w)=(yx,yx')$.
We have $X'=\overline X' = 1$ and $\overline Z_z = \overline Z_{z'}$ since $T_{1}$ is a path.
Also, $\overline X < X$, because $\ell=k-1 > 0$. Therefore,
$X \overline X' \overline Z_{z'} > \overline X X' \overline Z_{z}$, and hence $F(\rho_1(T)) < F(T)$
by Lemma~\ref{lem-rotating-edge}.
\end{proof}
\begin{lem} \label{lem-case2-gr}
Suppose that $T_{1}$ is not a path and that (C\ref{type1}) holds for some $i\in\{2,\dots, k\}$.
Then the rotation $\rho_2$ is good for $T$.
\end{lem}
\begin{proof}
Since (C\ref{type1}) holds for some $i\in\{2,\dots, k\}$,
Lemma~\ref{lem-case1_2-mss} implies that $\alpha(T) = \alpha(\rho_2(T))$.
We prove that $F(\rho_2(T)) < F(T)$, using Lemma~\ref{lem-rotating-edge} again.
Let $x:=v_{1}, x':=w, y:=v$, and consider the notations
associated to the rotation $\rho_2=(vv_{1}, vw)=(yx,yx')$.
(Thus, in particular, $z$ and $z'$ are the neighbors of respectively $v_1$ and $w$ that are included in the unique $v_1w$-path.)
We have $X'=\overline X' = 1$ since $w$ is a leaf. Hence, by Lemma~\ref{lem-rotating-edge},
to prove $F(\rho_2(T)) < F(T)$ it suffices to show
that $\overline X \ \overline Z_z < X \overline Z_{z'}$.
If $v_{1}w \in E(T)$, then $T_z$ is empty and $\overline Z_z = \overline Z_{z'} = 1$. Moreover, $\l \ge1$ because $T_1$ is not a path,
implying $\overline X < X$ and $F(\rho_2(T)) < F(T)$.
Now, assume that $v_{1}w \notin E(T)$.
Since $v_{1}$ is a center of the {\tos} $T^+_1$, its distance to the leaf $w$ is odd, and hence at least $3$. This implies $z \neq z'$.
Also, the component of $T^+_1-v_{1}$ that includes $w, z$, and $z'$ is also a {\tos}; hence,
$T_{z}$ is either a {\tos}, or almost a {\tos}, or an odd path.
In the first two cases,
$z$ and $z'$ are respectively a leaf and a center of $T_{z}$. Furthermore, $z'$ is the {exposed center} of $T_{z}$
in the second case. It follows $Z > \phi \cdot \overline Z_z$ and $Z < \phi \cdot \overline Z_{z'}$
from respectively Lemmas~\ref{lem-ToS-leaf} and~\ref{lem-ToS-center}, implying $\overline Z_z < \overline Z_{z'}$.
This in turn implies $\overline X \ \overline Z_z < X \overline Z_{z'}$ since trivially $\overline X \le X$.
In the third case, we have that $\overline Z_z = \overline Z_{z'}$, because $T_{z} - z$ is isomorphic to
$T_{z} - z'$. Since $T_{1}$ itself is not a path, we also have $\ell = \deg_{T}(v_{1}) - 2 > 0$, implying
$\overline X < X$. It follows again $\overline X \ \overline Z_z < X \overline Z_{z'}$.
\end{proof}
\begin{lem} \label{lem-case3-gr}
Suppose that (C\ref{type2}) holds for every $i\in\{2,\dots, k\}$.
Then $T$ admits a {good rotation}.
\end{lem}
\begin{proof}
The proof involves two different rotations.
First, let $x:=v, x':=w, y:=v'$, and consider the rotation $\rho_{1}=(v'v, v'w)=(yx,yx')$ and the corresponding notations.
(Thus, $z=v_{1}$ and $z'$ is the unique neighbor $w'$ of $w$.)
We have $X'=\overline X' = 1$ and $\overline X < X$.
If $\overline Z_{z} \le \overline Z_{z'}$, then $F(\rho_1(T)) < F(T)$ by Lemma~\ref{lem-rotating-edge}.
Also, $\alpha(T) = \alpha(\rho_1(T))$ by Lemma~\ref{lem-case3-mss}, implying that
$\rho_{1}$ is a good rotation.
Thus, we may assume that $\overline Z_{z} > \overline Z_{z'}$, that is,
\begin{equation}
\label{eq:T1}
F(T_1 - \{w, v_1\}) > F(T_1 - \{w, w'\}).
\end{equation}
(Observe that this implies $w' \neq v_{1}$.)
\begin{figure}
\begin{center}
\includegraphics[scale=0.9]{fig-second-rot}
\caption{\label{fig-second-rot}
The rotation $\rho_3 = (vv_{1}, vw')$ in the proof of Lemma~\ref{lem-case3-gr} (note that possibly $z=z'$).
}
\end{center}
\end{figure}
Now, let $\rho_3$ be the rotation $\rho_3:= (vv_1,vw')$ (see Figure~\ref{fig-second-rot} for an illustration).
We will show that $\rho_3$ is a good rotation.
Similarly as before, let $x:=v_{1}, x':=w', y:=v$, and consider the notations associated to $\rho_3 = (vv_1,vw')=(yx,yx')$.
We have
$$
F(T_1-\{w,v_1\}) = X(X'Z + \overline X' \overline Z_{z'})
$$
and
$$
F(T_1-\{w,w'\}) = X'(XZ + \overline X \ \overline Z_{z}).
$$
Thus,~\eqref{eq:T1} implies that $X \overline X' \overline Z_{z'} > \overline X X' \overline Z_{z}$, and it follows from
Lemma~\ref{lem-rotating-edge} that $F(\rho_3(T)) < F(T)$.
Therefore, it remains to show that $\alpha(T) = \alpha(\rho_3(T))$.
Let $S$ be a {maximum stable set} of $T$. We may assume that $w' \notin S$. (Indeed, if not, consider the set $(S - \{w'\}) \cup \{w\}$ instead.)
The set $S$ is also a stable set of $\rho_3(T)$, showing $\alpha(T) \leq \alpha(\rho_3(T))$.
Now, let $S$ be a {maximum stable set} of $\rho_3(T)$.
If $v\notin S$ or $v_{1}\notin S$, then $S$ is a stable set of $T$, showing $\alpha(T) \geq \alpha(\rho_3(T))$,
and hence $\alpha(T) = \alpha(\rho_3(T))$. Thus, suppose that $v, v_{1} \in S$.
The set $\tilde S:=(S - \{v_{1}\}) \cap V(T^{+}_{1})$ is a stable set of the {\tos} $T^{+}_{1}$.
Moreover, $|\tilde S| < |S^{+}_{1}|$ since $z\notin \tilde S$, where $S^{+}_{1}:=V(T^{+}_{1}) - \C(T^{+}_{1})$ is
the unique {maximum stable set} of $T^{+}_{1}$ (see Lemma~\ref{lem-mss}). This implies that the set
$$
(S - V(T^{+}_{1})) \cup S^{+}_{1}
$$
has cardinality at least that of $S$, and furthermore is a stable set of $T$.
Again, it follows that $\alpha(\rho_3(T)) = \alpha(T)$.
\end{proof}
\subsection{Proof of Theorem~\ref{th-task1}}
Now, we may turn to the proof of Theorem~\ref{th-task1}.
\begin{proof}[Proof of Theorem~\ref{th-task1}]
Let $T$ be an extremal tree. Arguing by contradiction, assume that $T$ is neither a path nor a {\ts}.
(Recall that even paths are {\tss}.)
Root the tree $T$ at an arbitrary leaf $r$.
Let $T_{v}$ ($v\in V(T)$) be the subtree of $T$ rooted at vertex $v$.
To each leaf $u$ of $T$ we associate a corresponding {\DEF witness}, defined
as the highest ancestor $v$ of $u$ such that $T_{v}$ is a {\tos} {\em and} $v$ is a
leaf of $T_{v}$. Since $u$ itself satisfies these two conditions, the latter
vertex is well-defined.
Let us look at a few properties of witnesses: First, clearly $r$ is not a witness (for otherwise
$T$ would be a {\tos}). Also, if $v$ and $w$ are two distinct witnesses, then
$v$ is neither an ancestor nor a descendant of $w$. (In other words, the set of
witnesses forms an antichain in the partial order implied the rooted tree.)
We may assume that $r$ has been chosen so that it satisfies:
\begin{equation}
\label{eq-change-root}
\textrm{
If the neighbor of $r$ has degree $2$ in $T$, then $T - r$ is not a {\ts}.
}
\end{equation}
Indeed, let $v$ be the neighbor of $r$, and
suppose $\deg_{T}(v)=2$ and that $T - r$ is a {\tos}. Since $T - r$ is not a path,
there is a vertex $w$ with degree at least 3 in $T-r$, which is therefore a center of $T-r$.
Since $v$ is a leaf of $T-r$, it follows that the distance in $T$ between $r$ and $w$ is even.
Now, consider a component of $T-w$ that does not contain $r$, and select a leaf $r'$ of that component.
The tree $T-r'$ cannot be a {\ts}, because $w$ is at even distance from the leaf $r$ in $T-r'$.
Thus, we deduce that~\eqref{eq-change-root} holds if we root $T$ at $r'$ instead of $r$.
Let $u$ be a witness of maximum depth in $T$, and let $v_{1}$ be the parent of $u$.
It follows from~\eqref{eq-change-root} that $v_{1}\neq r$, hence $v_{1}$ has a parent $v$.
Let $T_{1}^{+}$ be the tree obtained from $T_{v_1}$ by adding the vertex $v$ and the edge $vv_1$.
We show:
\begin{equation}
\label{eq-T1}
\textrm{
$T_{1}^{+}$ is a {\tos}.
}
\end{equation}
This is trivially true if $\deg(v_{1})=2$, hence assume $\deg(v_{1})\geq 3$,
and let $u_{1}, \dots, u_{\ell}$ denote the children of $v_{1}$ distinct from $u$.
Let $j\in \{1, \dots, \ell\}$, and
consider a leaf $z$ of $T$ which is contained in $T_{u_{j}}$.
In $T$, the witness $w$ of $z$ cannot be higher than $u_{j}$ (since
otherwise $w$ would be an ancestor of the witness $u$). Also, $w$ cannot be
a descendant of $u_{j}$, because it would contradict the fact that $u$ has maximum
depth among all witnesses. Thus $w=u_{j}$, and hence $T_{u_{j}}$ is a {\tos}
and $u_{j}$ is a leaf of $T_{u_{j}}$. It follows that
$T_{1}^{+}$ is also a {\tos}.
The vertex $v$ has at least two children (counting $v_{1}$), since
otherwise $T_{v} = T^{+}_{1}$, contradicting the fact that $u$ is
a witness. Observe that this implies $v \neq r$.
Let $v_{2}, \dots, v_{k}$ be the children of $v$ that are
distinct from $v_{1}$, and let $T_{i}^{+}$ ($i\in \{2, \dots, k\}$)
be the tree obtained from $T_{v_i}$ by adding the vertex $v$ and the edge $vv_i$.
Let $i\in \{2, \dots, k\}$, and consider any witness $u'$
contained in $V(T_{v_{i}})$ (observe that there must be at least one).
Since the depth of $u'$ is at most that of $u$, either $u'=v_{i}$ or $u'$ is a child of $v_{i}$.
In the first case, $T_{v_{i}}$ is a {\tos} and $v_{i}$ is a leaf of that tree,
by definition of a witness.
In the second one, since $u$ and $u'$ have the same depth, the proof
of~\eqref{eq-T1} directly shows that $T_{i}^{+}$
is a {\tos} (one just needs to replace $u$ by $u'$ and $v_{1}$ by $v_{i}$).
Let $T'$ be the component of $T - v$ containing the root $r$.
Also, let $T_{i} := T_{v_{i}}$ for $i \in \{1, \dots, k\}$.
Let us summarize the previous observations:
$T_{1}^{+}$ is a {\tos}, and for every $i \in \{2, \dots, k\}$, either
$T_i$ is a {\tos} and $v_{i}$ is a leaf of $T_{i}$, or $T_{i}^{+}$ is a {\tos}.
It follows that $T$ satisfies the requirements of Lemma~\ref{lem-gr}, and therefore
contains a good rotation, contradicting the fact that $T$ is extremal.
\end{proof}
\section{Extremal Trees of Stars are Balanced}
\label{sec-task2}
We have seen that every extremal tree is a {\ts} or an odd path (cf.\ Theorem~\ref{th-task1}).
In this section, we refine this result by showing that,
if an extremal tree $T$ is a {\ts}, then $T$ must be balanced.
(Recall that a {\ts} is balanced if the degrees of every two of its centers differ by at most $1$.)
\begin{thm}
\label{th-task2}
Every extremal tree is either a balanced {\ts} or an odd path.
\end{thm}
We again resort to edge rotations to prove Theorem~\ref{th-task2}. To this aim,
we need to introduce a few additional lemmas.
For an integer $k\geq 2$, let $f_{k}: \mathbb R{} \to \mathbb R{}$ be the function defined as
$$
f_{k}(x) := x^{k} - x^{k-1} + 2x - 1.
$$
\begin{lem}
\label{lem-f-increasing}
The function $f_{k}(\cdot)$ is strictly increasing on the interval $[0,1]$.
\end{lem}
\begin{proof}
We prove that the first derivative $f'_{k}(\cdot)$ is strictly positive on $[0,1]$,
by induction on $k$. Let $x\in [0,1]$. We have $f'_{k}(x) > 0$ when $k=2$, since $f'_{2}(x) = 2x + 1 > 0$.
For the inductive step, let $k \geq 3$,
and rewrite $f'_{k}(x)$ as follows:
\begin{align*}
f'_{k}(x) &= kx^{k-1} - (k-1)x^{k-2} + 2 \\
&= 2 + x\left( (k-1)x^{k-2} + x^{k-2} - (k-2)x^{k-3} - x^{k-3} \right) \\
&= 2 + x\left( f'_{k-1}(x) - 2 \right) + x^{k-1} - x^{k-2}.
\end{align*}
Since $f'_{k-1}(x) - 2 > -2$ by the induction hypothesis, we deduce
\begin{align*}
f'_{k}(x) &> 2 - 2x + x^{k-1} - x^{k-2}\\
&=(1-x)(2-x^{k-2}) \\
&\ge 0,
\end{align*}
as claimed.
\end{proof}
Since $f_{k}(1/2)= -1/2^{k} < 0 < 1 = f_{k}(1)$ and $f_{k}(\cdot)$ is continuous,
it follows from Lemma~\ref{lem-f-increasing}
that $f_{k}(\cdot)$ has a unique root in the open interval $(1/2, 1)$, which we denote by $R_{k}$.
\begin{lem}
\label{lem-Rk}
$R_{k} < 2^{-k/(k+1)}$ for every $k\geq 2$.
\end{lem}
\begin{proof}
Let
$$
g(k) := f_{k}\left(2^{-k/(k+1)}\right)
= 2^{- k^{2}/(k+1)} - 2^{ - k(k-1)/(k+1)}
+ 2 \cdot 2^{- k/(k+1)} - 1.
$$
By Lemma~\ref{lem-f-increasing}, it is enough to show $g(k) > 0$ for every integer $k \geq 2$.
A quick hand-on computation shows that this is true for $k=2$ and $3$.
Hence, we may assume $k\geq 4$.
We first rewrite $g(k)$ as follows:
\begin{align*}
g(k) &= 2^{- k^{2}/(k+1)} - 2^{ - k(k-1)/(k+1)}
+ 2 \cdot 2^{- k/(k+1)} - 1 \\
&= 2^{- k^{2}/(k+1)} - 2^{ k/(k+1) - k^{2}/(k+1)}
+ 2^{1/(k+1)} - 1 \\
&= 2^{1/(k+1)} - 1
- \frac{2^{k/(k+1)} - 1}{2^{k^{2}/(k+1)}}.
\end{align*}
Using $k^{2}/(k+1) > k - 1$ and $2^{k/(k+1)} < 2$, we then obtain
\begin{align*}
g(k) &> 2^{1/(k+1)} - 1 - \frac{2^{k/(k+1)} - 1}{2^{k-1}} \\
&> 2^{1/(k+1)} - 1 - 2^{-(k-1)}.
\end{align*}
For every real $y>0$, we have $e^{y} > 1 + y$, and hence $2^{y} > 1 + y\ln 2$,
since $2^{y} = e^{y\ln2}$. It follows
$$
g(k) > \frac{\ln 2}{k+1} - \frac{1}{2^{k-1} }
= \frac{1}{k+1}\left(\ln 2 - \frac{k+1}{2^{k-1}} \right).
$$
Since the function $\ln 2 - (k+1)/(2^{k-1})$ is increasing in $k$ and positive for $k=4$,
we deduce that $g(k) > 0$, as claimed.
\end{proof}
\begin{lem}
\label{lem-ToS-leaf-gen}
Let $k\geq 2$ be an integer, let $T$ be a {\ts} such that all its centers have degree at least $k$, and
let $v$ be a leaf of $T$. Then $F(T - v) < R_{k} \cdot F(T)$.
\end{lem}
Observe that $R_{2}=1 / \phi$; hence, this lemma generalizes Lemma~\ref{lem-ToS-leaf} in the case of {\tss}.
\begin{proof}[Proof of Lemma~\ref{lem-ToS-leaf-gen}]
The proof is by induction on $|T|$. The claim is true when $|T|=1$, since then
$F(T-v) / F(T) = 1/2 < R_{k}$.
For the inductive step, assume $|T| > 1$, and
let $w$ be the unique neighbor of $v$. Each component of $T-\{v,w\}$ is a {\ts};
let us denote these trees by $T_{1}, \dots, T_{\ell}$ (thus $\ell\geq k-1$). Let also $v_{i}$
be the leaf of $T_{i}$ that is adjacent to $w$ in $T$.
Letting
$$
\gamma := \prod_{i=1}^{\ell} \frac{F(T_{i} - v_{i})}{F(T_{i})},
$$
we deduce that
$$
\frac{F(T-v)}{F(T)} = \frac{\prod_{i=1}^{\ell} F(T_{i})
+ \prod_{i=1}^{\ell} F(T_{i}-v_{i})}{2\prod_{i=1}^{\ell} F(T_{i})
+ \prod_{i=1}^{\ell} F(T_{i}-v_{i})} =\frac{1 + \gamma}{2 + \gamma}.
$$
The induction hypothesis gives
$$
0 < \gamma < (R_{k})^{\ell} \leq (R_{k})^{k-1}.
$$
By definition of $R_{k}$, we have $(R_{k})^{k} - (R_{k})^{k-1} + 2R_{k} - 1=0$.
It then follows
$$
\frac{F(T-v)}{F(T)} = \frac{1 + \gamma}{2 + \gamma} <
\frac{1 + (R_{k})^{k-1}}{2 + (R_{k})^{k-1}} = R_{k},
$$
as desired.
\end{proof}
Now, we may prove Theorem~\ref{th-task2}.
\begin{proof}[Proof of Theorem~\ref{th-task2}]
Let $T$ be an extremal tree. We know by Theorem ~\ref{th-task1} that $T$ is a {\ts} or an odd path.
Arguing by contradiction, we assume that $T$ is a {\ts} which is not balanced.
We will show that $T$ admits a good rotation, which contradicts the fact that $T$ is extremal.
The tree $T$ has at least two centers, as otherwise $T$ is trivially balanced.
Let $x$ and $x'$ be distinct centers of $T$ maximizing the difference $\deg(x) - \deg(x')$.
Let also $k:=\deg(x)$ and $\ell:=\deg(x')$ (hence, $k \geq \ell + 2$).
Choose a neighbor $y$ of $x$ that is not on the unique $xx'$-path in $T$, and consider
the rotation $\rho:=(yx,yx')$ (see Figure~\ref{fig-rot-th-task2} for an illustration). Observe that $\rho(T)$ is also a {\ts}, and that $T$ and $\rho(T)$ have the same set of centers. In particular, $\alpha(\rho(T))=\alpha(T)$ by Lemma~\ref{lem-mss}.
Hence, to prove that $\rho$ is a good rotation, it remains to show that $F(\rho(T)) < F(T)$.
This will be done by combining Lemma~\ref{lem-rotating-edge}
with Lemmas~\ref{lem-Rk} and ~\ref{lem-ToS-leaf-gen}.
\begin{figure}[htb]
\begin{center}
\includegraphics[scale=0.9]{fig-rot-th-task2}
\caption{\label{fig-rot-th-task2}
The rotation $\rho$ used in the proof of Theorem~\ref{th-task2}.
}
\end{center}
\end{figure}
By our choice of $x$ and $x'$, every component of $T- \{x , x'\}$ is a {\ts}, every center
of which has degree at least $\ell$.
Consider Notation~\ref{nota-rot} with respect to the rotation $\rho=(yx,yx')$.
(Thus, $z$ and $z'$ are the neighbors of respectively $x$ and $x'$ that lie on the $xx'$-path in $T$,
and possibly $z=z'$.)
It follows from Lemma~\ref{lem-ToS-leaf-gen} that
$$
\frac{\overline Z_{z}}{Z} < R_{\ell}
$$
and
$$
\frac{\overline X}{X} < (R_{\ell})^{k-2} \leq (R_{\ell})^{\ell}.
$$
Also, we have that
$$
\frac{\overline Z_{z'}}{Z} \geq \frac{1}{2}
$$
and
$$
\frac{\overline X'}{X'} \geq \frac{1}{2^{\ell-1}},
$$
since $F(T') \leq 2F(T' - v)$
holds for every tree $T'$ and vertex $v$ of $T'$ (cf.\ Lemma~\ref{lem:basic}).
Combining these inequalities with Lemma~\ref{lem-Rk}, we obtain
\begin{align*}
X\overline X' \overline Z_{z'} &> \frac{\overline X}{(R_{\ell})^{\ell}}
\cdot \frac{X'}{2^{\ell-1}} \cdot \frac{\overline Z_{z}}{2R_{\ell}} \\
&= \overline X X' \overline Z_{z}\cdot \frac{(R_{\ell})^{-(\ell+1)}}{2^{\ell}} \\
&> \overline X X' \overline Z_{z}\cdot \frac{ \left ( 2^{-\ell/(\ell+1)} \right)^{-(\ell+1)}}{2^{\ell}} \\
&=\overline X X' \overline Z_{z}.
\end{align*}
Therefore, $F(\rho(T)) < F(T)$ by Lemma~\ref{lem-rotating-edge}. This completes the proof.
\end{proof}
\section{Further Results}
\label{sec-task3}
In Section~\ref{sec-task2}, it has been shown that every extremal tree $T$ that is not an odd path must be a balanced {\ts}. An open problem is to understand how these stars are linked together. A few results in this direction are given here, but the general problem is far from being solved.
Let $T$ be a {\ts}. The \emph{{center-tree}} of $T$ is the tree $C_T$ having the centers of $T$ as vertex set, and where two centers are adjacent
if they are at distance 2 in $T$ (see Figure~\ref{fig-ct} for an illustration).
\begin{figure}[!htb]
\begin{center}
\includegraphics[scale=0.85]{fig-ct}
\caption{\label{fig-ct}
A {\tos} (left) and its corresponding {center-tree} (right).}
\end{center}
\end{figure}
Let $T$ be a balanced {\tos} with $n$ vertices and stability number $\alpha$.
Since $T$ is balanced, every center of $T$ has degree
either $\left\lceil \frac{n-1}{n-\alpha} \right\rceil$
or $\left \lceil \frac{n-1}{n-\alpha} \right \rceil - 1$.
A center is said to be {\DEF heavy}
in the first case, {\DEF light} in the second.
It can be checked that the number of heavy centers is given by
$$
h(n,\alpha) :=
\left\{
\begin{array}{lll}
(n-1) \mod (n-\alpha) & & \text{if $(n-1) \mod (n-\alpha) \neq 0$} \\
n-\alpha & & \text{otherwise}. \\
\end{array}
\right.
$$
The number of light centers of $T$ is thus
$$
\l(n,\alpha) := n - \alpha - h(n,\alpha).
$$
The next theorem gives a simple characterization of the extremal trees when
$\l(n,\alpha) \leq 2$.
Let us remark that, if $\alpha = n/2$, then the path $P_{n}$
is the only tree with $n$ vertices
and stability number $\alpha$ that is extremal, as follows from
the result of Prodinger and Tichy~\cite{Prodinger82} mentioned in the introduction.
Hence, we assume $\alpha > n/2$ in what follows.
\begin{thm}
\label{th-ct-path}
Let $T$ be a tree with $n$ vertices and stability number $\alpha > n/2$, and
assume $\l(n,\alpha) \leq 2$. Then, $T$ is extremal if and only if
$T$ satisfies the following three conditions:
\begin{itemize}
\item $T$ is a balanced {\ts};
\item the {center-tree} of $T$ is isomorphic to a path $P$, and
\item each light center of $T$ (if any) is an endpoint of $P$.
\end{itemize}
\end{thm}
Our proof of Theorem~\ref{th-ct-path} is based on the following lemma.
\begin{lem} \label{lem-task3-a}
Let $T$ be a balanced {\ts}, and assume that a leaf $w$ of the {center-tree} $C_{T}$ is heavy.
If $T$ is extremal, then $C_{T}$ is a path, and all internal vertices of $C_{T}$ are heavy.
\end{lem}
\begin{proof}
Arguing by contradiction, suppose that $T$ is extremal and contains a center $v$
such that either $v$ is heavy with degree at least $3$ in $C_{T}$, or light with degree at least $2$ in $C_{T}$.
We may assume that $v$ has been chosen so that every inner vertex of the $vw$-path in $C_{T}$
is heavy and has degree $2$ in $C_{T}$.
Let $P$ be the unique $vw$-path in the tree $T$.
Let $y$ be a neighbor of $v$ in the {center-tree} $C_{T}$ such that $y \notin V(P)$. Let $x$ be the vertex of $T$
that is adjacent to both $v$ and $y$.
Let $x'$ be a leaf of $T$ that is adjacent to $w$. (Thus, $x' \notin V(P)$.)
Let $\rho$ be the rotation $\rho:=(yx,yx')$ (see Figure~\ref{fig-rot-lem-task3-a} for an illustration).
The tree $\rho(T)$ is a {\ts} and has the same number of centers as $T$, thus
$\alpha(\rho(T))=\alpha(T)$ by Lemma~\ref{lem-mss}. Hence, to reach a contradiction,
it is enough to show that $F(\rho(T)) < F(T)$.
\begin{figure}[htb]
\begin{center}
\includegraphics[scale=0.9]{fig-rot-lem-task3-a}
\caption{\label{fig-rot-lem-task3-a}
The rotation $\rho$ used in the proof of Lemma~\ref{lem-task3-a}.
}
\end{center}
\end{figure}
Consider Notation~\ref{nota-rot} with respect to the rotation $\rho$.
Thus $z = v$, $z' = w$, and $X = \overline X = X' = \overline X' = 1$.
By Lemma~\ref{lem-rotating-edge}, $F(\rho(T)) < F(T)$ if and only if $\overline Z_{z} < \overline Z_{z'}$.
Let $u$ and $u'$ be the neighbors of respectively $z$ and $z'$ in the path $P$.
Let $\widetilde T$ be the component of $T - \{z, z'\}$ that contains both $u$ and $u'$.
Since every center of $T$ included in $V(P) - \{z, z'\}$ is heavy and has degree $2$ in $C_{T}$, the trees
$\widetilde T - u$ and $\widetilde T - u'$ are isomorphic.
Let $v_{1}, \dots, v_{k}$ be the neighbors of $z$ ($=v$) in $T$ that are distinct from $u$ and $x$.
Let $T^{z}_{i}$ ($1 \leq i \leq k$) be the component of $T - \{z\}$ that includes $v_{i}$.
Let $\l:= \deg_{T}(z') - 2$. Thus, $\l = k$ or $k+1$, depending on whether $z$ is heavy or light.
Since the $\l$ neighbors of $z'$ that are distinct from $u'$ and $x'$ are all leaves of $T$, it follows that
\begin{align*}
\overline Z_{z} &= \left( \prod_{i=1}^k F(T^{z}_{i}) \right) \cdot \left(2^{\l}F(\widetilde T) + F(\widetilde T - u') \right), \\
\overline Z_{z'} &= 2^{\l} \cdot \left(F(\widetilde T)\prod_{i=1}^k F(T^{z}_{i}) + F(\widetilde T - u)\prod_{i=1}^k F(T^{z}_{i} - v_{i}) \right).
\end{align*}
Using that $F(\widetilde T - u) = F(\widetilde T - u')$, we obtain
$$
\overline Z_{z'} - \overline Z_{z} = F(\widetilde T - u) \cdot \left(2^{\l}\prod_{i=1}^k F(T^{z}_{i} - v_{i}) - \prod_{i=1}^k F(T^{z}_{i})\right).
$$
We have $F(T^{z}_{i}) \leq 2F(T^{z}_{i} - v_{i})$ for every $i \in \{1, \dots, k\}$, with strict inequality if $|T^{z}_{i}| > 1$.
If $z$ is light, then $\l = k + 1$, and
$$
2^{\l}\prod_{i=1}^k F(T^{z}_{i} - v_{i}) - \prod_{i=1}^k F(T^{z}_{i}) \geq
(2^{\l} - 2^{k})\prod_{i=1}^k F(T^{z}_{i} - v_{i}) > 0.
$$
Similarly, if $z$ is heavy, then $\l = k$. Moreover, $|T^{z}_{j}| > 1$ holds for some $j \in \{1, \dots, k\}$,
because $z$ has degree at least $3$ in $C_{T}$. This implies $F(T^{z}_{j}) < 2F(T^{z}_{j} - v_{j})$, and
$$
2^{\l}\prod_{i=1}^k F(T^{z}_{i} - v_{i}) - \prod_{i=1}^k F(T^{z}_{i}) >
(2^{\l} - 2^{k})\prod_{i=1}^k F(T^{z}_{i} - v_{i}) = 0.
$$
Thus, $\overline Z_{z} < \overline Z_{z'}$ holds in both cases, as claimed. This concludes the proof.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{th-ct-path}]
First, we observe that there is a unique tree (up to isomorphism) satisfying the three conditions
given in the statement of the theorem. Hence, it is enough to show that each of these three conditions
is necessary for $T$ to be extremal.
Thus, suppose that $T$ is extremal.
Then $T$ is a balanced {\ts} by Theorem~\ref{th-task2}
(note that $T$ cannot be an odd path since $\alpha > n/2$).
If the {center-tree} of $T$ is not a path, then it has
at least three leaves, and one of them is heavy.
But then Lemma~\ref{lem-task3-a} implies that $T$ is not extremal, a contradiction.
Hence, the {center-tree} is isomorphic to a path $P$. Furthermore, all internal vertices
of $P$ are heavy centers of $T$, by the same lemma. The theorem follows.
\end{proof}
When $\l(n,\alpha) \geq 3$ and $\alpha > n/2$, every extremal tree $T$ is
a balanced {\ts} by Theorem~\ref{th-task2}. However, the {center-tree} of $T$
is no longer necessarily a path.
This can be seen on Figure~\ref{fig-center-trees-n-24}, which provides a list
of the {center-tree s} of all extremal trees for $n=24$ and $\alpha > 12$ (these have been computed with
the system GraPHedron~\cite{Melot08}).
\begin{figure}
\begin{center}
\includegraphics[scale=0.9]{fig-center-trees-n-24}
\caption{\label{fig-center-trees-n-24}
The {center-tree s} of all {\tss} on $n=24$ vertices which are extremal.
Light centers are drawn in white and heavy centers in black.
}
\end{center}
\end{figure}
\section*{Acknowledgments}
The authors thank Guy Louchard for useful discussions, and
the two anonymous referees for their helpful comments on a previous version of the manuscript.
This work was supported by the {\em Communaut\'e Fran\c caise de Belgique (projet ARC)}.
|
\section{Introduction}\label{s1}
To describe the original Fredholm determinant result due to Jost and Pais \cite{JP51}, we need a few preparations. Denoting by $H_{0,+}^D$ and $H_{0,+}^N$ the one-dimensional Dirichlet and Neumann Laplacians in $L^2((0,\infty);dx)$, and assuming
\begin{equation}
V\in L^1((0,\infty);dx), \label{1.1}
\end{equation}
we introduce the perturbed Schr\"odinger operators
$H_{+}^D$ and $H_{+}^N$ in $L^2((0,\infty);dx)$ by
\begin{align}
&H_{+}^Df=-f''+Vf, \notag \\
&f\in \text{\rm{dom}}\big(H_{+}^D\big)=\{g\in L^2((0,\infty); dx) \,|\, g,g'
\in AC([0,R])
\text{ for all $R>0$}, \\
& \hspace*{4.8cm} g(0)=0, \, (-g''+Vg)\in L^2((0,\infty); dx)\}, \notag \\
&H_{+}^Nf=-f''+Vf, \notag \\
&f\in \text{\rm{dom}}\big(H_{+}^N\big)=\{g\in L^2((0,\infty); dx) \,|\, g,g'
\in AC([0,R])
\text{ for all $R>0$}, \\
& \hspace*{4.7cm} g'(0)=0, \, (-g''+Vg)\in L^2((0,\infty); dx)\}. \notag
\end{align}
Thus, $H_{+}^D$ and $H_{+}^N$ are self-adjoint if and only if $V$ is
real-valued, but since the latter restriction plays no special role in our results, we
will not assume real-valuedness of $V$ throughout this paper.
A fundamental system of solutions $\phi_+^D(z,\cdot)$,
$\theta_+^D(z,\cdot)$, and the Jost solution $f_+(z,\cdot)$ of
\begin{equation}
-\psi''(z,x)+V\psi(z,x)=z\psi(z,x), \quad z\in{\mathbb{C}}\backslash\{0\}, \;
x\geq 0, \label{1.4}
\end{equation}
are then introduced via the standard Volterra integral equations
\begin{align}
\phi_+^D(z,x)&=z^{-1/2}\sin(z^{1/2}x)+\int_0^x dx' \, z^{-1/2}\sin(z^{1/2}(x-x'))
V(x')\phi_+^D(z,x'), \\
\theta_+^D(z,x)&=\cos(z^{1/2}x)+\int_0^x dx' \, z^{-1/2}\sin(z^{1/2}(x-x'))
V(x')\theta_+^D (z,x'), \\
f_+(z,x)&=e^{iz^{1/2}x}-\int_x^\infty dx' \,
z^{-1/2}\sin(z^{1/2}(x-x')) V(x')f_+(z,x'), \label{1.7} \\
&\hspace*{3.85cm} z\in{\mathbb{C}}\backslash\{0\}, \; \Im(z^{1/2})\geq 0, \;
x\geq 0. \notag
\end{align}
In addition, we introduce
\begin{equation}
u=\exp(i\arg(V))\abs{V}^{1/2}, \quad v=\abs{V}^{1/2}, \, \text{ so
that } \, V=u\, v,
\end{equation}
and denote by $I_+$ the identity operator in $L^2((0,\infty); dx)$. Moreover,
we denote by
\begin{equation}
W(f,g)(x)=f(x)g'(x)-f'(x)g(x), \quad x \geq 0,
\end{equation}
the Wronskian of $f$ and $g$, where $f,g \in C^1([0,\infty))$. We
also use the standard convention to abbreviate (with a slight abuse of notation)
the operator of multiplication in $L^2((0,\infty);dx)$ by an element
$f\in L^1_{\text{\rm{loc}}}((0,\infty);dx)$ (and similarly in the higher-dimensional
context with $(0,\infty)$ replaced by an appropriate open set
$\Omega\subset{\mathbb{R}}^n$ later) by the same symbol $f$ (rather than $M_f$, etc.).
For additional notational conventions we refer to the paragraph at the end of this introduction.
Then, the following results hold (with ${\mathcal B}_1(\cdot)$ abbreviating the ideal of trace class operators):
\begin{theorem} \label{t1.1}
Assume $V\in L^1((0,\infty);dx)$ and let $z\in{\mathbb{C}}\backslash [0,\infty)$
with $\Im(z^{1/2})>0$. Then,
\begin{equation}
\overline{u\big(H_{0,+}^D-z I_+\big)^{-1}v}, \,
\overline{u\big(H_{0,+}^N-z I_+\big)^{-1}v} \in {\mathcal B}_1(L^2((0,\infty);dx))
\end{equation}
and
\begin{align}
\det\Big(I_+ +\overline{u\big(H_{0, +}^D-z I_+\big)^{-1}v}\,\Big) &=
1+z^{-1/2}\int_0^\infty dx\, \sin(z^{1/2}x)V(x)f_+(z,x) \notag \\
&= W(f_+(z,\cdot),\phi_+^D(z,\cdot)) = f_+(z,0), \label{1.11} \\
\det\Big(I_+ +\overline{u\big(H_{0, +}^N-z I_+\big)^{-1}v}\,\Big)
&= 1+ i z^{-1/2} \int_0^\infty
dx\, \cos(z^{1/2}x)V(x)f_+(z,x) \notag \\
&= - \frac{W(f_+(z,\cdot),\theta_+^D (z,\cdot))}{i z^{1/2}} =
\frac{f_+'(z,0)}{i z^{1/2}}. \label{1.12}
\end{align}
\end{theorem}
Equation \eqref{1.11} is the modern formulation of the celebrated result due to Jost and
Pais \cite{JP51}. Performing calculations similar to Section 4 in \cite{GM03} for the
pair of operators $H_{0,+}^N$ and $H_+^N$, one obtains the analogous result
\eqref{1.12}.
We emphasize that \eqref{1.11} and \eqref{1.12} exhibit a spectacular
reduction of a Fredholm determinant, that is, an infinite determinant (actually, a symmetrized perturbation determinant), associated with the trace class
Birmann--Schwinger kernel of a one-dimensional Schr\"odinger operator on the half-line $(0,\infty)$, to a simple Wronski determinant of ${\mathbb{C}}$-valued distributional solutions of
\eqref{1.4}. This fact goes back to Jost and Pais \cite{JP51} (see
also \cite{GM03}, \cite{Ne72}, \cite{Ne80}, \cite[Sect.\ 12.1.2]{Ne02},
\cite{Si00}, \cite[Proposition 5.7]{Si05}, and the extensive literature
cited in these references). The principal aim of this paper is to explore
the extent to which this fact may generalize to higher dimensions. While a
direct generalization of \eqref{1.11},
\eqref{1.12} appears to be difficult, we will next derive a formula for
the ratio of such determinants which indeed permits a natural extension to
higher dimensions.
For this purpose we introduce the boundary trace operators
$\gamma_D$ (Dirichlet trace) and $\gamma_N$ (Neumann trace) which, in the
current one-dimensional half-line situation, are just the functionals,
\begin{equation}
\gamma_D \colon \begin{cases} C([0,\infty)) \to {\mathbb{C}}, \\
\hspace*{1.3cm} g \mapsto g(0), \end{cases} \quad
\gamma_N \colon \begin{cases}C^1([0,\infty)) \to {\mathbb{C}}, \\
\hspace*{1.43cm} h \mapsto - h'(0). \end{cases}
\end{equation}
In addition, we denote by $m_{0,+}^D$, $m_+^D$, $m_{0,+}^N$, and $m_+^N$
the Weyl--Titchmarsh $m$-functions corresponding to $H_{0,+}^D$,
$H_{+}^D$, $H_{0,+}^N$, and $H_{+}^N$, respectively, that is,
\begin{align}
m_{0,+}^D(z) &= i z^{1/2}, \qquad m_{0,+}^N (z)= -\frac{1}{m_{0,+}^D(z)}
= i z^{-1/2}, \label{1.14} \\
m_{+}^D(z) &= \frac{f_+'(z,0)}{f_+(z,0)}, \quad m_{+}^N (z)=
-\frac{1}{m_{+}^D(z)} = -\frac{f_+(z,0)}{f_+'(z,0)}. \label{1.15}
\end{align}
In the case where $V$ is real-valued, we briefly recall the spectral theoretic significance of $m_{+}^D$: It is a Herglotz function (i.e., it maps the open complex upper half-plane ${\mathbb{C}}_+$ analytically into itself) and the measure $d\rho^D_+$ in its Herglotz representation is then the spectral measure of the operator $H_{+}^D$ and hence encodes all spectral information of
$H_{+}^D$. Similarly, $m_{+}^D$ also encodes all spectral information of $H_{+}^N$ since $-1/m_{+}^D = m_{+}^N$ is also a Herglotz function and the measure
$d\rho^N_+$ in its Herglotz representation represents the spectral measure of the operator $H_{+}^N$. In particular, $d\rho^D_+$ (respectively, $d\rho^N_+$) uniquely determine $V$ a.e.\ on $(0,\infty)$ by the inverse spectral approach of Gelfand and Levitan \cite{GL55} or Simon \cite{Si99}, \cite{GS00} (see also Remling \cite{Re03} and Section 6 in the survey \cite{Ge07}).
Then we obtain the following result for the ratio of the perturbation determinants in
\eqref{1.11} and \eqref{1.12}:
\begin{theorem} \label{t1.2}
Assume $V\in L^1((0,\infty);dx)$ and let $z\in{\mathbb{C}}\backslash\sigma(H_+^D)$
with $\Im(z^{1/2})>0$. Then,
\begin{align}
& \frac{\det\Big(I_+ +\overline{u\big(H_{0, +}^N-z I_+\big)^{-1} v}\,\Big)}
{\det\Big(I_+ +\overline{u\big(H_{0, +}^D-z I_+\big)^{-1}v}\,\Big)}
\notag \\
&\quad = 1 - \big(\overline{\gamma_N(H_+^D-z I_+)^{-1}V
\big[\gamma_D(H_{0,+}^N-\overline{z}I_+)^{-1}\big]^*}\,\big) \label{1.16} \\
& \quad = \frac{W(f_+(z),\phi_+^N(z))}{i z^{1/2}W(f_+(z),\phi_+^D(z))} =
\frac{f'_+(z,0)}{i z^{1/2}f_+(z,0)} = \frac{m_+^D(z)}{m_{0,+}^D(z)} =
\frac{m_{0,+}^N(z)}{m_+^N(z)}. \label{1.17}
\end{align}
\end{theorem}
The proper multi-dimensional generalization to Schr\"odinger operators in
$L^2(\Omega;d^n x)$ corresponding to an open set $\Omega \subset {\mathbb{R}}^n$ with compact, nonempty boundary $\partial\Omega$ then involves the operator-valued generalization of the Weyl--Titchmarsh function $m_+^D(z)$, the Dirichlet-to-Neumann map denoted by
$M_{\Omega}^D(z)$. In particular, we will derive the following multi-dimensional extension of \eqref{1.16} and \eqref{1.17} in Section \ref{s4}:
\begin{theorem} \label{t1.3}
Assume Hypothesis \ref{h2.6} and let
$z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{\Omega}^D\big)\cup \sigma\big(H_{0,\Omega}^D\big) \cup
\sigma\big(H_{0,\Omega}^N\big)\big)$. Then,
\begin{align}
& \frac{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}\,\Big)}
{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big)} \notag \\
& \quad = \det{}_2\Big(I_{{\partial\Omega}} - \overline{\gamma_N\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1} V
\big[\gamma_D(H_{0,\Omega}^N-\overline{z}I_{\Omega})^{-1}\big]^*}\,\Big)
e^{\text{\rm{tr}}(T_2(z))} \label{1.18} \\
& \quad =
\det{}_2\big(M_{\Omega}^{D}(z)M_{0,\Omega}^{D}(z)^{-1}\big)
e^{\text{\rm{tr}}(T_2(z))}. \label{1.19}
\end{align}
\end{theorem}
Here, ${\det}_2(\cdot)$ denotes the modified Fredholm determinant in
connection with Hilbert--Schmidt perturbations of the identity, $T_2(z)$ is given by
\begin{equation}
T_2(z)=\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1} V
\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1} V
\big[\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*},
\end{equation}
and $I_{\Omega}$ and $I_{\partial\Omega}$ represent the identity operators in
$L^2(\Omega; d^n x)$ and $L^2(\partial\Omega; d^{n-1} \sigma)$, respectively (with
$d^{n-1}\sigma$ the surface measure on $\partial\Omega$).
For pertinent comments on the principal reduction of (a ratio of) modified Fredholm determinants associated with operators in $L^2(\Omega; d^n x)$ on the left-hand side of \eqref{1.18} to a modified Fredholm determinant associated with operators in
$L^2(\partial\Omega; d^{n-1} \sigma)$ on the right-hand side of \eqref{1.18} and especially, in \eqref{1.19}, we refer to Section \ref{s4}.
Finally, we briefly list some of the notational conventions used throughout this
paper. Let $T$ be a linear operator
mapping (a subspace of) a Banach space into another, with
$\text{\rm{dom}}(T)$ and $\text{\rm{ran}}(T)$ denoting the domain and range of $T$.
The closure of a closable operator $S$ is denoted by $\overline S$.
The kernel (null space) of $T$ is denoted by $\ker(T)$. The
spectrum and resolvent set of a closed linear operator in
a separable complex Hilbert space ${\mathcal H}$ (with scalar product
denoted by $(\cdot,\cdot)_{{\mathcal H}}$, assumed to be linear in the second factor)
will be denoted by $\sigma(\cdot)$ and $\rho(\cdot)$. The Banach
spaces of bounded and compact linear operators in ${\mathcal H}$ are
denoted by ${\mathcal B}({\mathcal H})$ and ${\mathcal B}_\infty({\mathcal H})$, respectively.
Similarly, the Schatten--von Neumann (trace) ideals will
subsequently be denoted by ${\mathcal B}_p({\mathcal H})$, $p\in{\mathbb{N}}$. Analogous
notation ${\mathcal B}({\mathcal H}_1,{\mathcal H}_2)$, ${\mathcal B}_\infty ({\mathcal H}_1,{\mathcal H}_2)$, etc.,
will be used for bounded, compact, etc., operators between two
Hilbert spaces ${\mathcal H}_1$ and ${\mathcal H}_2$. In addition, $\text{\rm{tr}}(T)$
denotes the trace of a trace class operator $T\in{\mathcal B}_1({\mathcal H})$ and
$\det_{p}(I_{{\mathcal H}}+S)$ represents the (modified) Fredholm
determinant associated with an operator $S\in{\mathcal B}_p({\mathcal H})$,
$p\in{\mathbb{N}}$ (for $p=1$ we omit the subscript $1$). Moreover,
${\mathcal X}_1 \hookrightarrow {\mathcal X}_2$ denotes the continuous embedding
of the Banach space ${\mathcal X}_1$ into the Banach space ${\mathcal X}_2$.
For general references on the theory of modified Fredholm determinants we
refer, for instance, to \cite[Sect.\ XI.9]{DS88}, \cite[Chs.\ IX, XI]{GGK00},
\cite[Sect.\ IV.2]{GK69}, \cite{Si77}, and \cite[Ch.\ 9]{Si05}.
\section{Schr\"odinger Operators with Dirichlet and Neumann boundary conditions}
\label{s2}
In this section we primarily focus on various properties of Dirichlet,
$H^D_{0,\Omega}$, and Neumann, $H^N_{0,\Omega}$, Laplacians in
$L^2(\Omega;d^n x)$ associated with open sets $\Omega\subset {\mathbb{R}}^n$, $n=2,3$, introduced in Hypothesis \ref{h2.1} below. In particular, we study mapping properties of $\big(H^{D,N}_{0,\Omega}-zI_{\Omega}\big)^{-q}$, $q\in [0,1]$,
($I_{\Omega}$ the identity operator in $L^2(\Omega; d^n x)$) and trace ideal properties of the maps $f \big(H^{D,N}_{0,\Omega}-zI_{\Omega}\big)^{-q}$, $f\in L^p(\Omega; d^n x)$, for appropriate $p\geq 2$, and
$\gamma_N \big(H^{D}_{0,\Omega}-zI_{\Omega}\big)^{-r}$, and
$\gamma_D \big(H^{N}_{0,\Omega}-zI_{\Omega}\big)^{-s}$, for appropriate
$r>3/4$, $s>1/4$, with $\gamma_N$ and $\gamma_D$ being the Neumann and Dirichlet boundary trace operators defined in \eqref{2.2} and \eqref{2.3}.
At the end of this section we then introduce the Dirichlet and Neumann
Schr\"odinger operators $H^D_{\Omega}$ and $H^N_{\Omega}$ in
$L^2(\Omega;d^n x)$, that is, perturbations of the Dirichlet and Neumann Laplacians $H^D_{0,\Omega}$ and $H^N_{0,\Omega}$ by a potential $V$ satisfying
Hypothesis \ref{h2.6}.
We start with introducing our assumptions on the set $\Omega$:
\begin{hypothesis} \label{h2.1}
Let $n=2,3$ and assume that $\Omega\subset{{\mathbb{R}}}^n$ is an open
set with a compact, nonempty boundary $\partial\Omega$. In addition,
we assume that one of the following three conditions holds: \\
$(i)$ \, $\Omega$ is of class $C^{1,r}$ for some $1/2 < r <1$; \\
$(ii)$ \hspace*{.0001pt} $\Omega$ is convex; \\
$(iii)$ $\Omega$ is a Lipschitz domain satisfying a {\it uniform exterior ball condition}.
\end{hypothesis}
The class of domains described in Hypothesis \ref{h2.1} is a subclass of all Lipschitz domains with compact nonempty boundary. We also note that while ${\partial\Omega}$ is assumed to be compact, $\Omega$ may be unbounded
(e.g., an exterior domain) in connection with conditions $(i)$ or $(iii)$. For more details in this context, in particular, for the precise definition of the uniform exterior ball condition, we refer to \cite[App.\ A]{GLMZ05} and \cite[App.\ A]{GMZ06} (and the references cited therein, such as \cite[Ch.\ 1]{Gr85}, \cite{GI75}, \cite{JK81}, \cite{Ka64}, \cite[Ch.\ 3]{Mc00},
\cite{Mi96}, \cite{Mi01}, \cite[p.\ 189]{St70}, \cite{Ta65}, \cite{Tr02}, and
\cite[Sect.\ I.4.2]{Wl87}).
First, we introduce the boundary trace operator $\gamma_D^0$
(Dirichlet trace) by
\begin{equation}
\gamma_D^0\colon C(\overline{\Omega})\to C({\partial\Omega}), \quad \gamma_D^0 u = u|_{\partial\Omega} .
\end{equation}
Then there exists a bounded, linear operator $\gamma_D$ (cf.
\cite[Theorem 3.38]{Mc00}),
\begin{align}
\begin{split}
& \gamma_D\colon H^{s}(\Omega)\to H^{s-(1/2)}({\partial\Omega}) \hookrightarrow L^2(\dOm;d^{n-1}\si),
\quad 1/2<s<3/2, \label{2.2} \\
& \gamma_D\colon H^{3/2}(\Omega)\to H^{1-\varepsilon}({\partial\Omega}) \hookrightarrow L^2(\dOm;d^{n-1}\si),
\quad \varepsilon \in (0,1),
\end{split}
\end{align}
whose action is compatible with that of $\gamma_D^0$. That is, the two
Dirichlet trace operators coincide on the intersection of their
domains. We recall that $d^{n-1}\sigma$ denotes the surface measure on
${\partial\Omega}$ and we refer to \cite[App.\ A]{GLMZ05} for our notation in
connection with Sobolev spaces (see also \cite[Ch.\ 3]{Mc00}, \cite{Tr02}, and
\cite[Sect.\ I.4.2]{Wl87}).
Next, we introduce the operator $\gamma_N$ (Neumann trace) by
\begin{align}
\gamma_N = \nu\cdot\gamma_D\nabla \colon H^{s+1}(\Omega)\to L^2(\dOm;d^{n-1}\si), \quad
1/2<s<3/2, \label{2.3}
\end{align}
where $\nu$ denotes the outward pointing normal unit vector to
$\partial\Omega$. It follows from \eqref{2.2} that $\gamma_N$ is also a
bounded operator.
Given Hypothesis \ref{h2.1}, we introduce the self-adjoint and nonnegative Dirichlet and
Neumann Laplacians $H_{0,\Omega}^D$ and $H_{0,\Omega}^N$ associated
with the domain $\Omega$ as follows,
\begin{align}
H^D_{0,\Omega} = -\Delta, \quad \text{\rm{dom}}\big(H^D_{0,\Omega}\big) = \{u\in H^{2}(\Omega)
\,|\, \gamma_D u = 0\}, \label{2.4}
\\
H^N_{0,\Omega} = -\Delta, \quad \text{\rm{dom}}\big(H^N_{0,\Omega}\big) = \{u\in H^{2}(\Omega)
\,|\, \gamma_N u = 0\}. \label{2.5}
\end{align}
A detailed discussion of $H^D_{0,\Omega}$ and $H^N_{0,\Omega}$ is provided in
\cite[App.\ A]{GLMZ05} (cf.\ also \cite[Sects. X.III.14, X.III.15]{RS78}).
\begin{lemma} \label{l2.2}
Assume Hypothesis \ref{h2.1}. Then the operators $H_{0,\Omega}^D$
and $H_{0,\Omega}^N$ introduced in \eqref{2.4} and \eqref{2.5} are
nonnegative and self-adjoint in $L^2(\Om;d^nx)$ and the following boundedness
properties hold for all $q\in [0,1]$ and $z\in{\mathbb{C}}\backslash[0,\infty)$,
\begin{align}
\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-q},\, \big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-q}
\in{\mathcal B}\big(L^2(\Om;d^nx),H^{2q}(\Omega)\big). \label{2.6}
\end{align}
\end{lemma}
The fractional powers in \eqref{2.6} (and in subsequent analogous
cases) are defined via the functional calculus implied by the
spectral theorem for self-adjoint operators.
As explained in \cite[Lemma A.2]{GLMZ05} (based on results in \cite{JK95},
\cite[Thm.\ 4.4, App.\ B]{Mc00}, \cite{MMT01}, \cite{MT00}, \cite[Chs.\ 1, 2]{Ta91},
\cite[Props.\ 4.5, 7.9]{Ta96}, \cite[Sect.\ 1.3, Thm.\ 1.18.10, Rem.\ 4.3.1.2]{Tr78}, \cite{Ve84}), the key ingredients in proving
Lemma \ref{l2.2} are the inclusions
\begin{equation}
\text{\rm{dom}}\big(H_{0,\Omega}^D\big) \subset H^2(\Omega), \quad
\text{\rm{dom}}\big(H^N_{0,\Omega}\big) \subset H^2(\Omega)
\end{equation}
and real interpolation methods.
The next result is a slight extension of \cite[Lemma 6.8]{GLMZ05} and provides an explicit discussion of the $z$-dependence of the constant $c$ appearing in estimate (6.48) of \cite{GLMZ05}. For a proof we refer to \cite{GMZ06}.
\begin{lemma} [\cite{GMZ06}] \label{l2.3}
Assume Hypothesis \ref{h2.1} and let $2\leq p$, $n/(2p)<q\leq1$,
$f\in L^p(\Omega;d^nx)$, and $z\in{\mathbb{C}}\backslash[0,\infty)$. Then,
\begin{align} \label{2.7}
f\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-q}, \, f\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-q}
\in{\mathcal B}_p\big(L^2(\Om;d^nx)\big),
\end{align}
and for some $c>0$ $($independent of $z$ and $f$\,$)$
\begin{align}
\begin{split}
&\big\| f\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-q}\big\|_{{\mathcal B}_p(L^2(\Om;d^nx))}^2
\\
& \qquad \leq
c\bigg(1+\frac{\abs{z}^{2q}+1}{\text{\rm{dist}}\big(z,\sigma\big(H_{0,\Omega}^D\big)\big)^{2q}}\bigg)
\|(\abs{\,\cdot\,}^2-z)^{-q}\|_{L^p({\mathbb{R}}^n;d^nx)}^2
\|f\|_{L^p(\Omega;d^nx)}^2,
\\
&\big\| f\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-q}\big\|_{{\mathcal B}_p(L^2(\Om;d^nx))}^2
\\
& \qquad \leq
c\bigg(1+\frac{\abs{z}^{2q}+1}{\text{\rm{dist}}\big(z,\sigma\big(H_{0,\Omega}^N\big)\big)^{2q}}\bigg)
\|(\abs{\,\cdot\,}^2-z)^{-q}\|_{L^p({\mathbb{R}}^n;d^nx)}^2
\|f\|_{L^p(\Omega;d^nx)}^2.
\end{split} \label{2.8}
\end{align}
\end{lemma}
(Here, in obvious notation, $(\abs{\,\cdot\,}^2-z)^{-q}$ denotes the function
$(\abs{x}^2-z)^{-q}$, $x\in{\mathbb{R}}^n$.)
Next we recall certain boundedness properties of powers of the resolvents of Dirichlet and Neumann Laplacians multiplied by the Neumann and Dirichlet boundary trace operators, respectively:
\begin{lemma} \label{l2.4}
Assume Hypothesis \ref{h2.1} and let $\varepsilon > 0$,
$z\in{\mathbb{C}}\backslash[0,\infty)$. Then,
\begin{align}
\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-\frac{3+\varepsilon}{4}},
\gamma_D\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-\frac{1+\varepsilon}{4}} \in
{\mathcal B}\big(L^2(\Om;d^nx),L^2(\dOm;d^{n-1}\si)\big). \label{2.22}
\end{align}
\end{lemma}
As in \cite[Lemma 6.9]{GLMZ05}, Lemma \ref{l2.4} follows from Lemma \ref{l2.2}
and from \eqref{2.2} and \eqref{2.3}.
\begin{corollary} \label{c2.5}
Assume Hypothesis \ref{h2.1} and let $f_1\in L^{p_1}(\Omega;d^nx)$,
$p_1\geq 2$, $p_1>2n/3$, $f_2\in L^{p_2}(\Omega;d^nx)$, $p_2 > 2n$,
and $z\in{\mathbb{C}}\backslash[0,\infty)$. Then, denoting by $f_1$ and
$f_2$ the operators of multiplication by functions $f_1$ and
$f_2$ in $L^2(\Om;d^nx)$, respectively, one has
\begin{align}
\overline{\gamma_D\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}f_1} &\in
{\mathcal B}_{p_1}\big(L^2(\Omega;d^nx),L^2({\partial\Omega};d^{n-1}\sigma)\big), \label{2.25}
\\
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}f_2} &\in
{\mathcal B}_{p_2}\big(L^2(\Omega;d^nx),L^2({\partial\Omega};d^{n-1}\sigma)\big) \label{2.26}
\end{align}
and for some $c_j(z)>0$ $($independent of $f_j$$)$, $j=1,2$,
\begin{align}
\Big\| \, \overline{\gamma_D\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}f_1}\, \Big\|_
{{\mathcal B}_{p_1}(L^2(\Omega;d^nx),L^2({\partial\Omega};d^{n-1}\sigma))} &\leq c_1(z)
\norm{f_1}_{L^{p_1}(\Omega;d^nx)}, \label{2.27}
\\
\Big\| \, \overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}f_2}\, \Big\|_
{{\mathcal B}_{p_2}(L^2(\Omega;d^nx),L^2({\partial\Omega};d^{n-1}\sigma))} &\leq c_2(z)
\norm{f_2}_{L^{p_2}(\Omega;d^nx)}. \label{2.28}
\end{align}
\end{corollary}
As in \cite[Corollary 6.10]{GLMZ05}, Corollary \ref{c2.5} follows from Lemmas \ref{l2.3} and \ref{l2.4}.
Finally, we turn to our assumptions on the potential $V$ and the corresponding definition of Dirichlet and Neumann Schr\"odinger operators $H^D_{\Omega}$ and $H^N_{\Omega}$ in $L^2(\Omega; d^n x)$:
\begin{hypothesis} \label{h2.6}
Suppose that $\Omega$ satisfies Hypothesis \ref{h2.1} and assume that
$V\in L^p(\Omega;d^nx)$ for some $p$ satisfying $4/3
< p \leq 2$, in the case $n=2$, and $3/2 < p \leq 2$, in the case
$n=3$.
\end{hypothesis}
Assuming Hypothesis \ref{h2.6}, we next introduce the perturbed operators
$H_{\Omega}^D$ and $H_{\Omega}^N$ in $L^2(\Om;d^nx)$ by alluding to abstract perturbation results
due to Kato \cite{Ka66} (see also Konno and Kuroda \cite{KK66}) as
summarized in \cite[Sect.\ 2]{GLMZ05}: Let $V$, $u$, and $v$ denote
the operators of multiplication by functions $V$,
$u=\exp(i\arg(V))\abs{V}^{1/2}$, and $v=\abs{V}^{1/2}$ in
$L^{2}(\Omega;d^nx)$, respectively, such that
\begin{equation}
V=uv. \label{2.29}
\end{equation}
Since $u,v\in L^{2p}(\Omega;d^nx)$, Lemma \ref{l2.3} yields
\begin{align}
u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1/2}, \,
\overline{\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1/2}v} &\in{\mathcal B}_{2p}\big(L^2(\Om;d^nx)\big),
\quad z\in{\mathbb{C}}\backslash [0,\infty), \label{2.31}
\\
u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1/2}, \,
\overline{\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1/2}v} &\in{\mathcal B}_{2p}\big(L^2(\Om;d^nx)\big),
\quad z\in{\mathbb{C}}\backslash [0,\infty), \label{2.32}
\end{align}
and hence, in particular,
\begin{align}
& \text{\rm{dom}}(u)=\text{\rm{dom}}(v) \supseteq \text{\rm{dom}}\Big(\big(H^N_{0,\Omega}\big)^{1/2}\Big) =
H^{1}(\Omega) \supset H^{2}(\Omega) \supset \text{\rm{dom}}\big(H^N_{0,\Omega}\big), \\
& \text{\rm{dom}}(u)=\text{\rm{dom}}(v) \supseteq H^{1}(\Omega) \supset H^{1}_0(\Omega)
= \text{\rm{dom}}\Big(\big(H^D_{0,\Omega}\big)^{1/2}\Big)
\supset \text{\rm{dom}}\big(H^D_{0,\Omega}\big).
\end{align}
Moreover, \eqref{2.31} and \eqref{2.32} imply
\begin{equation}
\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}, \,
\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}
\in{\mathcal B}_p\big(L^2(\Om;d^nx)\big)\subset{\mathcal B}_2\big(L^2(\Om;d^nx)\big), \quad
z\in{\mathbb{C}}\backslash [0,\infty). \label{2.35}
\end{equation}
Utilizing \eqref{2.8} in Lemma \ref{l2.3}
with $-z>0$ sufficiently large, such that the ${\mathcal B}_{2p}$-norms
of the operators in \eqref{2.31} and \eqref{2.32} are less than
1, one concludes that the Hilbert--Schmidt norms of the operators in
\eqref{2.35} are less than 1. Thus, applying \cite[Thm.\ 2.3]{GLMZ05}, one
obtains the densely defined, closed operators $H_{\Omega}^D$ and
$H_{\Omega}^N$ (which are extensions of $H_{0,\Omega}^D+V$ defined on
$\text{\rm{dom}}\big(H_{0,\Omega}^D\big)\cap\text{\rm{dom}}(V)$ and $H_{0,\Omega}^N+V$ defined on
$\text{\rm{dom}}\big(H_{0,\Omega}^N\big)\cap\text{\rm{dom}}(V)$, respectively). In particular, the
resolvent of $H_{\Omega}^D$ (respectively, $H_{\Omega}^N$) is explicitly given by
\begin{align}
\big(H^D_{\Omega}-zI_\Omega\big)^{-1} &= \big(H^D_{0,\Omega}-zI_\Omega\big)^{-1} -
\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}v
\Big[I_\Omega+\overline{u\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}v}\,\Big]^{-1}
u\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}, \notag \\
& \hspace*{9.1cm} z\in{\mathbb{C}}\backslash\sigma\big(H^D_{\Omega}\big), \label{2.20} \\
\big(H^N_{\Omega}-zI_\Omega\big)^{-1} &= \big(H^N_{0,\Omega}-zI_\Omega\big)^{-1} -
\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}v
\Big[I_\Omega+\overline{u\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}v}\,\Big]^{-1}
u\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}, \notag \\
& \hspace*{9.1cm} z\in{\mathbb{C}}\backslash\sigma\big(H^N_{\Omega}\big). \label{2.21}
\end{align}
Here invertibility of $\Big[I_\Omega+\overline{u\big(H^{D,N}_{0,\Omega}-zI_\Omega\big)^{-1}v}\,\Big]$ for $z\in \rho\big(H^{D,N}_{\Omega}\big)$ is guaranteed by arguments discussed, for instance, in \cite[Sect.\ 2]{GLMZ05} and the literature cited therein.
Although we will not explicitly use the following result in this paper, we feel
it is of sufficient independent interest to be included at the end of this section:
\begin{lemma}\label{L-1}
Assume $\Omega$ satisfies Hypothesis \ref{h2.1} with $n=2,3$ replaced by
$n\in{\mathbb{N}}$, $n\geq 2$, suppose that $V\in L^{n/2}(\Omega; d^nx)$, and let
\begin{equation}\label{SSS}
s\in\left\{
\begin{array}{l}
(0,1)\mbox{ if }n=2,
\\[6pt]
[0,{\textstyle{\frac{3}{2}}})\mbox{ if }n=3,
\\[6pt]
[0,2)\mbox{ if }n=4,
\\[6pt]
[0,2]\mbox{ if }n\geq 5.
\end{array}
\right.
\end{equation}
Then the operator
\begin{equation}\label{V-2}
V:H^s(\Omega)\rightarrow \big(H^{2-s}(\Omega)\big)^*
\end{equation}
is well-defined and bounded, in fact, it is compact.
\end{lemma}
\begin{proof} The fact that the operator (\ref{V-2}) is
well-defined along with the estimate
\begin{equation}\label{V-3}
\|V\|_{{\mathcal{B}}(H^s(\Omega),(H^{2-s}(\Omega))^*)} \leq
C(n,\Omega)\|V\|_{L^{n/2}(\Omega)},
\end{equation}
are direct consequences of standard embedding results (cf.\ \cite[Sect.\ 3.3.1]{Tr83} for smooth domains and \cite{Tr02} for arbitrary (bounded or unbounded) Lipschitz domains). Once the boundedness of (\ref{V-2}) has been established, the compactness follows from the fact that if $V_j\in C^\infty_0(\Omega)$ is a
sequence of functions with the property that
$V_j \underset{j\uparrow\infty}{\rightarrow} V$ in
$L^{n/2}(\Omega)$, then $V_j \underset{j\uparrow\infty}{\rightarrow} V$ in
${\mathcal{B}}\big(H^s(\Omega),\big(H^{2-s}(\Omega)\big)^*\big)$ by (\ref{V-3})
and each operator $V_j:H^s(\Omega)\to \big(H^{2-s}(\Omega)\big)^*$ is
compact, by Rellich's selection lemma (cf.\ \cite{ET89} for smooth domains and
\cite{Tr02} for arbitrary Lipschitz domains). Thus, the operator in (\ref{V-2}) is
compact as the operator norm limit of a sequence of compact operators.
\end{proof}
\section{Dirichlet and Neumann boundary value problems \\
and Dirichlet-to-Neumann maps} \label{s3}
This section is devoted to Dirichlet and Neumann boundary value problems associated with the Helmholtz differential expression $-\Delta - z$ as well as the corresponding differential expression $-\Delta + V - z$ in the presence of a potential $V$, both in connection with the open set $\Omega$. In addition, we provide a detailed discussion of Dirichlet-to-Neumann, $M^D_{0,\Omega}$,
$M^D_{\Omega}$, and Neumann-to-Dirichlet maps, $M^N_{0,\Omega}$, $M^N_{\Omega}$, in
$L^2(\partial\Omega; d^{n-1}\sigma)$.
Denote by
\begin{equation}
\widetilde\gamma_N : \big\{u\in H^1(\Omega) \,\big|\, \Delta u \in
\big(H^1(\Omega)\big)^*\big\} \to H^{-1/2}({\partial\Omega}) \label{3.0}
\end{equation}
a weak Neumann trace operator defined by
\begin{align} \label{3.1a}
\langle\widetilde\gamma_N u, \phi\rangle = \int_\Omega d^n x \, \nabla
u(x)\cdot\nabla \Phi(x) + \langle\Delta u, \Phi\rangle
\end{align}
for all $\phi\in H^{1/2}({\partial\Omega})$ and $\Phi\in H^1(\Omega)$ such that
$\gamma_D\Phi = \phi$. We note that this definition is
independent of the particular extension $\Phi$ of $\phi$, and that
$\widetilde\gamma_N$ is an extension of the Neumann trace operator
$\gamma_N$ defined in \eqref{2.3}. For more details we refer to
\cite[App.\ A]{GLMZ05}.
We start with a basic result on the Helmholtz Dirichlet and Neumann boundary value problems:
\begin{theorem} [\cite{GMZ06}] \label{t3.1}
Assume Hypothesis \ref{h2.1}. Then for every
$f \in H^1({\partial\Omega})$ and $z\in{\mathbb{C}}\big\backslash\sigma\big(H_{0,\Omega}^D\big)$ the
following Dirichlet boundary value problem,
\begin{align} \label{3.1}
\begin{cases}
(-\Delta - z)u_0^D = 0 \text{ on }\, \Omega, \quad u_0^D \in H^{3/2}(\Omega), \\
\gamma_D u_0^D = f \text{ on }\, {\partial\Omega},
\end{cases}
\end{align}
has a unique solution $u_0^D$ satisfying $\widetilde\gamma_N u_0^D \in
L^2(\dOm;d^{n-1}\si)$. Moreover, there exist constants $C^D=C^D(\Omega,z)>0$ such that
\begin{equation}
\|u_0^D\|_{H^{3/2}(\Omega)} \leq C^D \|f\|_{H^1(\partial\Omega)}. \label{3.3a}
\end{equation}
Similarly, for every $g\inL^2(\dOm;d^{n-1}\si)$ and
$z\in{\mathbb{C}}\backslash\sigma\big(H_{0,\Omega}^N\big)$ the following Neumann
boundary value problem,
\begin{align} \label{3.2}
\begin{cases}
(-\Delta - z)u_0^N = 0 \text{ on }\,\Omega,\quad u_0^N \in H^{3/2}(\Omega), \\
\widetilde\gamma_N u_0^N = g\text{ on }\,{\partial\Omega},
\end{cases}
\end{align}
has a unique solution $u_0^N$ satisfying $\gamma_D u^N_0 \in H^1({\partial\Omega})$. Moreover, there exist constants $C^N=C^N(\Omega,z)>0$ such that
\begin{equation}
\|u_0^N\|_{H^{3/2}(\Omega)} \leq C^N \|g\|_{L^2(\dOm;d^{n-1}\si)}. \label{3.4a}
\end{equation}
In addition, \eqref{3.1}--\eqref{3.4a} imply that the following maps are bounded
\begin{align}
\big[\gamma_N\big(\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^* &: H^1({\partial\Omega}) \to
H^{3/2}(\Omega), \quad z\in{\mathbb{C}}\big\backslash\sigma\big(H^D_{0,\Omega}\big), \label{3.4b}
\\
\big[\gamma_D \big(\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^* &: L^2(\dOm;d^{n-1}\si) \to
H^{3/2}(\Omega), \quad z\in{\mathbb{C}}\big\backslash\sigma\big(H^N_{0,\Omega}\big). \label{3.4c}
\end{align}
Finally, the solutions $u_0^D$ and $u_0^N$ are given by the formulas
\begin{align}
u_0^D (z) &= -\big(\gamma_N \big(H_{0,\Omega}^D-\overline{z}I_\Omega\big)^{-1}\big)^*f,
\label{3.3}
\\
u_0^N (z) &= \big(\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_\Omega\big)^{-1}\big)^*g.
\label{3.4}
\end{align}
\end{theorem}
A detailed proof of Theorem \ref{t3.1} will appear in \cite{GMZ06}.
We temporarily strengthen our hypothesis on $V$ and introduce the following assumption:
\begin{hypothesis} \label{h3.2}
Suppose the set $\Omega$ satisfies Hypothesis \ref{h2.1} and assume that
$V\in L^2(\Om;d^nx)\cap L^p(\Omega;d^nx)$ for some $p>2$.
\end{hypothesis}
By employing a perturbative approach, we now extend Theorem \ref{t3.1} in connection with the Helmholtz differential expression $-\Delta - z$ on
$\Omega$ to the case of a Schr\"odinger differential expression $-\Delta + V - z$ on $\Omega$.
\begin{theorem} \label{t3.3}
Assume Hypothesis \ref{h3.2}. Then for every $f \in
H^1({\partial\Omega})$ and
$z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^D\big)$
the following Dirichlet boundary value problem,
\begin{align} \label{3.9}
\begin{cases}
(-\Delta + V - z)u^D = 0 \text{ on }\, \Omega, \quad u^D \in H^{3/2}(\Omega), \\
\gamma_D u^D = f \text{ on }\, {\partial\Omega},
\end{cases}
\end{align}
has a unique solution $u^D$ satisfying $\widetilde\gamma_N u^D \in
L^2(\dOm;d^{n-1}\si)$. Similarly, for every $g\inL^2(\dOm;d^{n-1}\si)$ and
$z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^N\big)$
the following Neumann boundary value problem,
\begin{align} \label{3.10}
\begin{cases}
(-\Delta + V - z)u^N = 0 \text{ on }\, \Omega,\quad u^N \in H^{3/2}(\Omega), \\
\widetilde\gamma_N u^N = g\text{ on }\, {\partial\Omega},
\end{cases}
\end{align}
has a unique solution $u^N$. Moreover, the solutions $u^D$ and
$u^N$ are given by the formulas
\begin{align}
u^D (z) &= -\big[\gamma_N \big(\big(H_{\Omega}^D-zI_\Omega\big)^{-1}\big)^*\big]^*f,
\label{3.11}
\\
u^N (z) &= \big[\gamma_D \big(\big(H_{\Omega}^N-zI_\Omega\big)^{-1}\big)^*\big]^*g.
\label{3.12}
\end{align}
\end{theorem}
\begin{proof}
We temporarily assume that
$z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^D\big)\cup\sigma\big(H_{\Omega}^D\big)\big)$ in the case of the Dirichlet problem and
$z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^N\big)\cup\sigma\big(H_{\Omega}^N\big)\big)$ in the context of the Neumann problem. \\
Uniqueness follows from the fact that $z\notin\sigma(H_\Omega^D)$
and $z\notin\sigma(H_\Omega^N)$, respectively.
Next, we will show that the functions
\begin{align}
u^D (z) &= u_0^D (z) - \big(H_\Omega^D-zI_\Omega\big)^{-1} V u_0^D (z), \label{3.13}
\\
u^N (z)&= u_0^N (z) - \big(H_\Omega^N-zI_\Omega\big)^{-1} V u_0^N (z), \label{3.14}
\end{align}
with $u_0^D, u_0^N$ given by Theorem \ref{t3.1},
satisfy \eqref{3.11} and \eqref{3.12}, respectively. Indeed, it
follows from Theorem \ref{t3.1} that $u_0^D,u_0^N\in H^{3/2}(\Omega)$
and $\widetilde\gamma_N u_0^D \in L^2(\dOm;d^{n-1}\si)$. Using the Sobolev embedding theorem
$H^{3/2}(\Omega) \hookrightarrow L^q(\Omega;d^nx)$, $q\geq2$, and the
fact that $V\in L^p(\Omega;d^nx)$, $p>2$, one concludes that $Vu_0^D,
Vu_0^N\inL^2(\Om;d^nx)$, and hence \eqref{3.13} and \eqref{3.14} are
well-defined. Since one also has $V\inL^2(\Om;d^nx)$, it follows from Lemma
\ref{l2.3} that $V\big(H_{0,\Omega}^D-zI_\Omega\big)^{-1}$ and
$V\big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}$ are Hilbert--Schmidt, and hence
\begin{align}
\big[I+V\big(H_{0,\Omega}^D-zI_\Omega\big)^{-1}\big]^{-1} \in {\mathcal B}\big(L^2(\Om;d^nx)\big), \quad
z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^D\big)\cup\sigma\big(H_{\Omega}^D\big)\big),
\label{3.15}
\\
\big[I+V\big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}\big]^{-1} \in {\mathcal B}\big(L^2(\Om;d^nx)\big), \quad
z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^N\big)\cup\sigma\big(H_{\Omega}^N\big)\big).
\label{3.16}
\end{align}
Thus, by \eqref{2.4} and \eqref{2.5},
\begin{align}
\big(H_\Omega^D-zI_\Omega\big)^{-1} V u_0^D &=
\big(H_{0,\Omega}^D-zI_\Omega\big)^{-1}\big[I+V\big(H_{0,\Omega}^D-zI_\Omega\big)^{-1}\big]^{-1}Vu_0^D
\in H^2(\Omega),
\\
\big(H_\Omega^N-zI_\Omega\big)^{-1} V u_0^N &=
\big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}\big[I+V\big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}\big]^{-1}Vu_0^N
\in H^2(\Omega),
\end{align}
and hence $u^D,u^N\in H^{3/2}(\Omega)$ and $\widetilde\gamma_N u^D \in L^2(\dOm;d^{n-1}\si)$.
Moreover,
\begin{align}
(-\Delta+V-z)u^D &= (-\Delta-z)u_0^D + Vu_0^D -
(-\Delta+V-z)\big(H_{\Omega}^D-zI_\Omega\big)^{-1}Vu_0^D \notag
\\ &=
Vu_0^D - I_\Omega Vu_0^D = 0,
\\
(-\Delta+V-z)u^N &= (-\Delta-z)u_0^N + Vu_0^N -
(-\Delta+V-z)\big(H_{\Omega}^N-zI_\Omega\big)^{-1}Vu_0^N \notag
\\ &=
Vu_0^N - I_\Omega Vu_0^N = 0,
\end{align}
and by \eqref{2.4}, \eqref{2.5} and \eqref{3.15}, \eqref{3.16} one
also obtains,
\begin{align}
\gamma_D u^D &= \gamma_D u_0^D - \gamma_D\big(H_{\Omega}^D-zI_\Omega\big)^{-1}Vu_0^D \notag
\\&=
f - \gamma_D \big(H_{0,\Omega}^D-zI_\Omega\big)^{-1}
\big[I+V\big(H_{0,\Omega}^D-zI_\Omega\big)^{-1}\big]^{-1} Vu_0^D=f,
\\
\widetilde\gamma_N u^N &= \widetilde\gamma_N u_0^N -
\widetilde\gamma_N\big(H_{\Omega}^N-zI_\Omega\big)^{-1}Vu_0^N \notag
\\&=
g - \gamma_N \big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}
\big[I+V\big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}\big]^{-1} Vu_0^N=g.
\end{align}
Finally, \eqref{3.11} and \eqref{3.12} follow from \eqref{3.3},
\eqref{3.4}, \eqref{3.13}, \eqref{3.14}, and the resolvent
identity,
\begin{align}
u^D (z) &= \big[I_\Omega - \big(H_\Omega^D-zI_\Omega\big)^{-1} V\big] \big[-\gamma_N
\big(\big(H_{0,\Omega}^D-zI_\Omega\big)^{-1}\big)^*\big]^*f \notag
\\ &=
-\big[\gamma_N \big(\big(H_{0,\Omega}^D- zI_\Omega\big)^{-1}\big)^*
\big[I_\Omega - \big(H_\Omega^D-zI_\Omega\big)^{-1} V\big]^*\big]^*f \notag
\\ &=
-\big[\gamma_N \big(\big(H_\Omega^D-zI_\Omega\big)^{-1}\big)^*\big]^*f,
\\[1mm]
u^N (z) &= \big[I_\Omega - \big(H_\Omega^N-zI_\Omega\big)^{-1} V\big] \big[\gamma_D
\big(\big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}\big)^*\big]^*g \notag
\\ &=
\big[\gamma_D \big(\big(H_{0,\Omega}^N-zI_\Omega\big)^{-1}\big)^*
\big[I_\Omega - \big(H_\Omega^N-zI_\Omega\big)^{-1} V\big]^*\big]^*g \notag
\\ &=
\big[\gamma_D \big(\big(H_\Omega^N-zI_\Omega\big)^{-1}\big)^*\big]^*g.
\end{align}
Analytic continuation with respect to $z$ then permits one to remove the additional condition $z \notin \sigma\big(H_{0,\Omega}^D\big)$ in the case of the Dirichlet problem, and
the additional condition $z \notin \sigma\big(H_{0,\Omega}^N\big)$ in the context of the Neumann problem.
\end{proof}
Assuming Hypothesis \ref{h3.2}, we now introduce the
Dirichlet-to-Neumann maps, $M_{0,\Omega}^{D}(z)$ and $M_\Omega^{D}(z)$,
associated with $(-\Delta-z)$ and $(-\Delta+V-z)$ on $\Omega$, as follows,
\begin{align}
M_{0,\Omega}^{D}(z) \colon
\begin{cases}
H^1({\partial\Omega}) \to L^2(\dOm;d^{n-1}\si),
\\
\hspace*{10mm} f \mapsto -\widetilde\gamma_N u_0^D,
\end{cases} \quad z\in{\mathbb{C}}\big\backslash\sigma\big(H_{0,\Omega}^D\big),
\end{align}
where $u_0^D$ is the unique solution of
\begin{align}
(-\Delta-z)u_0^D = 0 \,\text{ on }\Omega, \quad u_0^D\in
H^{3/2}(\Omega), \quad \gamma_D u_0^D = f \,\text{ on }{\partial\Omega},
\end{align}
and
\begin{align}
M_\Omega^{D}(z) \colon
\begin{cases}
H^1({\partial\Omega}) \to L^2(\dOm;d^{n-1}\si),
\\
\hspace*{10mm} f \mapsto -\widetilde\gamma_N u^D,
\end{cases} \quad z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^D\big),
\end{align}
where $u^D$ is the unique solution of
\begin{align}
(-\Delta+V-z)u^D = 0 \,\text{ on }\Omega, \quad u^D \in
H^{3/2}(\Omega), \quad \gamma_D u^D= f \,\text{ on }{\partial\Omega}.
\end{align}
In addition, still assuming Hypothesis \ref{h3.2}, we introduce the Neumann-to-Dirichlet maps, $M_{0,\Omega}^{N}(z)$ and $M_\Omega^{N}(z)$,
associated with $(-\Delta-z)$ and $(-\Delta+V-z)$ on $\Omega$, as
follows,
\begin{align}
M_{0,\Omega}^{N}(z) \colon \begin{cases} L^2(\dOm;d^{n-1}\si) \to H^1({\partial\Omega}),
\\
\hspace*{20.5mm} g \mapsto \gamma_D u_0^N, \end{cases} \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{0,\Omega}^N\big),
\end{align}
where $u_0^N$ is the unique solution of
\begin{align}
(-\Delta-z)u_0^N = 0 \,\text{ on }\Omega, \quad u_0^N\in
H^{3/2}(\Omega), \quad \widetilde\gamma_N u_0^N = g \,\text{ on }{\partial\Omega},
\end{align}
and
\begin{align}
M_\Omega^{N}(z) \colon \begin{cases}
L^2(\dOm;d^{n-1}\si) \to H^1({\partial\Omega}),
\\
\hspace*{20.5mm} g \mapsto \gamma_D u^N,
\end{cases} \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^N\big),
\end{align}
where $u^N$ is the unique solution of
\begin{align}
(-\Delta+V-z)u^N = 0 \,\text{ on }\Omega, \quad u^N \in
H^{3/2}(\Omega), \quad \widetilde\gamma_N u^N= g \,\text{ on }{\partial\Omega}.
\end{align}
It follows from Theorems \ref{t3.1} and \ref{t3.3}, that under the
assumption of Hypothesis \ref{h3.2}, the operators
$M_{0,\Omega}^D(z)$, $M_{\Omega}^D(z)$, $M_{0,\Omega}^N(z)$, and
$M_{\Omega}^N(z)$ are well-defined and satisfy the following
equalities,
\begin{align}
M_{0,\Omega}^{N}(z) &= - M_{0,\Omega}^{D}(z)^{-1}, \quad
z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^D\big)\cup\sigma\big(H_{0,\Omega}^N\big)\big),
\label{3.28}
\\
M_{\Omega}^{N}(z) &= - M_{\Omega}^{D}(z)^{-1}, \quad
z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{\Omega}^D\big)\cup\sigma\big(H_{\Omega}^N\big)\big),
\label{3.29}
\intertext{and}%
M^{D}_{0,\Omega}(z) &= \widetilde\gamma_N\big[\gamma_N
\big(\big(H^D_{0,\Omega} - zI_\Omega\big)^{-1}\big)^*\big]^*, \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{0,\Omega}^D\big), \label{3.30}
\\
M^{D}_{\Omega}(z) &= \widetilde\gamma_N\big[\gamma_N \big(\big(H^D_{\Omega} -
zI_\Omega\big)^{-1}\big)^*\big]^*, \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^D\big),
\label{3.31}
\\
M^{N}_{0,\Omega}(z) &= \gamma_D\big[\gamma_D
\big(\big(H^N_{0,\Omega} - zI_\Omega\big)^{-1}
\big)^*\big]^*, \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{0,\Omega}^N\big), \label{3.32}
\\
M^{N}_{\Omega}(z) &= \gamma_D\big[\gamma_D
\big(\big(H^N_{\Omega} - zI_\Omega\big)^{-1}\big)^*\big]^*, \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^N\big).
\label{3.33}
\end{align}
The representations \eqref{3.30}--\eqref{3.33} provide a convenient point of departure for proving the operator-valued Herglotz property of
$M^{D}_{\Omega}$ and $M^{N}_{\Omega}$. We will return to this topic in a future paper.
Next, we note that the above formulas \eqref{3.30}--\eqref{3.33} may
be used as alternative definitions of the
Dirichlet-to-Neumann and Neumann-to-Dirichlet maps. In
particular, we will next use \eqref{3.31} and \eqref{3.33} to extend
the above definition of the operators $M^{D}_{\Omega}(z)$ and
$M^{N}_{\Omega}(z)$ to the more general situation governed by Hypothesis \ref{h2.6}:
\begin{lemma} [\cite{GMZ06}] \label{l3.4}
Assume Hypothesis \ref{h2.6}. Then the operators
$M^{D}_{\Omega}(z)$ and $M^{N}_{\Omega}(z)$ defined by equalities
\eqref{3.31} and \eqref{3.33} have the following boundedness
properties,
\begin{align}
& M^{D}_{\Omega}(z) \in {\mathcal B}\big(H^{1}(\partial\Omega),L^2(\dOm;d^{n-1}\si)\big), \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^D\big),
\label{3.42a} \\
& M^{N}_{\Omega}(z) \in {\mathcal B}\big(L^2(\dOm;d^{n-1}\si),H^{1}(\partial\Omega)\big), \quad
z\in{\mathbb{C}}\big\backslash\sigma\big(H_{\Omega}^N\big).
\label{3.43a}
\end{align}
\end{lemma}
A detailed proof of Lemma \ref{l3.4} will be provided in \cite{GMZ06}.
Weyl--Titchmarsh operators, in a spirit close to ours, have recently been discussed by Amrein and Pearson \cite{AP04} in connection with the interior and exterior of a ball in ${\mathbb{R}}^3$ and real-valued potentials $V\in L^\infty({\mathbb{R}}^3;d^3x)$. For additional literature on Weyl--Titchmarsh operators, relevant in the context of boundary value spaces (boundary triples, etc.), we refer, for instance, to \cite{ABMN05}, \cite{BL06},
\cite{BMN06},
\cite{BMN00}, \cite{BMN02}, \cite{BM04}, \cite{DM91}, \cite{DM95}, \cite{GKMT01},
\cite[Ch.\ 3]{GG91}, \cite{MM06}, \cite{Ma04}, \cite{MPP07}, \cite{Pa87}, \cite{Pa02}.
Next, we prove the following auxiliary result, which will play a crucial role in Theorem \ref{t4.2}, the principal result of this paper.
\begin{lemma} \label{l3.5}
Assume Hypothesis \ref{h2.6}. Then the following identities hold,
\begin{align}
M_{0,\Omega}^D(z) - M_\Omega^D(z) &= \overline{\widetilde\gamma_N
\big(H^D_{\Omega}-zI_\Omega\big)^{-1} V \big[\gamma_N
\big(\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*}, \notag
\\
&\hspace*{3.1cm}
z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^D\big)\cup\sigma\big(H_{\Omega}^D\big)\big),
\label{3.35}
\\
M_\Omega^D(z) M_{0,\Omega}^D(z)^{-1} &= I_{\partial\Omega} - \overline{\widetilde\gamma_N
\big(H^D_{\Omega}-zI_\Omega\big)^{-1} V \big[\gamma_D
\big(\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*}, \notag
\\
&\hspace*{2.45cm}
z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^D\big)\cup\sigma\big(H_{\Omega}^D\big)
\cup\sigma\big(H_{0,\Omega}^N\big)\big). \label{3.36}
\end{align}
\end{lemma}
\begin{proof}
Let $z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^D\big)\cup\sigma\big(H_{\Omega}^D\big)\big)$.
Then \eqref{3.35} follows from \eqref{3.30}, \eqref{3.31}, and
the resolvent identity
\begin{align}
M_{0,\Omega}^D(z) - M_\Omega^D(z) &=
\widetilde\gamma_N\big[\gamma_N\big(\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1} -
\big(H^D_{\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^* \notag
\\ &=
\overline{\widetilde\gamma_N\big[\gamma_N \big(
\big(H^D_{\Omega}-zI_\Omega\big)^{-1}V\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1} \big)^*\big]^*}
\\ &=
\overline{\widetilde\gamma_N \big(H^D_{\Omega}-zI_\Omega\big)^{-1} V \big[\gamma_N
\big(\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*}. \notag
\end{align}
Next, if
$z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{0,\Omega}^D\big)\cup\sigma\big(H_{\Omega}^D\big)
\cup\sigma\big(H_{0,\Omega}^N\big)\big)$, then it follows from \eqref{3.28},
\eqref{3.32}, and \eqref{3.35} that
\begin{align} \label{3.40}
M_\Omega^D(z) M_{0,\Omega}^D(z)^{-1} &= I_{\partial\Omega} + \big(M_\Omega^D(z) -
M_{0,\Omega}^D(z)\big)M_{0,\Omega}^D(z)^{-1} \notag
\\ &=
I_{\partial\Omega} + \big(M_{0,\Omega}^D(z) - M_{\Omega}^D(z)\big)M_{0,\Omega}^N(z)
\notag
\\ &=
I_{\partial\Omega} + \overline{\widetilde\gamma_N \big(H^D_{\Omega}-zI_\Omega\big)^{-1} V \big[\gamma_N
\big(\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*}
\\ &\quad \times
\gamma_D\big[\gamma_D
\big(\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*. \notag
\end{align}
Let $g\inL^2(\dOm;d^{n-1}\si)$. Then by Theorem \ref{t3.1},
\begin{align}
u=\big[\gamma_D\big(\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*g
\label{3.41}
\end{align}
is the unique solution of
\begin{align}
(-\Delta-z)u = 0 \,\text{ on }\Omega, \quad u\in H^{3/2}(\Omega),
\quad \widetilde\gamma_N u = g \,\text{ on }{\partial\Omega}.
\end{align}
Setting $f=\gamma_D u \in H^1({\partial\Omega})$ and utilizing Theorem
\ref{t3.1} once again, one obtains
\begin{align}
u &= -\big[\gamma_N \big(H_{0,\Omega}^D-\overline{z}I_\Omega\big)^{-1}\big]^*f \notag
\\ &=
-\big[\gamma_N \big(\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*
\gamma_D\big[\gamma_D
\big(\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*g. \label{3.43}
\end{align}
Thus, it follows from \eqref{3.41} and \eqref{3.43} that
\begin{align}
\big[\gamma_N \big(\big(H^D_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*
\gamma_D\big[\gamma_D
\big(\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^* =
-\big[\gamma_D\big(\big(H^N_{0,\Omega}-zI_\Omega\big)^{-1}\big)^*\big]^*.
\label{3.44}
\end{align}
Finally, insertion of \eqref{3.44} into \eqref{3.40} yields
\eqref{3.36}.
\end{proof}
It follows from \eqref{4.24}--\eqref{4.29} that $\widetilde \gamma_N$ can be replaced by $\gamma_N$ on the right-hand side of \eqref{3.35} and \eqref{3.36}.
\section{A Multi-Dimensional Variant of a Formula due to Jost and Pais}
\label{s4}
In this section we prove our multi-dimensional variants of the Jost and Pais formula as discussed in the introduction.
We start with an elementary comment on determinants which, however, lies at the heart of the matter of our multi-dimensional variant of the one-dimensional Jost and Pais result. Suppose $A \in {\mathcal B}({\mathcal H}_1, {\mathcal H}_2)$, $B \in {\mathcal B}({\mathcal H}_2, {\mathcal H}_1)$ with $A B \in {\mathcal B}_1({\mathcal H}_2)$ and $B A \in {\mathcal B}_1({\mathcal H}_1)$. Then,
\begin{equation}
\det (I_{{\mathcal H}_2}-AB) = \det (I_{{\mathcal H}_1}-BA). \label{4.0}
\end{equation}
In particular, ${\mathcal H}_1$ and ${\mathcal H}_2$ may have different dimensions. Especially, one of them may be infinite and the other finite, in which case one of the two determinants in \eqref{4.0} reduces to a finite determinant. This case indeed occurs in the original one-dimensional case studied by Jost and Pais \cite{JP51} as described in detail in
\cite{GM03} and the references therein. In the proof of the next theorem, the role of
${\mathcal H}_1$ and ${\mathcal H}_2$ will be played by $L^2(\Omega; d^n x)$ and
$L^2(\partial\Omega; d^{n-1} \sigma)$, respectively.
We start with an extension of a result in \cite{GLMZ05}:
\begin{theorem} \label{t4.1}
Assume Hypothesis \ref{h2.6} and let
$z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{\Omega}^D\big)\cup \sigma\big(H_{0,\Omega}^D\big) \cup
\sigma\big(H_{0,\Omega}^N\big)\big)$. Then,
\begin{align}
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}V \big(H_{\Omega}^D-zI_{\Omega}\big)^{-1}V
\big[\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*}
&\in{\mathcal B}_1\big(L^2(\dOm;d^{n-1}\si)\big), \label{4.1} \\
\overline{\gamma_N\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1}V
\big[\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*}
&\in{\mathcal B}_2\big(L^2(\dOm;d^{n-1}\si)\big), \label{4.2}
\end{align}
and
\begin{align}
&
\frac{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}\,\Big)}
{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big)} \notag \\
&\quad = \det{}_2\Big(I_{{\partial\Omega}} -
\overline{\gamma_N\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1}V
\big[\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*} \, \Big)
\label{4.3}\\
&\quad\quad \times \exp\Big(\text{\rm{tr}}\Big(\,
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}V \big(H_{\Omega}^D-zI_{\Omega}\big)^{-1}V
\big[\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*} \,\Big)\Big). \notag
\end{align}
\end{theorem}
\begin{proof}
From the outset we note that the left-hand side of \eqref{4.3} is
well-defined by \eqref{2.35}. Let
$z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{\Omega}^D\big)\cup\sigma\big(H_{0,\Omega}^D\big) \cup
\sigma\big(H_{0,\Omega}^N\big)\big)$ and
\begin{align}
u(x) &= \exp(i\arg(V(x)))\abs{V(x)}^{1/2},\quad
v(x)=\abs{V(x)}^{1/2},
\\
\widetilde u(x) &= \exp(i\arg(V(x)))\abs{V(x)}^{p/p_1},\quad \widetilde
v(x)=\abs{V(x)}^{p/p_2},
\end{align}
where
\begin{align} \label{4.6}
p_1=\begin{cases} 3p/2,&n=2, \\ 4p/3,&n=3,\end{cases} \qquad
p_2=\begin{cases}3p,&n=2, \\ 4p,& n=3 \end{cases}
\end{align}
with $p$ as introduced in Hypothesis \ref{h2.6}.
Then it follows that $\f1{p_1}+\frac{1}{p_2}=\f1p$, in both cases
$n=2,3$, and hence $V=uv=\widetilde u \widetilde v$.
Next, we introduce
\begin{equation}
K_D(z)=-\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}, \quad
K_N(z)=-\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}
\end{equation}
and note that
\begin{align}
[I_{\Omega}-K_D(z)]^{-1} \in{\mathcal B}\big(L^2(\Om;d^nx)\big), \quad
z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{\Omega}^D\big)\cup\sigma\big(H_{0,\Omega}^D\big)\big).
\end{align}
Thus, utilizing the following facts,
\begin{align}
[I_{\Omega} - K_D(z)]^{-1} &= I_{\Omega} + K_D(z)[I_{\Omega} - K_D(z)]^{-1}
\end{align}
and
\begin{align}
1& = \det{}_2\big([I_{\Omega}-K_D(z)][I_{\Omega} - K_D(z)]^{-1}\big)
\\
&=
\det{}_2\big(I_{\Omega}-K_D(z)\big)\det{}_2\big([I_{\Omega} - K_D(z)]^{-1}\big)
\exp\big(\text{\rm{tr}}\big(K_D(z)^2[I_{\Omega} - K_D(z)]^{-1}\big)\big), \notag
\end{align}
one obtains
\begin{align}
&\det{}_2\big([I_{\Omega}-K_N(z)][I_{\Omega}-K_D(z)]^{-1}\big) \notag \\
&\quad = \det{}_2\big(I_{\Omega}-K_N(z)\big)
\det{}_2\big([I_{\Omega}-K_D(z)]^{-1}\big)
\exp\big(\text{\rm{tr}}\big(K_N(z)K_D(z)[I_{\Omega}-K_D(z)]^{-1}\big)\big) \\
&\quad =
\frac{\det{}_2\big(I_{\Omega}-K_N(z)\big)}{\det{}_2\big(I_{\Omega}-K_D(z)\big)}
\exp\big(\text{\rm{tr}}\big((K_N(z)-K_D(z))K_D(z)[I_{\Omega}-K_D(z)]^{-1}\big)\big).
\notag
\end{align}
At this point, the left-hand side of \eqref{4.3} can be rewritten
as
\begin{align}
&\frac{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}\,\Big)}
{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big)} =
\frac{\det{}_2\big(I_{\Omega}-K_N(z)\big)}{\det{}_2\big(I_{\Omega}-K_D(z)\big)}
\notag \\
&\quad = \det{}_2\big([I_{\Omega}-K_N(z)][I_{\Omega}-K_D(z)]^{-1}\big)
\notag \\
&\quad\quad \, \times
\exp\big(\text{\rm{tr}}\big((K_D(z)-K_N(z))K_D(z)[I_{\Omega}-K_D(z)]^{-1}\big)\big)
\notag \\
&\quad =
\det{}_2\big(I_{\Omega}+(K_D(z)-K_N(z))[I_{\Omega}-K_D(z)]^{-1}\big)
\label{4.12}
\\
&\quad\quad \, \times
\exp\big(\text{\rm{tr}}\big((K_D(z)-K_N(z))K_D(z)[I_{\Omega}-K_D(z)]^{-1}\big)\big).
\notag
\end{align}
Next, temporarily suppose that $V\in L^p(\Omega;d^nx)\cap
L^\infty(\Omega;d^nx)$. Using \cite[Lemma A.3]{GLMZ05} (an extension of a
result of Nakamura \cite[Lemma 6]{Na01}) and \cite[Remark A.5]{GLMZ05},
one finds
\begin{align}
\begin{split}
K_D(z)-K_N(z) &=
-\overline{u\big[\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}- \big(H_{0,\Omega}^N
-zI_{\Omega}\big)^{-1}\big]v}
\\ &=
-\overline{u\big[\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*}\,
\overline{\gamma_N \big(H_{0,\Omega}^D -zI_{\Omega}\big)^{-1}v}
\\ &=
-\Big[\,\overline{\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}\,\Big]^* \,
\overline{\gamma_N \big(H_{0,\Omega}^D -zI_{\Omega}\big)^{-1}v}.
\end{split}\label{4.13}
\end{align}
Thus, inserting \eqref{4.13} into \eqref{4.12} yields,
\begin{align} \label{4.14}
&\frac{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}\,\Big)}
{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big)} \notag
\\
&\quad = \det{}_2\Big(I_{\Omega} - \Big[\,\overline{\gamma_D
\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}\,\Big]^* \overline{\gamma_N
\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v} \notag
\Big[I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big]^{-1}\Big)
\notag
\\
&\quad\quad \times \exp\Big(\text{\rm{tr}}\Big(\Big[\,\overline{\gamma_D
\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}\,\Big]^*
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v} \notag
\\
& \hspace*{2.6cm} \times
\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\Big[I_{\Omega}+\overline{u\big(H_{0,\Omega}^D
-zI_{\Omega}\big)^{-1}v}\,\Big]^{-1}\Big)\Big).
\end{align}
Then, utilizing Corollary \ref{c2.5} with $p_1$ and $p_2$ as in
\eqref{4.6}, one finds,
\begin{align}
\overline{\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}
&\in{\mathcal B}_{p_1}\big(L^2(\Om;d^nx),L^2(\dOm;d^{n-1}\si)\big),
\\
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}
&\in{\mathcal B}_{p_2}\big(L^2(\Om;d^nx),L^2(\dOm;d^{n-1}\si)\big),
\end{align}
and hence,
\begin{align}
\Big[\,\overline{\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}\,\Big]^*
\overline{\gamma_N \big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v} &\in {\mathcal B}_p\big(L^2(\Om;d^nx)\big)
\subset{\mathcal B}_2\big(L^2(\Om;d^nx)\big),
\\
\overline{\gamma_N \big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v} \Big[\,\overline{\gamma_D
\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}\,\Big]^* &\in
{\mathcal B}_p\big(L^2(\dOm;d^{n-1}\si)\big) \subset{\mathcal B}_2\big(L^2(\dOm;d^{n-1}\si)\big).
\end{align}
Moreover, using the fact that
\begin{align}
\Big[I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big]^{-1} \in
{\mathcal B}\big(L^2(\Om;d^nx)\big), \quad z\in{\mathbb{C}}\big\backslash
\big(\sigma\big(H_{\Omega}^D\big)\cup\sigma\big(H_{0,\Omega}^D\big)\big),
\end{align}
one now applies the idea expressed in formula \eqref{4.0} and rearranges the terms in \eqref{4.14} as follows:
\begin{align} \label{4.20}
&\frac{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}\,\Big)}
{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big)} \notag
\\
&\quad = \det{}_2\Big(I_{{\partial\Omega}} - \overline{\gamma_N
\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}
\Big[I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big]^{-1}
\Big[\, \overline{ \gamma_D
\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}\,\Big]^*\Big) \notag
\\
&\quad\quad \times \exp\Big(\text{\rm{tr}}\Big(\,\overline{\gamma_N
\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1} v} \; \overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}
v} \notag
\\
& \hspace*{2.45cm} \times \Big[I_{\Omega}+
\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1} v}\,\Big]^{-1} \Big[\,\overline{\gamma_D
\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{u}}\,\Big]^*\Big)\Big) \notag
\\
&\quad = \det{}_2\Big(I_{{\partial\Omega}} - \overline{\gamma_N \big(H_{0,\Omega}^D
-zI_{\Omega}\big)^{-1}\widetilde v} \Big[I_{\Omega}+\overline{\widetilde
u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v}\,\Big]^{-1} \Big[\,\overline{\gamma_D
\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{\widetilde u}}\,\Big]^*\Big) \notag
\\
&\quad\quad \times
\exp\Big(\text{\rm{tr}}\Big(\,\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v} \;
\overline{\widetilde u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v}
\\
& \hspace*{2.45cm} \times \Big[I_{\Omega}+\overline{\widetilde
u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v}\,\Big]^{-1} \Big[\,\overline{\gamma_D
\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\overline{\widetilde u}}\,\Big]^*\Big)\Big).
\notag
\end{align}
In the last equality we employed the following simple
identities,
\begin{align}
& V = u v = \widetilde u \widetilde v, \\
& v\Big[I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big]^{-1}u =
\widetilde v \Big[I+\overline{\widetilde u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde
v}\,\Big]^{-1} \widetilde u.
\end{align}
Utilizing \eqref{4.20} and the following resolvent identity,
\begin{align} \label{4.23}
\overline{\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v} =
\overline{\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v} \Big[I_{\Omega}+\overline{\widetilde
u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v}\,\Big]^{-1},
\end{align}
one arrives at \eqref{4.3}, subject to the extra assumption $V\in
L^p(\Omega;d^n x)\cap L^\infty(\Omega;d^n x)$.
Finally, assuming only $V\in L^p(\Omega;d^n x)$ and utilizing \cite[Thm.\ 3.2]{GLMZ05}, Lemma \ref{l2.3}, and Corollary \ref{c2.5} once again, one obtains
\begin{align}
\Big[I_{\Omega}+\overline{\widetilde u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1} \widetilde
v}\,\Big]^{-1} &\in {\mathcal B}\big(L^2(\Om;d^nx)\big), \label{4.24}
\\
\widetilde u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-p/p_1} &\in
{\mathcal B}_{p_1}\big(L^2(\Om;d^nx)\big), \label{4.25}
\\
\widetilde v\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-p/p_2} &\in
{\mathcal B}_{p_2}\big(L^2(\Om;d^nx)\big), \label{4.26}
\\
\overline{\gamma_D \big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}\widetilde u} &\in
{\mathcal B}_{p_1}\big(L^2(\Om;d^nx),L^2(\dOm;d^{n-1}\si)\big), \label{4.27}
\\
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v} &\in
{\mathcal B}_{p_2}\big(L^2(\Om;d^nx),L^2(\dOm;d^{n-1}\si)\big), \label{4.28}
\end{align}
and hence
\begin{equation}
\overline{\widetilde u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}\widetilde v} \in
{\mathcal B}_p\big(L^2(\Om;d^nx)\big)\subset{\mathcal B}_2\big(L^2(\Om;d^nx)\big). \label{4.29}
\end{equation}
Relations \eqref{4.23}--\eqref{4.29} prove \eqref{4.1} and
\eqref{4.2}, and hence, the left- and the right-hand sides of
\eqref{4.3} are well-defined for $V\in L^p(\Omega;d^nx)$. Thus, using
\eqref{2.8}, \eqref{2.27}, \eqref{2.28}, the continuity of
$\det{}_2(\cdot)$ with respect to the Hilbert--Schmidt norm
$\|\cdot\|_{{\mathcal B}_2\big(L^2(\Om;d^nx)\big)}$, the continuity of $\text{\rm{tr}}(\cdot)$ with
respect to the trace norm $\|\cdot\|_{{\mathcal B}_1\big(L^2(\Om;d^nx)\big)}$, and an
approximation of $V\in L^p(\Omega;d^nx)$ by a sequence of potentials
$V_k \in L^p(\Omega;d^nx)\cap L^\infty(\Omega;d^nx)$, $k\in{\mathbb{N}}$, in the
norm of $L^p(\Omega;d^nx)$ as $k\uparrow\infty$, then extends the
result from $V\in L^p(\Omega;d^nx)\cap L^\infty(\Omega;d^nx)$ to $V\in
L^p(\Omega;d^nx)$, $n=2,3$.
\end{proof}
Given these preparations, we are now ready for the principal result of this paper, the multi-dimensional analog of Theorem \ref{t1.2}:
\begin{theorem} \label{t4.2}
Assume Hypothesis \ref{h2.6} and let
$z\in{\mathbb{C}}\big\backslash\big(\sigma\big(H_{\Omega}^D\big)\cup \sigma\big(H_{0,\Omega}^D\big) \cup
\sigma\big(H_{0,\Omega}^N\big)\big)$. Then,
\begin{equation}
M_{\Omega}^{D}(z)M_{0,\Omega}^{D}(z)^{-1} - I_{\partial\Omega} =
- \overline{\gamma_N\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1} V
\big[\gamma_D(H_{0,\Omega}^N-\overline{z}I_{\Omega})^{-1}\big]^*}
\in {\mathcal B}_2\big(L^2(\partial\Omega; d^{n-1}\sigma)\big)
\end{equation}
and
\begin{align}
& \frac{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}\,\Big)}
{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big)} \notag \\
& \quad = \det{}_2\Big(I_{{\partial\Omega}} - \overline{\gamma_N\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1} V
\big[\gamma_D(H_{0,\Omega}^N-\overline{z}I_{\Omega})^{-1}\big]^*}\,\Big)
e^{\text{\rm{tr}}(T_2(z))} \label{4.30} \\
& \quad =
\det{}_2\big(M_{\Omega}^{D}(z)M_{0,\Omega}^{D}(z)^{-1}\big)
e^{\text{\rm{tr}}(T_2(z))}, \label{4.31}
\end{align}
where
\begin{equation}
T_2(z)=\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1} V
\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1} V
\big[\gamma_D \big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*}
\in {\mathcal B}_1\big(L^2(\partial\Omega; d^{n-1}\sigma)\big).
\end{equation}
\end{theorem}
\begin{proof}
The result follows from combining Lemma \ref{l3.5} and Theorem \ref{t4.1}.
\end{proof}
A few comments are in order at this point.
The sudden appearance of the exponential term $\exp(\text{\rm{tr}}(T_2(z)))$ in \eqref{4.30} and \eqref{4.31}, when compared to the one-dimensional case, is due to the necessary use of the modified determinant ${\det}_p(\cdot)$ in Theorems \ref{t4.1} and \ref{t4.2}.
The multi-dimensional extension \eqref{4.30} of
\eqref{1.16}, under the stronger hypothesis $V\in L^2(\Omega; d^n x)$, $n=2,3$, first appeared in \cite{GLMZ05}. However, the present results in Theorem \ref{t4.2} go decidedly beyond those in \cite{GLMZ05} in the sense that the class of domains
$\Omega$ permitted by Hypothesis \ref{h2.1} is greatly enlarged as compared to \cite{GLMZ05} and the conditions on $V$ satisfying Hypothesis \ref{h2.6} are nearly optimal by comparison with the Sobolev inequality (cf.\ Cheney \cite{Ch84}, Reed and Simon \cite[Sect.\ IX.4]{RS75}, Simon \cite[Sect.\ I.1]{Si71}). Moreover, the multi-dimensional extension \eqref{4.31} of \eqref{1.17} invoking Dirichlet-to-Neumann maps is a new result.
The principal reduction in Theorem \ref{t4.2} reduces (a ratio of) modified Fredholm determinants associated with operators in $L^2(\Omega; d^n x)$ on the left-hand side of \eqref{4.30} to modified Fredholm determinants associated with operators in $L^2(\partial\Omega; d^{n-1} \sigma)$ on the right-hand side of \eqref{4.30} and especially, in \eqref{4.31}. This is the analog of the reduction described in the one-dimensional context of Theorem \ref{t1.2}, where $\Omega$ corresponds to the half-line $(0,\infty)$ and its boundary $\partial\Omega$ thus corresponds to the one-point set $\{0\}$.
In the context of elliptic operators on smooth $k$-dimensional manifolds, the idea of reducing a ratio of zeta-function regularized determinants to a calculation over the
$(k-1)$-dimensional boundary has been studied by Forman \cite{Fo87}. He also pointed out that if the manifold consists of an interval, the special case of a pair of boundary points then permits one to reduce the zeta-function regularized determinant to the determinant of a finite-dimensional matrix. The latter case is of course an analog of the one-dimensional Jost and Pais formula mentioned in the introduction (cf.\ Theorems \ref{t1.1} and \ref{t1.2}). Since then, this topic has been further developed in various directions and we refer, for instance, to Burghelea, Friedlander, and Kappeler \cite {BFK91}, \cite {BFK92}, \cite {BFK93}, \cite{BFK95}, Carron \cite{Ca02}, Friedlander \cite{Fr05}, M\"uller \cite{Mu98}, Park and Wojciechowski \cite{PW05}, and the references therein.
\begin{remark} \label{c4.3} The following observation yields a simple
application of formula \eqref{4.30}. Since by the Birman--Schwinger principle (cf., e.g., the discussion in \cite[Sect.\ 3]{GLMZ05}), for any
$z\in{\mathbb{C}}\big\backslash \big(\sigma\big(H_{\Omega}^D\big)\cup
\sigma\big(H_{0,\Omega}^D\big) \cup \sigma\big(H_{0,\Omega}^N\big)\big)$, one has
$z\in\sigma\big(H_{\Omega}^N\big)$ if and only if $\det{}_2\Big(I_{\Omega}+ \overline{u\big(H_{0,\Omega}^N -zI_{\Omega}\big)^{-1}v}\,\Big)=0$, it follows from \eqref{4.30} that
\begin{align}
\begin{split}
&\text{for all } \, z\in{\mathbb{C}}\big\backslash \big(\sigma\big(H_{\Omega}^D\big)\cup
\sigma\big(H_{0,\Omega}^D\big) \cup \sigma\big(H_{0,\Omega}^N\big)\big), \text{ one has }
z\in\sigma\big(H_{\Omega}^N\big) \\
&\quad \text{if and only if } \, \det{}_2\Big( I_{{\partial\Omega}} -
\overline{\gamma_N\big(H_{\Omega}^D-zI_{\Omega}\big)^{-1}V
\big[\gamma_D\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*}\, \Big)=0.
\end{split}
\end{align}
One can also prove the following analog of \eqref{4.30}:
\begin{align}
&\frac{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}v}\,\Big)}
{\det{}_2\Big(I_{\Omega}+\overline{u\big(H_{0,\Omega}^N-zI_{\Omega}\big)^{-1}v}\,\Big)} \notag \\
&\quad = \det{}_2\Big(I_{{\partial\Omega}} +
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}V
\big[\gamma_D\big((H_{\Omega}^N-z I_{\Omega}\big)^{-1})^*\big]^*}\,\Big)
\label{4.37} \\
&\quad\quad \times \exp\Big(-\text{\rm{tr}}\Big(\,
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}V
\big(H_{\Omega}^N-zI_{\Omega}\big)^{-1}V
\big[\gamma_D\big(H_{0,\Omega}^N-\overline{z}I_{\Omega}\big)^{-1}\big]^*}
\,\Big)\Big). \notag
\end{align}
Then, proceeding as before, one obtains
\begin{align}
& \text{for all } \, z \in {\mathbb{C}} \big\backslash
\big(\sigma\big(H_{\Omega}^N\big)\cup\sigma\big(H_{0,\Omega}^N\big)\cup\sigma\big(H_{0,\Omega}^D\big)\big),
\text{ one has } z\in\sigma\big(H_{\Omega}^D\big) \\
&\quad \text{if and only if }\, \det{}_2\Big(I_{{\partial\Omega}} +
\overline{\gamma_N\big(H_{0,\Omega}^D-zI_{\Omega}\big)^{-1}V
\big[\gamma_D\big(\big(H_{\Omega}^N-z I_{\Omega}\big)^{-1}\big)^*\big]^*} \,\Big)=0. \notag
\end{align}
\end{remark}
\noindent {\bf Acknowledgments.}
We are indebted to Yuri Latushkin and Konstantin A. Makarov for numerous
discussions on this topic. We also thank the referee for a careful reading of
our manuscript and his constructive comments.
Fritz Gesztesy would like to thank all
organizers of the international conference on Operator Theory and Mathematical Physics (OTAMP), and especially, Pavel Kurasov, for their kind invitation, the stimulating atmosphere during the meeting, and the hospitality extended to him during his stay in Lund in June of 2006. He also gratefully acknowledges a research leave for the academic year 2005/06 granted by the Research Council and the Office of Research of the University of Missouri--Columbia.
|
\section{Introduction}
Phylogenetic methods have recently been rediscovered in several
interesting areas among which immunodynamics, epidemiology and many
branches of evolutionary dynamics. The reconstruction of phylogenetic
trees belongs to a general class of inverse problems whose relevance
is now well established in many different disciplines ranging from
biology to linguistics and social
sciences~\cite{Gray2003,css_2009,pybus_2009,warnow_science_2009}. In
a generic inverse problem one is given with a set of data and has to
infer the most likely dynamical evolution process that presumably
produced the given data set. The relevance of inverse problems has
been certainly triggered by the fast progress in data-revealing
technologies. In molecular biology, for instance, a great amount of
genomes data are available thanks to the new high-throughput methods
for genome analysis~\cite{ragousis_2009}. In historical
linguistics~\cite{renfrew_2000} a remarkable effort has been recently
done for the compilation of corpora of homologous features (lexical,
phonological, syntactic) or characters for many different languages.
Although phylogenetic reconstruction is not a novel topic, dealing
with not purely tree-like processes and identifying the possible
sources of non-additivity and their effects in a given dataset is
still an open and challenging
problem~\cite{felsenstein2004,GASCUEL_2007}.
Here we focus on distance-based
methods~\cite{cavalli_edwards67,fitch_margoliash67} and investigate
how deviations from additivity affect their performances. In
distance-based methods only distances between leaves are considered,
and all the information possibly encoded in the combinatorial
structure of the character states is lost. Despite their simplicity,
distance-based methods are still widely used thanks to their
computational efficiency, but a solid theoretical understanding on the
limitation of their applicability is still lacking. One of the most
popular distance-based reconstruction algorithm, Neighbor-Joining
(NJ)~\cite{NJ}, was proposed in the late {\em 80's}, but it is only
recently that its theoretical background was put on a more solid
basis~\cite{Atteson97, Gascuel2006, Mihaescu2007}. Another step
toward a better understanding of distance-based methods was obtained
thanks to an interesting property of additive distances, the Pauplin's
formula~\cite{pauplin2000}. This property has been used in the
formulation of a novel algorithmic strategy with improved performances
(FastME)~\cite{fastme_2002}. In parallel, another fundamental
property of additive trees, the {\em four-points}
condition~\cite{Gusfield97}, has been extensively exploited in
distance-based phylogenetic reconstruction
methods~\cite{erdos_1998,snir_2008}. Both the Pauplin's formula and
the {\em four-points} condition will be discussed in details below.
Here we propose a new approach that combines the {\em four-points}
condition and the Pauplin's formula in a Stochastic Local Search (SLS)
scheme that we name Stochastic Big-Quartet Swapping (SBiX) algorithm.
SLS~\cite{booksls} algorithms transverse the search space of a given
problem in a systematic way, allowing for a sampling of low cost
configurations. SLS algorithms start from a randomly chosen initial
condition. Subsequently the elementary step is a move from a
configuration to a neighboring one. Each move is determined by a
decision based on local knowledge only. Typically the decision is
taken combining, with a given a priori probability, a greedy step
(i.e., a step that reduces the local cost contribution), with a random
one where the local cost is not taken into account. SLS algorithms
have been widely used in solving complex combinatorial optimization
problems such as Satisfiability, Coloring, MAX-SAT, and Traveling
Salesman Problem~\cite{booksls}.
At the heart of our new algorithmic scheme (named SBiX) there is the
notion of {\em quartet frustration}, a quantitative measure of how
good a given configuration is, in the space of trees. Following a
concept already introduced in~\cite{snir_2008}, we weight the
different quartets according to their length, in order to reduce the
effect of those which are more likely to undergo double mutations. The
strength of our approach comes from a combination of this strategy
with a Pauplin's like one, weighting each quartet according to a
purely topological property.
We tested the performances of the proposed reconstruction algorithm,
i.e., the ability to reconstruct the true topology, in the presence of
high levels of deviation from additivity due to both horizontal
transfer and back-mutation processes. We use both a very simple model
to generate artificial phylogenies of binary sequences, and the more
realistic Kimura two-parameters model, considering sequences with
$q$-state sites, with $q=4$. We have evidence that the performance of
our algorithm does not rely on the particular evolutive process giving
rise to the phylogeny, nor on the particular representation of the
taxa. We find that when the lack of additivity arises from high
mutation rates (and consequently high probability of back mutations),
our algorithm significantly outperforms the state-of-art
distance-based algorithms. When the lack of additivity arises from
high rate of horizontal transfer events, our algorithm performs better
than the algorithms we considered as competing ones.
We show results both for a greedy and a simulated-annealing-like
strategy of our algorithm, the former being significantly faster than
the latter and with comparable performances. The SBiX algorithm has a
complexity of $O(N^4)$, which is higher that the one of the
distance-based algorithms used as competitors for comparing the
performances. Nevertheless, the prefactor of the greedy version is so
low that our algorithm is fast enough to reconstruct large phylogenies
e.g., of a few thousands taxa, in a time remarkably slower with
respect to any character based reconstruction algorithm. A comparison
of the running time and of the performances of our algorithm with a
popular character based one, MrBayes~\cite{MrBayes}, is reported
respectively in the Appendix II and in the section Results. We show
that, while the performances of the two algorithms are comparable, the
running time of MrBayes greatly exceed ours, becoming comparable when
the reconstruction becomes in practise unfeasible, that is for running
times of the order of some thousand of years, and number of taxa of
the order of one hundred thousand.
\section{Methods}
\label{sec:add}
\subsection{Additivity and the four-points condition}
A distance matrix is said to be additive if it can be constructed as
the sum of a tree's branches lengths. Two fundamental and widely used
properties of additive distances are the Pauplin's
formula~\cite{pauplin2000} and the {\em four-points}
condition~\cite{Gusfield97}. For the sake of clarity, we recall them
here.
Given any quadruplet of taxa $a,b,c,d$, let $D_1=d(a,b)+d(c,d)$,
$D_2=d(a,c)+d(b,d)$, and $D_3=d(a,d)+d(b,c)$ be the three possible
pairs of distances between the four taxa. A matrix $M$ is additive if
and only if $D_1<D_2=D_3$ or $D_2<D_1=D_3$ or
$D_3<D_1=D_2$\footnote{It is important to remark here that the
four-points condition is an equivalent definition of additivity.
That is: a distance matrix is additive if and only if the
four-points condition is satisfied.}. When considering experimental
data, additivity is almost always violated and so is the {\em
four-points} condition. In order to set up a robust method for
phylogeny reconstruction based on the {\em four-points} condition, we
need to relax the notion of additivity and to quantify violations in a
suitable way. For any four taxa $a,b,c,d$ such that $a,b$ are on one
side of the tree and $c,d$ on the other (as in fig.~\ref{fig:tree}),
the quartet $(ab:cd)$ is said to satisfy the {\em weak four-points}
condition if $D_1=\min(D_1,D_2,D_3)$ (where $D_1,D_2,D_3$ are defined
as above). It is easy to prove that if the distance matrix $M$ is
additive a unique tree exists in which all quartets satisfy the {\em
weak four-points} condition and this tree is the correct one. Many
algorithms have been proposed that exploit this {\em weak four-points}
condition, one of the most promising being, for instance, the
short-quartet method~\cite{erdos_1998,snir_2008}.
\begin{figure}
\centerline{\includegraphics[width=0.49\textwidth]{fig1}}
\caption{{\bf Quartet definition.} The quartet $a,b,c,d$ induces an
internal edge $x$ that divides the tree in two parts. All paths
joining any pair of sites sitting on opposite parts of the tree
pass through $x$.}
\label{fig:tree}
\end{figure}
\subsection{Pauplin's distance}
Another remarkable property of an additive tree is the possibility to
compute its total length $L$, defined as the sum of all its branches
lengths, through a formula, due to Pauplin~\cite{pauplin2000}, that
only uses distances between taxa:
\begin{equation}
\label{eq:pauplin}
L_P= \sum_{a<b} 2^{-t(a,b)}M(a,b),
\end{equation}
\noindent where $t(a,b)$ is the number of nodes on the path connecting
$a$ and $b$, i.e., their {\em topological distance}. For additive
trees $L\equiv L_P$, but even when the {\em four-points} condition is
no longer satisfied, $L_P$ is a particularly good approximation for
the tree length~\cite{fastme_2002}. Furthermore, it is recognized
that for distance matrices sufficiently ``close'' to
additivity~\cite{Atteson97} the correct phylogeny minimizes
$L_P$~\cite{Mihaescu2007,Bordewich_2009}. This principle is used in an implicit way
in Neighbor-Joining~\cite{NJ} and more explicitly in
a new generation distance-based algorithm, FastME~\cite{fastme_2002}.
When departure from additivity is too strong $L_P$ is no longer a good
functional to minimize in order to recover the correct tree.
\subsection{Violations of additivity}
Violations of additivity can arise both from experimental noise and
from properties of the evolutionary process the data come from. We
here consider two of the main sources of violations of the latter type
that can either occur together or singularly. (i) {\em back-mutation}:
in particularly long phylogenies, here the time-scale being set by the
mutation rate, a single character can experience multiple mutations.
In this case the distances between taxa are no longer proportional to
their evolutionary distances; we will use in the following the
expression {\em back-mutation} as synonymous of multiple mutation on
the same site; (ii) {\em horizontal transfer}: the reconstruction of a
phylogeny from data underlies the assumption that information flows
{\em vertically} from ancestors to offsprings. However, in many
processes information flows also {\em horizontally}. Horizontal (or
lateral) gene transfers~\cite{treeoflifepnas2005}
are often well known
confounding factors for a correct phylogenetic inference.
\subsection{The Stochastic Big-Quartet Swapping algorithm
\label{sec:algo}}
Here we describe the structure of our Stochastic Big-Quartet Swapping
(SBiX) algorithm. As already mentioned this algorithm crucially
exploits both the {\em weak four-points} condition and the Pauplin's
distance, featuring a larger robustness with respect to violations of
additivity if compared to algorithms based separately on the {\em weak
four-points} condition or on the minimization of the Pauplin's
length.
The general structure of the SBiX algorithm is as follows:
\begin{itemize}
\item[$1$] start with a tree topology for the given set of
taxa;
\item[$2$] update the tree topology by local elementary rearrangements
through sub-trees quartet swapping;
\item[$3$] repeat point $2$ till convergence is reached;
\end{itemize}
\begin{figure}
\centerline{\includegraphics[width=0.49\textwidth]{fig2}}
\caption{{\bf Elementary move of our Stochastic Local Search scheme.} Any internal edge defines the four
subtrees $A,B,C,D$ rooted respectively on
$\alpha,\beta,\gamma,\delta$. Here, the initial reference
configuration $((A,B),(C,D))$ displayed on the left, leads to two
possible rewiring: (i) swap the pair $B \leftrightarrow D$
(right-upper panel), (ii) swap the pair $B \leftrightarrow C $
(right-lower panel).}
\label{fig:mcstep}
\end{figure}
\subsection{Sub-trees (big-)quartets swaps}
We analyze the three points in details in the following:
\begin{itemize}
\item[$1$] In the simulated-annealing-like version of the algorithm we
start with a random topology. In the greedy version
we search for a local minimum (which can eventually be the global
one), so it is important to start with a meaningful topology. We
tested the algorithm starting both with the FastME and the NJ
reconstructed topologies.
\item[$2$] We sequentially consider all the internal edges of the
present topology. Each internal edge defines four subtrees (see
fig.~\ref{fig:mcstep}), say $A,B,C,D$, rooted respectively on the
four internal nodes $\alpha,\beta,\gamma,\delta$. Referring to
fig.~\ref{fig:mcstep}, let $((A,B),(C,D))$ being the initial
configuration. We randomly choose one of the two possible local
rewirings: (i) swap the pair $B \leftrightarrow D $ getting the
configuration $((A,D),(B,C))$, (ii) swap the pair $B \leftrightarrow
C $ getting the configuration $((A,C),(B,D))$. In the
simulated-annealing version, the new configuration is accepted with
a probability proportional to the statistical weight $e^{-\beta
\mathrm{ sign}(\Delta E)}$ where $\beta$ is an inverse
temperature-like parameter that is set by a simulated
annealing-like~\cite{Kirkpatrick83} strategy and $\Delta E$ is the
difference of the two configurations local costs: for case (i)
$\Delta E = E_{((A,D),(B,C))} - E_{((A,B),(C,D))}$ while for case
(ii) $\Delta E = E_{((A,C),(B,D))} - E_{((A,B),(C,D))}$. In the
greedy (or zero temperature) version, the new configuration is
accepted if and only if $\Delta E < 0$.
\item[$3$] In the simulated-annealing version, we iterate point $2$ starting
with $\beta=0$ and increasing it at a constant rate at each sweep
(where we call {\em sweep} an update of all the internal edges) until
convergence is reached, i.e., until the algorithm gets stuck in a
fixed topology. In the greedy version, we iterate
point $2$ until the algorithm gets stuck in a fixed topology.
\end{itemize}
In order to define the configurational local cost, say
$E_{((A,B),(C,D))}$, we consider all the quartets $(ab:cd)$ such that
$a \in A$ ($a$ is a taxa of the subtree $A$) , $b \in B$, $c \in C$,
and $ d \in D$. For each quartet we define the {\em quartet
frustration} as:
\begin{equation}
\label{eq:fru}
f_{(ab :cd)}=\max\left(0,\frac{D_1-\min(D_2,D_3)}{(D_1+\min(D_2,D_3))^k}\right),
\end{equation}
\noindent where $D_1,D_2,D_3$ are the sums of distances already
defined ($D_1=d(a,b)+d(c,d)$, $D_2=d(a,c)+d(b,d)$,
$D_3=d(a,d)+d(b,c)$). The normalization factor in the right-hand side
of Eq.~\ref{eq:fru}, as already pointed out in the introduction, gives
a smaller weight to longer distances, typically affected by noise and
recombination. The parameter fixing strategy for the exponent $k$ will
be discussed in the next subsection.
The cost $E_{((A,B),(C,D))}$ of the configuration $((A,B),(C,D))$ is
thus defined as the sum of the costs ({\em frustrations}) of all
the considered quartets, each weighted with a factor borrowed from the
Pauplin's formula:
\begin{equation}
\label{eq:quart-paup}
E_{((A,B),(C,D))}= \sum_{(ab :cd)} f_{(ab :cd)}
2^{-t(a,\alpha)-t(b,\beta)-t(c,\gamma)-t(d,\delta)},
\end{equation}
\noindent where $t(a,\alpha)$ is the topological distance between the
taxa $a$ and the internal node $\alpha$, and analogously for the other
taxa.
\subsection{Remarks on the cost definition}
When $k=0$ in Eq.~\ref{eq:fru} our procedure is equivalent to the
minimization of the Pauplin's length (the proof of this statement will
be discussed in the Appendix IA). On the other hand, if the
Pauplin weights $2^{-t(a,\alpha)}$, $2^{-t(b,\beta)}$,
$2^{-t(c,\gamma)}$, $2^{-t(d,\delta)}$ were absent, the difference in
local costs between two configurations would be equal to the variation
of a global cost, defined as $E= \sum_{(abcd)} f_{(ab :cd)}$. Here the
sum defining the cost, $\sum_{(abcd)}$, is running over {\em all} the
quartets of the tree and not only on the quartets compatible with the
sub-trees $A,B,C,D$. We will refer to our algorithm with this form of
the cost configuration as the {\em Normalized Quartets} method (NQ).
Conversely, when one takes the complete form of the local cost as
defined in Eq.~\ref{eq:quart-paup}, with $k>0$ in Eq.~\ref{eq:fru},
the local cost differences do not correspond to any global cost
difference (the proof of this statement will be discussed in Appendix
IB). It is however an open question whether a global functional can be
defined whose variation between each pair of configurations is
compatible with the sign of our local cost difference.
The complexity of a sweep of our algorithm (i.e., $N$ configurations
updates, where $N$ is the number of leaves in the tree), has a leading
term $O(N^4)$. The number of quartets to be considered when
updating all the edges of the tree in case of a perfectly balanced
tree reads:
\begin{equation}
{\cal N}= \frac{85}{5376}N^4-\frac{N^2}{3}+\frac{4}{7}N \,.
\end{equation}.
\noindent We show the results of numerical simulations for the running
time of the greedy version of our algorithm in Appendix IIB.
If considering the {\em Normalized Quartets} method, a na\"ive
minimization of the corresponding functional would lead to a
complexity of $O(N^5)$, while a Montecarlo sweep for a na\"ive
minimization of the Pauplin length is only $O(N^3)$. Despite the
$O(N^4)$ complexity of the SBiX algorithm, the greedy version has an
extremely low prefactor, making the algorithm suitable for trees with
a large number of taxa (see Appendix IIB).
\section{Results}
\subsection{Artificial phylogenies}
To test the performances of our algorithm, we consider artificially
generated phylogenies following one of the simplest evolutionary model
that takes into account both mutational events and horizontal
transfer. Each taxon is represented by a binary sequence of length
$l$. We start with one sequence, for instance the sequence with all
the bits equal to $0$. Then at each time step we perform the following
operations: (i) we randomly extract one of the already existing leaf
sequences, say $\bar{s}$; (ii) with probability $\tau$ a randomly
extracted portion of length $l/4$ of $\bar{s}$ is replaced with the
corresponding portion of another randomly chosen
sequence~\footnote{The choice of $l/4$ is arbitrary but does not bring
loss of generality. Choosing randomly in the interval $[0,l/4]$ the
length of the part of the sequence horizontally transferred does not
alter the qualitative behaviour of the reconstructing algorithms
(results reported in Appendix IIA). This last procedure is adopted
in the $4$-state two parameters Kimura model (see below).}; (iii)
$\bar{s}$ generates two clones as descendants; (iv) each site of the
two new sequences is independently flipped with probability $m/l$,
where $m$ is extracted from an exponential distribution with average
$\mu$ (average number of mutations per sequence per time step). To
ensure that at least on site mutates at each branching event, we
randomly choose a site to mutate if no site mutated. We iterate this
procedure until the desired number of taxa is obtained.
We here consider as distance between two taxa the {\em correct hamming
distance}~\cite{felsenstein2004}, defined as:
\begin{equation}
d_{corr}= -\frac{1}{2} \ln(1- 2 h),
\label{correct}
\end{equation}
\noindent where $h$ is the {\em hamming distance}, defined as the
fraction of sites in which the sequences differ\footnote{In all the
results reported in this paper we let the algorithm infer the
correct phylogeny by using the {\em correct hamming distance}
$d_{corr}$. Even though the defined correction has its theoretical
justification only in absence of horizontal gene transfer, we
checked (data not shown) that using the {\em hamming distance} $h$
all the considered algorithms show the same relative behavior as in
the reported results, but the absolute performances are remarkably
poorer.}.
Although the evolutionary model described above is a toy model for
describing evolution, it allows to control and to tune noise as well
as horizontal transfer events. We also test our algorithm on
phylogenies constructed following more realistic model of evolution,
such us the standard $4$-states two parameters Kimura
model~\cite{Kimura_1980}. In particular, we follow the same steps
described above for the $2$-state model, but we now consider sequences
of nucleotides, with an alphabet of $4$ letters, and different rates
of transitions ($\alpha$) and transversions ($2 \beta$). We consider
in this case as distance between two taxa the {\em correct hamming
distance} for the Jukes-Cantor model~\cite{felsenstein2004} that is
the limit of the Kimura model when $\alpha=\beta$, which reads:
\begin{equation}
d^{\mathrm{JC}}_{corr}= -\frac{3}{4} \ln(1- \frac{4}{3} h).
\label{correct_K2P}
\end{equation}
\subsection{Robinson-Foulds measure}
In order to assess the performances of the different algorithms to
reconstruct the true phylogeny, we consider the standard
Robinson-Foulds measure (RF)~\cite{Robinson_Foulds_1981}, which counts
the number of bipartitions on which the inferred tree differs from the
true one. A bipartition is a split of the leaves in two sets realized
through a cut of a tree edge. We recall that it exists a one-to-one
correspondence between the bipartitions of the tree and the set of its
edges, so that each tree is uniquely characterized by the set of
bipartitions it induces. Note that, since we consider only binary
trees, the number of true positive bipartitions equals the number of
false positive bipartitions and both are equal to the RF measure.
\subsection{Competing algorithms}
In order to assess the performances of our algorithm, both in its
simulated annealing and greedy version, we compare it with the
Neighbor-Joining (NJ)~\cite{NJ} and the FastME~\cite{fastme_2002}
algorithms. In addition we implemented the Pauplin's length
minimization (from now onwards referred as PAUPLIN) by making use of
our SBiX algorithm in its form with $k=0$ (see above the section about
the algorithm's description). This in order to directly investigate
the effectiveness of a non-greedy minimization of the Pauplin's length
in reconstructing trees. Finally we implemented the version of our
algorithm without the Pauplin's weights (NQ) (as discussed above). We
also show a comparison with the performances of a state-of-the-art
character based algorithm, MrBayes~\cite{MrBayes}.
\begin{figure}[bp]
\begin{center}
\includegraphics[width=0.80\textwidth]{fig3}
\end{center}
\caption{{\bf Performances comparison as a function of the mutation
rate.} Left: Robinson-Foulds distance between the reconstructed
and the true trees as a function of the mutation rate per site of
the generative evolutionary model. The horizontal transfer rate
$\tau$ is here kept $\tau=0$. We compare the performances of the
SBiX algorithm with $k=5$ with: Neighbor-Joining (NJ), FastME, the
Pauplin length minimization (PAUPLIN), and Normalized-Quartets
(NQ). Right: dependence of the SBiX algorithm on the parameter $k$.
The best performances are very stable in the range of $k$ between
$5$ and $10$. In both figures results are averaged over $100$
independent realizations for each reported mutation rate. The error
bars are standard errors. All the trees generated have $N=60$
leaves and the sequences have fixed length $l=1000$.}
\label{fig:rumandk}
\end{figure}
\begin{figure}[ht]
\centerline{\includegraphics[width=0.80\textwidth]{fig4}}
\caption{{\bf System-size dependence.} Behavior of the normalized
Robinson-Foulds distance for the SBiX algorithm and FastME for
different system sizes, i.e., different values $N$ of the number
of leaves. Here normalized means the Robinson-Foulds distance
divided by its maximal value $N-3$. In all the cases curves for
different values of $N$ collapse as a function of $\mu^2
\lambda(N)$ (see text for details) where $\lambda(N)$ is the
average distance between two leaves in a tree with $N$ leaves. In
both the analysis the horizontal transfer rate $\tau$ is kept
$\tau=0$. We use both the true value of $\lambda(N)$ in the
simulated phylogenies {\bf (a)}, and the value analytically
calculated in the case of perfectly balanced trees
$\lambda_{\mathrm{theo}}(N)$ {\bf (b)}. In the inset of {\bf (b)}
we report the behavior of $\lambda$ and $\lambda_{\mathrm{theo}}$
as a function of tree size $N$. The experimental values
$\lambda(N)$ are systematically larger than
$\lambda_{\mathrm{theo}}$ putting in evidence a slight deviation
of the generated trees from a perfectly balance condition.}
\label{fig:collapse}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.80\textwidth]{fig5}
\end{center}
\caption{{\bf Performances comparison as a function of the
horizontal transfer rate.} Robinson-Foulds measure for the
reconstructed trees as a function of the horizontal transfer rate
of the generative model. We compare the performances of our SBiX
algorithm (with $k=5$) with those of the following ones:
Neighbor-Joining (NJ), FastME, the Pauplin length minimization
algorithm (PAUPLIN) and the Normalized-Quartet (NQ)
algorithm. Results are averages over $100$ independent
realizations for each reported horizontal transfer rate. The error bars are
standard errors. All the trees generated have $N=60$ leaves. {\bf
(a)}: mutation rate per site $\mu/l=0.003$ while the sequences
have fixed length $l=10000$. {\bf (b)}: mutation rate per site of
$\mu/l=0.14$ while the sequences have fixed length $l=1000$.}
\label{fig:res_hor}
\end{figure}
\subsection{Performances of the different algorithms}
In this section we compare the performances of all the considered
algorithms as a function of the mutation rate and of the horizontal
transfer rate in the underlying evolutionary process described in
the previous subsection.
In fig.~\ref{fig:rumandk} we show the Robinson-Foulds curves for
different algorithms (Left) as a function of the mutation rate for a
fixed tree size ($N=60$) and $k=5$. In the whole range of values of
the mutation rate all the versions of our algorithm (PAUPLIN, NQ and
SBiX) outperforms both NJ and FastME. In particular SBiX outperforms
all the other algorithms. Differences between the global minimization
of the Pauplin's length (PAUPLIN) and FastME arise for very high
mutation rates, where the global Pauplin's length minimization
outperforms FastME. This is probably due to the fact that FastME is
time optimized and therefore less able of our Stochastic Local Search
scheme to find the global minimum of the functional for very high
mutation rates (for a discussion on the consistency of greedy local
moves based on the balanced minimum evolution principle
see~\cite{Bordewich_2009}). In fig.~\ref{fig:rumandk} we report the
dependence of the SBiX performances (Right) on the value of the
parameter $k$. It is evident the existence of a range of values
between $k=5$ and $k=10$ where the algorithm features the best results
in a stable way. In the following, unless otherwise stated, we will
consider the $k=5$ case.
Up to this point we have characterized the performance of the
different algorithms for fixed value of the number of leaves, i.e.,
for a given system size. We are now interested in the robustness of
our results at different number $N$ of leaves of the tree. Defining
$\lambda(N)$ as the mean topological distance between any couple of
leaves, we empirically found that each algorithm can be characterized
by a reference curve obtained by plotting the normalized
Robinson-Foulds distance as a function of $\mu^2 \lambda(N)$. This
scaling can be understood by considering that the relevant quantity
for the tree reconstruction is not the bare mutation rate but the
amount of back mutation events, that can be estimated as $\mu^2
\lambda(N)$.
The scaling of the normalized Robinson-Foulds distance when
reconstructing trees of different sizes is shown in
fig.~\ref{fig:collapse}, where for the sake of clarity we only report
the curves for FastME and the SBiX algorithm. Each of the two
algorithms is characterized by a different reference curve, and the
interesting point here is that the SBiX algorithm is systematically
better than FastME at all mutation rates and sizes. We use both the
measured value of $\lambda(N)$ in the simulated phylogenies
(fig.~\ref{fig:collapse}-(a)), and the value analytically calculated in
the case of perfectly balanced rooted trees
(fig.~\ref{fig:collapse}-(b)) that reads:
\begin{equation}
\lambda_{\mathrm{theo}}(N)=\frac{2 N (\log_2{N}+1)-4 N +2}{N-1}.
\end{equation}
We now consider the ability of the different algorithms in recovering
the correct tree in the presence of horizontal transfer events. The
Robinson-Foulds curves at a fixed tree size are shown as a function of
the probability $\tau$ for each sequence to receive a borrowing (with
the mechanism defined above). In fig.~\ref{fig:res_hor}-(a) results at low
mutation rate are reported, when deviation from additivity is almost
exclusively due to the horizontal transfer events. In
fig.~\ref{fig:res_hor}-(b), instead, results are reported at high
mutation rate, when both back mutations and the horizontal transfer
events are responsible for deviations from additivity. For horizontal
transfer events co-occurring with a low mutation rate, our algorithm is
the most suitable to recover the correct tree, at each rate $\tau$.
The {\em Normalized Quartets} method, conversely, shows a performance
lower than that of NJ. When a high mutation rate is considered
jointly with horizontal transfer events, our SBiX algorithm
significantly outperforms the others when the probability $\tau$ of
horizontal transfer is not too high, while in the high $\tau$ region
the performance of our algorithm becomes comparable to the
minimization of the Pauplin's length (PAUPLIN).
It is interesting to compare the performance of the different
algorithms in the case of a more realistic data generator. In
fig.~\ref{fig:res_K2P} we show the analogous of fig.~\ref{fig:rumandk}
for the two parameter Kimura model: we present the Robinson-Foulds
curves obtained for different algorithms (NJ, FastMe, SBiX greedy from
FastME, SBiX greedy from NJ, SBiX simulated annealing, and MrBayes)
for $N=60$ as a function of different per-site mutation rates $\mu/l$.
The details of the simulation for MrBayes are presented in Appendix
IIB, together with a comment on the computational complexity of the
whole simulation.
The first evidence is that SBiX, in its different flavors, clearly
outperforms the other two distance based algorithms (NJ, and
FastME). The improvement is even more evident compared with the binary
characters case displayed in fig~\ref{fig:rumandk}. The three variants
of SBiX perform similarly in the low mutation rate regime ($\mu/l
<0.2$), and the results show a moderate improvement of the simulated
annealing version only for high level of mutation rate, whereas the
difference between the two greedy versions of SBiX seem to be
statistically irrelevant in all the mutation rate interval analyzed.
The comparison with MrBayes is somehow surprising. After a very low
mutation rate regime ($\mu/l \leq 0.024$), where all algorithms show
analogous accuracies, we enter in an intermediate regime ($ 0.048 \leq
\mu/l \leq 0.24$) where MrBayes outperforms SBiX. One has to notice
that the improvement of SBiX vs. both NJ and FastME is more evident
compared with that of MrBayes vs. SBiX. In the high mutation rate
regime ($\mu/l > 0.24$) MrBayes shows results compatible with two
greedy versions of SBiX, while it is slightly outperformed by the
simulated annealing version of SBiX, only at the larger mutation rate
available ($\mu/l=0.4$).
One should be careful to draw conclusions about the performance of
MrBayes in the high mutation rate regime where the issue of reaching
the Monte Carlo Markov Chain steady state becomes really problematic.
We can not exclude that, upon doubling the simulation times in this
regime, one could improve MrBayes results. In this regime, as more
thoroughly discussed in Appendix IIB, the computational time for a
single sample is already of the order of five hours, whereas for both
greedy versions of SBiX is of the order of $10^{-2}$ seconds, and for
the simulated annealing version is around 5 minutes.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.40\textwidth]{fig6}
\end{center}
\caption{{\bf Performances comparison as a function of the per-site
mutations rate.}
The evolutionary model used to generate
phylogenies is the Kimura two parameters model with transition
rate $\alpha=0.4$ and transversion rate $2 \beta=0.6$.
Robinson-Foulds measure for the reconstructed trees as a function of
the mutation rate per site. We compare the performances of our SBiX
algorithm (with $k=5$) with those of the following ones:
Neighbor-Joining (NJ), FastME, SBiX greedy from NJ, SBiX greedy from
FastME, SBiX simulated annealing, and MrBayes. Results are averages
over $100$ independent realizations for each reported mutation rate.
The error bars are standard errors. All the trees generated have
$N=60$ leaves, and the sequence $l=1000$.}
\label{fig:res_K2P}
\end{figure}
\section{Discussion and conclusions}
In this paper we have introduced a new algorithmic scheme for
phylogeny reconstruction. Belonging to the family of Stochastic Local
Search algorithms, our scheme crucially exploits two known properties
of additive distance matrices, the {\em four-points condition} and the
so-called Pauplin's length. We proposed in particular a stochastic
scheme where the correct topology is inferred through a series of
swapping of the tree topology. When tested on artificially generated
phylogenies our algorithmic scheme significantly outperforms
state-of-art distance-based algorithms in cases of deviation from
additivity due to high rate of back mutations. A significant
improvement is also observed with respect to the state-of-art
algorithms in case of high rate of horizontal transfer.
Such good performances are due to the way we differentially weight the
different quartets contributions with a term inversely proportional to
their length, and thus to their probability to be affected by back
mutations. On the other hand further work is needed for a complete
theoretical understanding of the algorithm. In particular, despite
many attempts, we are, at present, unable to formulate the update
strategy in terms of a state functional. Beside the interest in
itself, this would open the way to analytic treatments as well as to
algorithmic optimization strategies possibly more efficient than the
Stochastic Local Search one.
As for the comparison of our algorithmic scheme with state-of-the art
algorithms it is fair to observe that SBiX features a definitely
larger computational complexity but, in practice, it is fast enough
for reconstructing phylogenies up to a few thousands of leaves.
Though SBiX outperforms all competitors also in presence of horizontal
transfers, the method is especially suited for dealing with
non-additivity originated by double mutations. The issue of horizontal
transfer is however central in many fields~\cite{Doolittle_2008}, and
we believe that formulating effective strategies to dealing with it,
considering both phylogenetic trees and networks, is an open challenge
for the next generation reconstruction algorithms and will be the aim
of further studies.
It is worth mentioning how the applicability of phylogenetic
algorithms has recently widened its scope. Many different fields have
arisen in the last few years where a correct reconstruction of
phylogenetic trees may reveal underlying relevant dynamical processes.
For instance phylodynamics is a new field at the crossroad of
immunodynamics, epidemiology and evolutionary biology, that explores
the diversity of epidemiological and phylogenetic patterns observed in
RNA viruses of vertebrates~\cite{grenfell_2004}; phylogeography is
the study of the historical processes that may be responsible for the
contemporary geographic distributions of individuals as well as of
languages or viruses~\cite{avise_2000}. In all these cases a strong
effort is being devoted to the collection of comprehensive datasets
and efficient and reliable algorithms are needed especially when
deviations from perfect phylogenies become relevant.
\section{Appendix I}
\label{app_I}
\subsection{The equivalence with the Pauplin length minimization}
We show here that our SBiX algorithm minimizes the Pauplin's length
when $k=0$ in the Eq.~\ref{eq:fru}. In order to see this, we
explicitly calculate the cost difference between, say, the
configurations $((A,B),(C,D))$ and $((A,C),(B,D))$. We note that the
sum on the considered quartets can be divided in three parts, in which
one of the three distances $D_1$, $D_2$ and $D_3$ is respectively
minimal (where, as already defined in the text, $D_1=d(a,b)+d(c,d)$,
$D_2=d(a,c)+d(b,d)$ and $D_3=d(a,d)+d(b,c)$). After a little algebra
one gets:
\begin{equation}
\Delta E = \sum \left( D_2-D_1 \right) 2^{-t(a,\alpha)-t(b,\beta)-t(c,\gamma)-t(d,\delta)},
\end{equation}
\noindent where the sum is again over the $a \in A, b \in B, c \in C,
d \in D$ and $\Delta E \equiv E_{((A,C),(B,D))}-E_{((A,B),(C,D))}$.
Making use of the relation:
\begin{equation}
\sum_i 2^{-t(i,r)}=1,
\end{equation}
\noindent and of the equivalences:
$2^{-t(a,b)}=2^{-t(a,\alpha)-t(b,\beta)}/2$ in the configuration
$((A,B),(C,D))$ and $2^{-t(a,b)}=2^{-t(a,\alpha)-t(b,\beta)}/4$ in the
configuration $((A,C),(B,D))$ (and the analogous relations for the
other pairs of taxa), it is easy to prove that it holds:
\begin{equation}
\Delta E= 4 \Delta L_P,
\end{equation}
\noindent where $L_P$ is the Pauplin's length and $\Delta L_P \equiv
L_{P,((A,C),(B,D))}-L_{P,((A,B),(C,D))}$ is the difference of the
Pauplin's length between the two configurations.
\subsection{Locality of the SBiX configuration cost}
We give here an argument to prove that differences in the local cost
of our SBiX method cannot be written as differences of a functional on
the whole tree. If this was the case, a functional could be defined as
$F(x)=F(x_0)+ \sum\Delta E_i$, where $F(x_0)$ is the value taken by
the functional in a reference configuration $x_0$, and $\Delta E_i$
are the cost differences along a path from $x_0$ to $x$. Moving in the
space of tree's topologies, we should obtain the same value of $F$
each time we visit the same topology, i.e., the difference of cost
between two states does not depend on the path. This is not the case,
as we explicitly checked, when the cost is defined as in
Eq.~\ref{eq:quart-paup} and $k \neq 0$.
\section{Appendix II}
\label{app_II}
\subsection{Discussion on the horizontal gene transfer modeling}
In order to assure that the procedure of keeping the length of the
horizontal transfer fixed to $l/4$, as for the results shown in the
main text, does not alter the qualitative behaviour of the compared
performances, we here show results where this restriction is relaxed.
The Figure~\ref{fig:hor_nofix} is the analogous of Figure 5 (a) in the
main text, with the difference that here the proportion of horizontal
gene transfer is allowed to randomly vary in the interval [0,1/4].
\begin{figure}[htb]
\centerline{\includegraphics[width=0.40\textwidth]{transfer_nofix}}
\caption{{\bf Performances comparison as a function of the horizontal
transfer rate.} Robinson-Foulds measure for the reconstructed
trees as a function of the horizontal transfer rate of the
generative model. We compare the performances of our SBiX algorithm
(with $k=5$) with those of the following ones: Neighbor-Joining
(NJ), FastME, the Pauplin length minimization algorithm (PAUPLIN)
and the Normalized-Quartet (NQ) algorithm. Results are averages over
$100$ independent realizations for each reported horizontal transfer
rate The error bars are standard errors. All the trees generated
have $N=60$ leaves, the mutation rate per site is kept fixed to
$\mu/l=0.003$ while the sequences have fixed length $l=10000$.}
\label{fig:hor_nofix}
\end{figure}
\subsection{Discussion on the algorithmic computational complexity}
\subsubsection{Generalities}
Given the encouraging performance of the SBiX compared to other
distance based algorithm we have run a detailed comparison with
character based methods such as MrBayes \cite{MrBayes}. For all our
comparison we used {\tt MrBayes v3.1.2} on a standard linux platform,
Intel Core(TM)2 CPU $@$ $2.40 GHz$ and a cache
size of $4096 KB$. In the simulation we kept the default setting of
the program ( {\tt Nruns = 2, Datatype = DNA, Nucmodel = 4by4, Nst =
1, Covarion = No, \# States = 4, Rates = Equal} ) and we did not
specify any details of the model generator. The only parameters we
tuned were: (i) the number {\tt Ngen} of tree generated by the Monte
Carlo Markov Chain (MCMC), (ii) the number {\tt Burnin} of tree
configurations discarded in the analysis. Parameter fixing depends
critically on data, and in particular, the more the generative model
departs from additivity, the longer the MCMC must be run in order to
achieve convergence and obtaining a fair sampling of the likelihood
landscape of the problem. A useful proxy for MCMC convergence is the
average standard deviation of the partition split frequencies in the
{\tt Nrun} (in our case {\tt Nrun=2}) MCMCs. One expects that this
quantity, after a transient, approaches to zero for a sufficiently
long number of iterations, reflecting the fact that the two Markov
Chains become increasingly similar, sampling the same subspace of the
problem.
\subsubsection{Computational complexity}
Assessing precisely the computational complexity of MrBayes goes
definitely beyond the scope of our work. We performed however a
detailed analysis on a particular case for giving a fair comparison of
the computational cost of the two strategies.
As a benchmark we used the two parameters Kimura generator (with
transition rate $\alpha=0.4$ and transversion rate $2 \beta=0.6$)
discussed in the paper. We generated 100 realization of phylogenies,
with sequence length $l=1000$, with a per-site mutation rate $\mu/l =
0.1$, at different values of the number of leaves ranging from $N=15$
to $N=480$.
As a lower bound on the computational cost of MrBayes we decided
to stop the simulation as soon as the average standard deviation of
the partition split frequencies became lower than $0.05$. We intend
this as a lower bound on the true computational cost since this
can be considered as a good estimation only of the length of the
transient regime. To ensure a fair {\em equilibrium sampling} one
should check that the 2 Markov Chains maintains split values lower
than our threshold for a number of generations large enough. For each
of our samples we recorded the running time in seconds and we averaged
over the 100 sample realizations.
\begin{table}[htb]
\begin{tabular}{|*{8}{c|}}
\hline
$N$ & $\overline{t}_{\mathrm{mb}}$ & var$(t)_{\mathrm{mb}}$& samples
(mb) & $\overline{t}_{\mathrm{SBiX}}$ & var$(t)_{\mathrm{SBiX}}$ &
samples (SBiX) & ratio $\overline{t}_{\mathrm{mb}} /\overline{t}_{\mathrm{SBiX}}$ \\
\hline
&&& &\\[-1.3em]
\hline
15 & 11 & 1 & 50 & NA & NA & NA & NA\\[-0.4em]
30 & 60 & 3 & 50 & 0.004 & 0.04 & 100 & $1.5 \times 10^4$\\[-0.4em]
60 & 297 & 28 & 50 & 0.04 & 0.01 & 100 & $7.4 \times 10^3$\\[-0.4em]
120 & 2277 & 258 & 50 & 0.64 & 0.02 & 100 & $3.5 \times 10^3$ \\[-0.4em]
240 & 17106 & 692 & 10 & 10.72 & 0.03 & 100 &$1.5 \times 10^3$ \\[-0.4em]
480 & NA & NA & NA &200.67 & 0.03 & 100 & NA\\[-0.4em]
960 & NA & NA & NA &5096.7 & 0.3 & 100 & NA\\
\hline
\end{tabular}
\protect\caption{In this table we display the number of leaves $N$,
the average running time in seconds $\overline{t}$, its variance,
and the number of samples, both for MrBayes and SBiX algorithms.
The last column reports the ratio between the average running times
of MrBayes and SBiX.}
\label{tab:times}
\end{table}
On the same dataset we run the greedy version of SBiX starting from
the configuration predicted by Neighbor Joining, and recorded the
running time in seconds. In Table~\ref{tab:times} we displays the
comparison of the two algorithms. We did not display results for SBiX
at $N=15$, the execution time being so tiny (order of milliseconds)
that the measure would be strongly unreliable and basically dominated
by input-output operating system tasks. We also do not present results
for MrBayes at $N=480$, since each run would require a time larger
than 32 hours.
\begin{figure}[htb]
\centerline{\includegraphics[width=0.55\textwidth]{plot_time_sbix_vs_mb}}
\caption{Plot of the average time of computation per sample (in
seconds) as a function of the number of leaves of the phylogeny. All
points represent an average over the respective number of samples as
indicated in Table~\ref{tab:times}}
\label{fig:sbix_vs_mb}
\end{figure}
In Figure~\ref{fig:sbix_vs_mb} we plot the average running time per
sample as a function of $N$ for both algorithm. We fitted MrBayes
result with a simple power-law $ a + b x^c $ obtaining as best fit
$a=4(1)$ , $b=0.003(1)$, $c=2.8(1)$. The SBiX curve can be fitted with
a simple fourth power behavior $ A x ^4 $, and the best fit value is
$A = 5.4(4) \cdot 10^{-9}$. Such a small prefactor allows SBiX to be
faster than MrBayes to $N \sim 10^5$, despite the larger exponent (4
vs. 2.9) . Note that: (i) the running times at the crossing point are
of the order of the thousand of years, and (ii) hardly a $10^5$ leaves
genealogy could be of any interest.
\subsubsection{Computational cost as a function of the mutation rate}
In the previous section we outlined the analysis of the computational
complexity as a function of the number of taxa in the genealogy. In
the paper we presented the accuracy of the different algorithms fixing
the number of taxa ($N=60$) and varying the mutation rate per site
$\mu/l$. While the running time of other algorithms (NJ, FastME, SBiX
greedy) do not depend on the mutation rate, the running time required
by MrBayes depends critically on it. In order to obtain a fair
sampling of the space of trees we have run a number of preliminary
simulations in order to have a quantitative estimate of the
convergence time of the algorithm (the so-called {\em burn-in}
time). Then we considered a measuring time either $50000$ or $100000$
Monte Carlo steps depending on the mutation rate (and on the burn-in
time). In Table~\ref{tab:mc} we display the burn-in and measure times
for the different values of the mutation rate. For each of the point
we analyzed 100 different samples. The convergence time in the high
mutation rate regime ($\mu/l >= 0.2$) might be still too small, and
for some of the samples the MCMC might not have reached
convergence. In order to guarantee convergence for all samples one
should double the burn-in time, requiring a time too long for our
computational resources.
\begin{table}[!htb]
\begin{tabular}{|*{4}{c|}}
\hline
$\mu/l$ & burn-in & measure & sec. \\
\hline
&&&\\[-1.3em]
\hline
0.003 & 50000 & 50000 & 1300 \\[-0.40em]
0.006 & 50000 & 50000 & 1300 \\[-0.40em]
0.012 & 50000 & 50000 & 1300 \\[-0.40em]
0.024 & 50000 & 50000 & 1300 \\[-0.40em]
0.048 & 50000 & 50000 & 1300 \\[-0.40em]
0.070 & 50000 & 50000 & 1300 \\[-0.40em]
0.100 & 100000 & 100000 & 2600 \\[-0.40em]
0.140 & 100000 & 100000 & 2600 \\[-0.40em]
0.200 & 300000 & 100000 & 5200 \\[-0.40em]
0.280 & 300000 & 100000 & 5200 \\[-0.40em]
0.350 & 500000 & 100000 & 7800 \\[-0.40em]
0.400 & 700000 & 100000 & 12400 \\
\hline
\end{tabular}
\protect\caption{In this table we display, for the different value of
the per-site mutation rate $\mu/l$, the burn-in and the measure time in
elementary Monte Carlo steps unit, together with an average estimate
of running time in seconds. }
\label{tab:mc}
\end{table}
\section{Acknowledgments}
The authors wish to thank Simone Pompei, Francesca Colaiori, Ciro
Cattuto and Dami\'an H. Zanette for many interesting discussions and
suggestions. This research has been partly supported by the TAGora and
ATACD projects funded by the European Commission under the contracts
IST-34721 and 043415.
|
\section{Introduction}
\meta{General points:
\begin{enumerate}
\item should we discuss the expected regularity of the solution? For
instance, if $B$ is discontinuous, what regularity do we get for
$\rho$? See the marginal comment below.
\begin{G}
I added a comment just before Theorem 1; instead of the
regularization effect in $L^p$ (which, at least locally, always holds by
[7]), I just observed that even the invariant measure could be
non-smooth.
\end{G}
\item Should we note that the comparison property
(Theorem~\ref{thm:crucial}) is proved for smooth data, but should
hold much more generally?
{I thought that $C^{2,1}_{\mathrm b}$ would be a natural assumption for
applying classical maximum principle; what kind of regularity do
you have in mind?}\\
Anything that one can approximate pointwise by solutions of the smooth problem.
I added a suggestion in the text below.
\mathscr{G}{I just removed ``unbounded'' referred to solutions, since I am not sure
that you can apply the method and prove uniqueness in that
case. My doubts come from the fact that when you argue by approximation, you get a
uniqueness result just for the class of \emph{approximable
solution}, which could be a proper subclass of all the
reasonable solutions of the equation}
\item
\mathscr{G}{I restored variable-doubling}
\end{enumerate}}
The aim of this paper is to obtain new uniqueness and contractivity results for
nonnegative measure-valued solutions to the Fokker-Planck equation
\begin{equation} \label{eq:1}
\partial_t \rho-\Delta \rho-\div(\rho B)=0,\quad \rho\restr{t=0}=\rho_0,
\end{equation}
where $B:{\R^d}\to{\R^d}$ is a Borel $\lambda$-monotone operator, $\lambda\in {\mathbb R}$, i.e.
\begin{equation}
\label{eq:5}
\langle B(x)-B(y),x-y\rangle\ge \lambda \,\big|x-y\big|^2\quad\text{for every }x,y\in {\R^d}.
\end{equation}
Here we consider a weakly continuous family of probability measures
$(\mu_t)_{t\ge0} \subset \mathcal{P}({\R^d})$ satisfying
the equation \eqref{eq:1} in the sense of distributions
\begin{equation} \label{eq:2}
\int_0^{+\infty}\int_{\R^d} \Big(\partial_t\zeta+\Delta\zeta-B\cdot
\nabla \zeta\Big)\,\d\mu_t\d t=0\quad\forall\, \zeta\in
C^\infty_{\mathrm c}({\R^d}\times (0,+\infty)),
\end{equation}
with the initial datum $\mu_0$.
Equations of this type are the subject of several papers by
\textsc{Bogachev}, \textsc{Da Prato}, \textsc{Krylov},
\textsc{R\"ockner}, and \textsc{Stannat},
{who consider a very general situation where the Laplacian is replaced by a second order elliptic operator with variable coefficients and $B$ is locally bounded.}
Existence of solutions has been proved by \cite[Cor.~3.3]{Bogachev-DaPrato-Rockner08},
uniqueness has been considered in
\cite{Bogachev-DaPrato-Rockner-Stannat07} under general
growth-coercivity conditions on $B$,
and regularity has been investigated by \cite{Bogachev-Krylov-Rockner01}: in particular, it has been shown that
$\mu_t$ is absolutely continuous {with respect to the Lebesgue measure} for $\Leb{1}$-a.e.~$t$.
When $B$ is Lipschitz continuous, uniqueness can be obtained by
standard duality arguments, see
e.g.~\cite[Sec.~3]{Ambrosio-Savare-Zambotti09}.
Here we want to obtain a more precise stability estimate on the
solutions of \eqref{eq:1}, only assuming
monotonicity of $B$ without any growth condition.
To achieve this aim, we adopt the point of view of optimal
transportation.
\paragraph{The Wasserstein approach to Fokker-Planck equation in the
gradient case.}
When $B$ is the gradient of a $\lambda$-convex function $V:{\R^d}\to{\mathbb R}$
then \eqref{eq:1} can be considered as the \emph{gradient flow} of
the perturbed entropy functional
\begin{equation}
\label{eq:6}
\mathscr{H}(\rho):=\int_{\R^d} u(x)\log u(x)\,\d x+\int_{\R^d}
V(x)\,\d\rho(x)\quad
\rho=u\Leb d
\end{equation}
in the space $\mathcal{P}_2({\R^d})$ of probability measures with finite
quadratic moments endowed with the so called $L^2$-Kantorovich-Rubinstein-Wasserstein
distance $W_2(\cdot,\cdot)$. This distance can be defined
by
\begin{equation}
\label{eq:7}
\begin{aligned}
W_2(\rho^1,\rho^2):=\min\Big\{&\int_{{\R^d}\times{\R^d}}\big|x_1-x_2\big|^2\,\d{\mbox{\boldmath$\rho$}}(x_1,x_2):{\mbox{\boldmath$\rho$}}\in
\mathcal{P}({\R^d}\times{\R^d})\\&
{\mbox{\boldmath$\rho$}}\text{ is a coupling between $\rho^1$ and
$\rho^2$
\Big\}
\end{aligned}
\end{equation}
in terms of couplings, i.e.\ measures ${\mbox{\boldmath$\rho$}}$ in the product space
${\R^d}\times{\R^d}$ whose marginals are $\rho^1$ and $\rho^2$
respectively, so that ${\mbox{\boldmath$\rho$}}(E\times{\R^d})=\rho^1(E)$ and
${\mbox{\boldmath$\rho$}}({\R^d}\times E)=\rho^2(E)$ for every Borel subset $E\subset
{\R^d}$. It is possible to prove that \emph{optimal couplings}
realizing the minimum in \eqref{eq:7} always exist.
This remarkable interpretation found in
\cite{Jordan-Kinderlehrer-Otto98} gave rise to a series of studies on the
relationships between certain classes of diffusion equations and
distances between probability measures induced by optimal transport
problems (see e.g. the general overviews of
\cite{Villani03,Ambrosio-Gigli-Savare08,Villani09}).
One of the strengths of this approach is a new geometric insight
(developed in \cite{Otto01}) in
the evolution process: in the case of \eqref{eq:1} the
$\lambda$-convexity of the potential $V$ reflects a
$\lambda$-convexity property (also called \emph{displacement
convexity}) of the functional $\mathscr{H}$ along the geodesics of
$\mathcal{P}_2({\R^d})$. This nice feature, discovered by \cite{McCann97},
suggests that one can adapt some typical basic existence,
approximation, and regularity results for
gradient flows of convex functionals in Euclidean spaces or
Riemannian manifolds
to the measure-theoretic setting of $\mathcal{P}_2({\R^d})$. This program has
been carried out (see e.g.\ \cite{Ambrosio-Gigli-Savare08}) and,
among the most interesting estimates, it provides the
$\lambda$-contraction property
\begin{equation}
\label{eq:8}
W_2(\rho^1_t,\rho^2_t)\le {\mathrm e}^{-\lambda
t}\,W_2(\rho_0^1,\rho_0^2)\qquad\text{for every }t\ge 0,
\end{equation}
where $\rho^i_t$, $i=1,2$, are the solutions to \eqref{eq:1}
starting from the initial data $\rho_0^i\in \mathcal{P}_2({\R^d})$.
\paragraph{Two strategies for the derivation of the
contraction estimate \eqref{eq:8} in the gradient case.}
In order to prove \eqref{eq:8} in the gradient case $B=\nabla V$, essentially
two basic strategies have been proposed:
\begin{enumerate}
\item A first approach, developed by
\cite{Carrillo-McCann-Villani06} for smooth evolutions and by
\cite{Ambrosio-Gigli-Savare08} in a measure-theoretic setting,
starts from equation \eqref{eq:1} written in the form
\begin{equation}
\label{eq:10}
\partial_t\rho+\div(\rho{\mbox{\boldmath$v$}})=0,\quad {\mbox{\boldmath$v$}}=-\Big(\frac{\nabla
u}u+\nabla V\Big),\quad \rho=u\,\Leb d,
\end{equation}
and it is based on two ingredients: the first one is the formula which evaluates the
derivative of the squared Wasserstein distance from a fixed
measure $\sigma$ along {the (absolutely continuous)} curve $\rho$
in $\mathcal{P}_2({\R^d})$
\begin{equation}
\label{eq:11}
\frac\d{\d t}\frac 12
W_2^2(\rho_t,\sigma)=\int_{{\R^d}\times{\R^d}}\langle
{\mbox{\boldmath$v$}}_t(x),y-x\rangle \,\d{\mbox{\boldmath$\rho$}}_t(x,y)\quad \text{for $\Leb 1$-a.e.\ $t>0$}
\end{equation}
where ${\mbox{\boldmath$\rho$}}_t$ is an optimal coupling between $\rho_t$ and
$\sigma$.
The second ingredient is the ``subgradient'' property of the vector
field ${\mbox{\boldmath$v$}}_t$ given by \eqref{eq:10}, related to the displacement
convexity of $\mathscr{H}$: in the case
$\lambda=0$ it reads as
\begin{equation}
\label{eq:12}
\int_{{\R^d}\times{\R^d}}\langle
{\mbox{\boldmath$v$}}_t(x),y-x\rangle \,\d{\mbox{\boldmath$\rho$}}_t(x,y)\le
\mathscr{H}(\sigma)-\mathscr{H}(\rho_t)\quad
\text{if }{\mbox{\boldmath$v$}}_t=-\Big(\frac{\nabla u_t}{u_t}+\nabla V\Big).
\end{equation}
Combination of \eqref{eq:11} and \eqref{eq:12} yields the so called
Evolution Variational Inequality
\begin{equation}
\label{eq:13}
\frac\d{\d t}\frac 12
W_2^2(\rho_t,\sigma)\le \mathscr{H}(\sigma)-\mathscr{H}(\rho_t)\quad
\text{for every }\sigma\in \mathcal{P}_2({\R^d})
\end{equation}
which easily yields \eqref{eq:8} for $\lambda=0$ by a variable-doubling\
argument (see \cite[Theorem 11.1.4]{Ambrosio-Gigli-Savare08}).
The main technical point here is that \eqref{eq:12} requires
${\mbox{\boldmath$v$}}_t\in L^2(\rho_t)$ and \eqref{eq:11} holds
if
{for every $0<t_0<t_1<+\infty$ }
\begin{equation}
\label{eq:14}
\int_{t_0}^{t_1} \int_{\R^d} |{\mbox{\boldmath$v$}}_t|^2\,\d\rho_t\,\d t= \int_{t_0}^{t_1} \int_{\R^d}
\left|\frac{\nabla u_t}{u_t}+\nabla V\right|^2\,\d\rho_t\,\d
t<+\infty,
\end{equation}
which should be imposed (in a suitable distributional sense) as an
\emph{a priori} regularity assumption on the
solution of \eqref{eq:1}. {We do not know if solutions to
\eqref{eq:2} exhibit a similar regularization effect.}
A second, even more difficult
point prevents a simple extension of
\eqref{eq:13} to the general non-gradient case: it is the lack of a
potential $V$ and therefore of an entropy-like functional $\mathscr{H}$
satisfying
an inequality similar to \eqref{eq:12}.
\item
A second approach has been proposed by
\cite{Otto-Westdickenberg05} and further developed in \cite{Daneri-Savare08,Carrillo-Lisini-Savare-Slepcev10}: it is based on the
\textsc{Benamou-Brenier} \cite{Benamou-Brenier00} representation formula for the
Wasserstein distance
\begin{equation}
\label{eq:15}
\begin{aligned}
&W_2^2(\mu_0,\mu_1)= \inf\Big\{\int_0^1 \int_{\R^d}
|{\mbox{\boldmath$v$}}_t |^2\,\d \rho_t\,\d t\,:\\
&\quad
\partial_t\rho_t+\nabla\cdot(\rho_t\mathbf{\mbox{\boldmath$v$}}_t)=0\ \text{in
${\R^d}\times (0,1)$,}\quad \mu_0=\rho\restr{t=0},\quad
\mu_1=\rho\restr{t=1}\Big\}
\end{aligned}
\end{equation}
and on a careful analysis of the effect of the evolution semigroup
generated by the equation on curves in $\mathcal{P}_2({\R^d})$ and its
Riemannian tensor $\int_{\R^d} |{\mbox{\boldmath$v$}}|^2\,\d\rho$.
This technique involves various repeated differentiations and
works quite well if a nice semigroup preserving smoothness and
strict positivity of the densities has already
been defined. Once contraction has been proved on smooth initial
data, the evolution can be extended to more general ones but
it seems hard to extend the uniqueness result to cover a general
distributional solution to the equation.
\end{enumerate}
\paragraph{Main result of the paper: contraction estimates for distributional
solutions.}
Our purpose is twofold:
\begin{itemize}\item
First of all we want to find a new approach working directly on
measure-valued solutions to \eqref{eq:1} just satisfying the usual
distributional formulation \eqref{eq:2}.\\
\begin{G}
We note that in general \eqref{eq:1} does not exhibit the same
regularization effect of the heat equation. Even in the
gradient case $B=\nabla V$, there exist solutions $\rho_t$ to
\eqref{eq:2} which are not of class $C^1({\R^d})$ for every
$t\ge0$: take, e.g., the invariant measure $\rho_t\equiv
Z^{-1}{\mathrm e}^{-V}$ for a suitable convex function $V\not\in
C^1({\R^d})$ with ${\mathrm e}^{-V}\in L^1({\R^d})$. Moreover,
distributional solutions are easily obtained by approximation
arguments, as regularization or splitting methods, and they
should be better suited to deal with the infinite dimensional
case, as in \cite{Ambrosio-Savare-Zambotti09}: a stability
result for such a weak class of solutions should be useful in
these cases.
\end{G}
\item Second, we want to cover the case of an arbitrary monotone field
$B$, without any growth restriction, and to extend contraction
estimates to more general transportation costs.
\end{itemize}
To this aim, let us first introduce the general cost functional
\begin{equation} \label{cost}
\begin{aligned}
\Cost h{\mu^1}{\mu^2} := \inf\Big\{&\int_{{\R^d}\times {\R^d}}h(|x_1-x_2|)\,\d\hbox{\boldmath$\mu$}(x_1,x_2):
\hbox{\boldmath$\mu$} \in \mathcal{P}({\R^d}\times{\R^d}), \\&
{\mbox{\boldmath$\rho$}}\text{ is a coupling between $\rho^1$ and
$\rho^2$}\Big\}.
\end{aligned}
\end{equation}
Throughout this paper we assume that
\begin{quote}
$h:[0,+\infty)\to [0,+\infty)$ is a continuous and \textbf{non-decreasing} function with $h(0)=0$.
\end{quote}
Among the possible interesting choices of $h$, the case
$h(r):=r^p$ is associated to the family of $L^p$ Wasserstein
distances (whose $L^2$-version has been introduced in
\eqref{eq:7})
on the space $\mathcal{P}_p({\R^d})$ of all the probability measures with
moment of order $p$.
When $h$ is a bounded concave function {satisfying $h(r)>0$ if
$r>0$, $d(x,y):= h(|x-y|)$
is a bounded and complete distance function on ${\R^d}$ inducing the usual
euclidean topology so that }
$\Cost h\cdot\cdot$ is a
complete metric on the space $\mathcal{P}({\R^d})$ whose topology coincides
with the usual weak one
(see, e.g., \cite[Proposition
7.1.5]{Ambrosio-Gigli-Savare08}).
Since we are not assuming any homogeneity on the general cost
function $h$, its rescaled versions
\begin{equation}
\label{eq:16}
h_s(r):=h(r\,{\mathrm e}^s)\quad s\in {\mathbb R},\ r\ge0
\end{equation}
will be useful.
Let us now state our main result:
\begin{teor} \label{teor:main}
If $\rho^1,\rho^2$ are two distributional solutions to \eqref{eq:2}
satisfying the sum\-mability condition
\begin{equation}
\label{Cond}
\int_{t_0}^{t_1}\int_{\R^d} |B(x)-\lambda \, x|\,\d\mu_t(x)\,\d t< +\infty\quad\text{for every }0<t_0<t_1<+\infty,
\end{equation}
then they satisfy
\begin{equation} \label{++}
\Cost {h_{\lambda t}}{\mu^1_t}{\mu^2_t} \le \Cost
h{\mu^1_0}{\mu^2_0}\quad\text{for every }t\ge0.
\end{equation}
In particular, if $\rho^1_0=\rho^2_0$ then $\rho^1$ and $\rho^2$
coincide for every time $t\ge0$.
\end{teor}
Let us make explicit some consequences of \eqref{++} according to
the different signs of $\lambda$ and the behaviour of $h$ near $0$ and $+\infty$:
\begin{cor} \label{cor:main}
Let $\rho^1,\rho^2$ be two distributional solutions to \eqref{eq:2}
satisfying \eqref{Cond}.
\begin{enumerate}[a)]
\item If $B$ is monotone, i.e.\ $\lambda \ge 0$, then
$$\Cost h{\mu^1_t}{\mu^2_t} \le \Cost h{\mu^1_0}{\mu^2_0}.$$
\item If $B$ is $\lambda$-monotone with $\lambda > 0$ and $h$
satisfies for some exponent $p>0$
\begin{equation}
\label{eq:9}
h(\alpha\, r)\ge \alpha^p h(r)\quad \text{for every $\alpha\ge 1$
and $r\ge0$}
\end{equation}
then
$$\Cost h{\mu^1_t}{\mu^2_t} \le {\mathrm e}^{-p\lambda \,t}\,\Cost h{\mu^1_0}{\mu^2_0}.$$
\item If $B$ is $\lambda$-monotone with $\lambda < 0$ and $h$
satisfies
for some exponent $p>0$
\begin{displaymath}
h(\alpha\, r)\ge \alpha^p h(r)\quad \text{for every $\alpha\le 1$
and $r\ge0$}
\end{displaymath}
then
$$\Cost h{\mu^1_t}{\mu^2_t} \le {\mathrm e}^{-p\lambda \,t}\,\Cost
h{\mu^1_0}{\mu^2_0}.$$
\end{enumerate}
In the particular case of the Wasserstein distance $W_p$, $p\ge1$, we have
\begin{equation} \label{8}
W_p (\mu^1_t,\mu^2_t ) \le {\mathrm e}^{-\lambda t}\, W_p (\mu^1_0,\mu^2_0 ).
\end{equation}
\end{cor}
Theorem \ref{teor:main} has a simple application to invariant
measures $\rho_\infty\in \mathcal{P}({\R^d})$, which are stationary solutions of
\eqref{eq:2} and therefore satisfy
\begin{equation}
\label{eq:66}
\int_{\R^d} \Big(\Delta\zeta-B\cdot
\nabla \zeta\Big)\,\d\mu_\infty=0\quad\forall\, \zeta\in
C^\infty_{\mathrm c}({\R^d}).
\end{equation}
\begin{cor}[Strongly monotone operators and invariant measures]
Let us suppose that $B$ is strongly monotone, i.e.\
$\lambda>0$. Then equation \eqref{eq:66}
has at most one solution $\rho_\infty\in \mathcal{P}({\R^d})$ satisfying the
integrability condition
\begin{equation}
\label{eq:67}
\int_{\R^d} |Bx-\lambda x|\,\d\rho_\infty(x)<\infty.
\end{equation}
For each solution $\rho_t$ to \eqref{eq:2}-\eqref{Cond} and each
cost $h$ satisfying \eqref{eq:9} we have
\begin{equation}
\label{eq:68}
\Cost h{\rho_t}{\rho_\infty}\le {\mathrm e}^{-p\lambda \,(t-t_0)}\Cost h{\rho_{t_0}}{\rho_\infty}.
\end{equation}
\end{cor}
Note that in the case $\lambda>0$ condition \eqref{eq:67} is weaker
than
$B\in L^1(\rho_\infty;{\R^d})$.
\begin{remark}[An equivalent formulation of the contraction estimate]
\label{rem:equivalent}
We can give an equivalent version of \eqref{++} by keeping fixed the
cost but rescaling the measures. In fact, we can associate to the
solutions $\rho^1,\rho^2$ of \eqref{eq:2} their rescaled versions
$\tilde\rho^1,\tilde\rho^2$ defined by
\begin{equation}
\label{eq:50}
\tilde\rho^j(E):=\rho^j({\mathrm e}^{-\lambda t} \,E)\quad\text{for every
Borel set }E\subset {\R^d},\ j=1,2.
\end{equation}
Then $\tilde\rho^j$ is the push-forward of $\rho^j$ through the map
$x\mapsto {\mathrm e}^{\lambda t}x$ and satisfies the change-of-variables
formula
\begin{equation}
\label{eq:51}
\int_{\R^d} \zeta(y)\,\d\tilde\rho^j(y)=\int_{\R^d} \zeta({\mathrm e}^{\lambda
t}\,x)\,\d\rho^j(x)\quad
\text{for every }\zeta\in C_{\mathrm b}({\R^d}).
\end{equation}
Inequality \eqref{++} is then equivalent to
\begin{equation}
\label{eq:52}
\Cost h{\tilde\rho^1_t}{\tilde\rho^2_t}\le \Cost
h{\rho^1_0}{\rho^2_0}\quad
\text{for every }t>0.
\end{equation}
\end{remark}
\paragraph{Strategy of the proof: Kantorovich duality and a
variable-doubling\ technique.}
In order to prove Theorem \ref{teor:main} we develop a new strategy,
generalizing
\cite{Portegies-Peletier08}.
It relies on the well-known dual Kantorovich formulation~\cite{Villani03} of
the transportation cost \eqref{cost}:
\begin{equation} \label{dualintro}
\begin{aligned}
\Cost h{\mu^1}{\mu^2}=\sup\Big\{& \int_{\R^d} \phi^1\,\d\mu^1+
\int_{\R^d} \phi^2\,\d\mu^2:\\&\phi^1,\phi^2\in C_{\rm b}({\R^d}),\
\phi^1(x_1)+\phi^2(x_2)\le h(|x_1-x_2|)\Big\}.
\end{aligned}
\end{equation}
This formula reduces the estimate of the cost $\Cost h{\rho^1_T}{\rho^2_T}$ of two
solutions of \eqref{eq:1} at a certain final time $T$ to the estimate
of
\begin{equation}
\label{eq:18}
\Sigma(\phi^1,\phi^2;T):=\int_{\R^d} \phi^1\,\d\mu^1_T+
\int_{\R^d} \phi^2\,\d\mu^2_T
\end{equation}
for an arbitrary pair of functions $\phi^1,\phi^2$ satisfying the
constraint
\begin{equation}
\label{eq:17}
\phi^1(x_1)+\phi^2(x_2)\le h(|x_1-x_2|)\quad\text{for every
}x_1,x_2\in {\R^d}.
\end{equation}
Assuming for the sake of simplicity that $B$ is monotone, bounded
and smooth, we can obtain an estimate of $\Sigma(\phi^1,\phi^2;T)$ by
solving the final-value problem for the adjoint equation
\begin{equation}
\label{eq:19}
\partial_t \phi^i+\Delta\phi^i-B\cdot\nabla\phi^i=0\quad\text{in
}{\R^d}\times(0,T),\quad
\phi^i(\cdot,T):=\phi^i
\end{equation}
since the distributional formulation \eqref{eq:2} yields
\begin{equation}
\label{eq:20}
\Sigma(\phi^1_T,\phi^2_T;T)=\Sigma(\phi^1_0,
\phi^2_0;0)
\end{equation}
The following crucial result, based on a
``variable-doubling\ technique'', provides the final step, showing that
$\phi^1_0,\phi^2_0$ still satisfy the constraint
\eqref{eq:17} so that $\Sigma(\phi^1_0,
\phi^2_0;0)\le \Cost h{\rho^1_0}{\rho^2_0}$.
\begin{teor}
\label{thm:crucialintro}
If $\phi^1,\phi^2\in C^{2,1}_{\mathrm b}({\R^d}\times [0,T])$ are solutions
of \eqref{eq:19} in the case when $B$ is monotone, bounded and smooth,
such that
\begin{equation*}
\phi^1(x_1,T)+\phi^2(x_2,T)\le h(|x_1-x_2|)\quad\forall\,x_1,x_2\in {\R^d},
\end{equation*}
then
\begin{equation*}
\phi^1(x_1,0)+\phi^2(x_2,0)\le h(|x_1-x_2|)\quad\forall\,x_1,x_2\in {\R^d}.
\end{equation*}
\end{teor}
\begin{remark}
While we prove Theorem~\ref{thm:crucialintro} for bounded and smooth
drifts $B$, and solutions $\phi^{1,2}\in C^{2,1}_{\mathrm b}({\R^d}\times
[0,T])$, the property clearly carries over to any pointwise limit of
such solutions. We therefore expect it to hold for a much larger class
of monotone drifts $B$ and \GGr{for unbounded} solutions.
\end{remark}
\paragraph{Plan of the paper}
In section 2, we collect some tools useful to our arguments: we
present a slightly refined version of Kantorovich duality, an
approximation technique of the cost functional, the construction of a
smooth and bounded approximation of the operator $B$, and a rescaling
trick which allows to consider $\lambda=0$ in the following arguments.
Section 3 is devoted to the proof of Theorem \ref{thm:crucialintro},
the last Section contains the proof of Theorem \ref{teor:main}.
\section{Preliminaries}
In this section we collect some preliminary and technical
regularization results which will turn to be useful in the sequel.
\subsection{{$C^\infty_{{\mathrm c}}({\R^d})$ functions in} Kantorovich duality}
Let us first show that we can assume $\phi^1,\phi^2$ are smooth and
compactly supported in the duality formula \eqref{dualintro}.
\begin{prop}\label{Kant+}
If the cost function $h$ is Lipschitz continuous and satisfies
$ \lim_{r\uparrow+\infty}h(r)=+\infty,$
then
\begin{equation} \label{Kdual[II]}
\begin{aligned}
\Cost h{\mu^1}{\mu^2}=\sup\Big\{& \int_{\R^d} \phi^1\,\d\mu^1+
\int_{\R^d} \phi^2\,\d\mu^2:\\&\phi^1,\phi^2\in C^\infty_{{\mathrm c}}({\R^d}),\
\phi^1(x_1)+\phi^2(x_2)\le h(|x_1-x_2|)\Big\}.
\end{aligned}
\end{equation}
\end{prop}
\begin{proof}
Let us recall that the $h$-transform of a given bounded function
$\zeta:{\R^d}\to{\mathbb R}$ is defined as
\begin{equation}
\label{eq:22}
\zeta^h(x):=\inf_{y\in {\R^d}}h(|x-y|)-\zeta(y),
\end{equation}
and it is a bounded and Lipschitz continuous function satisfying $\zeta(x)+\zeta^h(y)\le h(|x-y|)$.
Let us fix $c<\Cost h {\mu^1}{\mu^2}$ and admissible $\phi^1,\phi^2\in C_{{\mathrm b}}({\R^d})$ such that
\begin{equation}
\label{eq:24}
\int_{\R^d} \phi^1\,\d\mu^1+
\int_{\R^d} \phi^2\,\d\mu^2>c.
\end{equation}
By possibly replacing $\phi^2$ with $(\phi^1)^{h}\ge \phi^2$
and $\phi^1$ with $(\phi^1)^{hh}\ge \phi^1$, it is not restrictive to assume that
$\phi^1,\phi^2$ are also Lipschitz continuous. Adding to $\phi^1$ and subtracting from $\phi^2$
a suitable constant, we can also assume that $\phi^1\ge0$ and $\phi^2\le 0$.
Let us now consider a family of mollifiers
$\kappa_\eta$ and of cutoff functions ${\raise.3ex\hbox{$\chi$}}_R$ defined by
\begin{subequations}
\begin{gather}
\label{eq:32}
\kappa_\eta(x):=\eta^{-d}\kappa\big(x/\eta\big),\quad
{\raise.3ex\hbox{$\chi$}}_R(x):={\raise.3ex\hbox{$\chi$}}(x/R)\quad x\in{\R^d},\
\eta, R>0,
\intertext{where $\kappa,{\raise.3ex\hbox{$\chi$}}\in C^\infty_{{\mathrm c}}({\R^d})$ satisfy}
\label{eq:69}
\kappa\ge 0,\quad \int_{{\R^d}} \kappa(x)\, \d x =1,\qquad
0\le {\raise.3ex\hbox{$\chi$}}\le1,
\quad
{\raise.3ex\hbox{$\chi$}}(x)=0\text{ if }|x|\ge1,
\quad{\raise.3ex\hbox{$\chi$}}(x)=1\text{ if }|x|\le 1/2.
\end{gather}
\end{subequations}
\meta{Symmetry and decreasing property of $\kappa$ are relevant just in dimension 1\\
and could be confusing at this point since they are not required in the lemma.\\
I separated
the role of the mollifiers and of the cut-off\\ and
I re-stated the similar assumptions later on\marginpar{I fully agree!}}
\bigskip
We set $\phi^1_\eta:=\phi^1\ast \kappa_\eta$ and
$\phi^2_\eta:=\phi^2\ast \kappa_\eta-\delta_\eta$, where
\begin{displaymath}
\delta_\eta:=\sup(\phi^1\ast\kappa_\eta-\phi^1)^+
+\sup(\phi^2\ast\kappa_\eta -\phi^2)^+.
\end{displaymath}
The definition of $\delta_\eta$ yields
\begin{displaymath}
\phi^1_\eta(x_1)+\phi^2_\eta(x_2)\le
\phi^1\ast\kappa_\eta(x_1)-\phi^1(x_1)+\phi^2\ast\kappa_\eta(x_2)-\phi^2(x_2)-\delta_\eta+h(|x_1-x_2|)
\le h(|x_1-x_2|).
\end{displaymath}
Moreover, since $\phi^1,\phi^2$ are Lipschitz, $\phi^1_\eta$ and $\phi^2_\eta$ converge to $\phi^1,\phi^2$
uniformly as $\eta\downarrow0$, so that
$\phi^1_\eta$ and $\phi^2_\eta$ are a smooth admissible {pair} \Killed{couple} still satisfying \eqref{eq:24}
and the sign condition $\phi^1_\eta\ge0,\ \phi^2_\eta\le 0$.
Let us now choose $R_0>0$ such that
\begin{equation}
\label{eq:25}
h(r)\ge \sup \phi^1_\eta\quad\text{for every $r\ge R_0$}
\end{equation}
{Setting $\phi^1_{\eta,R}:=\phi^1_\eta \, {\raise.3ex\hbox{$\chi$}}_R\le \phi^1_\eta$} we easily have for $R\ge R_0$
\begin{displaymath}
\inf_{x_1\in {\R^d}} h(|x_1-x_2|)-\phi^1_{\eta,R}(x_1)\ge 0\quad\text{if }|x_2|\ge 2R\ge R+R_0.
\end{displaymath}
%
Since $\phi^2_{\eta,4R}:=\phi^2_\eta\,{\raise.3ex\hbox{$\chi$}}_{4R}$ satisfies
$\phi^2_{\eta,4R}(x_2)=\phi^2_\eta(x_2)$ if $|x_2|\le 2R$
and $\phi^2_{\eta,4R}(x_2)\le 0$ for every $x_2\in {\R^d},$
it follows that $\phi^1_{\eta,R},\phi^2_{\eta,4R}$
is an admissible couple in
$C^\infty_{\mathrm c}({\R^d})$, and,
for $R$ sufficiently large, it still satisfies \eqref{eq:24}.
\end{proof}
\subsection{Regularization of the cost function.}
\label{subsec:regularization_cost}
In this section we shall show that it is sufficient to consider
nonnegative, Lipschitz, and unbounded costs
(as those considered in Proposition \ref{Kant+})
in the proof of Theorem \ref{teor:main}.
\begin{lemma}
\label{le:cost_approximation}
If \eqref{++} holds for every nonnegative Lipschitz and
nondecreasing cost function $h$
with $\lim_{r\uparrow+\infty}h(r)=+\infty$, then it holds for every
continuous and nondecreasing cost $h$.
\end{lemma}
\begin{proof}
We first prove that it is sufficient to consider nonnegative
Lipschitz costs; in a second step, we deal with the asymptotic
requirement.
\noindent\underline{Step 1:} \emph{$h$ Lipschitz.}
Adding a suitable constant we can assume that $h(r)\ge h(0)=0$.
We can then approximate $h$ from below by the increasing sequence of nonnegative Lipschitz functions
\begin{equation*}
h^n(r):=\inf_{s\ge 0} h(s)+n|r-s|
\end{equation*}
which satisfies
\begin{equation*}
0=h^n(0)\le h^n(r)\le h(r),\quad
\lim_{n\uparrow+\infty} h^n(r)=h(r)\quad
\forall\, r\ge0,
\end{equation*}
the convergence being uniform on each compact interval of $[0,+\infty)$.
Applying Lemma \ref{le:stability} below we find
\begin{displaymath}
\Cost{h_{\lambda\,t}}{\rho^1_t}{\rho^2_t}\topref{eq:57}=
\lim_{n\uparrow+\infty} \Cost{h^n_{\lambda\,t}}{\rho^1_t}{\rho^2_t}
\topref{++}\le \liminf_{n\uparrow+\infty}\Cost{h^n}{\rho^1_0}{\rho^2_0}\topref{eq:57}=
\Cost{h}{\rho^1_0}{\rho^2_0}.
\end{displaymath}
\noindent\underline{Step 2:} \emph{$\lim_{r\uparrow+\infty} h(r)=+\infty$.}
{Let us set $\rho_0:=\rho^1_0+\rho^2_0$,}
let us introduce the function
\begin{displaymath}
m(r):=\rho_0({\R^d}\setminus r\, U),\qua
U:=\big\{x\in {\R^d}:|x|< 1\big\},
\end{displaymath}
and let us consider a sequence $r_n$ in $[0,+\infty)$ such that
\begin{displaymath}
r_0:=0, \quad
r_1:=1,\quad r_{n+1}-r_n\ge r_n-r_{n-1}\quad\text{and }m(r_{n+1})\le 2^{-n}.
\end{displaymath}
It is easy to check that $r_n$ is a diverging increasing sequence; if
$g$ is the piecewise linear function satisfying $g(r_n)=n$, i.e.
\begin{displaymath}
g(r):= n+\frac{r-r_n}{r_{n+1}-r_n}\quad \text{if }r\in [r_n,r_{n+1}],
\end{displaymath}
then $g$ is Lipschitz continuous, increasing, unbounded, concave, and
it satisfies {$g(0)=0$} and
\begin{align*}
G:=&\int_{\R^d} g(|x|)\,\d\rho_0(x)=
%
\int_{\R^d}\Big(\int_0^{|x|}g'(r)\,\d r\Big) \,\d\rho_0(x)=
\int_{\R^d}\Big(\int_0^{+\infty} g'(r)\mathds 1_{r\le |x|} \,\d r\Big) \,\d\rho_0(x)
\\=&
\int_0^{\infty} g'(r)\, m(r)\,\d r=\sum_{n=1}^{+\infty}\frac{1}{
r_n-r_{n-1}}\int_{r_{n-1}}^{r_{n}}m(r)\,\d r\le \sum_{n=0}^{+\infty} m(r_n)<+\infty.
\end{align*}
We can thus consider the perturbed cost
\begin{equation*}
h^{\varepsilon}(r):=h(r)+\varepsilon\, g(r)
\end{equation*}
which is Lipschitz, increasing, unbounded. Since $g$ is concave,
increasing, {and $g(0)=0$}, we
have
\begin{equation}
\label{eq:54}
g(|x_1-x_2|)\le g(|x_1|+|x_2|)\le g(|x_1|)+g(|x_2|)\quad\text{for
every }x_1,x_2\in {\R^d},
\end{equation}
so that if {${\mbox{\boldmath$\rho$}}_0$} \Killed{${\mbox{\boldmath$\rho$}}$}
is an optimal coupling between {$\rho^1_0$ and $\rho^2_0$}
\Killed{$\rho^1$ and $\rho^2$}
for the cost $h$ (we can assume that the initial cost is finite), then
\begin{align*}
\Cost h{\rho^1_0}{\rho^2_0}&\le \Cost{h^\varepsilon}{\rho^1_0}{\rho^2_0}\le
\Cost h{\rho^1_0}{\rho^2_0}+\varepsilon\int_{{\R^d}\times{\R^d}}
g(|x_1-x_2|)\,\d{\mbox{\boldmath$\rho$}}_0(x_1,x_2)\\&\topref{eq:54}\le
\Cost h{\rho^1_0}{\rho^2_0}+\varepsilon\int_{{\R^d}\times{\R^d}}
\Big(g(|x_1|)+g(|x_2|)\Big)\,\d{\mbox{\boldmath$\rho$}}_0(x_1,x_2)=
\Cost h{\rho^1_0}{\rho^2_0}+\varepsilon G.
\end{align*}
Therefore, if Theorem \ref{teor:main} holds for $h^\varepsilon$ we have
\begin{displaymath}
\Cost h{\rho^1_t}{\rho^2_t}\le
\Cost {h^\varepsilon}{\rho^1_t}{\rho^2_t}\le
\Cost {h^\varepsilon}{\rho^1_0}{\rho^2_0}\le
\Cost h{\rho^1_0}{\rho^2_0}+\varepsilon G.
\end{displaymath}
Passing to the limit as $\varepsilon\down0$ we conclude.
\end{proof}
The following result provides a variant of well known stability properties
of transportation costs (see \cite[Theorem 3]{Schachermayer-Teichmann09}, \cite[Theorem 5.20]{Villani09})
and holds the much more general setting of optimal transportation in Radon metric spaces
\cite[Chapter 6]{Ambrosio-Gigli-Savare08}.
\begin{lemma}[Lower semicontinuity of the cost functional w.r.t.\ local uniform convergence of $h$]
\label{le:stability} \ \\
Let $h:[0,+\infty)\to [0,+\infty)$ be a continuous cost function and let $h^n:[0,+\infty)\to [0,+\infty)$
be a sequence of lower semicontinuous functions converging to $h$ locally uniformly in $[0,+\infty)$.
For every couple $\rho^1,\rho^2\in \mathcal{P}({\R^d})$ we have
\begin{equation}
\label{eq:56}
\liminf_{n\uparrow+\infty}\Cost{h^n}{\rho^1}{\rho^2}\ge \Cost{h}{\rho^1}{\rho^2}.
\end{equation}
In particular, if $h^n\le h$ for every $n\in {\mathbb N}$ then
\begin{equation}
\label{eq:57}
\lim_{n\to+\infty}\Cost{h^n}{\rho^1}{\rho^2}=\Cost h{\rho^1}{\rho^2}.
\end{equation}
\end{lemma}
\begin{proof}
Let us set $H^n(x_1,x_2):=h^n(|x_1-x_2|)$ and observe that $H^n$ converges to
$H(x_1,x_2):=h^n(|x_1-x_2|)$ uniformly on compact sets of ${\R^d}\times{\R^d}$.
If ${\mbox{\boldmath$\rho$}}_n\in \mathcal{P}({\R^d}\times{\R^d})$ is an optimal coupling between $\rho^1,\rho^2$
with respect to the cost $h^n$ then
\begin{displaymath}
\Cost{h^n}{\rho^1}{\rho^2}=
\int_{[0,+\infty)} z\,\d \mu_n(z),\quad\text{where } \mu_n=(H^n)_\# {\mbox{\boldmath$\rho$}}_n.
\end{displaymath}
Since the marginals of ${\mbox{\boldmath$\rho$}}_n$ are fixed, the sequence $({\mbox{\boldmath$\rho$}}_n)_{n\in {\mathbb N}}$ is tight and up to the
extraction of a suitable subsequence (still denoted by ${\mbox{\boldmath$\rho$}}_n$)
we can suppose that ${\mbox{\boldmath$\rho$}}_n$ converge to
to some limit coupling ${\mbox{\boldmath$\rho$}}$ between $\rho^1,\rho^2$ in $\mathcal{P}({\R^d}\times{\R^d})$.
Since $\mu_n$ weakly converge to $\mu=H_\# {\mbox{\boldmath$\rho$}}$ by
\cite[Lemma 5.2.1]{Ambrosio-Gigli-Savare08}, standard lower semicontinuity
of integrals with nonnegative continuous integrands \cite[Lemma 5.1.7]{Ambrosio-Gigli-Savare08}
yields
\begin{displaymath}
\liminf_{n\to+\infty} \int_{[0,+\infty)} z\,\d \mu_n(z)\ge
\int_{[0,+\infty)} z\,\d \mu(z)=\int_{{\R^d}\times{\R^d}}H(x_1,x_2)\,\d{\mbox{\boldmath$\rho$}}(x_1,x_2)\ge \Cost h{\rho^1}{\rho^2}.
\end{displaymath}
\end{proof}
\subsection{Bounded, smooth approximations of a monotone operator}
\label{subsec:Bregular}
If $A:{\R^d}\to{\R^d}$ is a monotone operator then there exists \cite[Corollary 2.1]{Brezis73} a
maximal monotone {multivalued} extension
${\sf A}:{\R^d}\rightrightarrows{{\R^d}}$ {(thus taking values in $2^{{\R^d}}$)}
such that $A(x)\in
{\sf A}(x)$ for every $x\in {\R^d}$. We denote by ${\sf A}^\circ(x)$ the element of
minimal norm in (the closed convex set) ${\sf A}(x)$.
\Killed{Applying} \cite[Corollary 1.4]{Alberti-Ambrosio99}
{shows that the set ${\sf A}(x)\subset {\R^d}$ reduces to the singleton
$\{A(x)\}$
$\Leb d$-almost everywhere}: in fact it satisfies
\begin{equation}
\label{eq:26}
{\sf A}(x)=\{{\sf A}^\circ(x)\}=\{A(x)\}\quad\text{for $\Led$-a.e.\ $x\in {\R^d}$},\quad
{\sf A}(x)=\mathrm{conv}\big\{\lim_{n\to\infty}A(x_n)\ \text{for
some }x_n\to x\big\}
\end{equation}
We recall the following important approximation result
\cite[Theorem 4.1]{Fitzpatrick-Phelps92}:
we denote by $U$ the open unit ball in ${\R^d}$.
\begin{tteor}[Fitzpatrick-Phelps] \label{Fitz}
For every maximal monotone operator ${\sf A}:{\R^d}\rightrightarrows{\R^d}$, there
exists a sequence of maximal monotone operators
${\sf A}_n:{\R^d}\rightrightarrows{\R^d}$ such that,
for each $x \in {\R^d}$ and all n,
\begin{equation} \label{Fitz:eq}
{\sf A}(x)\cap n\, U \subset {\sf A}_n(x)\subset n\,\overline U,
\quad {\sf A}_n(x) \setminus {\sf A}(x) \subset n \,\partial U\quad\text{for
every }x\in {\R^d}.
\end{equation}
\end{tteor}
Notice that \eqref{Fitz:eq} yields in particular
\begin{equation}
\label{eq:27}
|{\sf A}_n^\circ(x)|= \min\big(|{\sf A}^\circ(x)|,n\big)\quad\text{for every }x\in {\R^d}.
\end{equation}
\begin{teor} \label{approxB}
Let ${\sf A}:{\R^d}\rightrightarrows{\R^d}$ be a maximal monotone operator
and
$(\beta_n)_{n\in {\mathbb N}}$ a vanishing sequence of positive real numbers.
There exists a sequence of smooth, globally Lipschitz, and bounded monotone operators
$A_n:{\R^d}\to{\R^d}$ such that
\begin{equation} \label{B approx}
\mathrm{Lip}(A_n)\le n,\quad
|A_n(x)|\le \min\big(|{\sf A}^\circ(x)|,n\big)+\beta_n,\quad
\lim_{n\to+\infty}A_n(x)={\sf A}^\circ(x)\quad
\mbox{for every } x\in {\R^d}.
\end{equation}
\end{teor}
\begin{proof}
Let ${\sf A}_n$ be a sequence of maximal monotone operators satisfying
\eqref{Fitz:eq}
and let $Y_n:{\R^d}\to{\R^d}$ be the Moreau-Yosida approximation of ${\sf A}_n$ of
parameter $n^{-1}$ \cite[Proposition 2.6]{Brezis73}
\begin{displaymath}
Y_n(x):=n\Big(x-(I+n^{-1}\,{\sf A}_n)^{-1} x\Big)
\end{displaymath}
Note that $Y_n$ is a $n$-Lipschitz monotone map satisfying
\begin{equation}
\label{eq:28}
|Y_n(x)|\le
|{\sf A}_n^\circ(x)|\topref{eq:27}=\min\big(|{\sf A}^\circ(x)|,n\big)\quad\text{for every
}x\in {\R^d}
\end{equation}
Let us fix $x\in {\R^d}$ and let $x_n\in {\R^d}$ be the unique solution of
\begin{equation}
\label{eq:29}
x_n+n^{-1}\,{\sf A}_n(x_n)\ni x\quad\text{so that}\quad
Y_n(x)=n(x-x_n)\in {\sf A}_n(x_n).
\end{equation}
If $n>|{\sf A}^\circ(x)|$ then \eqref{eq:28} yields $Y_n(x)\notin n\,\partial
U$;
applying \eqref{Fitz:eq} and \eqref{eq:28} again we get
\begin{equation}
\label{eq:30}
Y_n(x)\in {\sf A}(x_n),\quad
|Y_n(x)|\le |{\sf A}^\circ(x)|,\quad
|x-x_n|\le n^{-1}\,|{\sf A}^\circ(x)|\quad\text{for every }n>|{\sf A}^\circ(x)|.
\end{equation}
Since the graph of ${\sf A}$ is closed, any accumulation point $y$ of the
bounded sequence $Y_n(x)$ satisfies
\begin{equation}
\label{eq:31}
y\in {\sf A}(x),\quad |y|\le |{\sf A}^\circ(x)|.
\end{equation}
We thus conclude that $\lim_{n\uparrow+\infty}Y_n(x)={\sf A}^\circ(x)$ for
every $x\in {\R^d}$.
To conclude the proof we need to regularize $Y_n$: to this aim we
consider the family of mollifiers $\kappa_\eta$ as in \eqref{eq:32}
and we set
\begin{equation}
\label{eq:33}
A_n:=Y_n\ast \kappa_\eta \quad\text{with}\quad
\eta:= (n\, k)^{-1}\beta_n \text{ where } k:=\int_{\R^d}
|x|\kappa(x)\,\d x,
\end{equation}
so that
\begin{displaymath}
|A_n(x)-Y_n(x)|\le \eta\,k\, \mathrm{Lip}(Y_n)\le n\,\eta\,k\le \beta_n.
\end{displaymath}
\end{proof}
We consider now a radial smoothing:
\begin{prop}
\label{approxBn}
Let $A_n:{\R^d}\to{\R^d}$ be smooth, Lipschitz, and bounded monotone
operators satisfying
\eqref{B approx}.
For every $m\in{\mathbb N}$ there exists bounded, smooth,
Lipschitz, and monotone operators $A_{n,m}$ such that
\begin{gather}
\label{eq:34}
\mathrm{Lip}(A_{n,m})\le n,\quad
\sup_{x\in {\R^d}}|A_{n,m}(x)|\le n+\beta_n,\quad
\sup_{x\in{\R^d}} \big|\mathbb{D}\,A_{n,m}(x)\cdot x\big|\le 2m\,(n+\beta_n)\\
\label{eq:62}
\lim_{m\uparrow+\infty}A_{n,m}(x)=A_n(x)\quad\text{for every }x\in {\R^d}.
\end{gather}
\end{prop}
\begin{proof}
%
We consider a family of mollifiers $\kappa_\eta=\eta^{-1}\kappa(\cdot/\eta)\in C^\infty_{\mathrm c}({\mathbb R})$,
where $\kappa$ satisfies
\begin{equation}
\label{eq:71}
\supp(\kappa)\subset [0,2],\quad 0\le \kappa\le \kappa(1)=1,\quad
(1-x)\kappa'(x)\ge 0,\quad \int_{\mathbb R}\kappa(x)\,\d x=1,
\end{equation}
and the function $\vartheta\in C^\infty_{\rm
c}(0,+\infty)$ defined by $\vartheta(r):=\kappa(-\log r)$, $r>0$.
\Killed{where $\kappa$ is the mollifier of \eqref{eq:32} in dimension $d=1$.}
We set
\begin{equation}
\label{eq:35}
A_{n,m}(x):=m\int_0^{+\infty}
A(rx)\vartheta(r^{m})\,\frac{\d r}r
\end{equation}
The change of variable $r=e^{-z}$ shows that
\begin{displaymath}
A_{n,m}(x)=
m\int_{\mathbb R} A_n(x\,{\mathrm e}^{-z})\,\kappa(m\,z)\,\d z
=A_n^x\ast \kappa_{1/m} (0),\quad\text{where }
A_n^x(z):=A_n(x\,e^{z})\ \text{for } z\in {\mathbb R}.
\end{displaymath}
It is then easy to check that \Killed{ if $A_n$ has bounded derivatives then}
{$|\mathbb{D} A_{n,m}|\le n$} \Killed{enjoys the same property}
%
since
\begin{displaymath}
|D A_{n,m}(x)|\le
m\int_{\mathbb R} \big|\mathbb{D} A_n(x\,{\mathrm e}^{-z})\big|{\mathrm e}^{-z}\,\kappa(m\,z)\,\d z
\topref{B approx}\le
n\int_{\mathbb R} {\mathrm e}^{-y/m}\,\kappa(y)\,\d y
\topref{eq:71}\le n,
\end{displaymath}
and {$A_{n,m}$} \Killed{it} converges pointwise to $A_n$ as
$m\uparrow+\infty$.
Concerning the second bound of \eqref{eq:34} we easily have
\begin{align*}
\mathbb{D}\,A_{n,m}(x)\cdot x &=
m\int_0^{+\infty}
\mathbb{D}\,A_n(rx)\cdot x\vartheta(r^{m})\,{\d r}=
m\int_0^{+\infty}
\frac\d{\d r}\Big(A_n(rx)\Big)\vartheta(r^{m})\,{\d r}\\&=
-m^2 \int_0^{+\infty}
A_n(rx)\tilde \vartheta(r^{m})\,\frac{\d r}r\qquad \text{where }\tilde\vartheta(r):=r\vartheta'(r),
\end{align*}
so that the inequality follows choosing $\kappa$ even and
nondecreasing in $[0,+\infty$), so that $\int_0^{+\infty}
|\vartheta'(r)|\,\d r=2$.
\end{proof}
\subsection{$\lambda$-monotonicity and rescaling}
\label{subsec:rescaling}
We show here a simple rescaling argument (inspired by
\cite{Carrillo-Toscani00}, where the rescaling technique has been
applied to a wide class of diffusion equations), which is useful to
deduce the estimates in the general $\lambda$-monotone case to
the simpler case of a monotone operator.
We therefore assume that $\lambda\neq 0$,
and we introduce the time rescaling functions
\begin{gather}
\label{eq:36}
{\sf s}(t):= \int_0^t {\mathrm e}^{2\lambda r}\,\d r=\frac1{2\lambda}\big({\mathrm e}^{2\lambda t}-1\big),\quad
{\sf t}(s):= \frac1{2\lambda} \log(1+2\lambda s)\quad s\in [0,S_\infty)
\intertext{where}
\label{eq:61}
S_\infty:=
\begin{cases}
+\infty&\text{ if $\lambda>0$},\\
-1/(2\lambda)&\text{ if $\lambda<0$.}
\end{cases}
\end{gather}
We associate to a family of probability measures $\rho_t$, $t\in [0,T]$,
their rescaled versions $\sigma_s$, $s\in [0,S_\infty)$, defined by
\begin{equation}
\label{eq:40}
\sigma_s(E):=\rho_{{\sf t}(s)}({\mathrm e}^{-\lambda {\sf t}(s)}\,E)\quad\text{for every Borel set }E\subset {\R^d}.
\end{equation}
If $B:{\R^d}\to{\R^d}$ is a $\lambda$-monotone Borel map we set $A:=B-\lambda
I$ and
\begin{equation}
\label{eq:41}
\tilde B(y,s):={\mathrm e}^{-\lambda {\sf t}(s)} B({\mathrm e}^{-\lambda {\sf t}(s)}\,y),\quad
\tilde A(y,s)=
{\mathrm e}^{-\lambda {\sf t}(s)} A({\mathrm e}^{-\lambda {\sf t}(s)}\,y)
\quad\text{for }y\in {\R^d},\ s\in {\mathbb R}.
\end{equation}
Notice that if $B$ is $\lambda$-monotone, then $A$ and $\tilde
A(\cdot,s)$, $s\in [0,S_\infty)$, are monotone.
\begin{prop}
\label{prop:rescaled}
A continuous family $\rho_t\in \mathcal{P}({\R^d})$ is a distributional
solution of \eqref{eq:2} if and only if the rescaled measures
$\sigma_s$ defined by \eqref{eq:40} and \eqref{eq:36} satisfy
\begin{equation}
\label{eq:39}
\int_{0}^{S_\infty}\int_{\R^d} \Big(\partial_s\varphi+\Delta\varphi-\tilde A(\cdot,s)\cdot
\nabla \varphi\Big)\,\d\sigma_s\,\d s=0\quad\forall\, \varphi\in
C^\infty_{\mathrm c}({\R^d}\times (0,S_\infty)).
\end{equation}
If $\rho$ satisfies \eqref{Cond} then
\begin{equation}
\label{eq:55}
\int_{s_0}^{s_1}\int_{\R^d} |\tilde A(x,s)|\,\d\sigma_s\,\d s<+\infty\quad\text{for every }0<s_0<s_1<S_\infty,
\end{equation}
and in this case $\sigma$ satisfies
\begin{equation} \label{eq:3}
\int_{\R^d} \varphi(\cdot,s_1)\,\d\sigma_{s_1}-\int_{\R^d} \varphi(\cdot,s_0)\,\d\sigma_{s_0} =
\int_{s_0}^{s_1} \int_{\R^d} \Big(\partial_s\varphi+ \Delta\varphi-
\tilde A(y,s)\cdot \nabla \varphi\Big)\,\d\mu_s\,\d s.
\end{equation}
for every test function $ \varphi\in C^{2,1}_{\mathrm
b}({\R^d}\times[s_0,s_1])$
with bounded first and second derivatives.
\end{prop}
\begin{proof}
We introduce the change of variable map
${\sf X}(x,t):=({\mathrm e}^{\lambda t}x,{\sf s}(t))$ and for a given smooth function
$\varphi\in C^\infty_{\mathrm c}({\R^d}\times(0,s_\infty))$ we set
$\zeta(x,t):=\varphi({\mathrm e}^{\lambda {\sf s}(t)},{\sf s}(t))=\varphi\circ {\sf X}$.
Denoting by $(y,s)\in {\R^d}\times[0,s_\infty )$ the new variables, easy calculations show that
in ${\R^d}\times (0,+\infty)$ we have
\begin{align*}
\partial_t\zeta&={\sf s}'\Big(\partial_s\varphi+\lambda {\mathrm e}^{-2 \lambda t}\nabla_y\varphi\cdot y\Big)\circ {\sf X},\quad
&
\nabla_x\zeta&={\mathrm e}^{\lambda t}\,\nabla_y\varphi\circ{\sf X}\\
\Delta_x\zeta&={\mathrm e}^{2\lambda t}\,\Delta_y\varphi\circ{\sf X}&
B\cdot \nabla_x\zeta&={\mathrm e}^{2\lambda t}\,\Big(\tilde B(y,s)\cdot \nabla_y\varphi\Big)\circ{\sf X},
\end{align*}
where we used the fact that $B={\mathrm e}^{\lambda t}\tilde B\circ{\sf X}$.
In particular we have
\begin{displaymath}
\label{eq:43}
\partial_t\zeta- B\cdot \nabla_x\zeta={\sf s}'\Big(\partial_s\varphi-
\tilde A(y,s)\cdot \nabla_y\varphi\Big)\circ{\sf X}
\end{displaymath}
We thus have
\begin{align*}
\int_{\R^d}\Big(\partial_t\zeta+\Delta_x\zeta-B\cdot \nabla_x\zeta\Big)\,\d\rho_t&=
{\sf s}'(t) \int_{\R^d}\Big(\partial_s\varphi+\Delta_y\varphi-\tilde A(y,s)\cdot \nabla_y\varphi\Big)\circ{\sf X}\,\d\rho_t\\&=
{\sf s}'(t) \int_{\R^d}\Big(\partial_s\varphi+\Delta_y\varphi-\tilde A(y,s)\cdot \nabla_y\varphi\Big)\,\d\sigma_{{\sf s}(t)}
\end{align*}
since $\sigma_{{\sf s}(t)}(E)=\rho_t({\mathrm e}^{-\lambda t}\,E)$ for every Borel set $E\subset {\R^d}$.
Eventually we obtain
\begin{displaymath}
\int_0^{+\infty}\int_{\R^d}\Big(\partial_t\zeta+\Delta_x\zeta-B\cdot \nabla_x\zeta\Big)\,\d\rho_t\,\d t=
\int_0^{s_\infty} \int_{\R^d} \Big(\partial_s\varphi+\Delta_y\varphi-\tilde A(y,s)\cdot \nabla_y\varphi\Big)\,\d\sigma_s\,\d s
\end{displaymath}
\eqref{eq:55} follows by a simple application of the change of
variable formula \eqref{eq:51}, since for every $t>0$
\begin{align*}
\int_{\R^d} \big|\tilde A(y,s)\big|\,\d\sigma_s(y)&\topref{eq:41}=
{\mathrm e}^{-\lambda{\sf t}(s)}\int_{\R^d} \big|A({\mathrm e}^{-\lambda
{\sf t}(s)}y)\big|\,\d\sigma_{s}(y)
\\&\topref{eq:40}=
{\mathrm e}^{-\lambda{\sf t}(s)}\int_{\R^d} \big|A(x)\big|\,\d\rho_{{\sf t}(s)}(x)=
{\mathrm e}^{-\lambda{\sf t}(s)}\int_{\R^d} \big|B(x)-\lambda x\big|\,\d\rho_{{\sf t}(s)}(x).
\end{align*}
Since ${\sf t}'(s)={\mathrm e}^{-\lambda {\sf t}(s)}$ we eventually get for
$t_i={\sf t}(s_i)$
\begin{align*}
\int_{s_0}^{s_1}\int_{\R^d} |\tilde A(x,s)|\,\d\sigma_s\,\d s&=
\int_{s_0}^{s_1}\Big(\int_{\R^d} \big|B(x)-\lambda
x\big|\,\d\rho_{{\sf t}(s)}(x)\Big)\,{\sf t}'(s)\,\d s\\&=
\int_{t_0}^{t_1} \int_{\R^d} \big|B(x)-\lambda
x\big|\,\d\rho_t(x)\,\d t\topref{Cond}<+\infty.
\end{align*}
\eqref{eq:3} follows from \eqref{eq:39} when $\varphi$ belongs to
$C^\infty_{{\mathrm c}}({\R^d}\times [s_0,s_1])$.
If $\varphi\in C^{2,1}_{{\mathrm b}}({\R^d}\times[s_0,s_1])$
via a standard convolution and truncation argument we find an
approximation sequence $\varphi_k \in C^\infty_{\mathrm
c}({\R^d}\times [s_0,s_1])$
such that $\varphi_k,\partial_t\varphi_k,\nabla\varphi_k,\Delta\varphi_k$
remains
uniformly bounded and converge pointwise to
$\varphi,\partial_t\varphi,\nabla\varphi,\Delta\varphi$
respectively.
By \eqref{eq:55} we can apply the Lebesgue Dominated
Convergence theorem to pass to the limit in \eqref{eq:3} written
for $\varphi_k$, thus obtaining the same identity for $\varphi$.
\end{proof}
We conclude this section by a simple remark combining the
regularization technique of Section~\ref{subsec:Bregular} and the time
rescaling \eqref{eq:41}.
\begin{lemma}
\label{le:bound}
Let $A:=B-\lambda I$ be a monotone operator,
let us consider a sequence $A_{n,m}$,
$n,m\in {\mathbb N}$, of smooth monotone operators given by Theorem
\ref{approxB} and Proposition \ref{approxBn}, and let us set
\begin{equation}
\label{eq:63}
\tilde A_{n,m}(y,s):={\mathrm e}^{-\lambda{\sf t}(s)}
A_{n,m}({\mathrm e}^{-\lambda{\sf t}(s)}y)\quad y\in {\R^d},\ s\in [0,S_\infty)
\end{equation}
defined as in \eqref{eq:41}, \eqref{eq:36}.
Then $\tilde A_{n,m}$ are Lipschitz in ${\R^d}\times[0,S]$ for every $S\in
[0,S_\infty)$.
\end{lemma}
\begin{proof}
We just have to check that $|\partial_s \tilde A_{n,m}(\cdot,s)|$ is
uniformly bounded in ${\R^d}\times[0,S]$:
sine ${\sf t}'(s)={\mathrm e}^{-\lambda {\sf t}(s)}$ a simple calculation yields
\begin{displaymath}
\partial_s \tilde A_{n,m}(y,s)=-\lambda
{\mathrm e}^{-\lambda{\sf t}(s)}\, \tilde
A_{n,m}(y,s)-\lambda{\mathrm e}^{-2\lambda {\sf t}(s)}
\mathbb{D} A_{n,m}({\mathrm e}^{-\lambda{\sf t}(s)}y)\cdot y =
-\lambda{\mathrm e}^{-\lambda{\sf t}(s)}\tilde Q_{n,m}(y,s)
\end{displaymath}
where
\begin{displaymath}
Q_{n,m}(x)=A_{n,m}(x)+\mathbb{D} A_{n,m}(x)\cdot x,\quad x\in {\R^d}.
\end{displaymath}
Since ${\mathrm e}^{-\lambda{\sf t}(s)}$ is uniformly bounded with all its
derivative in each compact interval $[0,S]$, $S<\infty$,
\eqref{eq:34} show that $Q_{n,m}$ is bounded and therefore $\tilde
A_{n,m}$ is Lipschitz with respect to $s$.
\end{proof}
\section{A comparison result for the backward equation}
In this section we give the proof of Theorem
\ref{thm:crucialintro} in a slightly more general form, in order to be
applied to (a suitably regularized version of) the rescaled
formulation considered in Proposition~\ref{prop:rescaled}.
Let us suppose that $\tilde A:(y,s)\in {\R^d}\times[0,S_\infty)\to \tilde
A(y,s)\in{\R^d}$ is a smooth
vector field
satisfying
\begin{gather}
\label{eq:38}
\sup_{{\R^d}\times [0,S]}|\tilde A_s|+|\partial_s \tilde A|+|\mathbb{D}\tilde
B|<+\infty\quad\text{for every $S\in [0,S_\infty)$,}\\
\label{eq:45}
\tilde A(\cdot,s)\text{ is monotone for every }s\in
[0,S_\infty).
\end{gather}
We denote by $\Oper\cdot$ the differential operator defined by
\begin{equation}
\label{eq:46}
\Oper\varphi(y,s):=\Delta_y \varphi(y,s)-\tilde A(y,s)\cdot
\nabla_y\varphi(y,s)\quad
\varphi(\cdot,s)\in
C^2({\R^d}),\quad
(y,s)\in {\R^d}\times [0,S_\infty).
\end{equation}
Thanks to \eqref{eq:38} and \eqref{eq:45}, we can apply the existence result
\cite[Theorem~3.2.1]{Stroock-Varadhan06} and for every $S\in [0,S_\infty)$
and $\phi\in C^\infty_{\mathrm c}({\R^d})$ we can find a solution
$\varphi\in C^{2,1}_{\mathrm b}({\R^d}\times [0,S))$ of the backward evolution equation
\begin{equation} \label{PB:back}
\partial_s\varphi + \Oper\varphi = 0\quad\text{in } {\R^d}\times[0,S],
\quad
\varphi(\cdot,S)=\phi(\cdot).
\end{equation}
We have
\begin{teor}
\label{thm:crucial}
Let $h:[0,+\infty)\to {\mathbb R}$ be a continuous and non-decreasing function.
Let $\varphi^1,\varphi^2\in C^{2,1}_{\mathrm b}({\R^d}\times [0,S])$
be solutions of the ``backward'' inequality
\begin{equation} \label{eq:23}
\partial_s \varphi+\Oper{\varphi}\ge0\quad
\text{in }{\R^d}\times [0,S]
\end{equation}
such that
\begin{equation*}
\varphi^1(y_1,S)+\varphi^2(y_2,S)\le h(|y_1-y_2|)\quad\text{for every }y_1,y_2\in {\R^d}.
\end{equation*}
Then
\begin{equation*}
\varphi^1(y_1,0)+\varphi^2(y_2,0)\le h(|y_1-y_2|)\quad\text{for every
}y_1,y_2\in {\R^d}.
\end{equation*}
\end{teor}
\begin{proof}
By approximating $h$ from above, it is not restrictive to assume
that $h \in C^1[0,+\infty)$ with $h'(0)=0$; in particular the map
$H(y_1,y_2):= h(|y_1-y_2|)$ is of class $C^1$ in ${\R^d}\times{\R^d}$ and
satisfies
\begin{gather}
\label{eq:47}
\nabla_{y_1}H(y_1,y_2)=-\nabla_{y_2}H(y_1,y_2)=g(y_1,y_2)(y_1-y_2),
\intertext{where }
\label{eq:48}
0\le g(y_1,y_2)=g(y_2,y_1):=
\begin{cases}
\frac{h'(|y_1-y_2|)}{|y_1-y_2|}&\text{if }y_1\neqy_2,\\
0&\text{if }y_1=y_2.
\end{cases}
\end{gather}
The argument combines a variable-doubling\ technique and a classical variant of the maximum principle.
Let us first show that if $\varphi^1,\varphi^2$ satisfy the \emph{strict} inequality
\begin{equation} \label{clash<0}
\partial_s \varphi^j+\Oper{\varphi^j}>0\quad
\text{in }{\R^d}\times[0,S),\quad j=1,2.
\end{equation}
then the function
\begin{equation*}
f(y_1,y_2,s) := \varphi^1(y_1,s)+\varphi^2(y_2,s)-H(y_1,y_2)
\end{equation*}
cannot attains a (local) maximum in a point $(\bar y_1,\bar y_2,\bar s)$ with
$\bar s<S$.
We argue by contradiction and we suppose that $(\bar y_1,\bar y_2,\bar s)$ is
a local maximizer of $f$ with $\bar s<S$; we thus have
\begin{equation*}
\partial_s f(\bar y_1,\bar y_2,\bar s) \le 0,
\quad\nabla_{y_1} f(\bar y_1,\bar y_2,\bar s) = 0,
\quad\nabla_{y_2} f(\bar y_1,\bar y_2,\bar s) = 0;
\end{equation*}
so that
\begin{equation} \label{dt<0}
\begin{aligned}
\partial_t \varphi^1(\bar y_1,\bar s) + \partial_t \varphi^2(\bar y_2,\bar s)\le 0
\end{aligned}
\end{equation}
\begin{equation*}
\begin{aligned}
\nabla_{y_1} \varphi^1(\bar y_1,\bar s)
&= \nabla_{y_1} H(\bar y_1,\bar y_2) \topref{eq:47}=
g(\bar y_1,\bar y_2)(y_1-y_2)\\
\nabla_{y_2} \varphi^2(\bar y_2,\bar s)
&= \nabla_{y_2} H(\bar y_1,\bar y_2) \topref{eq:47}=
g(\bar y_1,\bar y_2)(y_2-y_1).
\end{aligned}
\end{equation*}
It follows that
\begin{equation} \label{Bn<0}
\begin{aligned}
\tilde A(\bar y_1,\bar s)\cdot\nabla_{y_1}\varphi^1(\bar y_1,\bar s)+
&\tilde A(\bar y_2,\bar s)\cdot \nabla_{y_2}\varphi^2(\bar y_2,\bar s) \\&=
g(\bar y_1,\bar y_2)\big(\tilde A(\bar y_1,\bar s)-\tilde A(\bar y_2,\bar
s)\big)\cdot (\bar y_1-\bar y_2)
\topref{eq:48}\ge0
\end{aligned}
\end{equation}
On the other hand, since $H(\bar y_1+z,\bar y_2+z) = H(\bar y_1,\bar y_2)$, the function
\begin{displaymath}
{\R^d}\ni z\mapsto \varphi^1(\bar y_1+z,\bar s)+\varphi^2(\bar y_2+z,\bar s)-H(\bar y_1,\bar y_2)=
f(\bar y_1+z,\bar y_2+z,\bar s)
\end{displaymath}
has a local maximum at $z=0$ so that
\begin{equation} \label{Delta<0}
\begin{aligned}
\Delta_{y_1}\varphi^1(\bar y_1,\bar s)+\Delta_{y_2}\varphi^2(\bar y_2,\bar s)\le 0.
\end{aligned}
\end{equation}
Combining \eqref{dt<0},\eqref{Bn<0}, and \eqref{Delta<0} we obtain
\begin{equation*}
(\partial_s \varphi^1+\Oper{\varphi^1})(\bar y_1,\bar s)+(\partial_s\varphi^2+\Oper\varphi^2)(\bar y_2,\bar s)\le 0,
\end{equation*}
which contradicts \eqref{clash<0}.
Suppose now that $\varphi^1,\varphi^2$ satisfy the inequality \eqref{eq:23} and let us set for $\varepsilon,\delta>0$
\begin{equation*}
\varphi^j_{\varepsilon,\delta}(y_j,s):= \varphi^j(y_j,s)-\delta(S-s)-\varepsilon
{\mathrm e}^{-s}|y_j|^2
\quad j=1,2.
\end{equation*}
We easily get
\begin{align*}
\partial_s\varphi^j_{\varepsilon,\delta}&=\partial_s \varphi^j+\delta +\varepsilon {\mathrm e}^{-s}|y_j|^2\\
\Oper{\varphi^j_{\varepsilon,\delta}}&=\Oper{\varphi^j}-{\mathrm e}^{-s}\big(d\varepsilon+2\varepsilon
\tilde A(y_j,s)\cdot y_j\big)\\
\partial_s \varphi^j_{\varepsilon,\delta}+\Oper{\varphi^j_{\varepsilon,\delta} }
&\ge\delta +\varepsilon {\mathrm e}^{-s}\big(|y_j|^2-d-C_n|y_j|\big),
\end{align*}
where $C_n=\sup_{y,s} |\tilde A_n(y,s)|<+\infty$.
It follows that for every $\delta>0$ there exists a coefficient
$\varepsilon>0$ sufficiently small such that
$\varphi^1_{\varepsilon,\delta},\varphi^2_{\varepsilon,\delta}$ satisfy
\eqref{clash<0}.
On the other hand, the continuous function
\begin{equation*}
(y_1,y_2,s)\mapsto f_{\varepsilon,\delta}(y_1,y_2,s):=\varphi^1_{\varepsilon,\delta}(y_1,s)+\varphi^2_{\varepsilon,\delta}(y_2,s)-
h(|y_1-y_2|)\quad y_1,y_2\in {\R^d},\ s\in [0,S],
\end{equation*}
attains its maximum at some point $(\bar y_1,\bar y_2,\bar s)\in
{\R^d}\times{\R^d}\times [0,S]$;
by the previous argument, we conclude that $\bar s=S$ and therefore
for every $y_1,y_2\in {\R^d}$
\begin{align*}
\varphi^1_{\varepsilon,\delta}(y_1,0)&+\varphi^2_{\varepsilon,\delta}(y_2,0)-h(|y_1-y_2|)\le
f_{\varepsilon,\delta}(\bar y_1,\bar y_2,S)\le \varphi^1(\bar y_1,S)+\varphi^2(\bar y_2,S) - h(|\bar y_1-\bar y_2|)\le 0.
\end{align*}
Passing to the limit as $\varepsilon,\delta\downarrow 0$ we conclude.
\end{proof}
We conclude this section by recalling two well known estimates:
\begin{lemma}[Uniform estimates] \label{Grad}
Let $\varphi\in C^{2,1}_{\mathrm b}({\R^d}\times[0,S])\cap
C^\infty({\R^d}\times (0,S))$ be the solution of
\eqref{PB:back}.
Then
\begin{equation}
\label{eq:49}
\sup_{{\R^d}\times[0,S]} |\varphi|\le \sup_{ {\R^d}} |\phi|
,\qquad
\sup_{{\R^d}\times[0,S]} |\nabla\varphi|\le \sup_{{\R^d}} |\nabla\phi|.
\end{equation}
\end{lemma}
\begin{proof}
The first inequality is direct application of
the maximum principle (see
e.g.~\cite[Theorem~3.1.1]{Stroock-Varadhan06}.
By differentiating the equation with respect to $y$ we obtain
\begin{equation*}
\partial_s \,\mathbb{D}\varphi+\Oper{\mathbb{D}\varphi}-\mathbb{D} \tilde A\,\mathbb{D}\varphi=0
\end{equation*}
and then
\begin{equation*}
\frac 12\partial_t |\mathbb{D} \varphi|^2+
\frac 12 \Oper{|\mathbb{D}\varphi|^2} - \mathbb{D} \tilde A\mathbb{D}\varphi\cdot\mathbb{D}\varphi - |\mathbb{D}^2 \varphi|^2=0.
\end{equation*}
Since $\tilde A$ is monotone the quadratic form associated to
$\mathbb{D}\tilde A$ is nonnegative and therefore
\begin{equation*}
\partial_t |\mathbb{D} \varphi|^2+
\Oper{|\mathbb{D}\varphi|^2} \ge0.
\end{equation*}
A further application of the maximum principle yields \eqref{eq:49}.
\end{proof}
\section{Proof of Theorem \ref{teor:main}}
We split the proof in various steps.
Just to fix some notation, we consider a family $A_{n,m}$ of smooth, bounded,
Lipschitz, and monotone operators approximating $A:=B-\lambda I$ as
in Proposition
\ref{approxBn} and their rescaled version $\tilde A_{n,m}$ defined by
\eqref{eq:63}.
$\Opernm{n,m}\cdot$ are the associated differential operators
\begin{equation}
\label{eq:46bis}
\Opernm{n,m}\varphi(y,s):=\Delta_y \varphi(y,s)-\tilde A_{n,m}(y,s)\cdot
\nabla_y\varphi(y,s)\quad
\varphi(\cdot,s)\in
C^2({\R^d}),\quad
(y,s)\in {\R^d}\times [0,S_\infty),
\end{equation}
as in \eqref{eq:64}. Lemma \ref{le:bound} show that $\tilde A$
satisfy \eqref{eq:38}.
\noindent\underline{Step 1:} \emph{reduction to the monotone case
}$\lambda=0$.
When $\lambda\neq 0$ we apply the rescaling argument of section \ref{subsec:rescaling}:
we thus introduce the time rescaling ${\sf t}(s)$ defined by
\eqref{eq:36}
and the corresponding
measures $\sigma^i_s=\tilde\rho^i_{{\sf t}(s)}$ as in \eqref{eq:40},
which satisfy \eqref{eq:55} and \eqref{eq:3} for the rescaled
operators $\tilde A$ of \eqref{eq:41}.
Taking into account Remark \ref{rem:equivalent} and the fact that
$\sigma^i_s=\tilde\rho^i_{{\sf t}(s)}$, the thesis follows if we show
that
\begin{equation}
\label{eq:59}
\Cost h{\sigma^1_s}{\sigma^2_s}\le \Cost
h{\sigma^1_0}{\sigma^2_0}\quad\text{for every }s\in [0,S_\infty),
\end{equation}
(see \eqref{eq:61} for the definition of $S_\infty$).
\noindent\underline{Step 2:} {\em If
\begin{equation}
\label{eq:60}
\Cost h{\sigma^1_{s_1}}{\sigma^2_{s_1}}\le \Cost
h{\sigma^1_{s_0}}{\sigma^2_{s_0}}\quad\text{for every }0<s_0<s_1<S_\infty,
\end{equation}
then \eqref{eq:59} holds.}
When $h$ is bounded, \eqref{eq:60} implies \eqref{eq:59} by taking a
simple limit as $s_0\downarrow0$ and using the fact that the map
$(\sigma^1,\sigma^2)\mapsto \Cost h{\sigma^1}{\sigma^2}$ is continuous
with respect to weak convergence in $\mathcal{P}({\R^d})\times\mathcal{P}({\R^d})$.
If \eqref{eq:59} holds for every bounded Lipschitz cost, then it holds
for every {continuous and nondecreasing} \Killed{admissibl}e cost by Lemma \ref{le:cost_approximation}.
\noindent\underline{Step 3:} We claim the following:
{\em Let $\phi^1,\phi^2\in
C^\infty_{\mathrm c}({\R^d})$
be satisfying the constraint $\phi^1(y_1)+\phi^2(y_2)\le h(|y_1-y_2|)$
Then
\begin{equation}
\label{eq:64}
\int_{\R^d} \phi^1\,\d\sigma^1_{s_1}+\int_{\R^d} \phi^2\sigma^2_{s_1}\le
\Cost h{\sigma^1_{s_0}}{\sigma^2_{s_0}}+
\ell\, K_{n,m}
\end{equation}
where $\ell:=\sup_{\R^d} |\nabla\phi^1|+\sup_{\R^d} |\nabla\phi^2|$ and
\begin{equation*}
K_{n,m} := \int_{s_0}^{s_1}\int_{\R^d} |\tilde A_{n,m}-\tilde A|
\,\d \sigma^1_s\,\d s+
\int_{s_0}^{s_1}\int_{\R^d} |\tilde A_{n,m}-\tilde A|
\,\d \sigma^2_s\,\d s.
\end{equation*}}
Indeed, applying
\cite[Theorem~3.2.1]{Stroock-Varadhan06}
we can introduce the solutions $\varphi^1_{n,m},\varphi^2_{n,m}\in C^{2,1}_{\mathrm b}({\R^d}\times[s_0,s_1])$ of the backward equations
\begin{equation*}
\partial_s \varphi^j_{n,m}+\Opernm{n,m}\varphi^j=0\quad\text{in }
{\R^d}\times [s_0,s_1],\quad \varphi^j_{n,m}(\cdot,s_1)=
\phi^j(\cdot)\quad\text{in }{\R^d}.
\end{equation*}
Identity~\eqref{eq:3} shows that, for $j=1,2$,
\begin{align*}
\int_{\R^d} \varphi^j(\cdot,s_1)\,\d\sigma^j_{s_1}-\int_{\R^d} \varphi^j(\cdot,s_0)\,\d\sigma^j_{s_0} &=
\int_{s_0}^{s_1}\int_{\R^d} \big(\tilde A_{n,m}-\tilde A\big)\cdot
\nabla\varphi^j_{n,m}\,\d \sigma^j_s\,\d s
\\&\topref{eq:49}\le\ell \int_{s_0}^{s_1}\int_{\R^d} \big|\tilde A_{n,m}-\tilde A\big|
\,\d \sigma^j_s\,\d s
\end{align*}
Summing up the these equation for $j=1,2$ we obtain
\begin{equation} \label{eq:4}
\int_{\R^d} \varphi^1\,\d \sigma^1_{s_1}+
\int_{\R^d} \varphi^2\,\d\sigma^2_{s_1}\le
\int_{\R^d} \varphi^1_{n,m}(\cdot,s_0)\,\d \sigma^1_{s_0}+
\int_{\R^d} \varphi^2_n(\cdot,s_0)\,\d\sigma^2_{s_0}+\ell K_{n,m}
\end{equation}
Theorem \ref{thm:crucial} yields
$\varphi^1_{n,m}(y_1,s_0)+\varphi^2_{n,m}(y_2,s_0) \le h(|y_1-y_2|)$
which implies \eqref{eq:64}.
\noindent\underline{Step 4:}
\begin{equation}
\label{eq:65}
\limsup_{n\uparrow+\infty}\Big(\limsup_{m\uparrow+\infty} K_{n,m}\Big)=0.
\end{equation}
Let us first notice that setting $t_i:={\sf t}(s_i)$ and recalling that
${\sf t}'(s)={\mathrm e}^{-\lambda {\sf t}(s)}$ we have
\begin{align*}
\int_{s_0}^{s_1}\int_{\R^d} |\tilde A_{n,m}-\tilde A|
\,\d \sigma^1_s\,\d s&=
\int_{{\sf s}(t_0)}^{{\sf s}(t_1)}{\sf t}'(s)\int_{\R^d}
|A_{n,m}-A|\,\d\rho^i_{{\sf t}(s)}\,\d s=
\int_{t_0}^{t_1}\int_{\R^d}
|A_{n,m}-A|\,\d\rho^i_{t}\,\d t
\end{align*}
so that
\begin{align*}
K_{n,m}=K_{n,m}^1+K_{n,m}^2,\quad
K^j_{n,m}:=\int_{t_0}^{t_1}\int_{\R^d} |A_{n,m}-A|\,\d\rho^j_{t}\,\d
t\quad j=1,2.
\end{align*}
We can estimate $K^j_{n,m}$ by
\begin{align*}
K^j_{n,m}\le \int_{t_0}^{t_1}\int_{\R^d} |A_{n,m}-A_n|\,\d\rho^j_{t}\,\d t +
\int_{t_0}^{t_1}\int_{\R^d} |A_{n}-A|\,\d\rho^j_{t}\,\d t,
\end{align*}
observing that
by \eqref{eq:34}, \eqref{eq:62}, and the Lebesgue Dominated Convergence
Theorem we get
\begin{displaymath}
\lim_{m\uparrow+\infty}K^j_{n,m}=
\int_{t_0}^{t_1}\int_{\R^d} |A_{n}-A|\,\d\rho^j_{t}\,\d t.
\end{displaymath}
{Since $|A_n(x)|\le |{\sf A}^\circ(x)|\le |A(x)|=|B(x)-\lambda x|$ for every $x\in
{\R^d}$, the integrability assumption \eqref{Cond},} a further application of the Lebesgue Theorem, and \eqref{B approx}
yield
\begin{equation}
\label{eq:72}
\lim_{n\uparrow+\infty}\Big(\lim_{m\uparrow+\infty} K_{n,m}\Big)=
\int_{t_0}^{t_1}\int_{\R^d} |{\sf A}^\circ-A|\,\d\rho^j_{t}\,\d t.
\end{equation}
This last integrand is $0$ if $A$ coincides with the minimal
selection of ${\sf A}$, in particular when $A$ is continuous. In the
general case, \Killed{applying} the regularity result of
\cite{Bogachev-Krylov-Rockner01}
shows that $\rho^j_t\ll \Leb d$ for $\Leb 1$ a.e.\ $t\in (0,+\infty)$
and \eqref{eq:26} says that ${\sf A}^\circ=A$ $\Led$-a.e.\ in ${\R^d}$;
therefore the last integral of \eqref{eq:72} vanishes and we get \eqref{eq:65}.
\noindent\underline{Step 5:} \emph{conclusion.}
Thanks to \eqref{eq:65}, passing to the limit in \Killed{\eqref{eq:4}} {\eqref{eq:64}} we obtain
\begin{equation*}
\int_{\R^d} \phi^1\,\d \sigma^1_{s_1}+\int_{\R^d} \phi^2\,\d\sigma^2_{s_1}\le\Cost h{\sigma^1_{s_0}}{\sigma^2_{s_0}}.
\end{equation*}
Taking the supremum with respect to $\phi^1,\phi^2\in
C^\infty_{\mathrm c}({\R^d})$ and recalling Proposition \ref{Kant+}
we obtain \eqref{eq:60}.
\begin{remark}
\upshape
As it appears from the final argument of the previous step 4, in the case
when $A=B-\lambda I$ is the minimal selection ${\sf A}^\circ$ of ${\sf A}$ (in
particular when $B$ is continuous), we do not need to invoke the
regularity result of \cite{Bogachev-Krylov-Rockner01} to conclude
our proof.
\end{remark}
\begin{proof}[Proof of Corollary \ref{cor:main}]
For (a), it is sufficient to observe that $e^{\lambda t} \ge 1$; this implies $h(r) \le h_{\lambda t}(r)$ and so
\begin{equation*}
\Cost h{\mu^1_t}{\mu^2_t} \le \Cost {h_{\lambda t}}{\mu^1_t}{\mu^2_t}
\topref{++}\le \Cost h{\mu^1_0}{\mu^2_0}.
\end{equation*}
Similarly, for (a) and (b)
\begin{equation*}
e^{p\lambda t}\Cost h{\mu^1_t}{\mu^2_t}\le\Cost {h_{\lambda t}}{\mu^1_t}{\mu^2_t}
\topref{++}\le \Cost h{\mu^1_0}{\mu^2_0}.
\end{equation*}
We conclude recalling that
\begin{equation*}
W_p(\mu^1,\mu^2) = \Cost h{\mu^1}{\mu^2}^{1/p} \mbox{ with } h(r) = {|r|}^p
\end{equation*}
and applying (a) and (b).
\end{proof}
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
|
\section{Introduction} \label{sec1}
The majority of extrasolar planets that have been discovered by ground-based transit surveys to date are so-called ``Hot Jupiters'', with radii and masses similar to Jupiter and orbital periods less than $10$ days~(d). Ground-based transit surveys can be divided into two main types: (i) shallow surveys that use telescopes with apertures $\sim$10cm to monitor $\sim$50\,deg${}^2$ field of views (FOVs) over a magnitude range $V \lesssim 13$ \citep[eg.][]{2004ApJ...613L.153A,2007ApJ...656..552B,2007MNRAS.379..773S}, and (ii) deep surveys that use $\sim$1m telescopes to monitor $\sim$1\,deg${}^2$ FOVs over a magnitude range $V \lesssim 19$ \citep[eg.][]{2003AcA....53..133U,2009IAUS..253..333B,2009AJ....137.4368B}.
Most transit searches have adopted the shallow survey approach, largely because the required equipment is less expensive and more readily available than that needed to conduct a deep survey. Another appeal of shallow surveys is that $\sim$1m class telescopes can be used to perform follow-up spectroscopy on their bright targets, with signal-to-noise~ratios~(S/N) that are sufficiently high to detect radial velocity variations caused by an unseen Jupiter-mass companion. The fainter targets of deep surveys, on the other hand, require follow-up spectroscopy to be performed on $\sim$6-10m telescopes in order to achieve the same level of precision. Certainly, spectroscopic investigations of planetary atmospheres are only currently possible for the bright systems monitored by the shallow surveys \citep[eg.][]{2002ApJ...568..377C,2007Natur.447..183K}.
Nevertheless, shallow surveys do have a number of disadvantages compared to deep surveys. Perhaps most significantly, late-type main sequence stars are much rarer over the magnitude range covered by shallow surveys than they are in deeper surveys. More stars must therefore be monitored overall in a shallow survey than in a deep survey to have the same probability of observing a Hot Jupiter transit a late-type main sequence parent star.
Another drawback of ground-based shallow surveys is the lower spatial resolutions that they achieve, typically limited by point-spread functions (PSFs) with full-width-at-half-maximum~(FWHM) sizes~$\sim$20$''$, whereas deep surveys typically have much smaller PSFs of~$\sim$2$''$. Spatial resolution is important in transit surveys, because blending with physically unassociated stars can dilute any actual planetary signal or cause a transit by an eclipsing binary~(EcB) system to be mistaken for a planetary system. Although the number of stars per square degree of sky is much higher at the faint magnitudes covered by deep surveys, we find that the difference in PSF sizes is the dominant effect, making blending more prevalent in shallow surveys~(Table~\ref{tab:nblends} and \S{\ref{sec:embf}}).
The planetary yield of any photometric transit survey is far outweighed by the number of detected EcBs \citep[eg.][]{2008AJ....135..649W,2008AJ....135..850D}. Generally, the light curve produced by an EcB is easy to distinguish from one produced by a Hot Jupiter transiting its parent star. The transit durations are typically longer and the primary eclipse depths are typically deeper for EcBs than they are for transiting Hot Jupiters. Also, because both components of an EcB are luminous, a detectable secondary eclipse is usually produced. Unless both components of the EcB have very similar radii and luminosities, the secondary eclipse will not have the same depth as the primary eclipse, immediately revealing the non-planetary nature of the signal.
While most EcBs can be identified using one or more such diagnostics, in certain cases an EcB can mimic the photometric signal produced by a transiting Hot Jupiter. There are three main situations in which this can occur: (i) a late-type dwarf star eclipsing a larger, brighter dwarf star; (ii) two dwarfs undergoing grazing eclipses; (iii) two dwarfs eclipsing each other blended with the light of one or more unresolved stars. Meanwhile, detached EcBs with a giant component do not exist for periods less than $10$ days due to Roche lobe overflow \citep{2004A&A...421..241S}, and hence are not able to mimic the signal of a transiting Hot Jupiter.
It has not been uncommon for the rate of false planetary detections due to EcBs to be over an order of magnitude higher than the rate of true planet detections \citep[eg.][]{2006ApJ...644.1237O}. The identification and rejection of such objects can be far from trivial and poses a significant challenge to searches for transiting planets.
It is instructive to consider the yields of transit surveys in terms of what is already known about the properties of multiple star systems and Hot Jupiter systems from published empirical studies. \cite{2008ApJ...686.1302B} have developed a methodology to predict, \emph{a priori}, the number of planetary detections that a user-specified survey will make. Such estimates are helpful both in designing surveys and in evaluating the effectiveness of existing surveys. On the other hand, \cite{2003ApJ...593L.125B} made \emph{a priori} estimates of the rate of false detections in two synthetic transit surveys, one based on the shallow STARE survey and the other based on the then-planned Kepler space-based survey. A bottom-up understanding of the types of EcBs that are likely to cause false detections and the rates at which they occur is valuable both for interpreting the objects detected by transit surveys and for refining follow-up approaches used to reject non-planetary signals.
In this paper, we describe a method analogous to those of \cite{2008ApJ...686.1302B} and \cite{2003ApJ...593L.125B} to estimate the expected rates of false detections and Hot Jupiter detections in two synthetic surveys: (i) a ground-based shallow survey; and (ii) a ground-based deep survey. For each synthetic survey, the types of EcB configurations that are predicted to cause false detections are examined in some detail. These results are then compared with the reported experience of real transit surveys surveys, and the general implications for ground-based transit surveys are discussed.
Primarily, our study builds upon that of Brown (2003), by making more rigorous a number of the simplifying assumptions contained in that earlier work. For instance, unlike Brown, we explore the potential for genuine Hot Jupiter detections to be combined with faint blends, which can affect the estimated planetary radius (\S\ref{sec:configs}), whereas he only investigated unblended Hot Jupiter detections. We also do not ignore EcB secondary eclipses and use a more realistic S/N criterion in our definition of what constitutes a detection (\S\ref{sec:crit}). We have also attempted to account for the possibility that blended neighbors fainter than the ``formal'' upper magnitude limit of the survey can contribute in a non-negligible way to the observed photometric signals of stars within the formal magnitude range (\S\ref{sec:embf}). Furthermore, Brown only considered single-star blends in his calculations, but we have allowed for multiple-star blends (\S\ref{sec:phifb}). Our experience with Lupus-TR-3b in the SuperLupus survey \citep{2009IAUS..253..333B,2009AJ....137.4368B}, which we describe in \S\ref{sec2}, suggests that such cases might not be especially rare.
\section{Astrophysical false positives} \label{sec2}
Survey teams often report that false detections due to EcBs outnumber true planet detections by over an order of magnitude \citep[eg.][]{2006ApJ...644.1237O}. It should be noted that reports from shallow surveys dominate the literature, and this ratio may in general be lower for deeper surveys, as has been the experience of the SuperLupus survey \citep{2009IAUS..253..333B,2009AJ....137.4368B}. In any case, astrophysical false positives are a ubiquitous feature of any transit survey and the process of identifying and rejecting them has proven to be a time-consuming and, in many cases, subtle task.
Hot Jupiter transit signals can be imitated by: (i)~a late-type main sequence star eclipsing a larger main sequence star, (ii)~two main sequence stars undergoing grazing eclipses, or (iii)~two eclipsing main sequence stars blended with one or more other stars inside a photometric confusion radius, determined by either the instrumental PSF or the CCD pixel scale. For this reason, survey teams must follow up all promising candidates with additional observations to rule out the possibility that the photometric signal is caused by an EcB. Such observations include:
\begin{itemize}
\item Multi-color photometry: Apart from high-order stellar limb-darkening effects, planetary transit signals are color independent. Signals due to EcBs, on the other hand, typically have a stronger color dependence that can be detected by making photometric measurements in more than one passband.
\item High spatial resolution imaging: Often a candidate that was originally presumed to be a single star will be revealed to consist of two or more fainter stars by higher resolution imaging. By monitoring the flux changes of each component separately, it is sometimes possible to identify astrophysical false positives caused by blended EcBs by recognizing that one of the resolved components undergoes an eclipse that is too deep to be caused by a planetary transit.
\item Spectral typing: Spectroscopy can be used to identify the spectral type and luminosity class of the star that is being transited. If the star turns out to be a giant, then the photometric signal cannot be caused by a transiting planet since it would produce a flux change too small to be detected. Instead, the signal must be due to a blended EcB.
\item Radial velocity measurements: For a solar-type star, variations in radial velocity caused by an orbiting stellar companion will be $\sim$10\,km\,s${}^{-1}$, whereas the oscillations will only be $\sim$0.1\,km$\,$s$
{}^{-1}$ for a Jupiter-mass companion. Radial velocity measurements with a precision of $\sim$1\,km$\,$s${}^{-1}$ can therefore be used to identify cases in which the companion is stellar.
\end{itemize}
\cite{2003ASPC..294..409L}, \cite{2004AIPC..713..151C}, and \cite{2006ApJ...644.1237O,2007ApJ...662..658O} describe astrophysical false positives encountered as part of the Vulcan and TrES surveys, detailing how the non-planetary natures of candidate objects were revealed using one or more of the follow-up techniques listed above.
An example of a more challenging false positive, and the resource-intensive processes that were employed to reject it, has been detailed by \cite{2005ApJ...621.1061M}. In this case, no color dependence was observed in the eclipse and initial spectral typing showed that the primary, GSC 01944-02289, was an F dwarf. Radial velocity measurements indicated that the companion had a mass of $\sim$30$\,M_J$. All of these observations were consistent with a brown dwarf transiting the primary. However, careful measurement of asymmetries in the spectral lines and extensive modelling of possible blend scenarios ultimately showed that the observed signal was caused by a faint EcB blended with another unresolved star rather than a transiting brown dwarf.
The experience of the SuperLupus survey \citep{2009IAUS..253..333B,2009AJ....137.4368B} has also reinforced how difficult it can be to determine the precise nature of candidates, especially for the faint objects targeted by deep surveys. One promising candidate that has been identified as part of this survey, Lupus-TR-3b, appears to be a Hot Jupiter with radius $0.9R_J$ and mass $0.8M_J$ based on its light curve and follow-up spectroscopy performed at the Magellan~II telescope. However, high resolution imaging data obtained with the Magellan~I telescope, combined with image deconvolution analysis, has revealed that the host ``star'' in fact consists of up to seven separate stars that are unresolved in the SuperLupus photometry \citep{2009IAUS..253...55S}. While four of these stars are too faint to be responsible for the observed transit signal, it is unclear which of the remaining three stars is undergoing eclipse. Therefore, while unlikely, Lupus-TR-3b can still not be completely ruled out as being a blended EcB rather than a transiting planet. At the very least, if Lupus-TR-3b is a planet, it will have different properties to those listed above, which were derived under the assumption that the signal was caused by a planet transiting a single star.
These two examples illustrate the particular challenge posed to transit surveys by blending, both in terms of the potential for false detections to be made and the effect that it can have on estimated planetary properties. The \emph{a posteriori} processes used to identify blended neighbors, such as spectral line asymmetry analysis and iterative image deconvolution, are typically time-consuming and resource-intensive tasks, albeit important ones. Considering the problem in the forward direction, starting with well-understood statistics of stellar crowding and property distributions of multiple stars and Hot Jupiters provides another perspective on the problem, and is the approach taken in this work.
\section{Methodology} \label{sec3}
\subsection{The synthetic surveys}
We investigate the rates of astrophysical false detections (i.e.\ EcBs that mimic a Hot Jupiter signal) and genuine Hot Jupiter detections separately for a synthetic deep survey, based on the SuperLupus survey \citep{2009IAUS..253..333B,2009AJ....137.4368B}, and a synthetic shallow survey, loosely based on surveys such as HAT-Net~\citep{2007ApJ...656..552B} and TrES~\citep{2004ApJ...613L.153A}. The important properties of each survey are listed in Table~\ref{tab:survs}. For simplicity, and ease of comparison, the field of views~(FOVs) of both synthetic surveys are taken to be centered on the same Galactic coordinates as the SuperLupus survey $l=331.5^\circ$, $b=11^\circ$. The magnitude range of the deep survey is taken to be $15<V<19$, the same as SuperLupus, while for the shallow survey a magnitude range of $10<V<13$ is used. Using these parameters, expected star counts are binned according to luminosity class, spectral type, and apparent~$V$ magnitude for each synthetic survey using the Besan\c{c}on model of the Milky Way Galaxy \citep{2003A&A...409..523R}. Main sequence stars are found to comprise approximately $75$\% of all stars in our synthetic deep survey, and about $20$\% in our synthetic shallow survey.
\subsection{Estimating detection rates}
We estimate the rate of false detections~$\Omega_{FD}$ due to EcBs by integrating a probability density function~$\Phi_{FD}$ of the form:
\begin{eqnarray}
\frac{d^6 \, \Omega_{FD}}{d\Psi\cdot dV_p\cdot dP\cdot dq\cdot d(\cos i)\cdot df_B} &=& \Phi_{FD}(\Psi,V_p,P,q,\cos i,f_B) \label{eq:omfd}
\end{eqnarray}
where $\Psi$~is the spectral type of the primary, $V_p$~is the apparent magnitude of the primary, $P$~is the orbital period, $q$~is the secondary-to-primary mass ratio, $i$~is the orbital inclination, and~$f_B$ is the total blended flux. These properties determine the observed primary transit depth~$\delta_p$, observed secondary transit depth~$\delta_s$, and transit duration~$\tau$ (see \S\ref{sec:philc}). Hence in effect, the integrated density function~$\Phi_{FD}$ gives the probability that any given star monitored will be an EcB consisting of two dwarf components with light curve properties $\{\delta_p,\delta_s,\tau\}$ that satisfy our criteria for a false detection, which are presented in \S{\ref{sec:crit}. A detailed description of the integration process is given in \S{\ref{sec:integrating}}.
The rate of Hot Jupiter detections~$\Omega_{HJ}$ is estimated in exactly the same way, using a probability density function of the form:
\begin{eqnarray}
\frac{d^6 \, \Omega_{HJ}}{d\Psi \cdot dV_p\cdot dP\cdot dR\cdot d(\cos i)\cdot df_B} &=& \Phi_{HJ}(\Psi,V_p,P,R,\cos i,f_B), \ \ \label{eq:omhj}
\end{eqnarray}
where the binary mass ratio~$q$ is replaced with the planetary radius~$R$.
\subsection{EcB and Planet-Star configurations considered} \label{sec:configs}
Ground-based transit surveys are overwhelmingly sensitive to transiting planets with periods $P\lesssim 10$~days, due to observational window effects as well as the low geometric probability of a transit occurring for longer periods. Therefore, we only consider EcBs and Hot Jupiters with periods in this range. This allows us to ignore EcBs with a giant component, as detached giant binaries do not exist for periods $P<10$~days \citep{2004A&A...421..241S}. We also only consider Hot Jupiters transiting main sequence stars, since the photometric signal caused by a Hot Jupiter transiting a giant star is too faint to be detected \citep{2003ApJ...585.1056G}.
We divide both false detections and Hot Jupiter detections into two main types: blended and unblended. Blended detections are those for which the eclipsing system is accompanied by one or more unresolved neighbor stars. The resolution limit~$r_c$ is taken to be approximately equal to the size of the PSF or the pixel scale, whichever is largest. For the deep survey we use a value of $1.5''$, which is roughly equal to the spatial resolution achieved by the SuperLupus survey, and for the shallow survey we use a value of $20''$, which is typical of the PSFs in actual shallow surveys \citep[eg.][]{2004PASP..116..266B}. Blending has the effect of reducing the measured transit depths~$\delta_m$, according to:
\begin{eqnarray}
\delta_m &=& \frac{f_u}{f_u+f_B}\delta_u \label{eq:deltam}
\end{eqnarray}
where $f_u$~is the flux of the unblended EcB, $f_B$~is the total flux of the blended neighbors, and $\delta_u$ is the transit depth that would be measured if the EcB was unblended.
With regard to tertiary companions of EcBs, in reality some would fall within the confusion radius $r_c$ and some would fall outside. However, we chose not to estimate the true fractions of resolved and unresolved tertiary components, and instead perform the calculations separately under two limiting assumptions: (i) all tertiary companions are resolved, i.e.~fall outside the confusion radius~$r_c$ of the survey photometry; and (ii) all tertiary companions are unresolved. In the latter case, the tertiary companion is treated as an additional blended flux in Eq.~\ref{eq:deltam}. Higher order companions are not accounted for.
We do not take stellar multiplicity into account when calculating Hot Jupiter detection rates. That is, we treat every star as single and ignore possible complications associated with detecting Hot Jupiter transits in multiple systems such as blending by binary companions. \cite{2008ApJ...686.1302B} also made this simplification, and estimated that it could result in an overestimation of planet detection rates by at most $30\%$. However, we do consider blended cases of Hot Jupiter detections (i.e. planet $+$ host star $+$ physically unassociated stars), in addition to unblended Hot Jupiter detections.
\subsection{Criteria for a detection} \label{sec:crit}
We define an astrophysical false positive as an EcB that meets all of the following criteria:
\begin{enumerate}
\item Both stellar components are dwarfs with masses greater than $0.1 M_{\Sun}$, which approximately corresponds to the minimum hydrogen burning mass \citep{2000ARA&A..38..337C}. We do not include transits by brown dwarf companions in our false detection counts, since brown dwarfs orbiting within 5AU of main sequence stars are known to be extremely rare (the ``brown dwarf desert'') \citep{2000PASP..112..137M}.
\item The observed primary transit depth~is $\delta_p<0.05$ and has S/N$>\gamma$, where $\gamma$~is some minimum detection threshold. While it is possible for a Hot Jupiter transiting a late M dwarf to cause a fractional flux change greater than $0.05$, such cases are rare, and generally disregarded by survey teams due to the high probability that they are caused by an EcB.
\item The observed secondary transit depth~$\delta_s$ either has S/N$<\gamma$ or $\delta_p-\delta_s < \epsilon$, where $\epsilon$ is the measurement noise, i.e.\ it is either too shallow to be detected or it is indistinguishable from the primary eclipse.
\item The transit duration $\tau$ is less than $0.25$\,d, which approximately corresponds to the time taken for a Hot Jupiter with a period less than 10\,d to undergo a central transit of the disk of an F dwarf. Hot Jupiter transits of dwarfs earlier than F, which have larger radii and hence longer transit durations, will generally be too shallow for detection by ground-based surveys \citep{2003ApJ...585.1056G}. EcBs with periods $P<10\,$d, on the other hand, are much more likely to have transit durations $\tau>0.25\,$d due to the large relative disk sizes of the eclipsing bodies.
\item The effective magnitude~$V_{eff}$ of the EcB combined with any blended neighbors is within the magnitude limits of the survey.
\end{enumerate}
With the exception of the mass constraint on the secondary, our criteria for a Hot Jupiter detection are exactly the same as those listed above for an EcB false detection. Of course, the secondary eclipse that occurs when the planet passes behind the star does not produce an observable signal, so for a transiting Hot Jupiter condition (3) will always be satisfied. We also note that Hot Jupiter transits with primary transit depths $\delta_p>0.05$ or transit durations $\tau > 0.25$\,d are possible, but since such signals are far more likely to be caused by EcBs, survey teams generally discard such candidates if they arise.
\subsection{The extended magnitude range} \label{sec:embf}
In addition to the ``formal'' survey magnitude range, which we denote by~$V_b<V<V_f$ (i.e.\ $10<V<13$ for our synthetic shallow survey, and $15<V<19$ for our synthetic deep survey), we define an extended magnitude range~$V_b<V<V_e$ for each survey, where $V_f<V_e$. This is to allow for the possibility that blended stars with apparent magnitudes fainter than the formal survey cut-off~$V_f$ can contribute to the observed photometric signal. The faint limit~$V_e$ of the extended magnitude range is taken to correspond to the flux~$f_e \sim \sigma_m$, where $\sigma_m$ is the lowest root-mean square~(rms) scatter~$\sigma$ in the photometry achieved over the formal magnitude range. That is, as long as a star has flux at least at the level of the noise, we account for its contribution in the total measured flux.
To obtain analytic expressions for the fractional rms~$\sigma$ as a function of apparent magnitude~$V$, we fit parabolic curves of the form:
\begin{eqnarray}
\log_{10}\sigma &=& aV^2+bV+c \label{eq:rms}
\end{eqnarray}
by eye to the scatter in light curves as a function of $V$ in the SuperLupus survey~\citep[Fig.~2,][]{2009AJ....137.4368B} and the TrES survey~\citep[Fig.~2.2,][]{odonovanphdthesis}, corresponding to our synthetic deep and shallow surveys, respectively. Values for $a$, $b$, and $c$ for each survey are listed in Table~\ref{tab:survs}. Using Eq.~\ref{eq:rms}, we find $V_e=22$ for the deep survey and $V_e=16$ for the shallow survey.
In order for any star with magnitude $V_f<V<V_e$ to contribute to a photometric measurement, however, it must be blended with at least one other star that has magnitude $V_b<V<V_f$, i.e.\ within the formal survey magnitude limits. To simplify the analysis, we ignore those cases in which two or more unresolved stars each have a brightness fainter than the faint limit of the survey, but a total brightness greater than the faint limit. Such ``borderline'' cases will be relatively rare and should not have a significant effect on the final calculated result.
Supposing we have a FOV containing a total of $N$~stars, the probability~$p_n$ that any one of those stars will be blended with precisely~$n$ other stars is given by
\begin{eqnarray}
p_n &=& \frac{(N-1)!}{n!\,(N-n-1)!\,M^n}\left( 1-\frac{1}{M} \right)^{N-n-1} \label{eq:pn}
\end{eqnarray}
where $M=A_{FOV}/\pi r_c^2$ is the number of resolution elements in a FOV of area~$A_{FOV}$, given a confusion radius~$r_c$.
By letting $N$~in Eq.~\ref{eq:pn} be equal to the total number of stars~$\alpha$ in the FOV (both resolved and unresolved) with magnitude $V_b<V<V_f$, we obtain the fraction $\tilde{p}_n$ of stars that are blended with exactly $n$~other stars with magnitudes in the formal magnitude range. This means that if $\beta$ is the number of stars with magnitude $V_f<V<V_e$ between the faint limit $V_f$ of the formal magnitude range and the faint limit $V_e$ of the extended magnitude range, then only $(1-\tilde{p}_0)\beta$ of them will be able to contribute to the photometric measurements, i.e.\ those that are blended with at least one other star within magnitude in the formal range. The effective total star count~$N_\Sigma$ over the extended magnitude range is therefore:
\begin{eqnarray}
N_\Sigma &=& \alpha + (1-\tilde{p}_0)\beta \label{eq:nsig}
\end{eqnarray}
and the number of resolved objects~$N_R$ is given by:
\begin{eqnarray}
N_R &=& \alpha \, \sum_n{\frac{\tilde{p}_n}{n+1}} \label{eq:nr}
\end{eqnarray}
In \S\ref{sec:rates}, we use Eq.~\ref{eq:nr} to quote our detection rates as the number of detections per $N_R=10,000$ resolved stars.
Values for the crowding probability~$p_n$ in each synthetic survey computed using Eq.~\ref{eq:pn} and~\ref{eq:nsig} are listed in Table~\ref{tab:nblends}. The low blending probabilities~$p_n$ for $n \geq 1$ we obtain for the synthetic deep survey could mean that the case of Lupus-TR-3b, which consists of up to seven blended stars (\S\ref{sec2}), is an unlikely single event. On the other hand, of the seven blended stars, at least four have magnitudes $V>22$, making them too faint to be accounted for in our analysis which imposes a strict cut-off at $V=22$. Two of the remaining blends may also be a wide physical binary pair \citep{2009IAUS..253...55S}. It is also possible that the star counts we obtain from the Besan\c{c}on model are underestimated. For instance, the actual interstellar extinction may be lower than the value of 0.7\,mag\,kpc${}^{-1}$ that we use for the Besan\c{c}on model input.
\subsection{Integrating the probability density function} \label{sec:integrating}
To estimate detection rates, we integrate the probability density functions given by Eq.~\ref{eq:omfd} and \ref{eq:omhj} over appropriate property ranges (see below). Due to the independence of parameters in these relations, the probability density functions $\Phi_{FD}$ and $\Phi_{HJ}$ can be broken into a number of component probabilities. For the false detections:
\begin{eqnarray}
\Phi_{FD}(\Psi,V_p,P,q,\cos i,f_B) &=& \Phi_{d}(\Psi,V_p) \cdot \Phi_{b}(P,q) \cdot \Phi_i(\cos i) \cdot \Phi_f(f_B) \nonumber \\
&& \hspace{1.0cm} \times \ \ \Phi_w(P) \cdot \Phi_{lc}(\Psi,V_p,P,q,\cos i,f_B) \label{eq:phifd}
\end{eqnarray}
where $\Phi_d(\Psi,V_p)$ is the probability that the star will be main sequence with spectral type~$\Psi$ and apparent magnitude~$V_p$; $\Phi_b(P,q)$ is the probability that a main sequence star will be the primary component of a multiple system with period~$P$ and mass ratio~$q$; $\Phi_i(\cos i)$ is the probability that the orbital inclination is~$i$; $\Phi_f(f_B)$~is the probability that the total flux due to blended companions is~$f_B$; $\Phi_w(P)$ is the probability that at least three transits will be observed given the observing window function; and $\Phi_{lc}(\Psi,V_p,P,q,\cos i,f_B)$ is the probability that the transit signal of a binary consisting of two dwarfs with properties~$\{\Psi,V_p,P,q,\cos i,f_B \}$ resembles that of a transiting Hot Jupiter. The assumption of independence between the spectral type~$\Psi$ and the other properties of the binary, namely, the period~$P$ and mass ratio~$q$, is supported by empirical studies~\citep{2003A&A...397..159H}.
Similarly, for the planet detections, we can divide the probability density function of Eq.~\ref{eq:omhj} into:
\begin{eqnarray}
\Phi_{HJ}(\Psi,V_p,P,R,\cos i,f_B) &=& \Phi_{d}(\Psi,V_p) \cdot \Phi_{b}(P,R) \cdot \Phi_i(\cos i) \cdot \Phi_f(f_B) \nonumber \\
&& \hspace{1.0cm} \times \ \ \Phi_w(P) \cdot \Phi_{lc}(\Psi,V_p,P,R,\cos i,f_B) \label{eq:phihj}
\end{eqnarray}
where $\Phi_{b}(P,R)$ refers to the probability that a main sequence star will host a Hot Jupiter companion with orbital period~$P$ and radius~$R$, and all other terms are the same as for the false detection case given by Eq.~\ref{eq:phifd}. Again, we assume independence between the spectral type of the primary~$\Psi$ and the companion properties, which in this case are the planet radius~$R$ and orbital period~$P$. While correlations are beginning to emerge between system parameters as more transiting planet systems are discovered, such as between stellar mass and planet mass \citep{2007ApJ...670..833J}, and planet mass and period \citep{2008ApJ...677.1324T}, we note that the observed scatter is still large\footnote{http://exoplanet.eu/}. Therefore, while our assumption of independence between the system parameters mentioned above is a simplification, it is reasonable for the purpose of this study and is unlikely to have a substantial effect on our final results.
For the EcB false detections, integration is performed numerically over the following ranges:
\begin{itemize}
\item All spectral types $\Psi \in \left[\mbox{O}0,\mbox{M}9 \right]$, with a step size $d\Psi$ of five spectral sub-types.
\item The extended magnitude range $V_p \in \left[ 15,22 \right]$ for the deep survey and $V_p \in \left[ 10,16 \right]$ for the shallow survey, with a step size $dV_p=0.5$ mag.
\item Periods $P \in \left[ 1,10 \right]$, with a step size $dP=0.5$ days.
\item Mass ratios $q \in \left[ 0,1 \right]$ with a step size $dq=0.1$.
\item Orbital inclinations $\cos i \in \left[ 0,1 \right]$ with a step size $d(\cos i) = 0.01$.
\item For the blended cases, fluxes $f_B \in \left[ f_f,f_b \right]$ where $f_f$~is the flux corresponding to the faint magnitude limit of the extended magnitude range and $f_b$~is the flux corresponding to the bright magnitude limit. A step size $df_B$ corresponding to half a magnitude is used. For the unblended cases, $f_B$ is not integrated over, i.e. $\Phi_f(0)=1$ in Eq.~\ref{eq:phifd} and~\ref{eq:phihj}.
\end{itemize}
For the Hot Jupiter detections, spectral types~$\Psi \in [\mbox{F}0,\mbox{M}9]$ are integrated over, since dwarfs of spectral type earlier than about $\mbox{F}0$ are too large for a Hot Jupiter transit to be detectable \citep{2003ApJ...585.1056G}. We integrate over Hot Jupiter radii~$R\in [0.8,1.6]R_J$ with a step size $dR=0.1R_J$. We choose this range of radii because it encompasses the vast majority of Hot Jupiters discovered via transits at the time of writing\footnote{http://exoplanet.eu/}. All other integration ranges are the same as for the EcB false detections listed above.
In the following subsections, we describe the observational data and assumptions that are used to define the individual components of the probability densities in Eq.~\ref{eq:phifd} and~\ref{eq:phihj}.
\subsubsection{Stellar property distribution $\Phi_d(\Psi,V_p)$}
We use star counts from the Besan\c{c}on model \citep{2003A&A...409..523R} to calculate the probability $\Phi_d(\Psi,V_p)\,d\Psi\,dV_p$ that a star will be main sequence with spectral type $\Psi$ and apparent magnitude~$V_p$:
\begin{eqnarray}
\Phi_d(\Psi,V_p) &=& \frac{N(\Psi,V_p)}{N_{\Sigma}}
\end{eqnarray}
where $N(\Psi,V_p)$ is the number of main sequence stars with spectral type~$\Psi$ and apparent magnitude~$V_p$, and $N_{\Sigma}$ is the number of stars of all luminosity classes, spectral types, and magnitudes in the extended magnitude range given by Eq.~\ref{eq:nsig}.
\subsubsection{Companion property distributions $\Phi_b(P,q)$ and $\Phi_b(P,R)$}
We assume that the period~$P$ of a stellar binary is independent of the mass ratio~$q$, i.e.
\begin{eqnarray}
\Phi_b(P,q) &=& \Phi_P(P)\cdot\Phi_q(q) \label{eq:phipq}
\end{eqnarray}
where $\Phi_P(P)$ is the probability that a main sequence star will be the primary of a binary system of period~$P$, and $\Phi_q(q)$ is the probability that a main sequence binary system will have mass ratio~$q$. The distributions for $\Phi(P)$ and $\Phi_q(q)$ are taken from the empirical study of spectroscopic binaries by \cite{2003A&A...397..159H}, who found no evidence for the period being dependent on the mass ratio over the range $0<P<10$\,d. This is consistent with our assumption of independence between these two properties (Eq.~\ref{eq:phipq}). When performing calculations for the limiting case of all tertiary companions being unresolved~(see \S\ref{sec:configs}), we take the fraction of binary systems with tertiary components as a function binary period from \cite{2006A&A...450..681T}.
Similarly, we assume that the period~$P$ and radius~$R$ of a Hot Jupiter are independent, i.e.
\begin{eqnarray}
\Phi_b(P,R) &=& \Phi_P(P)\cdot\Phi_R(R)
\end{eqnarray}
where $\Phi_P(P)$ is the probability that an F0-M9 main sequence star will host a Hot Jupiter of period~$P$, and $\Phi_R(R)$ is the probability that a Hot Jupiter will have radius~$R$. The distribution for $\Phi_P(P)$ is obtained from the empirical estimates of \cite{2007A&A...475..729F}, and for simplicity we assume a uniform distribution for~$\Phi_R(R)$ in the range of radii we consider here, namely 0.8-1.6\,$R_J$. We chose to take the period distribution $\Phi_P(P)$ of \cite{2007A&A...475..729F}, which was derived from the results of the OGLE transit survey, as opposed to analogous distributions obtained from radial velocity surveys \citep[eg.][]{2008PASP..120..531C} due to the inherent metallicity bias in the latter. See \S{2.2} of \cite{2008ApJ...686.1302B} for a fuller discussion of this issue.
\subsubsection{Inclination distribution $\Phi_i(\cos i)$}
Orbits are assumed to be orientated randomly in three-dimensional space, giving $\Phi_i(\cos i)=1$.
\subsubsection{Blending distribution $\Phi_f(f_B)$} \label{sec:phifb}
The probability~$\Phi_f(f_B)$ that a star is blended with one or more other stars with a total blended flux~$f_B$ is
\begin{eqnarray}
\Phi_f(f_B) &=& \sum_n{\tilde{\Phi}_f(n,f_B)} \label{eq:phifb1}
\end{eqnarray}
where $\tilde{\Phi}_f(n,f_B)$ is the probability that the blend consists of exactly $n$~other stars with total flux $f_B$, and is given by
\begin{eqnarray}
\tilde{\Phi}_f(n,f_B) &=& \sum_{\varphi}{\left[ \prod_{i=1}^n{N_i} \right] }\frac{1}{M^n}\left( 1-\frac{1}{M} \right)^{N_{\Sigma}-n-1} \label{eq:phifb2}
\end{eqnarray}
where $\varphi$ denotes all combinations of fluxes $f_1, f_2, \ldots, f_n$ that satisfy
\begin{eqnarray}
f_B &=& \sum_{i=1}^n{f_i}
\end{eqnarray}
$N_i$ is the number of stars in the field with flux $f_i$, $M$~is the same as in Eq.~\ref{eq:pn}, and $N_\Sigma$ is the effective total number of stars in the FOV as given by Eq.~\ref{eq:nsig}. Values for the blended flux probabilities $\Phi_f(f_B)$ computed using Eq.~\ref{eq:phifb1} and~\ref{eq:phifb2} are plotted for each synthetic survey in Fig.~\ref{fig:fblends}.
\subsubsection{Window function $\Phi_w(P)$} \label{subsec:window}
For both surveys, we use the window function~$\Phi_w(P)$ provided in Fig.~2 of \cite{2009IAUS..253..333B}. This function is an approximation of the real window function for the SuperLupus survey, and is calculated assuming 100 contiguous observing nights with weather statistics that match actual weather logs from the Siding Spring Observatory site.
\subsubsection{Light curve detection probabilities $\Phi_{lc}(\Psi,V_p,P,q,\cos i,f_B)$ and $\Phi_{lc}(\Psi,V_p,P,R,\cos i,f_B)$} \label{sec:philc}
Assuming that the orbit is circular, and given the properties $\{\Psi,V_p,P,q,\cos i,f_B\}$, we can determine the observed transit depths, $\delta_p$ and~$\delta_s$, and the transit duration~$\tau$ as follows. First, we use the spectral type~$\Psi$ of the primary to find its mass~$M_p$, radius~$R_p$, and luminosity~$L_p$ using a fit to the relations for main sequence stars tabulated by~\cite{2000asqu.book.....C}. The luminosity combined with the apparent magnitude~$V_p$ then provides the distance~$d$ to the system. The primary mass and the mass ratio~$q$ yield the secondary mass~$M_s$, and hence the secondary radius~$R_s$ and luminosity~$L_s$. The distance~$d$ and secondary luminosity~$L_s$ is next used to calculate the flux~$f_s$ from the secondary. The masses, radii, and period~$P$ provide the orbital semi-major axis~$a$ via Kepler's Third Law.
With these properties, we have enough information to compute the observed primary transit depths, $\delta_p$ and $\delta_s$, and the transit duration~$\tau$ using standard relations such as those set out in \cite{1999poss.conf..189S}. We assume a quadratic limb darkening law with solar coefficients taken from \cite{2000asqu.book.....C}, regardless of spectral type.
The S/N of the transit signature in each light curve is then calculated for each of $\delta_p$ and $\delta_s$ via:
\begin{eqnarray}
\mbox{S/N} &=& \frac{\delta_i}{\sigma}\sqrt{\mu} \label{eq:sn}
\end{eqnarray}
where~$\sigma$ is the measured out-of-transit photometric rms scatter given by Eq.~\ref{eq:rms}, and~$\mu$ is the number of in-transit data points, which we take to be:
\begin{eqnarray}
\mu &=& \frac{\eta h \upsilon}{\omega}\frac{\tau}{P}
\end{eqnarray}
where $\eta$~is the number of observing nights, $h$~is the observing time per night, $\upsilon$~is the fraction of usable nights, and $\omega$~is the measurement cadence. For simplicity, and ease of comparison, we use $\eta=100$ nights, $h=8$ hours, $\upsilon=0.67$, and $\omega=6$ minutes for both synthetic surveys, which is comparable with the actual experience of the SuperLupus survey \citep{2009IAUS..253..333B,2009AJ....137.4368B}.
Equipped with the transit depths $\delta_p$ and~$\delta_s$, the associated S/N ratios, the transit measurement noise~$\epsilon\equiv \sigma / \sqrt{\mu}$ and the transit duration, we are then able to determine if the EcB configuration satisfies the criteria for a detection given in \S{\ref{sec:crit}}. The probability is then simply:
\begin{eqnarray}
\Phi_{lc}(\Psi,V_p,P,q,\cos i,f_B)=\left\{
\begin{array}{cl}
1 & \ \ \ \textnormal{if the detection criteria are met} \\
& \\
0 & \ \ \ \textnormal{otherwise}
\end{array}
\right.\label{eq:thetap}
\end{eqnarray}
The process for the Hot Jupiter detections is directly analogous to that outlined above for false detections. The only difference is that instead of a mass ratio~$q$ being used to obtain the secondary mass, the Hot Jupiter mass~$M$ is obtained from the radius~$R$ by assuming all Hot Jupiters have the same density~$\rho_J$ as Jupiter, i.e. $M=4\pi R^3 \rho_J/3$. This is a simplification, considering the observed range in mean densities of the transiting planets discovered so far, from abnormally-inflated planets such as TrES-4 with only $\sim$0.1 times the mean density of Jupiter \citep{2007ApJ...667L.195M}, to objects such as XO-3b \citep{2008ApJ...677..657J} and WASP-18b \citep{2009Natur.460.1098H} with mean densities $\sim$10 times that of Jupiter. However, this simplifying assumption will only have a minor effect on the final results. This is because we only use the mass of the planet in the calculation of the orbital semi-major axis~$a$, which is then used with the orbital inclination $\cos i$ to calculate the transit latitude $a \cos i$. Since the semi-major axis is $a\sim \left( M_p+M_s \right)^{1/3}$ for a given orbital period~$P$ according to Kepler's Third Law, and $M_s << M_p$ in the case of a star-planet system, the uncertainty in the mass of the planet will only have a negligible effect on the calculated semi-major axis and hence on the calculated transit latitude.
It must be noted that Eq.~\ref{eq:sn} assumes Poisson statistics for the noise, i.e.\ that the noise decreases according to $\sim 1/\sqrt{\mu}$ (white noise). In practice, however, effects such as varying airmass, weather conditions, moon phase, and telescope tracking cause systematic trends in the measured light curves that can be correlated over the timescale of a planetary transit (red noise) \citep{2006MNRAS.373..231P}. This means that a measured light curve will in fact have a lower S/N than predicted by Eq.~\ref{eq:sn}. Rather than attempting to model the white noise and red noise components individually, as was done by \cite{2008ApJ...686.1302B}, we simply require that the underestimated S/N calculated using Eq.~\ref{eq:sn} exceeds a higher cut-off than a survey team would typically require. Specifically, we calculate separate detection rates for S/N cut-offs of 10, 20, and 30. By varying the S/N cut-off, we are also able to investigate the sensitivity of the yield to this component of the detection criteria.
\section{Results}
\subsection{Predicted rates} \label{sec:rates}
Calculated detection rates for the synthetic deep and shallow surveys obtained using S/N detection cut-offs of 10, 20, and 30 are reported in Tables~\ref{tab:tcrrates} and~\ref{tab:tcurates}, assuming separately that all tertiary components in multiple star systems are resolved and unresolved, respectively (\S\ref{sec:configs}). Values are expressed as the number of detections per $10,000$ resolved stars monitored with sufficient photometric precision to detect a Hot Jupiter transit. Recall that the number of resolved stars~$N_R$ in the survey field is given by Eq.~\ref{eq:nr}. Values without parentheses are those that have been calculated with a window function~(\S\ref{subsec:window}), while those values with parentheses have been calculated ignoring window effects. In both synthetic surveys, the inclusion of our window function does not significantly reduce the predicted number of detections. This reflects the fact that the window function we used~(\S\ref{subsec:window}) gives almost complete recovery for signals with periods $\lesssim5$~days.
The results in Tables~\ref{tab:tcrrates} and~\ref{tab:tcurates} show that the total false detection rates are hardly changed if we assume that all tertiary components are resolved or unresolved in the survey photometry. This is not surprising, considering that in our calculations, an unresolved tertiary component only has the effect of increasing the overall flux by an amount less than or equal to that of the primary (since we assume a mass ratio $q_3 \leq 1$, \S{\ref{sec:configs}). Therefore, some of the EcB transits that are too deep to be mistaken for a planetary signal if the diluting effect of the tertiary component is not accounted for will become shallow enough to cause a false detection when the additional flux is included (Eq.~\ref{eq:deltam}). On the other hand, some of the EcB transits that are only just deep enough to be detected when they are not blended with an unresolved tertiary component will instead fall below the noise when the additional flux of an unresolved tertiary component is included in the transit depth calculation. The fact that our overall false detection rates show hardly any change if we assume all tertiary components are resolved or unresolved indicates that these competing effects are roughly equal in magnitude.
The results in Table~\ref{tab:tcurates} suggest that hierarchical triple systems are likely to constitute a substantial fraction of false detections in transit surveys close to the galactic plane. For the limiting case that all tertiary components are unresolved, in both synthetic surveys we find that hierarchical triples cause $\sim$5 times more false detections than EcBs without tertiary components. This result can be explained by the fact that short period binaries are highly likely to have tertiary components \citep{2006A&A...450..681T}, and they also have a relatively high geometric probability of transiting \citep[see Eq.~2 of ][]{1999poss.conf..189S}.
Overall, we find $\sim$2-4 false detections and $\sim$1-3 Hot Jupiter detections per $10,000$ stars monitored in the deep synthetic survey, depending on the S/N cut-off used to define a detection (\S\ref{sec:philc}). False detections are divided roughly evenly between blended and unblended detections, whereas only about $20\%$ of planet detections are blended.
These results compare well with the SuperLupus survey in which a total of 13 planetary candidates have been identified from a sample of $\sim$25,000 stars monitored~(Bayliss 2009, private communication). Of these, however, at least five do not meet the all of the detection criteria used in the present study. They either have transit durations longer than $0.25$~days or secondary eclipses that were detected once greater phase coverage was obtained for the orbit. This leaves a total of eight candidates, six of which have not yet been followed up. Of the two candidates that have been followed up, one is most likely a transiting Hot Jupiter (Lupus-TR-3b, \S\ref{sec2}) and the other has been confirmed as a false positive.
In the synthetic shallow survey, we find $\sim$3-4 false detections, similar to the synthetic deep survey, but only $\sim$0.1-0.6 Hot Jupiter detections per $10,000$ stars monitored. In other words, one planet is detected for every $\sim$15,000-100,000 stars monitored, and for every planet detected we obtain $\sim$5-25 false detections, depending on the S/N cut-off used in the calculations. Approximately $90\%$ of the false detections in the shallow survey are caused by blended EcBs, whereas blended and unblended planet detections occur in roughly equal amounts.
These results seem broadly consistent with the commonly-reported experience of shallow surveys that astrophysical false positives outnumber planet detections by an order of magnitude. For instance, as part of the TrES survey, \cite{odonovanphdthesis} used the 10cm Sleuth telescope to monitor $\sim$150,000 stars across 19 fields with photometric precision $\lesssim 2\%$. From this, 67 candidates were identified but only 4 were confirmed as genuine planet detections. By comparison, our results predict for every $150,000$ stars monitored there will be $\sim$2-12 Hot Jupiter detections and $\sim$40-70 false detections, depending on the S/N cut-off used and whether or not a window function is included. The agreement to within a factor of $\sim$2 is reasonable, considering the idealized nature of our detection criteria, our assumption that full orbital phase coverage is obtained, and the fact that our calculations were made for a single field location whereas Sleuth monitored several fields spread over a range of galactic latitudes.
Another point worth noting about the results for the synthetic shallow survey listed in Tables~\ref{tab:tcrrates} and~\ref{tab:tcurates} is that the rate of false detections caused by unblended configurations remains constant within the numerical precision as the S/N cut-off is increased from 10 to 30. This is due to the fact that as the S/N cut-off is increased, less primary transits \emph{and} less secondary transits are detected above the noise. The former has the effect of decreasing the false detection rate, whereas the latter has the effect of increasing it. Since no net change is observed in the calculated detection rates, both opposing effects must be equal in magnitude for the unblended false detections in the shallow survey.
More specifically, the type of configuration that is likely to cause a false detection at low S/N cut-offs but not at higher S/N cut-offs are those for which: (i)~the primary transit signal is only just deep enough to be detected above the noise at the lower S/N cut-offs but not at the higher cut-offs; and (ii)~the secondary transit signal is either shallow enough to escape detection or is indistinguishable from the primary transit. Such configurations will lead to a reduction in false detections as the S/N cut-off is increased. At the same though, there will be configurations for which the primary transit is deep enough to be readily detectable (i.e.~S/N$\geq30$) but not so deep that it is ruled out as a planetary candidate (i.e. $\delta_p<0.05$), while the secondary transit is only shallow enough to evade detection at the higher SN cut-offs. Such cases will increasingly contribute to the number of false detections as the S/N cut-off is increased. Both of the situations described here can arise from grazing configurations of an F or G star and a smaller star such as a late K or M dwarf (see \S{\ref{sec:dtrends}}), the difference being that the transit latitudes will be lower in the former case than they are in the latter.
In the synthetic deep survey, however, the unblended false detection rate decreases as the S/N cut-off is increased from 10 to 30. The reason for this is the lower photometric precision attained in the deep survey compared to the shallow survey (Eq.~\ref{eq:rms} and Table~\ref{tab:survs}). As a result, there are relatively fewer configurations with a distinct secondary transit that is detectable at lower S/N cut-offs but not detectable at higher S/N cut-offs than in the shallow survey. Instead, most of the unblended false detections in the deep survey have primary transit depths that are only just detectable above the noise at the lower S/N cut-off, and secondary transits that are either too shallow to be detected or of a similar depth to the primary transit, as in the first case described above for the synthetic shallow survey.
For both synthetic surveys, the rate of blended false detections decreases as the S/N cut-off is increased. Again, this is due to configurations with primary transits that are detectable at the low S/N cut-offs becoming undetectable as the cut-off is increased, and a secondary transit that is either too shallow to be detected above the noise or not discernible from the primary transit. This mainly occurs for cases in which significantly more flux comes from the blended stars than from the EcB itself (\S{\ref{sec:dtrends}}), resulting in shallower transit depths. Meanwhile, since planetary occultations only exhibit a primary transit signal, the rates of blended and unblended planet detections also decrease as the S/N cut-off is increased in both surveys.
We emphasize that our current study is intended to be illustrative: it should not be assumed that the results reported here for a FOV centered on galactic latitude $b=11^\circ$ also apply to surveys conducted further away from the galactic plane, where the stellar population statistics and crowding probabilities will be different. A proper investigation of how detection rates are affected by the galactic latitude of the survey field is left to a future analysis.
\subsection{Predicted detection trends} \label{sec:dtrends}
The rate of false detections and Hot Jupiter detections in both the synthetic shallow and deep surveys are shown in Fig.~\ref{fig:res12} as a function of the spectral type of the primary star being eclipsed. These distributions closely follow the star counts returned by the Besan\c{c}on model. However, in both surveys, the planet detections are skewed towards late-type dwarfs. This is due to the deeper, and hence more readily detectable, eclipses that result when Hot Jupiters transit late-type dwarfs which have comparatively small radii and low luminosities. For the false detections, we found that increasing the S/N cut-off between 10 and 30 biases the relative fraction of detections slightly towards earlier spectral types (F5 and G0). This occurs because the photometric rms scatter decreases towards brighter magnitudes (\S{\ref{sec:embf}}) where the fraction of earlier spectral types increases relative to later spectral types. However, this effect is quite small, with the relative fractions of detections changing by $\lesssim 5$\% for a given spectral type, and so we have only included the plotted results for the S/N$\geq20$ cut-off scenario in Fig.~\ref{fig:res12}.
Fig.~\ref{fig:dres34} and~\ref{fig:sres34} show the rate of false detections and Hot Jupiter detections as a function of effective magnitude in the deep synthetic survey and shallow synthetic survey, respectively. Results are shown for the S/N$\geq10$ and S/N$\geq30$ cases to illustrate the effect of that varying the S/N cut-off has on the relative fractions of detections. In both synthetic surveys for the S/N$\geq10$ cut-off and S/N$\geq30$ cut-off, the rate of false detections and Hot Jupiter detections initially increases with increasing magnitude, simply reflecting the increasing numbers of stars at fainter magnitudes. A maximum is then reached, followed by a decline in detection rates over the faintest portion of the survey magnitude range. The decline in rates occurs due to the decrease in photometric precision at fainter magnitudes. For the same reason, Fig.~\ref{fig:dres34} and~\ref{fig:sres34} show that the peak in the relative fraction of detections is shifted towards successively brighter magnitudes as the S/N cut-off used to define a detection is increased from 10 to 30. An interesting feature of Fig.~\ref{fig:dres34} and~\ref{fig:sres34} is that the peak in the relative fraction of planet detection rates tends to occur at brighter magnitudes with higher photometric precisions than it does for false detections, especially as the S/N detection cut-off is increased. This is because the number of EcBs with $\delta_p\sim0.05$ is greater than for Hot Jupiter transits, with the latter tending to produce shallower eclipses that are harder to detect above the noise.
Fig.~\ref{fig:res56} shows how the number of blended detections varies with the fraction of blended light in both the synthetic deep and shallow surveys. We found that increasing the S/N detection cut-off slightly shifted the relative fraction of false detections towards fainter blends (i.e.~blends contributing less of the overall flux) and the relative fraction of Hot Jupiter detections towards brighter blends. However, as in the case of relative detection rates versus primary spectral type (Fig.~\ref{fig:res12}), this effect is small ($\lesssim 5$\%), so we have only plotted the results for S/N$\geq20$ in Fig.~\ref{fig:res56}. These results show that around $35\%$ of the blended false detections in the deep survey are due to bright blends; namely, those blended false detections in which the EcB contributes $<30\%$ of the total flux. In the shallow survey, roughly $80\%$ are due to bright blends. For blended planet detections in both surveys we find that blends of up to about $50\%$ can occur, which has implications for planetary parameter estimates (see \S\ref{sec:disc}).
The EcB configurations responsible for unblended and blended false detections are illustrated in Fig.~\ref{fig:res7} and~\ref{fig:res8}, respectively. The primary purpose of these plots is to qualitatively illustrate the most common false detection configurations. Results are only shown for the case of S/N$\geq20$ since no appreciable change is observed as the S/N detection cut-off is increased from 10 to 30. We see that the majority of unblended false detections are caused by grazing ``twins'', which have indistinguishable primary and secondary eclipses. The remainder of cases are made up of: (i) grazing transits by stars with different radii; and (ii) low-latitude transits of a primary by a much smaller companion. For the blended false detections, there are two broad categories: (i) bright blends with a low-latitude EcB; and (ii) fainter blends with a grazing EcB.
\section{Discussion and Implications} \label{sec:disc}
The estimated false detection rates listed in Tables~\ref{tab:tcrrates} and~\ref{tab:tcurates} are similar for the deep and shallow synthetic surveys, but the estimated planet detection rates are $\sim$5 times higher for the deep survey than they are in the shallow survey. There are two main reasons for this: (i) $\sim$75\% of all stars in the deep survey FOV are dwarfs, whereas only $\sim$20\% are dwarfs in the shallow survey FOV, according to the Besan\c{c}on model star counts; and (ii) it is more difficult to detect a blended transiting Hot Jupiter than an unblended one, and the blend probability is higher in the shallow survey due to the larger confusion radius (Table~\ref{tab:nblends}). In practice, however, the comparatively low ratio of false detections to planet detections in the deep survey is at least partially offset by the inevitable difficulty of conducting suitable follow-up observations to verify or reject their faint candidates.
Extrapolating from our synthetic surveys, $\sim$20\% of all false detections in shallow surveys and $\sim$65\% of all false detections in deep surveys close to the galactic plane will be due to EcBs that are either unblended or which contribute more than $30\%$ of the light in a blended configuration. For the bright targets of shallow surveys, such cases will generally display radial velocity variations that are readily detectable with $1$m-class telescopes. For the faint targets of deep surveys, however, acquiring follow-up spectroscopy with sufficient S/N will always be a challenge, but these configurations present the most accessible configurations to rule out.
We find that $\sim$65\% of all false detections in shallow surveys close to the galactic plane will be caused by bright blends with a physically unassociated giant. The equivalent fraction is $\sim$10\% in deep surveys close to the galactic plane, reflecting the lower frequency of giants over fainter magnitude ranges. Typically, such cases are relatively easy to rule out, either through spectral typing or reference to existing stellar catalogs, at least in the case of the bright shallow survey targets.
The other bright blend false detections, which constitute $\sim$15\% of all false detections in shallow surveys and $\sim25\%$ of all false detections in deep surveys close to the galactic plane, will be due to blending with physically unassociated dwarfs. It is this latter case of blended EcBs that pose the greatest challenge to follow-up observations. Multi-color photometry can often be useful to reveal a color-dependence in the eclipse depth. However, the task of unmasking such false positives would be considerably more difficult if the colors of the EcB components happen to be similar and the blended dwarf is a rapid rotator, in which case its broadened spectral lines may conceal those of the fainter EcB.
\section{Conclusion}
We have investigated the rates of both false detections caused by EcBs and genuine Hot Jupiter detections in ground-based transit surveys from a bottom-up perspective, using empirically-determined property distributions of multiple star systems and Hot Jupiter systems as our input. In particular, we find that in both deep and shallow surveys: (i) significant numbers of false detections can be caused by faint EcBs blended with the light of one or more foreground dwarfs, a configuration which experience shows can produce particularly insidious impostors, and (ii) for genuine Hot Jupiter detections, up to $\sim$50\% of the system's light can be due to unassociated blends, which will affect the estimated planetary parameters if they are not accounted for. We therefore reiterate the conclusion of \cite{2009IAUS..253...55S}: namely, that it is important to identify or rule out blended neighbors for all candidates, including blended neighbors that are up to $\sim$5 magnitudes fainter than the primary.
Lastly, we note that the space-based missions CoRoT \citep{2007AIPC..895..201B} and Kepler \citep{2003ASPC..294..427B} may be particularly susceptible to false detections or planet parameter misestimation caused by blending. Due to the high photometric precision of these instruments, blended neighbors \emph{up to 8 to 10 magnitudes fainter than the primary} can make non-negligible contributions to the measured light curves if the signal that is being sought is that of an Earth-sized object transiting a dwarf star. Furthermore, since these surveys target planets with masses down to about an Earth mass, and the monitored stars are quite faint ( $V \lesssim 15$-$16$), radial velocity follow-up will be of little use in confirming their low-mass candidates. Constraining the presence of faint blended neighbors will therefore be of particular importance if a terrestrial-sized planet is to be secured with confidence in these surveys.
|
\section{Introduction and Background}\label{int}
Let $\mathbb{F}_q=GF(q)$ be a Galois field with $q$ elements, where $q$ is a
prime power. A $q$-ary linear {\em code} $C$ of length $n$ is a subset of $\mathbb{F}_q^{\,n}$. If
$C$ is a $k$-dimensional linear subspace of $\mathbb{F}^{\,n}_q$, then $C$ is a
$q$-ary {\em linear} code, denoted by $[n,k,d]_q$, where $d$ is the
{\em minimum distance} between any pair of codewords (the distance, or the Hamming distance, between two vectors is the number of coordinates they disagree). Given ${\bf v}\in\mathbb{F}^n_q$, denote by $supp({\bf v})$ the {\em support} of ${\bf v}$, that is, the set of its nonzero coordinate positions. For $a \in \mathbb{F}_q$
and a given $C$ denote $a\,C = \{a\,{\bf c}: ~{\bf c} \in C\}$. For two codes
$A$ and $B$ of the same length let $A + B$ be its direct product, i.e.
$A + B = \{{\bf a} + {\bf b}:~{\bf a} \in A,~{\bf b} \in B\}$.
A matrix $M$ is called monomial if there is exactly one nonzero entry
in each row and column.
Let $C$ be a linear code of length $n$ over $\mathbb{F}_q$, a finite field of $q$ elements, where $q$ is a prime power. The automorphism
group $Aut(C)$ of $C$ usually consists of all
$n\times n$ monomial matrices $M$ over $\mathbb{F}_q$ such that ${\bf c} M \in C$
for all ${\bf c} \in C$. In those cases when $q$ is a power of a prime number, then $Aut(C)$ also
contains all the field automorphisms of $\mathbb{F}_q$ which preserve $C$. Note that, for
binary codes, the automorphism group $Aut(C)$ coincides with the subgroup of the
symmetric group $S_n$ consisting
of all $n!$ permutations of the $n$ coordinate positions which send $C$ into itself.
For a $q$-ary code $C$ with minimum distance $d$ denote by
$e = \lfloor (d-1)/2 \rfloor$ its {\em packing radius}.
Given any vector ${\bf v} \in \mathbb{F}^n_q$, its {\em distance} to the code
$C$ is
\[
d({\bf v},C)=\min_{{\bf x} \in C}\{ d({\bf v}, {\bf x})\}
\]
and the {\em covering radius} of the code $C$ is
\[
\rho=\max_{{\bf v} \in \mathbb{F}^n_q} \{d({\bf v}, C)\}.
\]
Clearly $e\leq\rho$ and $C$ is a {\em perfect} code, if
$e=\rho$.
For a given $q$-ary code $C$ with covering radius
$\rho=\rho(C)$ define
\[
C(i)~=~\{{\bf x} \in \mathbb{F}_q^{\,n}:\;d({\bf x},C)=i\},\;\;i=1,2,...,\rho.
\]
\begin{definition}\label{de:1.1} A code $C$ is completely regular,
if for all $l\geq 0$ every vector $x \in C(l)$ has the same
number $c_l$ of neighbors in $C(l-1)$ and the same number
$b_l$ of neighbors in $C(l+1)$. Also, define
$a_l = (q-1){\cdot}n-b_l-c_l$ and note that $c_0=b_\rho=0$.
Define the intersection array of $C$ as $(b_0, \ldots, b_{\rho-1}; c_1,\ldots, c_{\rho})$.
\end{definition}
The group $Aut(C)$ induces an action on the set of cosets of $C$ in the
following way: for all $\phi\in Aut(C)$ and for every vector
${\bf v} \in \mathbb{F}_q^{\,n}$ we have $\phi({\bf v} + C) = \phi({\bf v}) + C$.
In~\cite{Sole} it was introduced the concept of completely transitive
binary linear code and in~\cite{giudici} there is a generalization leading to the concept of coset-completely transitive code.
Let $C$ be a linear code over $\mathbb{F}_q$ with covering radius $\rho$, then
$C$ is coset-completely
transitive if $Aut(C)$ has $\rho +1$ orbits when acting on the cosets of $C$.
Since two cosets in the same orbit have the same weight distribution, it is clear that
any coset-completely transitive code is completely regular. Obviously, it may exist completely regular codes which are not coset-completely transitive.
To illustrate this last point we will show an easy example. Take the repetition code $C$ of length three over the field $\mathbb{F}_q$, where $q=2^4$. The code $C$ has 16 elements and it is not coset-completely transitive, but completely regular with intersection array $( 45, 28; 1, 6 )$. Clearly, the automorphism group $Aut(C)$ as defined before has order $|Aut(C)|=15\cdot6\cdot4=360$ and if all the cosets in $C(2)$ are in the same orbit of $Aut(C)$ it should be that the number of cosets in $C(2)$ is a divisor of $360$. But, this is not true, since the number of cosets in $C(2)$ is $210$ (indeed, $210=16^2-1-15\cdot 3$).
It has been conjectured for a long time that if $C$ is a
completely regular code and $|C|>2$, then $e \leq 3$ \cite{Neum}.
Hence, the existing completely regular codes have a small error correcting
capability. With respect to the covering radius, Sol\'{e} in
\cite{Sole} uses the direct sum of $\ell$ copies of fixed perfect
binary codes of length $n$, with $e=1$, to construct infinite families of
binary completely regular codes of length $n{\cdot} \ell$ with
covering radius $\rho =\ell$. Thus, the
covering radius of the resulting code is growing linearly to infinity with
the length of the code. In~\cite{Rif2} it was described a method of constructing of linear completely regular codes with arbitrary covering radius, which is constant when the length is growing. Now, in the present paper we describe a new method, based on lifting perfect (Hamming) $q$-ary codes up to a new finite field over $q^r$. This method allow us to construct new completely regular codes with arbitrary covering radius, growing to the infinity with the length of the code, but not in a linear way like in~\cite{Sole} but in a logarithmic way. We compute the intersection array for all these new completely regular codes. Finally, we show that the Hamming codes are the only codes that, after lift the ground field, give completely regular codes
\section{Lifting perfect $q$-ary codes}
Let $C$ be a $q$-ary linear nontrivial code of length $n$ over the ground field $\mathbb{F}_q$ (hence, $1 < |C| < q^{n-1}$) with minimum distance $d\geq 3$. Let $H$ be a parity check matrix of $C$ and consider the code $C_{r}$ of length $n$ over the field $\mathbb{F}_{q^r}$ with the same parity check matrix $H$. The code $C_{r}$ is a well defined code, since the values in $H$ are from $\mathbb{F}_q \subset \mathbb{F}_{q^r}$.
We will say that code $C_r$ is obtained by lifting $C$.
The next Proposition shows a representation for the code obtained by lifting.
\begin{proposition}\label{rep}
Let $C_r$ be the code of length $n$ over the field $\mathbb{F}_{q^r}$, obtained by lifting the code $C$ over $\mathbb{F}_q$. Then:
$$C_r = C + \alpha C + \alpha^2 C + \cdots + \alpha^{r-1} C,$$ where $\alpha$ is a primitive element in $\mathbb{F}_{q^r}$.
\end{proposition}
\begin{proof}
By construction, it is easy to see that $C_r \subset C + \alpha C + \alpha^2 C + \cdots + \alpha^{r-1} C$. The statement follows by computing the cardinality of $C_r$ and $C + \alpha C + \alpha^2 C + \cdots + \alpha^{r-1} C$.
\end{proof}
From~the previous Proposition it follows immediately that the minimum distance of the lifted code $C_{r}$ is the same that in code $C$ over $\mathbb{F}_q$. We can specify a little more in the following Lemma.
\begin{lemma}\label{lem:2.9}
Let $C_r$ be the code of length $n$ over the field $\mathbb{F}_{q^r}$, obtained by lifting the code $C$ over $\mathbb{F}_q$. Let ${\bf c}$ be a codeword of $C_r$ of minimum weight. Then ${\bf c} = \beta\,{\bf c}'$ where $\beta \in \mathbb{F}_{q^r}$ and ${\bf c}'$ is a codeword of minimum weight in $C$.
\end{lemma}
\begin{proof}
From~Proposition~\ref{rep} is easy to see that the minimum distance in $C_r$ is less or equal to the minimum distance in $C$. Vice versa, the argument is reduced to show that there are not linearly independent minimum weight vectors in $C$ with the same support. Hence, take two independent minimum weight vectors ${\bf v}, {\bf u} \in C$ with the same support. Now, notice that it is possible to compute the vector $a{\bf v} + b{\bf u}$ with appropriate values $a, b \in \mathbb{F}_q$ such that the support of the new vector $a{\bf v} + b{\bf u}$ is strictly less. Hence, the weight of $a{\bf v} + b{\bf u} \in C$ is less than the minimum weight of $C$, which is contradictory.
\end{proof}
Denote by $\pcm{m,q}$ the parity check matrix of a perfect Hamming
$[n,k,3]_q$-code over $\mathbb{F}_q$, where $n=(q^m-1)/(q-1)$.
Let $\{\xi_0=0,\xi_1=1,\ldots,\xi_{q-1}\}$ denote the elements of
$\mathbb{F}_q$. Then the matrix $\pcm{m,q}$ can be expressed, up to equivalence,
through the matrix $\pcm{m-1,q}$ as follows \cite{Sem3}:
\[
\pcm{m,q} = \left[
\begin{array}{cc|c|c|c|c}
&\,1\,&\,0\cdots 0\,&\,1\cdots 1\,&\,\cdots\,&\,\xi_{q-1}\cdots
\xi_{q-1}\,\\\hline
&\,{\bf 0}\, &\pcm{m-1,q}\,&\,\pcm{m-1,q}\,&\,\cdots \,&\,\pcm{m-1,q}\,\\
\end{array}\right],
\]
where ${\bf 0}$ is the zero column and where $\pcm{1,q} = [1]$.
Let $C$ be the Hamming code over $\mathbb{F}_q$ with parity check matrix
$\pcm{m,q}$. Denote by $C_{(m,r)}$ the lifting $C$,
i.e. $C_{(m,r)}$ is the code over $\mathbb{F}_{q^r}^n$ with parity check matrix $\pcm{m,q}$.
First we find the covering radius for $C_{(m,r)}$.
Take any vector ${\bf v}$ in the ambient space of $C_{(m,r)}$, so ${\bf v} \in \mathbb{F}_{q^r}^{\,n}$ and compute the syndrome ${\bf v} \pcm{m,q}^T \in \mathbb{F}_{q^r}^m$. We can represent this syndrome ${\bf v} \pcm{m,q}^T$ by a $r\times m$ matrix over $F_q$, taking as columns the representation of the elements in $\mathbb{F}_{q^r}$ using a fixed basis in $\mathbb{F}_q$. Call $S_{{\bf v}}$ this matrix and denote by $\mbox{\rm rank}(S_{{\bf v}})$ its rank over $F_q$.
\begin{lemma}\label{lem:2.1}
Let $C_{(m,r)}$ be the code with parity check matrix $\pcm{m,q}$. Then, for any vector ${\bf v} \in \mathbb{F}_{q^r}^{\,n}$ with weight $\mbox{\rm wt}({\bf v})$ we have that $\mbox{\rm rank}(S_{{\bf v}})\leq \mbox{\rm wt}({\bf v})$.
\end{lemma}
\begin{proof}
Assume that ${\bf v}$ is a nonzero vector and also that all the coordinates of ${\bf v}$ belong to the ground field $\mathbb{F}_q$. In this case, it is clear that $\syn{v}$ has rank $1$. In the general case, as $\pcm{m,q}$ is the parity check matrix of a Hamming code over the ground field $\mathbb{F}_q$, we find that the rank of $\syn{v}$ depends on the number of independent coordinates of the vector ${\bf v}$ over $\mathbb{F}_q$. When $r \leq m$ we have that this number $r$ is exactly the rank and, when $r>m$ the rank is $m$. In any case, $\mbox{\rm rank}(S_{{\bf v}})= min\{r,m\} \leq \mbox{\rm wt}({\bf v})$.
\end{proof}
\begin{proposition}\label{dist}
For any vector ${\bf v} \in \mathbb{F}_{q^r}^{\,n}$ we have that $d({\bf v},C_{(m,r)})=\mbox{\rm rank}(\syn{v})$.
\end{proposition}
\begin{proof}
Matrix $S_{{\bf v}}$ can be transformed into a diagonal matrix $P$ with only $l=\mbox{\rm rank}(S_{{\bf v}})$ nonzero diagonal values. The transformation can be carried out by multiplying by the corresponding nonsingular matrices over $\mathbb{F}_q$. Hence, we can write:
\begin{equation}\label{eq:p}
P=A\,S_{{\bf v}}B =A\,v\pcm{m,q}^TB,
\end{equation}
where $A$ is a $r\times r$ nonsingular matrix over $\mathbb{F}_q$ and $B$ is a $m\times m$ nonsingular matrix over $\mathbb{F}_q$.
We know that a Hamming parity check matrix is unique, up to permutations and scalar multiplications of columns. Hence, matrix $\pcm{m,q}^TB$ in~(\ref{eq:p}) represents a permutation of the $n$ coordinates, followed by scalar multiplications. Matrix $A$ makes a linear transformation in the vector space $\mathbb{F}_q^r$ over the ground field $\mathbb{F}_q$ (as we already mentioned, we use a representation of elements of $\mathbb{F}_{q^r}$ as elements of $\mathbb{F}_q^r$ by using a fixed basis). It is easy to see that the weight of the vectors ${\bf v} \in \mathbb{F}_{q^r}^n$ is not changed after doing these above transformations. Looking at the matrix $P$ in~(\ref{eq:p}), it is easy to find a vector ${\bf w} \in \mathbb{F}_{q^r}^n$ of weight $l=\mbox{\rm rank}(\syn{v})$ with syndrome equal to $P$, so $\syn{w}=P$. From Lemma~\ref{lem:2.1} we conclude that $d({\bf v},C_{(m,r)})=\mbox{\rm rank}(\syn{v})$.\end{proof}
\begin{corollary}\label{coro:2.3}
The covering radius of code $C_{(m,r)}$ is $\rho=min\{r,m\}$.
\end{corollary}
\begin{proof}
It is straightforward to see that $\rho=min\{r,m\}$, since the distance of any vector ${\bf v} \in \mathbb{F}_{q^r}^n$ to $C_{(m,r)}$ is given by $rank(S_{{\bf v}})$ and $S_{{\bf v}}$ is a $r\times m$ matrix over $F_q$.
\end{proof}
The second important goal is to prove the completely regularity for these codes over $\mathbb{F}_{q^r}$, which have $\pcm{m,q}$ as a parity check matrix.
To prove the condition of completely regularity we start by slightly modifying the usual definition of completely transitive codes~\cite{Sole,giudici}, and
introducing the concept of $(r,q)$-completely transitivity, which will be helpful throughout this paper. But, first, we give a different and convenient definition of the automorphism group of a code $C_r$ over $\mathbb{F}_{q^r}$, which we will denote by $Aut_{(r,q)}(C)$.
\begin{definition}
Define the automorphism group $Aut_{(r,q)}(C_r)$ of a code $C_r$ over $\mathbb{F}_{q^r}$ as all the
$n\times n$ monomial matrices $M$ over $\mathbb{F}_{q^r}$ such that ${\bf c} M \in C_r$
for any ${\bf c} \in C_r$; all the field automorphisms of $\mathbb{F}_{q^r}$ which preserve $C$ and, moreover, all the vector space morphisms of $\mathbb{F}_{q^r}$ over $\mathbb{F}_q$ which preserve $C_r$.
\end{definition}
Finally, we define the concept of $(r,q)$-completely transitivity.
\begin{definition}
Let $C_r$ be a $[n,k,d]_{q^r}$ code with covering radius $\rho$. We will say that $C_r$ is $(r,q)$-completely transitive if $Aut_{(r,q)}(C_r)$ has $\rho +1$ orbits when acting on the cosets of $C_r$.
\end{definition}
\begin{proposition}\label{prop:1.2}
If $C_r$ is a $(r,q)$-completely transitive code, then $C_r$ is completely regular.
\end{proposition}
\begin{proof}
It is enough to show that all the cosets in the same orbit have the same weight distribution. It is straightforward to see this, since an automorphism of $Aut_{(r,q)}(C_r)$ acts over the vectors by permuting coordinates, multiplying by a nonzero scalar and linearly modifying a coordinate. Indeed, each coordinate is an element of $\mathbb{F}_{q^r}$ and could be seen as a vector of $r$ coordinates in $\mathbb{F}_q$; after the modification we obtain another coordinate consisting in a change of basis in $\mathbb{F}_q^{\,r}$ as a vector space over $\mathbb{F}_q$. In all the cases a nonzero coordinate always gives a nonzero coordinate and the weight of the initial vector is maintained.
\end{proof}
Going again to the example we introduced in Section~\ref{int}, we see that now
$$|Aut_{(4,2)}(C)| = 360\cdot 20160 = 7257600,$$ where $20160=15\cdot 14\cdot 12\cdot 8$ comes from the number of vector space morphisms of $\mathbb{F}_{2^4}$ over $\mathbb{F}_2$ (or, equivalently, from the order of the general linear group $GL_4(\mathbb{F}_2)$ of degree $4$ over $\mathbb{F}_2$). As we will see, applying next Theorem, the code in that example is $(r,q)$-completely transitive, and so, completely regular.
\begin{theorem}\label{theo:2.1}
Let $C_{(m,r)}$ be the code of length $n$ over the field $\mathbb{F}_{q^r}$ with parity check matrix $\pcm{m,q}$, where $n=\frac{q^m-1}{q-1}$. The code $C_{(m,r)}$ is $(r,q)$-completely transitive and, hence, completely regular with covering radius $\rho = \min\{r,m\}$.
\end{theorem}
\begin{proof}
Taking into account Corollary~\ref{coro:2.3} and Proposition~\ref{prop:1.2},
it is enough to show that the code $C_{(m,r)}$ is $(r,q)$-completely transitive. Choose any two vectors ${\bf v}, {\bf w} \in \mathbb{F}_{q^r}^{\,n}$ at the same distance from $C_{(m,r)}$, and show that there is an automorphism
$\phi \in Aut_{(r,q)}(C_{(m,r)})$ such that ${\bf w}$ and $\phi({\bf v})$ are in the same coset. That is, ${\bf w} \in \phi({\bf v}) + C_{(m,r)}$ or, in other words, $\syn{{\bf w}} = S_{\phi({\bf v})}$.
From Proposition~\ref{dist} we have $\mbox{\rm rank}(\syn{{\bf v}}) = \mbox{\rm rank}(\syn{{\bf w}})$
and this means that there exist two nonsingular $r\times r$ and $m\times m$ matrices $A$ and $B$ over $\mathbb{F}_q$, respectively, such that $\syn{{\bf w}} = A\syn{{\bf v}}B$. Since $\pcm{m,q}$ consists of all nonzero mutually linearly independent vectors of length $m$ over $\mathbb{F}_q$, for any such matrix $B$,
there exists a monomial matrix $M \in Aut_{m,q}(C_{(m,r)})$ such that $\pcm{m,q}\,B = M\,\pcm{m,q}$. Hence, we obtain
$$
{\bf w}\pcm{m,q}^T = \syn{{\bf w}} = A\syn{{\bf v}}B = A{\bf v}\pcm{m,q}^TB = A{\bf v} M\pcm{m,q}^T = \phi({\bf v})\pcm{m,q}^T = S_{\phi({\bf v})},
$$
where $\phi \in Aut_{m,q}(C_{(m,r)})$.
\end{proof}
To compute the intersection array of the code $C_{(m,r)}$ we need
the following well known result.
\begin{lemma}[\cite{lan}]\label{lem:2.7}
The number of different $r\times m$ matrices over $F_q$, of rank $k\leq \min\{r,m\}$ is
$$
\frac{M_q(k,r)M_q(k,m)}{M_q(k,k)},
$$
where
$$
M_q(k,t) =(q^t-1)(q^t-q)\cdots (q^t-q^{k-1})
$$
represents, for $1\leq k\leq t$, the number of injective morphisms from a vector space of dimension $k$ to a vector space of dimension $t$, both over $\mathbb{F}_q$.
\end{lemma}
Now we can easily see that the number of injective morphisms from a vector space of dimension $k$ into a vector space of dimension $t$, when we fix the image of the first $k-1$ vectors in the basis of the first vector space, is given by $M'_q(k,t) = q^t-q^{k-1}$, i.e. the last factor in $M_q(k,t)$. We will say that this injective morphism has only one freedom degree.
Hence, the number of different $r\times m$ matrices over $F_q$, of rank $k\leq \min\{r,m\}$, but with only one freedom degree is:
\begin{equation}\label{gran}
\frac{M'_q(k,r)M'_q(k,m)}{M_q(1,1)}= \frac{(q^r-q^{k-1})(q^m-q^{k-1})}{(q-1)}.
\end{equation}
We need one more statement.
\begin{lemma}\label{lem:2.8}
Let $C_{(m,r)}$ be the code with parity check matrix $\pcm{m,q}$. Let $\mu_i$ be the number of cosets in $C_{(m,r)}(i)$. Then
\begin{equation}\label{ci}
\mu_i = \frac{M_q(i,r)M_q(i,m)}{M_q(i,i)}.
\end{equation}
\end{lemma}
\begin{proof}
Since Proposition~\ref{dist}, for all the vectors ${\bf v}$ in the same $C_{(m,r)}(i)$ we have that the corresponding $r\times m$ matrix $S_{{\bf v}}$ is of rank $i$. But, all the vectors in the same coset have the same syndrome and, hence, the number of different cosets in $C(i)$ is equal to the
number of different $r\times m$ matrices over $\mathbb{F}_q$ of rank $i$, which
is given by Lemma \ref{lem:2.7}.
\end{proof}
Now, it is easy to compute the intersection array for the lifted codes.
\begin{theorem}\label{theo:2.2}
Let $C_{(m,r)}$ be the code of length $n$ over the field $\mathbb{F}_{q^r}$ with parity check matrix $\pcm{m,q}$, where $n=\frac{q^m-1}{q-1}$.
\begin{itemize}
\item Code $C_{(m,r)}$ is a completely regular code with intersection array:
$$b_i = \frac{(q^r-q^{i})(q^m-q^{i})}{(q-1)}; \; c_i=q^{i-1}\frac{q^i-1}{q-1}.
$$
\item Codes $C_{(m,r)}$ and $C_{(r,m)}$ are, in general, different, but they have the same intersection array.
\end{itemize}
\end{theorem}
\begin{proof}
Directly from the definition of $b_i$ and $c_i$ we have:
\begin{equation}\label{bro}
\mu_ib_i = \mu_{i+1}c_{i+1}.
\end{equation}
For each index $i\in \{0,1,\cdots,\rho\}$ we know that $a_i+b_i+c_i=(q^r-1)n$ and so, for any index, it is only necessary to compute one of the values, for instance $b_i$, since the other two values comes from the last equality and~(\ref{bro}).
We begin computing $b_0$, so the number of vectors in $C_{(m,r)}(1)$ which are at distance one from one given vector in $C_{(m,r)}$. Without losing generality (since $C_{(m,r)}$ is completely regular) we can fix vector ${\bf 0}$ in $C_{(m,r)}$ and count how many vectors there are at distance one in $C_{(m,r)}(1)$. The answer is immediately
\[
n\,(q^r-1) = \frac{(q^r - 1)(q^m - 1)}{q-1}.
\]
Indeed, from Proposition~\ref{dist}, we need to count how many cosets have syndromes with rank one or, in another way, how many $r\times m$ matrices of rank one there are. From Lemma~\ref{lem:2.7} the result is $(q^r-1)(q^m-1)/(q-1)$.
Now, we know $\mu_1$ and $b_0$ and so, from~(\ref{bro}), it is easy to compute $c_1 =\displaystyle \frac{\mu_0b_0}{\mu_1}=1$ and $a_1=(q^r-1)n-b_1-c_1$.
In general, let $1 \leq i \leq \rho$. We use the same argumentation. The value $b_i$
is the number of different $r\times m$ matrices over $F_q$, of rank $i\leq \rho$, with exactly one freedom degree. We obtain from~(\ref{gran}), that
$$
b_i = \frac{M'_q(i+1,r)M'_q(i+1,m)}{M_q(1,1)}= \frac{(q^r-q^{i})(q^m-q^{i})}{(q-1)}\,.
$$
Now, using the expressions for $b_{i-1}$, $\mu_i$ and $\mu_{i-1}$ from Lemma \ref{lem:2.8}, we obtain
$$
c_i = \frac{\mu_{i-1}b_{i-1}}{\mu_i}= q^{i-1}\frac{(q^{i}-1)}{q-1}\,.
$$
Finally, we note that all the values $b_i, c_i$ of the above intersections arrays are symmetric for $r$ and $m$, so codes $C_{(r,m)}$ and $C_{(m,r)}$ are, in general, different, but they have the same intersection numbers $b_i, c_i$ (and different values for $a_i$ for the case $r \neq m$).
\end{proof}
It is interesting to notice that perfect codes are the only class of
nontrivial codes (i.e. $q$-ary codes with cardinality $1 < |C| < q^{n-1}$
and with minimum distance $d \geq 3$) whose codes, obtained by lifting the
ground field, are completely regular.
\begin{theorem}\label{theo:2.3}
Let $C$ be the nontrivial code of length $n$ over the field $\mathbb{F}_q$ with
minimum distance $d \geq 3$, with covering radius $\rho \geq 1$ and parity check matrix $H$. Let $C_r$ be the code over $\mathbb{F}_{q^r}$, whose
parity check matrix is $H$. Then $C_r$ is completely regular,
if and only if $C$ is a Hamming code.
\end{theorem}
\begin{proof}
Since Theorem \ref{theo:2.1}, it is enough to show that when $C$ is not a Hamming code, lifting the ground field does not give a
completely regular code $C_r$.
Take a vector ${\bf x}$ over $\mathbb{F}_q$ of weight two, which is covered by some codeword ${\bf v} \in C$ of weight $w \geq 4$ (this is possible also for any
code $C$ with $d=3$ and $n > 3$, since $C$ is not perfect). Then, from Lemma~\ref{lem:2.9}, the vector ${\bf x}'$ of weight two, having one nonzero coordinate from $\mathbb{F}_q$ and the other from $\mathbb{F}_{q^r} \setminus \mathbb{F}_q$ is not covered by any codeword of minimum weight. Now the two cosets $D = C_r - {\bf x}$
and $D' = C_r - {\bf x}'$ (both of weight two) have different weight distributions, which is impossible for completely regular code.
\end{proof}
\medskip
\begin{remark}
Any code $C_{(m,sr)}$ contains as a subcode the code
$C_{(m,r)}$ where $s,r \geq 1$ are arbitrary natural numbers. Hence, our construction induces, for any prime power $q$, an infinite family of
nested completely regular codes:
\[
C_{(m,r)} \subset C_{(m,ar)} \subset
C_{(m,a^2\,r)} \subset \ldots
\]
This nested family of codes induces, in turn,
infinite nested families of regular and completely regular partitions (see \cite[Sec. 11.1]{Bro2}) of completely regular codes into completely regular subcodes.
\end{remark}
\medskip
Finally, we end this paper with a comment about distance regular graphs.
\begin{remark}
It is well known that any linear completely regular code $C$ implies the existence of a coset distance-regular graph. From the completely regular codes $C_{(r,m)}$ described in this paper, we obtain distance-regular graphs with classical parameters (see \cite{Bro2}) which are distance-transitive since they come
from $(r,q)$-completely transitive codes. These graphs have $v = q^{r\,m}$
vertices, diameter $\rho = \min\{r,m\}$, and intersection array given by
$$
b_i = \frac{(q^r-q^{i})(q^m-q^{i})}{(q-1)}; \; c_i=q^{i-1}\frac{q^i-1}{q-1},
$$
where $0 \leq i \leq d$.
Notice that bilinear forms graphs~\cite[Sec. 9.5]{Bro2} have the same parameters and are distance-transitive too. These graphs are uniquely
defined by their parameters (see \cite[Sec. 9.5]{Bro2}). But, we did
not find in the literature (in particular in \cite{Dels}, where the
association schemes, formed by bilinear forms, have been introduced and
their application to coding theory have been considered) such a simple description of these graphs, as coset graphs of completely regular codes, constructed by lifting Hamming codes.
\end{remark}
|
\section{Introduction}
\seclabel{Intro}
This paper studies problems related to visibility and blocking in sets
of coloured points in the plane. A point $x$ \emph{blocks} two points
$v$ and $w$ if $x$ is in the interior of the line segment
$\overline{vw}$. Let $P$ be a finite set of points in the plane. Two
points $v$ and $w$ are \emph{visible} with respect to $P$ if no point
in $P$ blocks $v$ and $w$. The \emph{visibility graph} of $P$
has vertex set $P$, where two distinct points $v,w\in P$ are adjacent
if and only if they are visible with respect to $P$. A point set $B$ \emph{blocks}
$P$ if $P\cap B=\emptyset$ and for all distinct $v,w\in P$ there is a
point in $B$ that blocks $v$ and $w$. That is, no two points in $P$
are visible with respect to $P\cup B$, or alternatively, $P$ is an
independent set in the visibility graph of $P\cup B$.
A set of points $P$ is \emph{$k$-blocked} if each point in $P$ is
assigned one of $k$ colours, such that each pair of points $v,w\in P$
are visible with respect to $P$ if and only if $v$ and $w$ are
coloured differently. Thus $v$ and $w$ are assigned the same colour if
and only if some other point in $P$ blocks $v$ and $w$. A
\emph{$k$-set} is a multiset of $k$ positive
integers. For a $k$-set $\{n_1,\dots,n_k\}$, we say $P$ is
$\{n_1,\dots,n_k\}$-blocked if it is $k$-blocked and for some
labelling of the colours by the integers $[k]:=\{1,2,\dots,k\}$, the
$i$-th colour class has exactly $n_i$ points, for each
$i\in[k]$. Equivalently, $P$ is $\{n_1,\dots,n_k\}$-blocked if the
visibility graph of $P$ is the complete $k$-partite graph
$K(n_1,\dots,n_k)$. A $k$-set $\{n_1,\dots,n_k\}$ is
\emph{representable} if there is an $\{n_1,\dots,n_k\}$-blocked point
set. See \figref{K3333} for an example.
\Figure[!ht]{K3333}{\includegraphics{K3333}}{A $\{3,3,3,3\}$-blocked point set.}
The following fundamental conjecture regarding $k$-blocked point sets
is the focus of this paper.
\begin{conjecture}
\conjlabel{kBlocked} For each integer $k$ there is an
integer $n$ such that every $k$-blocked set has at most $n$ points.
\end{conjecture}
As illustrated in
\figref{23Blocked}, the following theorem is a direct consequence of the characterisation
of 2- and 3-colourable visibility graphs by \citet{KPW-DCG05}.
\begin{theorem}
\thmlabel{23}
$\{1,1\}$ and $\{1,2\}$ are the only representable 2-sets, and \\
$\{1,1,1\}$, $\{1,1,2\}$, $\{1,2,2\}$ and $\{2,2,2\}$ are the only representable 3-sets.
\end{theorem}
\Figure{23Blocked}{\includegraphics{23Blocked}}{The 2-blocked and
3-blocked point sets.}
In particular, every $2$-blocked point set has at most 3 points, and
every $3$-blocked point set has at most 6 points. This proves
\conjref{kBlocked} for $k\leq 3$.
This paper makes the following contributions. \secref{Background}
introduces some background motivation for the study of $k$-blocked
point sets, and observes that results in the literature prove
\conjref{kBlocked} for $k=4$. \secref{SmallColourClasses} describes
methods for constructing $k$-blocked sets from a given $(k-1)$-blocked
set. These methods lead to a characterisation of representable
$k$-sets when each colour class has at most three
points. \secref{Four} studies the $k=4$ case in more detail. In
particular, we characterise the representable 4-sets, and conclude
that the example in \figref{K3333} is in fact the largest
4-blocked point set. \secref{Midpoint} introduces a special class of
$k$-blocked sets (so-called midpoint-blocked sets) that lead to a
construction of the largest known $k$-blocked sets for infinitely many
values of $k$.
\section{Basic Properties}
\begin{lemma}
\lemlabel{SelfBlockingThree} At most three points are collinear in
every $k$-blocked point set.
\end{lemma}
\begin{proof}
Suppose that four points $p,q,r,s$ are collinear in this order. Thus $(p,q,r,s)$ is an induced path in the visibility graph. Thus $p$ is not adjacent to $r$ and not
adjacent to $s$. Thus $p$, $r$, and $s$ have the same colour. This
is a contradiction since $r$ and $s$ are adjacent. Thus no four
points are collinear.
\end{proof}
A set of points $P$ is in \emph{general position} if no three
points in $P$ are collinear.
\begin{lemma}
\lemlabel{GenPos} Each colour class in a $k$-blocked
point set is in general position.
\end{lemma}
\begin{proof}
Suppose on the contrary that three points from a single colour
class are collinear. Then no other points are in the same line by
\lemref{SelfBlockingThree}. Thus two of the three points are
adjacent, which is a contradiction.
\end{proof}
\section{Some Background Motivation}
\seclabel{Background}
Much recent research on blockers began with the following conjecture by
\citet{KPW-DCG05}.
\begin{conjecture}[Big-Line-Big-Clique Conjecture \citep{KPW-DCG05}]
\conjlabel{BigClique} For all integers $t$ and $\ell$ there is an
integer $n$ such that for every finite set $P$ of at least $n$
points in the plane:
\begin{itemize}
\item $P$ contains $\ell$ collinear points, or
\item $P$ contains $t$ pairwise visible points\\
(that is, the visibility graph of $P$ contains a $t$-clique).
\end{itemize}
\end{conjecture}
\conjref{BigClique} is true for $t\leq 5$, but is open for $t\geq6$ or
$\ell\geq4$; see \citep{PorWood-Blockers,EmptyPentagon}. Given that, in general, \conjref{BigClique} is challenging, Jan K{\'a}ra suggested the following weakening.
\begin{conjecture}[\citep{PorWood-Blockers}]
\conjlabel{BigChromaticNumber}
For all integers $t$ and $\ell$ there is an integer $n$ such that for every
finite set $P$ of at least $n$ points in the plane:
\begin{itemize}
\item $P$ contains $\ell$ collinear points, or
\item the chromatic number of the visibility graph of $P$ is at least $t$.
\end{itemize}
\end{conjecture}
\conjref{BigClique} implies \conjref{BigChromaticNumber} since every
graph that contains a $t$-clique has chromatic number at least $t$.
\begin{proposition}
\conjref{BigChromaticNumber} with $\ell=4$ and $t=k+1$ implies \conjref{kBlocked}.
\end{proposition}
\begin{proof}
Assume \conjref{BigChromaticNumber} holds for $\ell=4$ and
$t=k+1$. Suppose there is a $k$-blocked set $P$ of at least $n$
points. By \lemref{SelfBlockingThree}, at most three points are
collinear in $P$. Thus the first conclusion of
\conjref{BigChromaticNumber} does not hold. By definition, the
visibility graph of $P$ is $k$-colourable. Thus the second conclusion of
\conjref{BigChromaticNumber} does not hold. This contradiction
proves that every $k$-blocked set has less than $n$ points, and
\conjref{kBlocked} holds.
\end{proof}
Thus, since \conjref{BigClique} holds for $t\leq5$,
\conjref{kBlocked} holds for $k\leq4$. In \secref{Four} we take this
result much further, by characterising all representable 4-sets,
this concluding a tight bound on the size of a 4-blocked set.
\medskip Part of our interest in blocked point sets comes from the
following.
\begin{proposition}
\proplabel{TuranGraph} For all $k\geq3$ and $n\geq2$, the edge set
of the $k$-partite Tur\'an graph $K(n,n,\dots,n)$ can be partitioned
into a set of `lines', where:
\begin{itemize}
\item each line is either an edge or an induced path on three
vertices,
\item every pair of vertices is in exactly one line, and
\item for every line $L$ there is a vertex adjacent
to each vertex in $L$.
\end{itemize}
\end{proposition}
\begin{proof}
Let $\vvv{i,p}$ be the $p$-th vertex in the
$i$-th colour class for $i\in\ensuremath{\mathbb{Z}}_k$ and $p\in\ensuremath{\mathbb{Z}}_n$ (taken as additive
cyclic groups). We introduce three types of lines. First, for
$i\in\ensuremath{\mathbb{Z}}_k$ and distinct $p,q\in\ensuremath{\mathbb{Z}}_n$, let the triple
$\{\vvv{i,p},\vvv{i+1,p+q},\vvv{i,q}\}$ be a line. Second, for
$i\in\ensuremath{\mathbb{Z}}_k$ and $p\in\ensuremath{\mathbb{Z}}_n$, let the pair
$\{\vvv{i,p},\vvv{i+1,p+p}\}$ be a line. Third, for $p,q\in\ensuremath{\mathbb{Z}}_n$
and distinct non-consecutive $i,j\in\ensuremath{\mathbb{Z}}_k$, let
the pair $\{\vvv{i,p},\vvv{j,q}\}$ be a line. By construction each
line is either an edge or an induced path on three vertices.
Every pair of vertices in the same colour class are in exactly one
line (of the first type). Consider vertices $\vvv{i,p}$ and
$\vvv{j,q}$ in distinct colour classes. First suppose that $i$ and
$j$ are consecutive. Without loss of generality, $j=i+1$. If
$q\neq p+p$ then $\vvv{i,p}$ and $\vvv{j,q}$ are in exactly one line
(of the first type). If $q=p+p$ then $\vvv{i,p}$ and $\vvv{j,q}$
are in exactly one line (of the second type). If $i$ and $j$ are not
consecutive, then $\vvv{i,p}$ and $\vvv{j,q}$ are in exactly one
line (of the third type). This proves that every pair of vertices
is in exactly one line. Moreover, every edge of $K(n,n,\dots,n)$ is
in exactly one line, and the lines partition the edges set.
Since every line $L$ is contained in the union of two colour
classes, each vertex in neither colour class
intersecting $L$ is adjacent to each vertex in $L$.
\end{proof}
\propref{TuranGraph} is significant for \conjref{BigClique} because it
says that $K(n,\dots,n)$ behaves like a `visibility space' with no
four collinear points, but it has no large clique (for fixed
$k$). \conjref{kBlocked}, if true, implies that such a visibility
space is not `geometrically representable' for large $n$.
\medskip
Let $b(n)$ be the minimum integer such that some set of $n$ points in
the plane in general position is blocked by some set of $b(n)$
points. \citet{Matousek09} proved that $b(n)\geq 2n-3$.
\citet{DPT09} improved this bound to $b(n)\geq(\frac{25}{8}-o(1))n$.
Many authors have conjectured or stated as an open problem that
$b(n)$ is super-linear.
\begin{conjecture}[\citep{Matousek09,Pinchasi,DPT09,PorWood-Blockers}]
\conjlabel{Blockers}
$\frac{b(n)}{n}\rightarrow\infty$ as $n\rightarrow\infty$.
\end{conjecture}
\citet{PorWood-Blockers} proved that \conjref{Blockers}
implies \conjref{BigChromaticNumber}, and thus implies
\conjref{kBlocked}. That \conjref{kBlocked} is implied by a number of other
well-known conjectures, yet remains challenging, adds to its interest.
\section{$k$-Blocked Sets with Small Colour Classes}
\seclabel{SmallColourClasses}
We now describe some methods for building blocked point sets
from smaller blocked point sets.
\begin{lemma}
\lemlabel{BuildSelfBlocking} Let $G$ be a visibility graph. Let
$i\in\{1,2,3\}$. Furthermore suppose that if $i\geq2$ then
$V(G)\neq\emptyset$, and if $i=3$ then not all the vertices of $G$
are collinear. Let $G_i$ be the graph obtained from $G$ by adding
an independent set of $i$ new vertices, each adjacent to every
vertex in $G$. Then $G_1$, $G_2$, and $G_3$ are visibility graphs.
\end{lemma}
\begin{proof}
For distinct points $p$ and $q$, let $\overleftarrow{pq}$ denote the ray
that is (1) contained in the line through $p$ and $q$, (2) starting
at $p$, and (3) not containing $q$. Let $\mathcal{L}$ be the union
of the set of lines containing at least two vertices in $G$.
$i=1$: Since $\mathcal{L}$ is the union of finitely many lines,
there is a point $p\not\in \mathcal{L}$. Thus $p$ is visible from
every vertex of $G$. By adding a new vertex at $p$, we obtain a
representation of $G_1$ as a visibility graph.
$i=2$: Let $p$ be a point not in $\mathcal{L}$. Let $v$ be a vertex
of $G$. Each line in $\mathcal{L}$ intersects $\overleftarrow{vp}$
in at most one point. Thus
$\overleftarrow{vp}\setminus\mathcal{L}\neq\emptyset$. Let $q$ be a
point in $\overleftarrow{vp}\setminus\mathcal{L}$. Thus $p$ and $q$
are visible from every vertex of $G$, but $p$ and $q$ are blocked by
$v$. By adding new vertices at $p$ and $q$, we obtain a
representation of $G_2$ as a visibility graph.
$i=3$: Let $u,v,w$ be non-collinear vertices in $G$. Let $p$ be a
point not in $\mathcal{L}$ and not in the convex hull of $\{u,v,w\}$. Without loss
of generality, $\overline{uv}\cap\overline{pw}\neq\emptyset$. There
are infinitely many pairs of points $q\in\overleftarrow{up}$ and
$r\in\overleftarrow{vp}$ such that $w$ blocks $q$ and $r$. Thus
there are such $q$ and $r$ both not in $\mathcal{L}$. By
construction, $u$ blocks $p$ and $q$, and $v$ blocks $p$ and $r$. By
adding new vertices at $p$, $q$ and $r$, we obtain a representation
of $G_3$ as a visibility graph.
\end{proof}
Since no $(\geq3)$-blocked set is collinear, \lemref{BuildSelfBlocking} implies:
\begin{corollary}
\corlabel{Build} If $k\geq4$ and $\{n_1,\dots,n_{k-1}\}$ is
representable and $n_k\in\{1,2,3\}$, then
$\{n_1,\dots,n_{k-1},n_k\}$ is representable.
\end{corollary}
The representable $(\leq 3)$-sets were characterised in \thmref{23}. In
each case, each colour class has at most three vertices. Now we
characterise the representable $(\geq 4)$-sets, assuming that
each colour class has at most three vertices.
\begin{proposition}
\proplabel{SmallColourClasses}
$\{n_1,\dots,n_k\}$ is representable whenever $k\geq4$ and each $n_i\leq 3$, except for $\{1,3,3,3\}$.
\end{proposition}
\begin{proof}
We say the $k$-set $\{n_1,\dots,n_k\}$ \emph{contains}
the $(k-1)$-set $\{n_1,\dots,n_{i-1},n_{i+1},\dots,n_k\}$ for each $i\in[k]$.
\lemref{1333} below proves that $\{1,3,3,3\}$ is not representable. We
proceed by induction on $k$. If $\{n_1,\dots,n_k\}$ contains a
representable $(k-1)$-set, then \corref{Build} implies
that $\{n_1,\dots,n_k\}$ is also representable. Now assume that
every $(k-1)$-set contained in $\{n_1,\dots,n_k\}$ is not
representable. By induction, we may assume that $k\leq 5$.
Moreover, if $k=5$ then $\{n_1,\dots,n_5\}$ must contain $\{1,3,3,3\}$
(since by induction all other 4-sets are representable).
Similarly, if $k=4$ then $\{n_1,\dots,n_4\}$ must contain $\{1,1,3\}$,
$\{1,2,3\}$, $\{1,3,3\}$, $\{2,2,3\}$, $\{2,3,3\}$ or $\{3,3,3\}$ (since
$\{1,1,1\}$, $\{1,1,2\}$, $\{1,2,2\}$ and $\{2,2,2\}$ are
representable by \thmref{23}). The following table describes the required
construction in each remaining case.
\bigskip
\begin{tabular}{ll|ll}
\hline
$\{1,1,1,x\}$ & contains $\{1,1,1\}$ &
$\{1,1,2,x\}$ & contains $\{1,1,2\}$\\
$\{1,1,3,3\}$ & \figref{K3333} minus $\{r_1,g_3,r_3,g_1\}$&
$\{1,2,2,x\}$ & contains $\{1,2,2\}$\\
$\{1,2,3,3\}$ & \figref{K3333} minus $\{g_1,g_3,r_3\}$&
$\{2,2,2,x\}$ & contains $\{2,2,2\}$\\
$\{2,2,3,3\}$ & \figref{K3333} minus $\{g_3,r_3\}$ &
$\{2,3,3,3\}$ & \figref{K3333} minus $g_3$\\
$\{1,1,3,3,3\}$ & contains $\{1,1,3,3\}$ &
$\{1,2,3,3,3\}$ & contains $\{1,2,3,3\}$\\
$\{1,3,3,3,3\}$ & contains $\{3,3,3,3\}$\\
\hline
\end{tabular}
\newline
\end{proof}
\section{4-Blocked Point Sets}
\seclabel{Four}
As we saw in \secref{Background}, \conjref{kBlocked} holds for
$k\leq4$. In this section we study 4-blocked point sets in more
detail. First we derive explicit bounds on the size of 4-blocked sets
from other results in the literature. Then, following a more detailed
approach, we characterise all representable 4-sets, to conclude a
tight bound on the size of 4-blocked sets.
\begin{proposition}
\proplabel{4BlockedA} Every $4$-blocked set has less than $2^{790}$
points.
\end{proposition}
\begin{proof}
\citet{EmptyPentagon} proved that every set of at least
$\ES(\frac{(2\ell-1)^{\ell}-1}{2\ell-2} )$ points in the plane
contains $\ell$ collinear points or an empty convex pentagon, where
$\ES(k)$ is the minimum integer such that every set of at least
$\ES(k)$ points in general position in the plane contains $k$ points
in convex position. Let $P$ be a $4$-blocked set. The visibility
graph of $P$ is $4$-colourable, and thus contains no empty convex
pentagon. By \lemref{SelfBlockingThree}, at most three points in
$P$ are collinear. Thus $|P| \leq \ES(400)-1$ by the above result
with $\ell=4$. \citet{TothValtr05} proved that
$\ES(k)\leq\binom{2k-5}{k-2}+1$. Hence $|P|\leq
\binom{795}{398}<2^{790}$.
\end{proof}
\begin{lemma}
\lemlabel{SelfBlockingLargeColourClass} If $P$ is a blocked set of
$n$ points with $m$ points in the largest colour class, then
$n\geq 3m-3$ and $n\geq(\frac{33}{8}-o(1))m$.
\end{lemma}
\begin{proof}
If $S$ is the largest colour class then $P-S$ blocks $S$. By
\lemref{GenPos}, $S$ is in general position. By the results of
\citet{Matousek09} and \citet{DPT09} mentioned in
\secref{Background}, $n-m\geq 2m-3$ and
$n-m\geq(\frac{25}{8}-o(1))m$.
\end{proof}
\begin{proposition}
\proplabel{4BlockedB} Every $4$-blocked set has less than $2^{578}$
points.
\end{proposition}
\begin{proof}
Let $P$ be a $4$-blocked set of $n$ points. Let $S$ be the largest
colour class in $P$. Let $m: = |S| \geq\frac{n}{4}$. By
\lemref{SelfBlockingLargeColourClass},
$n\geq(\frac{33}{8}-o(1))m\geq(\frac{33}{32}-o(1))n$. Thus
$o(n)\geq\frac{n}{32}$. Hence $n$ is bounded. A precise bound is
obtained as follows. \citet{DPT09} proved that $S$ needs at least
$\frac{25m}{8}-\frac{25m}{2\ln m}-\frac{25}{8}$ blockers, which come
from the other colour classes. Thus $n-m \geq
\frac{25m}{8}-\frac{25m}{2\ln m}-\frac{25}{8}$, implying $n \geq
\frac{33m}{8}-\frac{25m}{2\ln m}-\frac{25}{8}$. Since
$\frac{n}{4}\leq m\leq n$, we have $\frac{25n}{2\ln n}+\frac{25}{8}
\geq \frac{n}{32}$. It follows that $n\leq 2^{578}$.
\end{proof}
The next result is the simplest known proof that every 4-blocked point
set has bounded size.
\begin{proposition}
\proplabel{4BlockedC} Every $4$-blocked set has at most $36$ points.
\end{proposition}
\begin{proof}
Let $P$ be a 4-blocked set. Suppose that $|P| \geq 37$. Let $S$ be
the largest colour class. Thus $|S|\geq10$. By \lemref{GenPos}, $S$
is in general position. By a theorem of \citet{Harborth78}, some
$5$-point subset $K\subseteq S$ is the vertex-set of an empty convex
pentagon $\conv(K)$. Let $T:= P \cap ( \conv(K) - K )$. Since
$\conv(K)$ is empty with respect to $S$, each point in $T$ is not in
$S$. Thus $T$ is 3-blocked. $K$ needs at least 8 blockers (5
blockers for the edges on the boundary of $\conv(K)$, and 3 blockers
for the chords of $\conv(K)$). Thus $|T| \geq 8$. But every
3-blocked set has at most 6 points, which is a contradiction. Hence
$|P| \leq 36$.
\end{proof}
We now set out to characterise all representable 4-sets. We need a
few technical lemmas.
\begin{lemma}
\lemlabel{New}
Let $A$ be a set of three monochromatic points in a
4-blocked set $P$. Then $P\cap\conv(A)$ contains a point from each colour class.
\end{lemma}
\begin{proof}
$P\cap\conv(A)$ contains at least three points in $A$. If
$P\cap\conv(A)$ contains no point from one of the three other colour
classes, then $P\cap\conv(A)$ is a $(\leq 3)$-blocked set with three
points in one colour class ($A$), contradicting \thmref{23}.
\end{proof}
\begin{lemma}
\lemlabel{1333}
$\{3,3,3,1\}$ is not representable.
\end{lemma}
\begin{proof}
Let $a*b*c$ mean that $b$ blocks $a$ and $c$, and let $a*b*c*d$ mean
that $a*b*c$ and $b*c*d$. This allows us to record the order in which points occur on a line.
Suppose that $P$ is a $\{3,3,3,1\}$-blocked set of points, with colour classes $A=\{a_1, a_2, a_3\}$, $B=\{b_1, b_2, b_3\}$, $C=\{c_1, c_2, c_3\}$, and $D=\{d\}$.
If $d$ does not block some monochromatic pair, then $P\setminus D$ is
a 3-blocked set of nine points, contradicting \thmref{23}. Therefore $d$
blocks some monochromatic pair, which we may call $a_1, a_2$. Now
$\overleftrightarrow{a_1a_2}$ divides the remaining seven points of
$P$ into two sets, $P_1$ and $P_2$, where WLOG $4\leq|P_1|\leq 7$ and $0\leq|P_2|\leq3$. If $|P_1|\ge6$, then $P_1\cup\{a_1\}$ is a 3-blocked set of more
than six points, contradicting \thmref{23}. Thus $|P_1|=4$ and $|P_2|=3$, or
$|P_1|=5$ and $|P_2|=2$. Consider the following cases, as illustrated in \figref{1333}.
\textbf{Case 1.} $|P_1|=4$ and $|P_2|=3$:
We have $a_3 \in P_1$ and $P_1\neq\{a_3,b_1,b_2,b_3\}$ and
$P_1\neq\{a_3,c_1,c_2,c_3\}$, as otherwise $P_1$ is a 2-blocked set of
four points, contradicting \thmref{23}. WLOG, $P_1=\{a_3, b_1,
b_2,c_1\}$ and $P_2=\{b_3, c_2, c_3\}$, where $c_2*b_3*c_3$. Some
point in $P_1$ blocks $b_1$ and $b_2$. If $b_1*a_3*b_2$, then neither
$b_1$ nor $b_2$ can block $\overline{a_3a_1}$ or $\overline{a_3a_2}$; thus $c_1$ blocks
$a_3$ from both $a_1$ and $a_2$, a contradiction. Therefore
$b_1*c_1*b_2$. Since two of these three points must block $\overline{a_3a_1}$
and $\overline{a_3a_2}$, we may assume that $a_1*b_1*a_3$, and either $a_2*b_2*a_3$ or
$a_2*c_1*a_3$.
\begin{figure}[ht]
\includegraphics[scale=0.61]{1333a.png}
\hfill
\includegraphics[scale=0.61]{1333b.png}
\hfill
\includegraphics[scale=0.61]{1333c.png}
\caption{\figlabel{1333} Cases 1.1, 1.2, and 2.}
\end{figure}
Case 1.1. $a_2*b_2*a_3$: Since $c_2*b_3*c_3$, the only possible blockers for $\overline{b_1b_3}$ are on $\overleftrightarrow{a_1a_2}$. We cannot have $b_3*a_1*b_1$, for then $b_3*a_1*b_1*a_3$. We cannot have $b_3*a_2*b_1$, for then $\overleftrightarrow{b_1b_3}$ separates $b_2$ from $a_1$ and $d$, one of which needs to block $\overline{b_2b_3}$. Therefore $b_1*d*b_3$. Similarly, $b_2*d*b_3$, a contradiction.
Case 1.2. $a_2*c_1*a_3$: Thus $a_2$ does not block
$\overline{c_1c_2}$ or $\overline{c_1c_3}$. Therefore $a_1$ and $d$
block $\overline{c_1c_2}$ and $\overline{c_1c_3}$. WLOG,
$c_2*a_1*c_1$ and $c_3*d*c_1$. Since $c_2*b_3*c_3$, the blocker for
$\overline{b_1b_3}$ is on the same side of
$\overleftrightarrow{c_1c_3}$ as $a_1$, and on the same side of
$\overleftrightarrow{b_1b_2}$ as $a_1$. The only such points are
$a_1$, $c_2$, and $b_3$. Thus $a_1$ or $c_2$ block
$\overline{b_1b_3}$. Now, $c_2$ does not block $\overline{b_1b_3}$, as
otherwise $b_1*c_2*b_3*c_3$. Similarly,
$a_1$ does not block $\overline{b_1b_3}$, as
otherwise $b_3*a_1*b_1*a_3$. This contradiction concludes this case.
\textbf{Case 2.} $|P_1|=5$ and $|P_2|=2$:
We have $a_3\in P_1$, as otherwise $P_1$ is 2-blocked, contradicting
\thmref{23}. WLOG,
$P_1=\{a_3, b_1, b_2, c_1, c_2\}$ and $P_2=\{b_3, c_3\}$. Note that
$a_3$ must block at least one of $\overline{b_1b_2}$,
$\overline{c_1c_2}$, because $P_1\setminus \{a_3\}$ is too large to be
2-blocked; however, it cannot block both, for then there would be no
valid blockers left for $a_3$. Thus WLOG $b_1*a_3*b_2$ and
$c_1*b_1*c_2$. Since $a_3\in \overline{b_1b_2}$, neither $b_1$ nor
$b_2$ block $\overline{a_3a_1}$ or $\overline{a_3a_2}$. Thus, WLOG, $a_1*c_1*a_3$ and $a_2*c_2*a_3$.
By \lemref{New}, $P\cap\conv\{c_1,c_2,c_3\}$ contains some member of $A$, WLOG $a_1$. We cannot have $a_1 \in \overline{c_1c_3}$, for then $c_3*a_1*c_1*a_3$. Therefore $a_1$ is on the same side of $\overleftrightarrow{c_1c_3}$ as $c_2$; consequently, $d$ and $a_2$ are, as well. It follows that $c_1*b_3*c_3$, and so $b_3$ sees $b_2$.
\end{proof}
\begin{lemma}
\lemlabel{K4221} Let $P$ be a 4-blocked set. Suppose that some
colour class $S$ of $P$ contains a subset $K$, such that $|K|=4$ and
$K$ is the vertex-set of a convex quadrilateral $\conv(K)$ that is
empty with respect to $S$. Then $P$ is $\{4,2,2,1)$-blocked.
\end{lemma}
\begin{proof}
Let $T:= P \cap (\conv(K) - K )$. Since $\conv(K)$ is empty with
respect to $S$, each point in $T$ is not in $S$. Thus $T$ is
3-blocked. $K$ needs at least 5 blockers (4 blockers for the edges
on the boundary of $\conv(K)$, and at least 1 blocker for the chords
of $\conv(K)$). The only representable 3-sets with at least 5
points are $\{2,2,1\}$ and $\{2,2,2\}$. Exactly four points in $T$ are
on the boundary of $\conv(T)$. Every $\{2,2,2\}$-blocked set contains
three points on the boundary of the convex hull. Thus $T$ is
$\{2,2,1\}$-blocked. Hence, as illustrated in
\figref{BradsTwistedGrid}, one point $c$ in $T$ is at the
intersection of the two chords of $\conv(K)$, and exactly one point
in $T$ is on each edge of the boundary of $\conv(K)$, such that the
points on opposite edges of $\conv(K)$ are collinear with $c$.
\begin{figure}[ht]
\includegraphics[scale=0.8]{BradsTwistedGrid}
\caption{\figlabel{BradsTwistedGrid} A $\{4,2,2,1\}$-blocked point
set.}
\end{figure}
We claim that no other point is in $P$. Suppose otherwise, and let
$p$ be a point in $P$ outside $\conv(K)$ at minimum distance from
$\conv(K)$. Let $x$ be a point in $\conv(K)$ receiving the same
colour as $p$. Thus $p$ and $x$ are blocked by some point $b$ in
$\conv(K)$. Thus $b$ and $x$ are collinear with no other point in
$P\cap \conv(K)$. Hence $x\neq c$ and $x\not\in K$. Thus $x$ is in the
interior of one of the edges of the boundary of $\conv(K)$. Let $y$
be the point in $\conv(K)$ receiving the same colour as $x$. Thus
$x$ and $y$ are on opposite edges of the boundary of $\conv(K)$.
Hence $p$ and $y$ receive the same colour, implying $p$ and $y$ are
blocked. Since $p$ is at minimum distance from $\conv(K)$, this
blocker must a point $z$ of $\conv(K)$. This implies that $p,z,y$
and one other point of $\conv(K)$ are four collinear points, which
contradicts \lemref{SelfBlockingThree}. Hence no other point is in $P$, and $P$ is
$\{4,2,2,1\}$-blocked.
\end{proof}
\lemref{K4221} has the following corollary (let $K:=S$).
\begin{corollary}
\corlabel{K4221} Let $P$ be a 4-blocked set. Suppose that some
colour class $S$ consists of exactly four points in convex
position. Then $P$ is $\{4,2,2,1\}$-blocked.
\end{corollary}
The next lemma is a key step in our characterisation of representable
4-sets.
\begin{lemma}
\lemlabel{FourInColourClass} Each colour class in a 4-blocked point
set has at most four points.
\end{lemma}
\begin{proof}
Suppose that some $4$-blocked point set $P$ has a colour class $S$
with at least five points. Esther Klein~\citep{ES35} proved that
every set of least five points in general position in the plane
contains an empty quadrilateral. By \lemref{GenPos}, $S$ is in
general position. Thus $S$ contains a subset $K$, such that $|K|=4$
and $K$ is the vertex-set of a convex quadrilateral $\conv(K)$ that
is empty with respect to $S$. By \lemref{K4221}, $P$ is
$\{4,2,2,1\}$-blocked, which is the desired contradiction.
\end{proof}
\begin{lemma}
\lemlabel{K4222} Let $P$ be a 4-blocked point set with colour classes
$A, B, C, D$. Suppose that no colour class consists of exactly
four points in convex position (that is, \corref{K4221} is not
applicable). Furthermore, suppose that some colour class $A$
consists of exactly four points in nonconvex position. Then $P$ is
$\{4,2,2,2\}$-blocked (as in \figref{K4222}).
\end{lemma}
\begin{figure}[ht]
\includegraphics{K4222}
\caption{\figlabel{K4222} A $\{4,2,2,2\}$-blocked point set.}
\end{figure}
\begin{proof}
By \lemref{FourInColourClass}, each colour class has at most four
points. By assumption, every 4-point colour class is in nonconvex
position. We may assume that $A$ is a minimal in the sense that no
other 4-point colour class is within $\conv(A)$.
Let $Q:=P\cap \conv(A)$. Thus $Q$ is 4-blocked, and one colour
class is $A$. By the minimality of $A$, each other colour class in
$Q$ has at most three points. We first prove that $Q$ is
$\{4,2,2,2\}$-blocked, and then show that this implies that $P=Q$.
Let $A=\{a_1, a_2, a_3, a_4\}$, where $a_4$ is the interior point of
$\conv(A)$. Note that the edges with points in $A$ divide
$\conv (A)$ into three triangles with disjoint interiors. By
\lemref{New}, each
colour class of $Q$ is represented in each of these triangles;
this requires at least two points of each colour (one of which
could sit on the edge shared by two triangles).
We name a point with reference to its colour class, such as $b_1\in
B$; or, we name a point with reference to its position. Let
$x_{ij}$ be the unique member of $Q':=Q\setminus A$ that blocks
$\overline{a_ia_j}$, for $1\le i<j \le 4$. This accounts for
exactly 6 points of $Q'$.
$Q$ is not $\{4,3,3,3\}$-blocked, as otherwise we could delete the three
outer members of $A$ to represent $\{3,3,3,1\}$, which contradicts
\lemref{1333}. Thus $Q$ is $\{4,2,2,2\}$-blocked,
$\{4,3,2,2\}$-blocked, or $\{4,3,3,2\}$-blocked.
First suppose that $Q$ is $\{4,3,2,2\}$-blocked. Then $Q'$ consists of
six points $x_{ij}$ and one additional point, $y$. We have three
cases:
\begin{figure}[ht]
\includegraphics[scale=0.69]{Nonconvex4c1.png}
\hspace*{-4mm}
\includegraphics[scale=0.69]{Nonconvex4c2.png}
\hspace*{-4mm}
\includegraphics[scale=0.69]{Nonconvex4c3.png}
\hspace*{-4mm}\\
\vspace*{-2ex}
\caption{Diagrams of $\{4,3,2,2\}$ cases 1, 2, and 3.}
\end{figure}
\begin{enumerate}
\item If $y$ blocks two points $x_{i4}$ and $x_{j4}$, then suppose
WLOG that $x_{24}*y*x_{34}$. Now $y$ is on one side or the other
of $\overleftrightarrow{x_{14}a_4}$, and therefore sees at least
one of $x_{12}$, $x_{13}$; WLOG we may say that $y$ sees $x_{12}$.
Now $x_{12}$, $x_{14}$, $x_{24}$, and $y$ are four mutually
visible points in $Q'$.
\item If $y$ is collinear with, but does not block, two points
$x_{i4}$ and $x_{j4}$, then suppose WLOG that $y*x_{24}*x_{34}$.
Now $x_{12}$ is on one side or the other of
$\overleftrightarrow{yx_{24}}$, and therefore sees either $x_{14}$
or $x_{23}$. Whichever one $x_{12}$ sees, that point is mutually
visible with $x_{12}$, $x_{24}$, and $y$; thus we have four
mutually visible points in $Q'$.
\item If $y$ is not collinear with two points $x_{i4}$ and $x_{j4}$,
then $y$ sees all such points; thus $x_{14}$, $x_{24}$, $x_{34}$,
and $y$ are four mutually visible points in $Q'$.
\end{enumerate}
In all cases, we have four mutually visible points in three colours,
which is impossible.
Now suppose that $Q$ is $\{4,3,3,2\}$-blocked. Then $Q'$ comprises six
points $x_{ij}$ and two additional points, $y_1$ and $y_2$. Let
$\delta = \{ x_{i4}:1\le i \le 3\}$ and let $\delta_2$ be the set of
segments with endpoints in $\delta$. How many of the points $y_i$
block members of $\delta_2$? Again, we have three cases:
\begin{figure}[ht]
\includegraphics[scale=0.69]{Nonconvex4c4.png}
\hfill
\includegraphics[scale=0.69]{Nonconvex4c5.png}
\hfill
\includegraphics[scale=0.69]{Nonconvex4c6.png}
\hspace*{-2mm}\\
\vspace*{-2ex}
\caption{Diagrams of $\{4,3,3,2\}$ cases 1, 2, and 3.}
\end{figure}
\begin{enumerate}
\item If neither $y_1$ nor $y_2$ blocks a segment in $\delta_2$, then
recall that $y_1$ and $y_2$ cannot both block each other. Thus,
WLOG, $y_2$ does not block $y_1$. Now $\delta$ and $y_1$ give us
four mutually visible points in $Q'$.
\item If $y_1$ blocks a segment in $\delta_2$ and $y_2$ does not,
then WLOG $x_{24}*y_1*x_{34}$. \0{ If $y_2$ and $x_{12}$ are on
opposite sides of $\overleftrightarrow{x_{14}y_1}$, then
$x_{12}$, $x_{14}$, $x_{24}$, and $y_1$ are four mutually
visible points in $Q'$. No: $a_4$ could block
$\overline{y_1x_{12}}$. On the other hand,}Also WLOG, suppose
that $a_2$ and $a_4$ are on the same side of
$\overleftrightarrow{x_{14}y_1}$, so $a_4 \notin \conv(x_{13},
x_{14}, x_{34}, y_1)$. Now $y_2 \in \conv(x_{13}, x_{14}, x_{34},
y_1)$, for otherwise these four points in $Q'$ are mutually
visible.
Also note that $\overline{y_1x_{12}}$ is the only edge with
points in $\{x_{12}, x_{14}, x_{24}, y_1\}$ that can be blocked
by $a_4$. To prevent these four points from being mutually
visible, we require $y_1*a_4*x_{12}$.
Thus $y_1$ and $x_{12}$ belong to the same colour class, which we
may suppose is $B$. Since $x_{12}$, $x_{14}$, and $x_{24}$ are
mutually visible, we may suppose that $x_{24}\in C$ and
$x_{14}\in D$.
Since $x_{24}*y_1*x_{34}$, we have $x_{34}\in C$. Since $x_{23}$
sees $x_{24}$ and $y_1$, we have $x_{23}\in D$.
Now $y_2$ sees $y_1 \in B$ and $x_{34}\in C$, and we conclude
that $y_2\in D$; however, $y_2$ sees either $x_{14}$ or $x_{23}$,
and colour class $D$ is not blocked.
\item If both $y_1$ and $y_2$ block segments in $\delta_2$, then WLOG
$x_{i4}*y_i*x_{34}$ for $i=1,2$. Now $\{x_{14}, x_{24}, a_4\}$ is
insufficient to block $x_{12}$ from $\{y_1, y_2\}$, for this would
imply some $x_{12}*x_{i4}*y_i*x_{34}$. Therefore some $y_i$ sees
$x_{12}$; now $x_{12}$, $x_{14}$, $x_{24}$, and $y_i$ are four
mutually visible points in $Q'$.
\end{enumerate}
The only remaining case is that $Q$ is $\{4,2,2,2\}$-blocked. This is
possible, as illustrated in \figref{K4222}. We now show that this
point set is essentially the only $\{4,2,2,2\}$-blocked set, up to
betweenness-preserving deformations. We have exactly enough points
in colour classes $B$, $C$, and $D$ to block all edges between
points in colour class $A$. As each of the three $A$-triangles
with point $a_4$ must contain a representative from each of the
other three colour classes, it follows that of the six $A$-edges,
$B$ blocks one interior edge and the opposite boundary edge, and
likewise for $C$ and $D$. WLOG: $b_1=x_{14}$, $b_2=x_{23}$,
$c_1=x_{24}$, $c_2=x_{13}$, $d_1=x_{34}$, $d_2=x_{12}$. Since $b_1$
blocks $a_4$, $a_4$ cannot block $b_1$. Since $\overline{b_1b_2}$
can be blocked only by an interior point of $\conv(A)$, it follows
that either $c_1$ or $d_1$ blocks $\overline{b_1b_2}$. As these
cases are symmetric, we may choose $c_1\in\overline{b_1b_2}$. Now
$b_1$ cannot block $c_1$, so $c_1$ must be blocked by $d_1$, which
must in turn be blocked by $b_1$.
Now we show that $P=Q$. (This basically says that the point set in
\figref{K4222} cannot be extended without introducing a new
colour.)\ Suppose to the contrary that some point $x$ is in $P
\setminus Q$; thus $x \notin \conv(A)$. Note that every point
outside $\conv(A)$ can see a vertex of $\conv(A)$, so $x$ must not
be in colour class $A$. WLOG, we suppose $x$ matches the points in
$B$. Recall that $b_2$ is on the supporting line
$\overleftrightarrow{a_2a_3}$, and will see $x$ unless $x$ is in the
same half-plane (as determined by this line) as the rest of $Q$.
Who blocks $\overline{b_1x}$? Not $a_2$ or $a_3$, for this would
put $x$ in the wrong half-plane. Not $a_1$, $a_4$, $d_1$, or $d_2$,
for $b_1$ blocks each of these. Not $c_1$, which is on
$\overline{b_1b_2}$. Therefore $\overline{b_1x}$ can be blocked
only by $c_2$. Since $c_1\in\overline{b_1b_2}$ and
$c_2\in\overline{b_1x}$, it follows that $b_2$ and $x$ (and any
blocker between them) are on the same side of
$\overleftrightarrow{c_1c_2}$. But the only other points of $Q$ in
that open half-plane are $a_2$ and $a_3$, which cannot block $b_2$.
Thus $x$ sees $b_2$, which is a contradiction. Thus $P=Q$ and $P$ is
$\{4,2,2,2\}$-blocked.
\end{proof}
We now prove the main theorem of this section.
\begin{theorem}
A 4-set $\{a,b,c,d\}$ is representable if and only if
\begin{itemize}
\item $\{a,b,c,d\} = \{4,2,2,1\}$, or
\item $\{a,b,c,d\} = \{4,2,2,2\}$, or
\item all of $a,b,c,d \leq 3$ except for $\{3,3,3,1\}$
\end{itemize}
\end{theorem}
\begin{proof}
\twofigref{BradsTwistedGrid}{K4222} respectively show
$\{4,2,2,1\}$-blocked and $\{4,2,2,2\}$-blocked point sets. When
$a,b,c,d \leq 3$, the required constructions are described in
\propref{SmallColourClasses}. Now we prove that these are the only
representable 4-sets. Let $P$ be a $4$-blocked point set. By
\lemref{FourInColourClass}, each colour class has at most four
points. Let $S$ be the largest colour class. If $|S|\leq 3$ then we
are done by \propref{SmallColourClasses}. Now assume that $|S|=4$.
If $S$ is in nonconvex position, then $P$ is $\{4,2,2,2\}$-blocked by
\lemref{K4222}. If $S$ is in convex position, then $P$ is
$\{4,2,2,1\}$-blocked by \corref{K4221}.
\end{proof}
\begin{corollary}
Every 4-blocked set has at most 12 points, and there is a
4-blocked set with 12 points.
\end{corollary}
Note that in addition to the $\{3,3,3,3\}$-blocked set shown in
\figref{K3333}, there is a different $\{3,3,3,3\}$-blocked point set, as
illustrated in \figref{3333b}.
\begin{figure}[!ht]
\includegraphics[scale=0.6]{3333b.png}
\vspace*{-2ex}
\caption{\figlabel{3333b} Another $\{3,3,3,3\}$-blocked point
set.}
\end{figure}
\section{Midpoint-Blocked Point Sets}
\seclabel{Midpoint}
A $k$-blocked point set $P$ is \emph{$k$-midpoint-blocked} if for each
monochromatic pair of distinct points $v,w\in P$ the midpoint of
$\overline{vw}$ is in $P$. Of course, the midpoint of $\overline{vw}$
blocks $v$ and $w$. A point set $P$ is
\emph{$\{n_1,\dots,n_k\}$-midpoint-blocked} if it is
$\{n_1,\dots,n_k\}$-blocked and $k$-midpoint-blocked.
For example, the
point set in \figref{K3333} is $\{3,3,3,3\}$-midpoint-blocked.
Another interesting example is the projection\footnote{If $G$ is the
visibility graph of some point set $P\subseteq\ensuremath{\mathbb{R}}^d$, then $G$ is the
visibility graph of some projection of $P$ to $\ensuremath{\mathbb{R}}^2$ (since a random
projection of $P$ to $\ensuremath{\mathbb{R}}^2$ is occlusion-free with probability
$1$).} of $[3]^d$. With $d=1$ this point set is $\{2,1\}$-blocked,
with $d=2$ it is $\{4,2,2,1\}$-blocked, and with $d=3$ it is
$\{8,4,4,4,2,2,2,1\}$-blocked. In general, each set of points with
exactly the same set of coordinates equal to $2$ is a colour
class, there are $2^{d-i}$ colour classes of points with exactly $i$
coordinates equal to $2$, and $[3]^d$ is $\{\binom{d}{i}\times
2^i:i\in[0,d]\}$-midpoint-blocked and $2^d$-midpoint-blocked.
We now prove \conjref{kBlocked} when restricted to
$k$-midpoint-blocked point sets. (Finally we have weakened
\conjref{BigClique} to something proveable!)\
\citet{HHUZ} introduced the following definition. Let $m(n)$ be the
minimum number of midpoints determined by some set of $n$ points in
general position in the plane. Since the midpoint of $\overline{vw}$
blocks $v$ and $w$, we have $b(n)\leq m(n)$. \citet{HHUZ} constructed
a set of $n$ points in general position in the plane that determine at
most $cn^{\log 3}$ midpoints for some contant $c$. (All logarithms
here are binary.)\ Thus $b(n)\leq
m(n)\leq cn^{\log3} =cn^{1.585\ldots}$. This upper bound was
improved by \citet{Pach-Midpoints-Geom03} (and later by
\citet{Matousek09}) to
$$b(n)\leq m(n)\leq nc^{\sqrt{\log n}}\enspace.$$
\citet{HHUZ} conjectured that $m(n)$ is super-linear, which was
verified by \citet{Pach-Midpoints-Geom03}; that is,
$\frac{m(n)}{n}\rightarrow\infty$ as $n\rightarrow\infty$.
\citet{PorWood-Blockers} proved the following more precise version:
For some constant $c>0$, for all $\epsilon>0$ there is an integer
$N(\epsilon)$ such that $m(n)\geq c n(\log n)^{1/(3+\epsilon)}$ for
all $n\geq N(\epsilon)$.
\begin{theorem}
For each $k$ there is an integer $n$ such that every
$k$-midpoint-blocked set has at most $n$ points. More precisely,
there is an absolute constant $c$ and for each $\epsilon>0$ there is
an an integer $N(\epsilon)$, such that for all $k$, every
$k$-midpoint-blocked set has at most
$k\max\{N(\epsilon),c^{(k-1)^{3+\epsilon}}\}$ points.
\end{theorem}
\begin{proof}
Let $P$ be $k$-midpoint-blocked set of $n$ points. If $n\leq k
N(\epsilon)$ then we are done. Now assume that $\frac{n}{k}>
N(\epsilon)$. Let $S$ be a set of exactly $s:=\ceil{\frac{n}{k}}$
monochromatic points in $P$. Thus $S$ is in general position by
\lemref{GenPos}. And for every pair of distinct points $v,w \in S$
the midpoint of $\overline{vw}$ is in $P-S$. Thus
$$ c \tfrac{n}{k}(\log\tfrac{n}{k})^{1/(3+\epsilon)}\leq m(s) \leq
n-s\leq n(1-\tfrac{1}{k})\enspace.$$ Hence
$(\log\tfrac{n}{k})^{1/(3+\epsilon)}\leq(k-1)/c$, implying
$n\leq k 2^{((k-1)/c)^{3+\epsilon}}$. The result follows.
\end{proof}
We now construct $k$-midpoint-blocked point sets with a `large' number
of points. The method is based on the following product of point sets
$P$ and $Q$. For each point $v\in P\cup Q$, let $(x_v,y_v)$ be the
coordinates of $v$. Let $P\times Q$ be the point set $\{(v,w):v\in
P,w\in Q\}$ where $(v,w)$ is at $(x_v,y_v,x_w,y_w)$ in $4$-dimensional
space. For brevity we do not distinguish between a point in $\ensuremath{\mathbb{R}}^4$ and
its image in an occlusion-free projection of the visibility graph of
$P\times Q$ into $\ensuremath{\mathbb{R}}^2$.
\begin{lemma}
\lemlabel{ProductLemma} If $P$ is a
$\{n_1,\dots,n_k\}$-midpoint-blocked point set and $Q$ is a
$\{m_1,\dots,m_{\ell}\}$-midpoint-blocked point set, then $P\times Q$
is $\{n_im_j:i\in[k],j\in[\ell]\}$-midpoint-blocked.
\end{lemma}
\begin{proof}
Colour each point $(v,w)$ in $P\times Q$ by the pair
$(\col(v),\col(w))$. Thus there are $n_im_j$ points for the
$(i,j)$-th colour class. Consider distinct points $(v,w)$ and
$(a,b)$ in $P\times Q$.
Suppose that $\col(v,w)=\col(a,b)$. Thus $\col(v)=\col(a)$ and
$\col(w)=\col(b)$. Since $P$ and $Q$ are midpoint-blocked,
$\ensuremath{\protect\tfrac{1}{2}}(v+a)\in P$ and $\ensuremath{\protect\tfrac{1}{2}}(w+b)\in Q$. Thus
$(\ensuremath{\protect\tfrac{1}{2}}(v+a),\ensuremath{\protect\tfrac{1}{2}}(w+b))$, which is positioned at
$(\ensuremath{\protect\tfrac{1}{2}}(x_v+x_a),\ensuremath{\protect\tfrac{1}{2}}(y_v+y_a),\ensuremath{\protect\tfrac{1}{2}}(x_w+x_b),\ensuremath{\protect\tfrac{1}{2}}(y_w+y_b))$, is
in $P\times Q$. This point is the midpoint of
$\overline{(v,w)(a,b)}$. Thus $(v,w)$ and $(a,b)$ are blocked by
their midpoint in $P\times Q$.
Conversely, suppose that some point $(r,s)\in P\times Q$ blocks
$(v,w)$ and $(a,b)$. Thus $x_r=\alpha x_v+(1-\alpha)x_a$ for some
$\alpha\in(0,1)$, and $y_r=\beta y_v+(1-\beta)y_a$ for some
$\beta\in(0,1)$, and $x_s=\delta x_w+(1-\delta)x_b$ for some
$\delta\in(0,1)$, and $y_s=\gamma y_w+(1-\gamma)y_b$ for some
$\gamma\in(0,1)$. Hence $r$ blocks $v$ and $a$ in $P$, and $s$
blocks $w$ and $b$ in $Q$. Thus $\col(v)=\col(a)\neq\col(r)$ in
$P$, and $\col(w)=\col(b)\neq\col(s)$ in $Q$, implying
$(\col(v),\col(w))=(\col(a),\col(b))$.
We have shown that two points in $P\times Q$ are blocked if and only
if they have the same colour. Thus $P\times Q$ is blocked. Since
every blocker is a midpoint, $P\times Q$ is midpoint-blocked.
\end{proof}
Say $P$ is a $k$-midpoint blocked set of $n$ points. By
\lemref{ProductLemma}, the $i$-fold product $P^i:=P\times\dots\times
P$ is a $k^i$-blocked set of $n^i=(k^i)^{\log_k n}$ points. Taking $P$
to be the $\{3,3,3,3\}$-midpoint-blocked point set in \figref{K3333}, we
obtain the following result:
\begin{theorem}
For all $k$ a power of $4$, there is a $k$-blocked set of $k^{\log_4
12}=k^{1.79\ldots}$ points.
\end{theorem}
This result describes the largest known construction of $k$-blocked or
$k$-midpoint blocked point sets. To promote further research, we make
the following strong conjectures:
\begin{conjecture}
\conjlabel{SizeBlocked} Every $k$-blocked point set has $O(k^2)$
points.
\end{conjecture}
\begin{conjecture}
\conjlabel{SizeColorClass} In every $k$-blocked point set there are
at most $k$ points in each colour class.
\end{conjecture}
\conjref{SizeColorClass} would be tight for the projection of $[3]^d$
with $k=2^d$. Of course, \conjref{SizeColorClass} implies
\conjref{SizeBlocked}, which implies \conjref{kBlocked}.
\section*{Acknowledgements} This research was initiated at The 24th
Bellairs Winter Workshop on Computational Geometry, held in February
2009 at the Bellairs Research Institute of McGill University in
Barbados. The authors are grateful to Godfried Toussaint and Erik
Demaine for organising the workshop, and to the other
participants for providing a stimulating working environment.
\def\cprime{$'$} \def\soft#1{\leavevmode\setbox0=\hbox{h}\dimen7=\ht0\advance
\dimen7 by-1ex\relax\if t#1\relax\rlap{\raise.6\dimen7
\hbox{\kern.3ex\char'47}}#1\relax\else\if T#1\relax
\rlap{\raise.5\dimen7\hbox{\kern1.3ex\char'47}}#1\relax \else\if
d#1\relax\rlap{\raise.5\dimen7\hbox{\kern.9ex \char'47}}#1\relax\else\if
D#1\relax\rlap{\raise.5\dimen7 \hbox{\kern1.4ex\char'47}}#1\relax\else\if
l#1\relax \rlap{\raise.5\dimen7\hbox{\kern.4ex\char'47}}#1\relax \else\if
L#1\relax\rlap{\raise.5\dimen7\hbox{\kern.7ex
\char'47}}#1\relax\else\message{accent \string\soft \space #1 not
defined!}#1\relax\fi\fi\fi\fi\fi\fi} \def\Dbar{\leavevmode\lower.6ex\hbox to
0pt{\hskip-.23ex\accent"16\hss}D}
|
\section{Introduction}
The discovery of ferro-pnictide superconductors has set a new and
unforseen direction in the search for high-temperature superconductivity.\cite{new_sc}
A common structural feature of these materials are
square lattices of iron atoms that are stacked up on top of each other.\cite{delacruz}
Density functional theory (DFT) calculations find that the iron 3d orbitals contribute most
of the electronic spectral weight near the Fermi surface.\cite{haule_08}$^{,}$\cite{cao_08}
This implies that the iron $3d$ electrons surprisingly play an active role in the superconductivity.
In contrast, DFT calculations also predict that
the iron moments order antiferromagnetically along the $a$ and along the $c$ axes
of the iron lattice at low temperature in the undoped parent compounds,\cite{dong_08}
with an ordered moment of approximately $2$ Bohr magnetons ($\mu_B$).\cite{ma_08}
Elastic neutron diffraction studies have confirmed the prediction of long-range
commensurate spin-density wave (cSDW) order,
but they find a much lower ordered moment that can be a small
fraction of the Bohr magneton ($0.36\,\mu_B$) in some cases.\cite{delacruz}
It was suggested early on that Heisenberg exchange between both nearest-neighbor ($J_1$)
and next-nearest neighbor ($J_2$) spins on a square lattice of iron atoms
could account for the low
ordered moment that is observed because of magnetic frustration.\cite{Si&A}
Recent calculations based on the linear spin-wave approximation\cite{j1j2_sw}
indicate that this effect requires fine tuning however,\cite{thalmeier_10}
and that it occurs only very close to the point of maximum frustration,
$J_2 = 0.5 \, J_1$.
A possibly more viable explanation is
that the low ordered cSDW moment commonly observed
in ferro-pnictide parent compounds is due to proximity to hidden magnetic order
among particular iron $3d$ orbitals such that the magnetic moments cancel out per iron atom.
The author and Rezayi have shown that such violations of Hund's rule
can occur in multi-orbital Heisenberg models with off-diagonal frustration.\cite{jpr_ehr_09}
Spin-wave excitations of the cSDW groundstate
have also been detected in parent compounds to ferro-pnictide
superconductors at low temperature
by inelastic neutron scattering.\cite{zhao_08}$^,$\cite{diallo_09}
Contrary to the prediction of
the conventional $J_1$-$J_2$ model over the square lattice however,\cite{oitmaa_10}
the measured spin-wave spectrum fails to soften at the wave number associated with N\'eel order,
$(\pi / a, \pi / a)$. Here $a$ denotes the square lattice constant.
It shows a local maximum there instead.\cite{zhao_09}
Fits to the measured spin-wave dispersion invoke
unphysically large anisotropy in the Heisenberg
spin exchange across nearest-neighbor links on the square lattice of iron atoms
in order to account for the dispersion at the Brillouin zone boundary.
More recently, low-energy magnetic excitations centered at the cSDW wave numbers
have been detected in a ferro-pnictide {\it superconductor}.\cite{hayden_10}
These disperse up in energy in an anisotropic fashion
that is consistent with the presence of some degree of cSDW order.
In this paper, we continue to explore the hidden magnetic order that is
predicted to exist in multi-orbital Heisenberg models over the square lattice when
the diagonal and the off-diagonal exchange coupling constants
are sufficiently different.\cite{jpr_ehr_09}
Two spin-1/2 orbitals per iron atom are assumed,
which implies spin-1 iron moments if Hund's rule is obeyed.
A quantum critical point\cite{sachdev_qcp}
is identified within the linear spin-wave approximation
that separates hidden ferromagnetic or hidden N\'eel order
at weak Hund's rule coupling from a cSDW at strong Hund's rule coupling.
Numerical exact diagonalizations of the two-orbital Heisenberg model
over a $4 \times 4$ lattice indicate
that the moment associated with cSDW order decays strongly as it enters hidden magnetic order
through the putative quantum critical point.
The linear spin-wave approximation, on the other hand,
finds observable low-energy spin-wave excitations at the quantum critical point
that disperse linearly from the cSDW wave numbers in an anisotropic fashion.
A spin gap begins to grow there upon entering hidden magnetic order.
Observable low-energy spin-wave excitations due to hidden order
that are centered at zero momentum
are also predicted by the linear spin-wave approximation,
but these have very small spectral weight.
Fits of the spin-wave spectrum observed by inelastic neutron scattering in
ferro-pnictide materials\cite{zhao_09}$^,$\cite{hayden_10}
to the critical linear spin-wave spectrum are achieved.
They notably account for
the dispersion of the low-energy spinwaves centered at the cSDW wave numbers
and for the local maximum that exists at the wavenumber associated with N\'eel order.
A critical Hund's rule coupling of order $100$ meV is extracted from these fits.
Below we discuss useful theoretical aspects of the
Heisenberg model for ferro-pnictide materials\cite{Si&A}
before we go on to analyze it.
\section {Multi-Orbital Heisenberg Model}
Recent low-temperature measurements of the optical conductivity
in the cSDW phase of parent compounds to ferro-pnictide high-$T_c$ superconductors
find a strong suppression of the integrated kinetic energy for the conduction electrons
compared to that predicted by electronic band-structure calculations.\cite{basov_09}
A low carrier density
in the cSDW phase of ferro-pnictide materials
is consistent with
quantum-oscillation experiments\cite{sebastian_08}$^,$\cite{analytis_09}
and with angle-resolved photoemission studies\cite{richard_10}
that find evidence for tiny 2D Fermi-surface ``spots'',
each on the order of 1\% of the Brillouin zone.
Also,
the dependence on temperature shown by the electrical resistance
of such parent compounds
indicates proximity to a metal-insulator transition.\cite{wang}
Undoped ferro-pnictide materials in the cSDW phase
then may very well lie near a Mott transition,
where conduction electrons experience moderately strong repulsive interactions.
By continuity
with the limit of strong inter-electron repulsion,
we therefore believe that a local-moment description of magnetism in parent compounds
to iron-based high-$T_c$ superconductors is a justifiable starting point at low temperature.
The absence of a Stoner continuum of incoherent magnetic excitations
reported in recent experimental studies of magnetic excitations
in the cSDW phase of a ferro-pnictide material
is consistent with this view.\cite{zhao_09}
Consider then the following
multi-orbital spin-1/2 Hamiltonian that
contains near-neighbor Heisenberg exchange among local iron moments
within isolated layers in addition to Hund's-rule coupling:\cite{Si&A}
\begin{equation}
H = {1\over 2} J_0 \sum_i \biggl[\sum_{\alpha} {\bf S}_i (\alpha)\biggr]^2 +
\sum_{\langle i,j \rangle} \sum_{\alpha , \beta} J_1^{\alpha,\beta}
{\bf S}_i (\alpha) \cdot {\bf S}_j (\beta) +
\sum_{\langle\langle i,j \rangle\rangle} \sum_{\alpha , \beta} J_2^{\alpha,\beta}
{\bf S}_i (\alpha) \cdot {\bf S}_j (\beta).
\label{j0j1j2}
\end{equation}
Above, ${\bf S}_i(\alpha)$ is the spin operator that acts on the spin-1/2 state of orbital
$\alpha$ in the iron atom at site $i$. The latter runs over the square lattice
of iron atoms that make up an isolated layer.
The application of Hund's rule is controlled by a local ferromagnetic
Heisenberg exchange constant $J_0 < 0$,
while nearest neighbor and next-nearest neighbor Heisenberg exchange
across the links $\langle i,j\rangle$ and $\langle\langle i,j\rangle\rangle$
is controlled by the tensor exchange constants $J_1^{\alpha,\beta}$ and $J_2^{\alpha,\beta}$, respectively.
The strength of the crystal field at each iron atom
determines the number of orbitals per iron atom above.
It can be as low as two for strong crystal fields compared to the Hund's-rule coupling,
in which case spin-1/2 moments exist on
the iron $3d_{xz}$ and $3d_{yz}$ orbitals.\cite{cao_08}$^,$\cite{li_08}
Four $3d$ orbitals per iron atom carry spin-1/2 moments
in the case of weak crystal fields compared to the Hund's-rule coupling,
on the other hand.\cite{haule_08}$^,$\cite{Si&A}
Last, the above model Hamiltonian reduces to the conventional
$J_1$-$J_2$ model over the square lattice
when Hund's rule is enforced,\cite{j1j2_sw}$^,$\cite{thalmeier_10}$^,$\cite{oitmaa_10}
$-J_0 \rightarrow \infty$,
where the spin at each site is given by half the number of orbitals, $S = 1$ or $S = 2$.
We now illustrate a useful transformation of the model Hamiltonians (\ref{j0j1j2})
over the square lattice
that relates eigenstates and eigenvalues with different nearest-neighbor Heisenberg exchange
coupling constants.
Let $p$ denote a permutation of the orbitals. Consider then the partial ``reflection''
$P_{A(B)}$:
\begin{equation}
\quad {\rm orbital}\quad \alpha \rightarrow p(\alpha)
\quad {\rm on\ \ the\ }A(B)\ {\rm sublattice}
\label{trans_orb}
\end{equation}
of the N\'eel state.
If $\Psi$ is an eigenstate of the original model Hamiltonian (\ref{j0j1j2})
with energy $E$,
the transformed state $\Psi^{\prime} = P_{A(B)} \Psi$ is clearly then
an eigenstate of the transformed Hamiltonian $H^{\prime} = P_{A(B)} H P_{A(B)}^{-1}$
with energy $E$.
Importantly, the Heisenberg exchange coupling constants transform under (\ref{trans_orb}) as
\begin{equation}
\quad J_1^{\alpha,\beta}\rightarrow J_1^{p(\alpha),\beta}
\quad {\rm and} \quad J_2^{\alpha,\beta}\rightarrow J_2^{p(\alpha),p(\beta)}
\label{trans_exc}
\end{equation}
when orbital $\alpha$ lies on the $A (B)$ sublattice
of the N\'eel state.
Both $J_2^{\alpha,\beta}$ on the $B (A)$ sublattice
and the Hund's rule coupling $J_0$ remain fixed.
Yet how does the spectrum of states transform under (\ref{trans_orb})?
Let $T_x$ and $T_y$ denote translations along the
$x$ axis and along the $y$ axis by one lattice constant $a$.
Because $T_x T_y$
moves each sublattice to itself,
it commutes with $P_A$ and with $P_B$.
If $\Psi$ has a crystal momentum $\hbar\bf k$,
we therefore have that
$T_x T_y\, \Psi^{\prime} = e^{ i (k_x+ k_y) a}\, \Psi^{\prime}$.
The crystal momentum then transforms under (\ref{trans_orb}) as
\begin{equation}
{\bf k} \rightarrow {\bf k} \ {\rm or} \ {\bf k} + (\pi / a, \pi / a) .
\label{trans_mmntm}
\end{equation}
Also, the spin of the new state $\Psi^{\prime}$ remains unchanged because $P_A$ and $P_B$
both commute
with ${\bf S}_i = \sum_{\alpha} {\bf S}_i (\alpha)$.
It is also useful to consider the special case where all nearest-neighbor and next-nearest-neighbor
exchange coupling constants are equal, respectively:
$J_1^{\alpha,\beta} = J_1$ and $J_2^{\alpha,\beta} = J_2$.
The model Hamiltonian (\ref{j0j1j2}) then reduces to
$H = {1\over 2} J_0 \sum_i {\bf S}_i \cdot {\bf S}_i +
J_1\sum_{\langle i,j \rangle} {\bf S}_i \cdot {\bf S}_j +
J_2\sum_{\langle\langle i,j \rangle\rangle} {\bf S}_i \cdot {\bf S}_j$.
Because ${\bf S}_i + {\bf S}_j$ commutes with ${\bf S}_i \cdot {\bf S}_i$,
the latter then commutes with the Hamiltonian.
The total spin at a given site $i$ is hence a good quantum number.
This means that the groundstate obeys Hund's rule in the classical limit
because states with maximum total spin at a given site
minimize both the Hund's-rule energy ($J_0 < 0$) and the
Heisenberg exchange energies in such a case.
A violation of Hund's rule will therefore
require some variation in the Heisenberg exchange coupling
constants among the different iron orbitals.
\section{Spin-1 Iron Atoms}
The iron 3d orbitals in ferro-pnictide materials experience a crystal field
due to the neigboring pnictide atoms that lie above/below the center of each unit
square of iron atoms.
Both DFT calculations and basic considerations indicate that
the crystal field splits these levels
so that the degenerate $3d_{xz}$ and $3d_{yz}$ orbitals
lie just below the highest-energy $3d$ orbital.\cite{cao_08}$^,$\cite{li_08}
The Pauli principle then implies that
spin $s = \hbar / 2$ moments exists
on both the $3d_{xz}$ and the $3d_{yz}$ orbitals
in the limit of weak Hund's rule coupling.
Henceforth,
we shall assume that only two spin-1/2 orbitals,
$a$ and $b$,
exist per iron atom in parent compounds to ferro-pnictide high-$T_c$ superconductors.
This is consistent with the ordered moment of $2\, \mu_B$ obtained by DFT calculations
for antiferromagnetic groundstates
of ferro-pnictide parent compounds.\cite{cao_08}$^,$\cite{dong_08}$^,$\cite{ma_08}
Let us next fix notation by assuming
unique diagonal and unique off-diagonal Heisenberg exchange coupling constants
over the nearest neighbor and the next-nearest neighbor links of
the square lattice of iron atoms:
\begin{equation}
J_{1(2)}^{a,a} = J_{1(2)}^{\parallel} = J_{1(2)}^{b,b}
\quad {\rm and} \quad
J_{1(2)}^{a,b} = J_{1(2)}^{\perp} = J_{1(2)}^{b,a}.
\label{jays}
\end{equation}
The author and Rezayi have recently shown that strong enough off-diagonal frustration,
$J_2^{\perp} > 0$,
can lead to either hidden ferromagnetic order or to hidden N\'eel order
among spin-1/2 moments that are governed by
the $J_0$-$J_1$-$J_2$ model Hamiltonian\cite{jpr_ehr_09} (\ref{j0j1j2}).
Figure \ref{hidden_order} displays the two types of hidden order,
both of which show no net magnetic moment.
Proximity of the collinear antiferromagnetic state,
which exists at $J_2 > |J_1|/2$ when Hund's rule is enforced,
to such hidden order can therefore
account for the low ordered moment that is observed by elastic neutron diffraction
in the cSDW state that exists at low temperature
in undoped ferro-pnictide materials.\cite{delacruz}
Balancing the energy of such hidden-order groundstates in the classical limit, $s\rightarrow\infty$,
against that of conventional magnetic order
leads to critical values
for the Hund's rule coupling that are listed in Table \ref{transitions}.
True magnetic order becomes unstable to hidden magnetic order
at weaker Hund's-rule coupling, $- J_0 < - J_{0c}$.
A true magnetically ordered groundstate therefore first becomes
unstable to the hidden-order groundstate
that has the larger of the two critical values for the Hund's rule coupling strength,
$- J_{0c}$.
By Table \ref{transitions}, for example,
the true cSDW state becomes unstable to hidden ferromagnetic order at
weak Hund's rule coupling only if the off-diagonal nearest-neighbor
Heisenberg exchange is larger than the corresponding diagonal exchange:
$J_1^{\perp} > J_1^{\parallel}$.
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{figure_1.ps}
\caption{Shown are two groundstates of a square lattice
of spin-1 iron atoms that violate Hund's rule, but that
exhibit hidden magnetic order.
[See Eq. (\ref{j0j1j2}) and Table \ref{transitions}.]}
\label{hidden_order}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
& True cSDW & True N\'eel & True Ferromagnet \\
\hline
Hidden Ferromagnet & $- J_{0c} = 2 (J_1^{\perp} - J_1^{\parallel}) - 4 J_2^{\parallel}$
& $- J_{0c} = 4 J_2^{\perp} - 4 J_1^{\parallel}$ & $- J_{0c} = 4 J_2^{\perp} + 4 J_1^{\perp}$ \\
Hidden N\'eel & $- J_{0c} = 2 (J_1^{\parallel} - J_1^{\perp}) - 4 J_2^{\parallel}$
& $- J_{0c} = 4 J_2^{\perp} - 4 J_1^{\perp}$ & $- J_{0c} = 4 J_2^{\perp} + 4 J_1^{\parallel}$ \\
\hline
\end{tabular}
\caption{Listed are the phase boundaries that separate true magnetic order
from hidden magnetic order. The critical Hund's rule coupling
is obtained by balancing the classical energies of the model Hamiltonian (\ref{j0j1j2}).}
\label{transitions}
\end{center}
\end{table}
{\it Linear Spin-Wave Theory.}
Study of the classical limit, $s\rightarrow\infty$, therefore indicates that
extremely low ordered moments are possible at weak enough
Hund's rule coupling, $J_0 < 0$,
when off-diagonal frustration exists:
$J_2^{\perp} > 0$.
The nature of spin-wave excitations about the hidden magnetic order
that is shown in Fig. \ref{hidden_order}
can also be obtained from the classical limit.
In such case,
each spin-1/2 moment obeys the following dynamical equation for precession:
\begin{equation}
\dot {\bf S}_i (\alpha) =
\Biggl[{\partial H \over{\partial {\bf S}_i (\alpha)}} - {\bf h}_i (\alpha)\Biggr] \times {\bf S}_i (\alpha).
\label{precess}
\end{equation}
Here we have applied a magnetic field to the system, ${\bf h}_i (\alpha)$,
that is in principle sensitive to
the orbital degree of freedom, $\alpha = a$ or $b$.
If the external magnetic field is a plane-wave with wavenumbers ${\bf k}$ and frequency $\omega$,
then its transverse component with respect to the spin axis for hidden magnetic order
generates a corresponding linear response in the
transverse components of the electronic spin.
It has the form
\begin{equation}
{\bf S}_{\perp} (\pm, {\bf k}, \omega) =
\chi_{\perp} (\pm, {\bf k},\omega) \, {\bf h}_{\perp} (\pm, {\bf k},\omega)
\label{diag_response}
\end{equation}
and
\begin{equation}
{\bf S}_{\perp} (\mp, {\bf k}, \omega) =
i {\bar\chi}_{\perp} ({\bf k},\omega) \, {\bf n}_{\parallel} \times {\bf h}_{\perp} (\pm, {\bf k},\omega)
\label{off_diag_response}
\end{equation}
in the case of the hidden ferromagnetic state.
Here we define true ($+$) and hidden ($-$) magnetic moments and magnetic fields by
${\bf S} (\pm) = {\bf S}(a) \pm {\bf S}(b)$ and by
${\bf h} (\pm) = [{\bf h}(a) \pm {\bf h}(b)]/2$,
while ${\bf n}_{\parallel}$ denotes the unit vector parallel to
the magnetization axis for hidden order.
Recall that the hidden ferromagnet is actually
an antiferromagnet over the two sublattices set by the two orbitals $a$ and $b$
(see Fig. \ref{hidden_order} and ref. \cite {af_sw_theory}).
The true magnetic susceptibility, $\chi_{\perp} (+, {\bf k},\omega)$,
the hidden magnetic susceptibility, $\chi_{\perp} (-, {\bf k},\omega)$,
and the conjugate susceptibility ${\bar\chi}_{\perp} ({\bf k},\omega)$
can then be obtained by solving the linear precession equation (\ref{precess})
that results in terms of the four transverse spin components per site,
${\bf S}_{\perp}(a)$ and ${\bf S}_{\perp}(b)$.
This yields
\begin{equation}
\chi_{\perp} (\pm, {\bf k},\omega) = s {\Omega_{\mp}\over{(\Omega_+ \Omega_-)^{1/2}}}
([\omega + (\Omega_+ \Omega_-)^{1/2}]^{-1} - [\omega - (\Omega_+ \Omega_-)^{1/2}]^{-1})
\label{chi}
\end{equation}
and
\begin{equation}
{\bar\chi}_{\perp} ({\bf k},\omega) = s {\omega\over{(\Omega_+ \Omega_-)^{1/2}}}
([\omega + (\Omega_+ \Omega_-)^{1/2}]^{-1} - [\omega - (\Omega_+ \Omega_-)^{1/2}]^{-1}),
\label{chi_bar}
\end{equation}
with frequencies given by
\begin{eqnarray*}
\Omega_- &=& s(J_1^{\perp} - J_1^{\parallel}) \sum_{n=x,y} (2\, {\rm sin}\, {1\over 2} k_n a )^2 +
s (J_2^{\perp} - J_2^{\parallel}) \sum_{n=+,-} (2\, {\rm sin}\, {1\over 2} k_n a)^2\\
\Omega_+ &=& 2sJ_0 + s J_1^{\perp} \sum_{n=x,y} (2\, {\rm cos}\, {1\over 2} k_n a )^2
- s J_1^{\parallel} \sum_{n=x,y} (2\, {\rm sin}\, {1\over 2} k_n a )^2 \\
&& + s J_2^{\perp} \sum_{n=+,-} (2\, {\rm cos}\, {1\over 2} k_n a)^2
- s J_2^{\parallel} \sum_{n=+,-} (2\, {\rm sin}\, {1\over 2} k_n a)^2
\end{eqnarray*}
in the case of the hidden ferromagnetic state.
Above, $k_\pm = k_x \pm k_y$,
$a$ denotes the square lattice constant,
and $s$ is the electron spin $\hbar / 2$.
Next observe that the hidden N\'eel state is related to the hidden ferromagnetic state
by the transformation (\ref{trans_orb}) mentioned in the previous section.
(See Fig. \ref{hidden_order}.)
By (\ref{trans_exc}),
the frequencies $\Omega_{\pm}$ in that case are therefore obtained from the ones listed above
for the hidden ferromagnet after making the exchange
of the nearest-neighbor Heisenberg coupling constants,
$J_1^{\parallel} \leftrightarrow J_1^{\perp}$.
Further,
the wavenumbers of the hidden spin ${\bf S} (-)$ and
of the hidden magnetic field ${\bf h} (-)$
that appear in the linear-response equations
(\ref{diag_response}) and (\ref{off_diag_response})
must be replaced by the wave numbers
on the antiferromagnetic sublattice,
${\bf k} \rightarrow {\bf k} - (\pi / a, \pi / a)$.
The wavenumbers of the true spin and of the true magnetic field that
appear there
remain unchanged, however.
The end result is that the true magnetic susceptibility
$\chi_{\perp} (+, {\bf k}, \omega)$
in the hidden N\'eel state is given by that in the hidden ferromagnetic phase with
the exchange
$J_1^{\parallel} \leftrightarrow J_1^{\perp}$ alone,
whereas the hidden magnetic susceptibility $\chi_{\perp} (-, {\bf k}, \omega)$ there
is given by that in the hidden ferromagnetic phase
with the previous exchange along with the replacement
${\bf k} \rightarrow {\bf k} + (\pi / a, \pi / a)$
in the wave numbers. [Cf. Eq. (\ref{trans_mmntm}).]
Equation (\ref{chi}) and the discussion above yield true static magnetic susceptibilities
$\chi_{\perp} (0) = (J_0 + 4 J_1^{\perp} + 4 J_2^{\perp})^{-1}$
for the hidden ferromagnetic state and
$\chi_{\perp} (0) = (J_0 + 4 J_1^{\parallel} + 4 J_2^{\perp})^{-1}$
for the hidden N\'eel state.
It also yields static susceptibilities for hidden magnetic order given by
$\chi_{\perp} (-, {\bf k}, 0) = 2s / \Omega_-$,
which correctly diverge as $\bf k$ approaches the wavenumber that
corresponds to either long-range ferromagnetic or N\'eel order.
Finally, the poles of the dynamical susceptibility
for hidden magnetic order (\ref{chi}), $\chi_{\perp} (-, {\bf k}, \omega)$,
represent two spin-wave excitations of momentum $\hbar {\bf k}$
that correspond to the two possible orientations of the
transverse component of the hidden moment,
${\bf S}_{\perp} (-, {\bf k}, \omega)$.
They both lie at an energy $\hbar (\Omega_+ \Omega_-)^{1/2}$ above the groundstate energy.
These spin-wave excitations are also reflected by
the poles in the dynamical susceptibility for
true magnetic order (\ref{chi}), $\chi_{\perp} (+, {\bf k}, \omega)$.
As mentioned above, however, the wave number ${\bf k}$ shifts
by the characteristic amount $(\pi/a, \pi/a)$
in the case of hidden N\'eel order.
Last, both spinwaves have spectral weights
\begin{equation}
A ({\pm, \bf k}) = \pi s (\Omega_{\mp} / \Omega_{\pm})^{1/2}
\label{weight}
\end{equation}
in the true ($+$) and the hidden ($-$) channels
that are read off from the imaginary part of the dynamical susceptibility:
\begin{equation}
{\rm Im}\, \chi_{\perp} (\pm, {\bf k}, \omega) =
A(\pm, {\bf k}) (\delta[\omega - (\Omega_+ \Omega_-)^{1/2}]
- \delta[\omega + (\Omega_+ \Omega_-)^{1/2}]).
\label{im_chi}
\end{equation}
The peaks in the dynamical structure function above
signal spin-wave excitations that
are observable by inelastic neutron scattering
only in the true magnetic ($+$) channel.
Recalling the discussion following
Eq. (\ref{chi}), the true dynamical susceptibility $\chi_{\perp} (+, {\bf k}, \omega)$
is common to a hidden ferromagnet and to a hidden N\'eel state that are related by
the transformation (\ref{trans_orb}).
Inelastic neutron scattering is therefore unable to discriminate between these two possible
types of hidden magnetic order!
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{fig2.ps}
\caption{Displayed is the linear spin-wave spectrum of the Hamiltonian (\ref{j0j1j2})
for the following set of Heisenberg exchange coupling constants:
$J_1^{\parallel (\perp )} = 0$, $J_1^{\perp (\parallel)} > 0$,
and $J_2^{\parallel} = 0.3\, J_1^{\perp (\parallel )} = J_2^{\perp}$.
The Hund's rule coupling at the quantum critical point that separates
hidden ferromagnetic (N\'eel) order
from a cSDW is set by the value listed in Table \ref{transitions}.
Spin-wave energies are given in units of $\hbar^2 J_1^{\perp (\parallel )}$
and the wavenumber ${\bf k}^{\prime}$ is associated with the antiferromagnetic sublattice
for hidden order. (See Fig. \ref{hidden_order}.)}
\label{sw}
\end{figure}
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{fig3.ps}
\caption{The spectral weight of observable (+) spinwaves at
the quantum critical point that separates hidden ferromagnetic (N\'eel) order from a cSDW
is shown in units of $\hbar$.
The following set of Heisenberg exchange coupling constants are used:
$J_1^{\parallel (\perp )} = 0$, $J_1^{\perp (\parallel )} > 0$,
and $J_2^{\parallel} = 0.3\, J_1^{\perp (\parallel )} = J_2^{\perp}$.
The critical Hund's rule exchange coupling is set to the value listed in Table \ref{transitions}.
Also shown are contours over the Brillouin zone for such spinwaves at a low fixed excitation energy.}
\label{cp_spctrm}
\end{figure}
Figure \ref{sw} depicts the linear spin-wave spectrum for
a cSDW with intervening hidden ferromagnetic (N\'eel) order
at two extreme values for the Hund's rule coupling: absent and critical.
Here, we have chosen Heisenberg exchange coupling constants
$J_1^{\parallel (\perp)} = 0, J_1^{\perp (\parallel)} > 0$,
and $J_2^{\parallel} = 0.3\, J_1^{\perp (\parallel)} = J_2^{\perp}$,
which lies near maximum frustration when Hund's rule is obeyed.
The top graph displays how the spectrum collapses linearly to the groundstate energy
at zero momentum in the absence of Hund's rule coupling.
In general,
the spin-wave velocity is given by
\begin{equation}
v_0 = 2 s a ([J_1^{\perp} - J_1^{\parallel} + 2 (J_2^{\perp} - J_2^{\parallel})]
\cdot [{1\over 2} J_0 + 2 J_1^{\perp} + 2 J_2^{\perp}])^{1/2}
\label{sw_vel}
\end{equation}
in the hidden ferromagnet,
and by the same expression after the exchange (\ref{trans_orb})
$J_1^{\parallel} \leftrightarrow J_1^{\perp}$
in the hidden N\'eel state.
It correctly collapses to zero at the phase boundary with the true ferromagnetic state
by Table \ref{transitions},
at which point the spin-wave excitation frequency disperses quadratically with wavenumber.
Notice, however, the mild depression in the top spectrum
that is centered at the cSDW wave number ${\bf k}_{cSDW} = (\pi/a, 0)$.
The bottom graph in Fig. \ref{sw} shows how it softens completely to zero energy
at the critical value for Hund's rule coupling, $- J_{0c} = 0.8\, J_1^{\perp (\parallel )}$,
in which case the hidden ferromagnetic (N\'eel) state
transits into the cSDW state by Table \ref{transitions}.
These special zero-energy
modes result from the collapse of the frequency
$\Omega_+ ({\bf k}_{cSDW})$ to zero there.
Evaluating $(\Omega_+ \Omega_-)^{1/2}$ at ${\bf k}_{cSDW}$ from the expressions for the frequencies
that are listed below Eq. (\ref{chi_bar}) yields the spin gap there:
\begin{equation}
\Delta_{cSDW} = 2s [(4 J_2^{\perp} - J_{0c}) (J_0 - J_{0c})]^{1/2}.
\label{sdw_gap}
\end{equation}
The former also implies a divergent
spectral weight (\ref{weight})
for such spinwaves at the quantum phase transition:
\begin{equation}
A(+, {\bf k}_{cSDW}) = \pi s [(4 J_2^{\perp} - J_{0c}) / (J_0 - J_{0c})]^{1/2}.
\label{weight_sdw}
\end{equation}
Because inelastic neutron scattering is proportional to
${\rm Im}\, \chi_{\perp} (+, {\bf k}, \omega)$,
low-energy spin-waves with momentum
near $\hbar {\bf k}_{cSDW}$ will dominate such measurements
near the transition between
the cSDW state and hidden magnetic order (see Fig. \ref{cp_spctrm}).
Study of the expressions for the frequencies $\Omega_{\pm}$
that are listed below Eq. (\ref{chi_bar})
yields that
the dominant spin-wave excitations centered at the cSDW wave numbers ${\bf k}_{cSDW}$
disperse up from the groundstate energy as
$[v_{l}^2 (k_{l} - \pi / a)^2 + v_{t}^2 k_{t}^2 + \Delta_{cSDW}^2]^{1/2}$
in general. (See Fig. \ref{sw}.)
Here $k_l$ and $k_t$ are the components of ${\bf k}$ that are respectively
parallel and perpendicular to ${\bf k}_{cSDW}$.
Near criticality, $\Delta_{cSDW} \rightarrow 0$,
the cSDW order is short range, with correlations lengths
$\xi_{cSDW}^{(l)} = v_l / \Delta_{cSDW}$ and
$\xi_{cSDW}^{(t)} = v_t / \Delta_{cSDW}$
in the longitudinal and transverse directions.
At criticality,
$\Delta_{cSDW} = 0$,
the longitudinal velocity $v_{l}$ coincides with the spin-wave velocity $v_0$
for hidden magnetic order (\ref{sw_vel}),
while the transverse velocity $v_{t}$ is
equal to the previous divided by the anisotropy parameter
\begin{equation}
{v_{l} \over{v_{t}}} =
\Biggl[{2(J_2^{\parallel} + J_2^{\perp}) + J_1^{\parallel} + J_1^{\perp}
\over{2(J_2^{\parallel} + J_2^{\perp}) - J_1^{\parallel} - J_1^{\perp}}}\Biggr]^{1/2}.
\label{anisotropy}
\end{equation}
The latter coincides with the anisotropy of low-energy spinwaves that is
predicted by the conventional $J_1$-$J_2$ model, in which case
$J_{1(2)} = [J_{1(2)}^{\parallel} + J_{1(2)}^{\perp}]/2$
by Hund's rule.
{\it Exact Diagonalization.}
The above semi-classical analysis indicates that large enough off-diagonal frustration
can induce a quantum phase transition into hidden magnetic order that is unfrustrated,
but that violates Hund's rule.
We shall now confirm this by obtaining
the exact low-energy spectrum of states for the
$J_0$-$J_1$-$J_2$ model (\ref{j0j1j2})
with two spin-1/2 moments per site
over a 4 by 4 square lattice with periodic boundary conditions.
The model Hamiltonian operator (\ref{j0j1j2}) acts on
a Hilbert space that is restricted to have 16 up spins and 16 down spins.
Next, translational invariance is exploited in order to
bring $H$ into block-diagonal form.
Each block of the Hamiltonian is labeled by the allowed momentum quantum numbers:
$(k_x a, k_y a) =
(0,0)$, $(\pi,0)$, $(\pi,\pi)$, $(\pi / 2, 0)$, $(\pi / 2, \pi / 2)$ and $(\pi, \pi / 2)$,
plus their symmetric counterparts.
Spin-flip symmetry
is also exploited to further block-diagonalize the Hamiltonian at
such momenta into two blocks
that are respectively even and odd under it.
The combination of translation and spin-flip symmetries reduces the dimension of each
block to a little under 19,000,000 states.
Next, we construct the Hamiltonian operator (\ref{j0j1j2}) over each of these subspaces.
Every term in Hamiltonian (\ref{j0j1j2}) represents a
Heisenberg spin exchange that permutes the Bloch-wave type states.
These permutations are stored in memory.
Hamiltonian matrix elements that involve Bloch waves
composed of configurations of spin up and spin down
that display absolutely no non-trivial translation invariance
up to a spin-flip
are also stored in memory.
They are calculated otherwise.
Because the vast majority of Bloch waves lie in the first category,
the Hamiltonian operator is stored in memory for all practical purposes.
This speeds up its application on a given state tremendously.
The application of the Hamiltonian $H$ on a given state is
accelerated further by
enabling shared-memory parallel computation through OpenMP directives.
Last, we apply the iterative Lanczos technique numerically over each subspace
labeled by momentum and by spin-flip quantum numbers.\cite{lanczos}
The ARPACK subroutine library is employed for this purpose.\cite{arpack}
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{figure_4.ps}
\caption{Shown is the low-energy spectrum for $4\times 4\times 2$ spin-1/2 moments
that experience the following set of Heisenberg exchange coupling constants:
$J_0 = 0$, $J_1^{\parallel} = 0$, $J_1^{\perp} > 0$,
and $J_2^{\parallel} = 0.3\, J_1^{\perp} = J_2^{\perp}$.
The lowest-energy spin-1 state at zero momentum is used as the reference
for the linear spin-wave approximation.
Hereafter, we set $\hbar \rightarrow 1$.}
\label{0_hund_spectra}
\end{figure}
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{figure_5.ps}
\caption{Shown is the low-energy spectrum for
$4\times 4\times 2$ spin-1/2 moments
that experience the following set of Heisenberg exchange coupling constants:
$J_1^{\parallel} = 0$, $J_1^{\perp} > 0$,
and $J_2^{\parallel} = 0.3\, J_1^{\perp} = J_2^{\perp}$.
The Hund's rule exchange coupling is set to $J_0 = - 1.35\, J_1^{\perp}$.}
\label{cp_hund_spectra}
\end{figure}
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{figure_6.ps}
\caption{The low-energy spectrum of the cSDW is obtained from exact diagonalization of
$4\times 4\times 2$ spin-1/2 moments. The following set of Heisenberg exchange coupling
constants are used: $J_1^{\parallel} = 0$, $J_1^{\perp} > 0$,
$J_2^{\parallel} = 0.3\, J_1^{\perp} = J_2^{\perp}$,
and $J_0 = - 15\, J_1^{\perp}$.}
\label{sdw_spectra}
\end{figure}
Figures \ref{0_hund_spectra}, \ref{cp_hund_spectra} and \ref{sdw_spectra}
show how the low-energy spectrum of the $J_0$-$J_1$-$J_2$ model (\ref{j0j1j2})
evolves with the strength of the Hund's rule coupling
in the presence of off-diagonal frustration.
The following Heisenberg exchange coupling constants were chosen:
$J_1^{\parallel} = 0$, $J_1^{\perp} > 0$, and $J_2^{\parallel} = 0.3\, J_1^{\perp} = J_2^{\perp}$.
A hidden ferromagnetic state is then expected in the absence of Hund's rule coupling,
while a cSDW is expected when Hund's rule in obeyed.
In the former case displayed by Fig. \ref{0_hund_spectra},
notice the coincidence
between the linear spin-wave approximation
about hidden-order, Eq. (\ref{im_chi}) and Fig. \ref{sw},
and the present exact-diagonalization results.
It indicates that the classical prediction of long-range hidden magnetic order
holds true for spin-1/2 moments per iron orbital.
Figure \ref{sdw_spectra} displays the spectrum
at strong Hund's rule coupling, $ - J_0 = 15\, J_1^{\perp}$, on the other hand.
The bulk of the groundstate energy is due to the Hund's rule coupling,
which is given by
$16\, J_0 = - 240\, J_1^{\perp}$ when Hund's rule is obeyed.
Notice the low-energy spin-1 state at the cSDW wave number $(\pi/a, 0)$,
as well as the softening of the spin-1 excited state
at the wave number $(\pi/a, \pi/a)$ that characterizes the N\'eel state.
The dispersion of these spin-1 excited states resembles
the spin-wave spectrum predicted for the cSDW in the conventional $J_1$-$J_2$ model,
where Hund's rule is obeyed.\cite{oitmaa_10}$^,$\cite{zhao_09}
Figure \ref{sdw_spectra} also displays that the groundstate is doubly degenerate,
which again is consistent with cSDW order over the square lattice.
Last, Fig. \ref{cp_hund_spectra} shows
the low-energy spectrum
for Hund's rule coupling $J_0 = - 1.35\, J_1^{\perp}$
near a possible quantum phase transition (cf. Fig. \ref{order_moment_1}).
The lowest-energy spin-1 excitations at momentum
$(0, 0)$ and $(\pi/a, 0)$ are degenerate, and their dispersion resembles the
prediction from linear spin-wave theory that is shown by Fig. \ref{sw}.
Notably,
the lowest-energy spin-1 excitation
does {\it not} ``dip down'' at momentum $(\pi/a, \pi/a)$.
Table \ref{transitions} predicts a transition into hidden order at
$J_0 = - 0.8\, J_1^{\perp}$ in the classical limit at large spin $s$, however,
which suggests that quantum effects renormalize up the critical Hund's rule coupling
in the physically relevant case of spin-1/2 iron $3d$ orbitals.
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{figure_7.ps}
\caption{Displayed is the autocorrelation of the order parameter (\ref{OP})
for true (+) and for hidden (-) magnetic order
on a $4 \times 4 \times 2$ lattice of spin-1/2 moments
as a function of Hund's rule coupling.
It is normalized to its value in the true ferromagnetic state (see text).}
\label{order_moment_1}
\end{figure}
Figure \ref{order_moment_1}
shows the evolution of relevant magnetic order parameters with Hund's rule coupling
under the previous set of Heisenberg exchange coupling constants.
Plotted on the vertical axis is the autocorrelation
$\langle {\bf O} ({\bf k})_\pm \cdot {\bf O} (-{\bf k})_\pm \rangle_0$
of the order parameter
\begin{equation}
{\bf O} ({\bf k})_\pm = \sum_i e^{i{\bf k}\cdot{\bf r}_i} [{\bf S}_i(a) \pm {\bf S}_i(b)]
\label{OP}
\end{equation}
over the groundstate normalized to its value in the true ferromagnetic state,
$\langle {\bf O} (0)_+ \cdot {\bf O} (0)_+ \rangle_0 = 16\cdot 17\,\hbar^2$.
The relation to the ordered magnetic moment per iron atom,
$\mbox{\boldmath$\mu$} (\pm, {\bf k})$,
given there is obtained from the identity
$| 16\, \mbox{\boldmath$\mu$} (\pm, {\bf k})|^2
=
(2\mu_B/\hbar)^2 \langle {\bf O} ({\bf k})_\pm \cdot {\bf O} (-{\bf k})_\pm \rangle_0$
that is valid over the 4 by 4 square lattice of spin-1 iron atoms.
The top graph in Fig. \ref{order_moment_1}
displays how the square of the ordered moment for true cSDW order (+)
decays with decreasing Hund's rule coupling.
It notably accounts for the low ordered moment that is observed
by elastic neutron diffraction in an undoped parent compound to the
recently discovered ferro-pnictide
high-$T_c$ superconductors.\cite{delacruz}
Also shown there is how hidden ferromagnetic order is established once cSDW order fades away.
Figure \ref{order_moment_1}
then provides evidence for a quantum phase transition that seperates a hidden ferromagnetic state
at weak Hund's rule coupling from a cSDW state at strong Hund's rule coupling.
Notice that the putative quantum phase transition at $J_0 \cong -1.35\, J_1^{\perp}$
(see Fig. \ref{cp_hund_spectra})
occurs at a Hund's rule exchange coupling that is almost a factor-of-2 times larger
than the classical prediction
$J_{0c} = - 0.8\, J_1^{\perp}$ listed in Table \ref{transitions}.
Last, the bottom graph in Fig. \ref{order_moment_1} shows the evolution of the relevant
order parameters in the case of true N\'eel order with intervening hidden N\'eel order.
Here, we have chosen Heisenberg exchange coupling constants
$J_1^{\parallel} > 0, J_1^{\perp} = 0$,
and $J_2^{\parallel} = 0.25\, J_1^{\parallel} = J_2^{\perp}$,
which lies at maximum frustration when Hund's rule is obeyed.
The common inflection point at $J_0 \cong - 1.2 J_1^{\parallel}$
is consistent with a quantum phase transition at a value of Hund's rule
coupling that is noticably larger than the classical prediction of $J_{0c} = - J_1^{\parallel}$
listed in Table \ref{transitions}.
\section {Discussion and Conclusions}
Recent inelastic neutron scattering measurements
on the parent compound to ferro-pnictide
superconductors CaFe$_2$As$_2$
have uncovered well-defined spin-wave excitations
in the cSDW phase
throughout virtually the entire Brillouin zone.\cite{zhao_09}
Instead of softening
at the wavenumber $(\pi/a, \pi/a)$ that corresponds to N\'eel order
as predicted by the conventional $J_1$-$J_2$ model over the square lattice,
they show a local maximum there.
Figure \ref{qcp_fit} displays a fit of the spin-wave spectrum obtained by inelastic neutron
scattering on CaFe$_2$As$_2$ to the prediction for the linear spin-wave dispersion of a
square lattice of spin-1 iron atoms, $(\Omega_+ \Omega_-)^{1/2}$,
at the quantum critical point that seperates
the cSDW phase from hidden ferromagnetic (N\'eel) order
[see Eq. (\ref{im_chi}) and Fig. \ref{sw}].
The spin per orbital is set to $s = 1/2$ and
the following Heisenberg exchange coupling
constants are used: $J_{0c} = - 57.0$ meV, $J_1^{\parallel (\perp)} = 0$,
$J_1^{\perp (\parallel)} = 115.8$ meV,
and $J_2^{\parallel} = 43.7\, {\rm meV}\, = J_2^{\perp}$.
These imply a ratio $J_2 / J_1 = 0.75$ between the next-nearest-neighbor and
the nearest-neighbor Heisenberg exchange constants deep inside the cSDW,
where Hund's rule is obeyed.
Notice that the measured spin-wave spectrum terminates before it reaches the predicted
Goldstone modes at zero momentum that correspond to hidden magnetic order.
This is consistent with Fig. \ref{cp_spctrm},
which displays how the spectral weight of low-energy spin-waves about zero momentum
is dwarfed by that (\ref{weight_sdw}) of low-energy spinwaves about momentum $(\pi/a,0)$
at the quantum critical point that seperates the cSDW phase from hidden magnetic order.
\begin{figure}
\includegraphics[scale=0.65, angle=-90]{fig8.ps}
\caption{The spin excitation spectra of two pnictide materials are fit to the linear spinwave
spectrum expected at a quantum critical point (QCP) that seperates a cSDW from ferromagnetic
(N\'eel) hidden order. [See Eq. (\ref{im_chi}) and Fig. \ref{sw}.]
The wavenumber in the direction perpendicular to the planes of iron atoms
corresponds to antiferromagnetic order
in both sets of data.
(See refs. \cite{zhao_09} and \cite{hayden_10}.)}
\label{qcp_fit}
\end{figure}
More recent inelastic neutron scattering measurements on the ferro-pnictide superconductor
BaFe$_{2-x}$Co$_x$As$_2$ have uncovered low-energy spin-wave excitations
that disperse {\it up} in energy in an anisotropic fashion
away from the wavenumbers that correspond to cSDW order;\cite{hayden_10} e.g. $(\pi/a, 0)$.
Paradoxically, no evidence for long-range magnetic order exists in this ferro-pnictide superconductor.
Figure \ref{qcp_fit} displays a fit of the spin-wave spectrum obtained by inelastic neutron
scattering on BaFe$_{2-x}$Co$_x$As$_2$ to the prediction for the linear spin-wave dispersion of a
square lattice of spin-1 iron atoms, $(\Omega_+ \Omega_-)^{1/2}$,
at the quantum critical point that seperates
the cSDW phase from hidden ferromagnetic (N\'eel) order
[see Eq. (\ref{im_chi}) and Fig. \ref{sw}].
The spin per orbital is again set to $s = 1/2$,
and the following Heisenberg exchange coupling
constants are used: $J_{0c} = - 105.4$ meV, $J_1^{\parallel (\perp)} = 0$,
$J_1^{\perp (\parallel)} = 147.2$ meV,
and $J_2^{\parallel} = 47.3\, {\rm meV}\, = J_2^{\perp}$.
This fit implies a ratio $J_2 / J_1 = 0.64$ between the next-nearest-neighbor and
the nearest-neighbor Heisenberg exchange constants deep inside the cSDW.
Although the conventional $J_1$-$J_2$ model over the square lattice
can also account for the linear and anisotropic
dispersion of the spinwaves that are
observed in superconducting\cite{hayden_10} BaFe$_{2-x}$Co$_x$As$_2$,
it cannot account for the absence of long-range magnetic order.
Both the semi-classical analysis about hidden magnetic order and
the exact diagonalization studies on a $4 \times 4$ lattice
described in the previous section
indicate, on the other hand,
that the moment associated with true cSDW order is extremely small
at the quantum critical point into hidden magnetic order.
(See Figs. \ref{hidden_order} and \ref{order_moment_1}.)
In conclusion,
a linear spin-wave analysis of the $J_0$-$J_1$-$J_2$ model Hamiltonian (\ref{j0j1j2})
that describes a square lattice of frustrated spin-1 iron atoms
yields a quantum critical point as a function of Hund's rule exchange coupling ($- J_0$)
that seperates a cSDW from hidden ferromagnetic or hidden N\'eel order.
The observable spectral weight of the Goldstone modes associated with hidden order
is predicted to be small in general
(see Fig. \ref{cp_spctrm}).
A strong variation between the diagonal and the off-diagonal
exchange couplings of the multi-orbital Heisenberg model (\ref{j0j1j2})
across nearest-neighbors {\it alone} is sufficient
for Hund's rule to give way to hidden magnetic order.
Recent DFT calculations obtain ferromagnetic direct exchange and
antiferromagnetic superexchange across nearest-neighbor iron moments in
the cSDW phase of ferro-pnictide materials.\cite{ma_08}
Their superposition potentially can result in
the required difference between $J_1^{\parallel}$ and $J_1^{\perp}$.
Fits of the critical linear spin-wave spectrum
to recent experimental studies of magnetic excitations
in ferro-pnictide materials\cite{zhao_09}$^,$\cite{hayden_10}
yield a Hund's rule exchange coupling constant on the order of $100$ meV.
Figure \ref{order_moment_1} displays
the dependence of the ordered magnetic moments on Hund's rule coupling
obtained from exact diagonalization of the model Hamiltonian (\ref{j0j1j2})
over a $4 \times 4$ lattice, with two spin-1/2 orbitals per site.
The top graph implies a critical Hund's rule coupling
at the transition between hidden ferromagnetic order and cSDW order
that is almost twice the
classical prediction listed in Table \ref{transitions}:
$- J_{0c} = 0.8\, J_1^{\perp}$.
Correcting for this discrepancy brings the value
of the Hund's rule coupling closer to the estimate
of 700 meV that is obtained for ferro-pnictides
by dynamical mean field theory.\cite{haule_09}
(Note that $-J_0$ is twice as big as the Hund's rule exchange coupling that
appears naturally in multi-orbital Hubbard models.)
The success of the theoretical fits to the inelastic neutron scattering data
displayed by Fig. \ref{qcp_fit}
indicates that a local-moment picture for magnetism in ferro-pnictide materials is valid.
Whether magnetic moments in
ferro-pnictide materials
are itinerant or localized is vigorously debated, however.
In particular,
the metallic temperature dependence shown by the electrical resistance
in the cSDW phase of ferro-pnictide materials
in addition to the absence of satellite peaks in inelastic X-ray scattering spectra
is consistent with itinerant magnetism.\cite{wang}$^,$\cite{yang_09}
The multi-orbital Heisenberg model Hamiltonian
(\ref{j0j1j2}) studied here
is also potentially consistent with the semi-metallic nature
of ferro-pnictide materials, on the other hand.
Figure \ref{hidden_order} indicates that the classical hidden ferromagnetic state
remains intact in the presence of mobile holes
that are prohibited from hopping between orbitals, for example.
Likewise, the classical hidden N\'eel state remains intact in the presence of mobile
holes that are restricted to hop between different orbitals at nearest-neighbor sites
of the square lattice of iron atoms.
Such limiting cases can result in low-energy electronic hole excitations near zero momentum,
which are consistent with the electronic structure revealed by
angle-resolved photoemission spectroscopy (ARPES)
in the cSDW phase of ferro-pnictide materials.\cite{kondo_10}
Resonant scattering of such hole excitations
with the quantum-critical spin-waves studied here
(Figs. \ref{sw} and \ref{cp_spctrm}) could also result
in low-energy single-particle excitations
at momenta near those associated with cSDW order.
Recent ARPES studies
of BaFe$_2$As$_2$
in the cSDW phase reveal the existence of low-energy electronic excitations
near such momenta.\cite{richard_10}
The measured electronic spectrum there forms a Dirac cone,
with a Fermi velocity of about $330$ meV\,\AA.
It is an order of magnitude smaller than in graphene,
and it is hence consistent with strong electronic correlations.\cite{harrison_09}
It also remarkably lies inside the range of spin-wave velocities
extracted from low-energy magnetic excitations near cSDW wave numbers
in the doped superconducting compound BaFe$_{2-x}$Co$_x$As$_2$:\cite {hayden_10}
$v_t = 230$ meV\,\AA\ and $v_l = 580$ meV\,\AA\
(see Fig. \ref{qcp_fit}).
The above observations therefore indicate that
injecting mobile holes into the states studied here
that display hidden magnetic order (Fig. \ref{hidden_order})
potentially can account for the nature of low-energy single-particle excitations
in ferro-pnictide materials.
They also
indicate that a description of the superconducting state of ferro-pnictide materials
in terms of a doped Mott insulator\cite{Si&A} is possible.
\begin{acknowledgments}
The author thanks Ed Rezayi, Radi Al Jishi and Elena Bascones for useful discussions.
Exact diagonalizations of the $J_0$-$J_1$-$J_2$ model (\ref{j0j1j2})
were carried out on the SGI Altix 4700 (Hawk)
at the AFRL DoD Supercomputer Resource Center.
This work was supported in part by the US Air Force
Office of Scientific Research under grant no. FA9550-09-1-0660.
\end{acknowledgments}
|
\section{Introduction}
The objective of this paper is to study the existence and uniqueness
of solutions to stochastic differential equations driven by
$G$-Brownian motion with integral-Lipschitz coefficients in the
framework of sublinear
expectation spaces.\\\\
Motivated by uncertainty problems, risk
measures and the superhedging in finance, Peng \cite{P2,P1,P3}
introduced $G$-Brownian motion. The expectation $\mathbb{E}[\cdot]$
associated with $G$-Brownian motion is a sublinear expectation which
is called $G$-expectation. The stochastic calculus with respect to the
$G$-Brownian motion has been established
\cite{P3}.\\\\
In this paper, we study the solvability of the following
stochastic differential equation driven by $G$-Brownian motion:
\begin{align*}
\bigg\{
\begin{aligned}
dX(s)&=b(s,X(s))ds+h(s,X(s))d\langle B,B\rangle_s+\sigma(s,X(s))dB_s;\\
X(0)&=x,
\end{aligned}
\end{align*}
or, more precisely,
\begin{equation}\label{eq 2}
X(t)=x+\int_0^tb(s,X(s))ds+\int_0^th(s,X(s))d\langle
B,B\rangle_s+\int_0^t\sigma(s,X(s))dB_s,
\end{equation}
where $t \in [0,T]$, the initial condition $x\in\mathbb{R}^n$ is given and
$(\langle B,B\rangle_t)_{t\geq0}$ is the quadratic variation process
of $G$-Brownian motion $(B_t)_{t\geq0}.$\\\\
It is well known that under a Lipschitz condition on the
coefficients $b$, $h$ and $\sigma$, the existence and uniqueness of the
solution to (\ref{eq 2}) has been obtained, see Peng \cite{P3} and Gao \cite{G}.\\\\
In this paper, we establish the existence and uniqueness
of the solution to equation (\ref{eq 2}) under the following
so-called integral-Lipschitz condition:
\begin{equation}\label{eq 3}
|b(t,x_1)-b(t,x_2)|^2+|h(t,x_1)-h(t,x_2)|^2+|\sigma(t,x_1)-\sigma(t,x_2)|^2\leq\rho(|x_1-x_2|^2),
\end{equation}
where $\rho:(0,+\infty)\rightarrow(0,+\infty)$ is a continuous,
increasing, concave function satisfying
$$\rho(0+)=0,\ \int_0^1\frac{dr}{\rho(r)}=+\infty.$$
A typical example of (\ref{eq 3}) is :
\begin{align*}
&|b(t,x_1)-b(t,x_2)|+|h(t,x_1)-h(t,x_2)|\\
+&|\sigma(t,x_1)-\sigma(t,x_2)|\leq|x_1-x_2|(\ln\frac{1}{|x_1-x_2|})^\frac{1}{2}.
\end{align*}
Under this condition,
the existence and uniqueness results for classical finite
dimensional stochastic differential equations can be found in
Watanabe-Yamada \cite{WY} and Yamada \cite{Y}, while the infinite
dimensional case can be found in Hu-Lerner \cite{HL}. In our paper,
in the $G$-expectation framework, under the condition (\ref{eq 3})
we will prove the existence and uniqueness of the solution to
(\ref{eq 2}) still
hold.\\\\
We also establish the existence and uniqueness of the solution to
equation (\ref{eq 2}) under a ``weaker" condition on $b$ and $h$,
i.e.,
\begin{equation}\label{eq plus}
|b(t, x_1)-b(t, x_2)|\leq\rho(|x_1-x_2|); \ |h(t, x_1)-h(t, x_2)|\leq\rho(|x_1-x_2|).
\end{equation}
A typical example of (\ref{eq plus}) is:
\begin{align*}
|b(t, x_1)-b(t, x_2)|&\leq|x_1-x_2|\ln\frac{1}{|x_1-x_2|}; \\
|h(t, x_1)-h(t, x_2)|&\leq|x_1-x_2|\ln\frac{1}{|x_1-x_2|}.
\end{align*}
In the classical case, the uniqueness result can be found in Watanabe-Yamada \cite{WY} and the existence can be found in Hu-Lerner \cite{HL}. In our paper, we obtain both the uniqueness and existence results in the $G$-expectation framework.\\\\
Yamada-Watanabe \cite{WY} and Hu-Lerner \cite{HL} have
also obtained the pathwise uniqueness result for the classical
one-dimensional stochastic differential equations. The reader
interested in the $G$-Brownian motion case is referred to Lin \cite{LQ}.
\\\\
This paper is organized as follows: Section 2 gives the
necessary preliminaries which include a short recall of some elements of
the $G$-stochastic calculus and some technique lemmas which will be used in what follows.
Section 3 proves the existence and uniqueness theorem for
$G$-stochastic differential equations, while Section 4 studies the
$G$-BSDE case.
\section{Preliminaries}
\subsection{$G$-Brownian motion and $G$-Capacity}
The aim of this section is to recall some basic definitions and
properties of $G$-expectations, $G$-Brownian motions and $G$-stochastic
integrals, which will be needed in the sequel. The reader interested
in a more detailed description of these notions is referred to
\cite{P3}.\\\\
Adapting Peng's approach in \cite{P3}, let
$\Omega$ be a given nonempty fundamental space and $\mathcal {H}$ be a linear space of real
functions defined on $\Omega$ such that :\\\\
$ \quad i)$ $1\in \mathcal {H}$. \\
$\quad ii)$ $\mathcal {H}$ is stable with respect to local Lipschitz
functions, i.e., for all $n\geq1$, and for all $X_1,\ldots,X_n\in
\mathcal {H}$, $\varphi \in C_{l,Lip}(\mathbb{R}^n),$ it holds also
$\varphi(X_1,\ldots,X_n)\in\mathcal {H}$.\\\\
Recall that $ C_{l,Lip}(\mathbb{R}^n)$ denotes the space of all
local Lipschitz functions $\varphi$ over $\mathbb{R}^n$ satisfying
$$|\varphi(x)-\varphi(y)|\leq C(1+|x|^m+|y|^m)|x-y|,\ x,y \in
\mathbb{R}^n,$$ for some $C>0,m\in\mathbb{N}$ depending on
$\varphi$. The set $\mathcal {H}$ is interpreted as the space of
random variables defined on $\Omega$.
\begin{definition}\label{de
2.1.1}A sublinear expectation $\mathbb{E}$ on $\mathcal {H}$ is a
functional $\mathbb{E}:\mathcal {H}\rightarrow\mathbb{R}$ with the following
properties : for
all $X,Y\in \mathcal {H}$, we have\\\\
$\quad i)$ \textbf{Monotonicity:} if $X\geq Y$, then $\mathbb{E}[X]\geq\mathbb{E}[Y];$\\
$\quad ii)$ \textbf{Preservation of constants:}\ $\mathbb{E}[c]=c,$ for all $c\in \mathbb{R}$;\\
$\quad iii)$\textbf{Sub-additivity}:\ $\mathbb{E}[X]-\mathbb{E}[Y]\leq\mathbb{E}[X-Y]$;\\
$\quad iv)$\textbf{Positive homogeneity:}\ $\mathbb{E}[\lambda
X]=\lambda\mathbb{E}[X]$, for all $\lambda\geq 0$.
\end{definition}
The triple $(\Omega,\mathcal {H}, \mathbb{E})$ is called a
sublinear expectation space. It generalizes the classical case of
the linear expectation $E[X]=\int_\Omega X dP,\ X \in
L^1(\Omega,\mathcal {F},\mathcal {P}),$ over a probability space
$(\Omega,\mathcal{F},\mathcal {P}).$
\begin{definition}\label{de 2.1.2} For arbitrary $n,m \geq 1,$ a random vector
$Y=(Y_1,\ldots,Y_n) \in \mathcal {H}^n\ (=\mathcal
{H}\times\ldots\times \mathcal {H})$ is said to be
independent of $X\in\mathcal {H}^m$ under $\mathbb{E}[\cdot]$ if for each test function $\varphi\in C_{l,Lip}(\mathbb{R}^{n+m}) $ we
have
$$\mathbb{E}[\varphi(X,Y)]=\mathbb{E}[\mathbb{E}[\varphi(x,Y)]_{x=X}].$$
\end{definition}
Let $X=(X_1,\ldots,X_n)\in\mathcal {H}^{ n}$ be a given random vector.
We define a functional on $C_{l,Lip}(\mathbb{R}^n)$ by
$$\mathbb{F}_X[\varphi]:=\mathbb{E}[\varphi(X)],\
\varphi\in
C_{l,Lip}(\mathbb{R}^n).$$
\begin{definition}\label{de 2.1.3}
Given two sublinear expectation spaces $(\Omega,\mathcal
{H},\mathbb{E})$ and $(\widetilde{\Omega},\widetilde{\mathcal
{H}},\widetilde{\mathbb{E}})$, two random vectors $X\in\mathcal
{H}^{n}$ and $Y\in\widetilde{\mathcal {H}}^{n}$ are said to be
identically distributed if for each test function $\varphi\in
C_{l,Lip}(\mathbb{R}^n)$
$${\mathbb{F}}_X[\varphi]=\widetilde{\mathbb{F}}_Y[\varphi].$$
\end{definition}
Now we begin to introduce the definition of $G$-Brownian motion and
$G$-expectation.
\begin{definition}\label{de 2.1.4}
A $d$-dimensional random vector $X$ in a sublinear expectation space
$(\Omega,\mathcal {H},\mathbb{E})$ is called $G$-normal distributed
if for each $\varphi \in C_{l,Lip} (\mathbb{R}^d)$,
$$
u(t,x):=\mathbb{E}[\varphi(x+\sqrt{t}X)],\ t\geq 0, x\in \mathbb{R}^d
$$
is the viscosity solution of the following PDE defined on
$[0,\infty)\times \mathbb{R}^d$:
$$
\frac{\partial u}{\partial t} -G(D^2 u)=0,\ u|_{t=0} = \varphi,
$$
where $G=G_X(A):\mathbb{S}^d \rightarrow \mathbb{R}$ is defined by
$$
G_X(A):=\frac{1}{2} \mathbb{E} [\langle AX,X \rangle], A
\in \mathbb{S}^d,
$$
and $D^2 u=(\partial^2_{x_i x_j} u)^d_{i,j=1}$.
\end{definition}
In particular, $\mathbb{E}[\varphi(X)]=u(1,0)$, and by Peng
\cite{P3} it is easy to check that, for a $G$-normal distributed random vector
$X$, there exists a bounded, convex and closed subset $\Gamma$ of
$\mathbb{R}^d$, which is the space of all $d\times d$ matrices, such
that for each $A\in \mathbb{S}^d$, $G(A)=G_X(A)$ can be represented
as
$$
G(A)=\frac{1}{2}
\sup_{\gamma\in\Gamma} tr[\gamma \gamma^T A].
$$
Consequencely, we can denote the $G$-normal distribution by
$N(0,\Sigma)$, where $\Sigma :=\{\gamma \gamma^T,
\gamma\in\Gamma\}$.\\\\
Let $\Omega$ denote the space of all $\mathbb{R}^d$-valued
continuous paths $(\omega_t)_{t\geq0}$ with $\omega_0 =0$, equipped
with the distance
$$
\rho(\omega^1,\omega^2):=\sum^\infty_{i=1} 2^{-i} [(\max_
{t\in[0,i]} | \omega^1_t-\omega^2_t|) \wedge 1],
$$
and we denote the canonical process by $B_t(\omega)=\omega_t$,
$t\geq0$, for each $\omega\in\Omega$. For each $T\geq0$, we set
$$
L_{ip}^0 (\mathcal{F}_T):=\{\varphi(B_{t_1}, \ldots, B_{t_n}):n\geq
1, t_1,\ldots, t_n \in [0, T], \varphi \in C_{l,Lip}
(\mathbb{R}^{d\times n})\}.
$$
Define
$$
L_{ip}^0 (\mathcal{F}):= \bigcup^\infty_{n=1} L_{ip}^0
(\mathcal{F}_n),
$$
it is clear that $L_{ip}^0 (\mathcal{F})$ is a vector lattices.
\begin{definition}\label{de 2.1.5}
Let $\mathbb{E}: L_{ip}^0(\mathcal{F}) \rightarrow \mathbb{R}$ be
a sublinear expectation on $L_{ip}^0(\mathcal{F})$, we call
$\mathbb{E}$ $G$-expectation if the d-dimensional canonical process
$(B_t)_{t\geq0}$ is a $G$-Brownian motion under
$\mathbb{E}$, that is, \\\\
$ \quad i)\ B_0(\omega)=0$;\\
$ \quad ii)$ For each $t,s\geq0$, the increment $B_{t+s} - B_t$ is
$N(0,s\Sigma)$-distributed and independent of
$(B_{t_1},\ldots,B_{t_n})$, for each $n\in\mathbb{N}$ and $0\leq
t_1\leq\ldots\leq t_n\leq t$, i.e., for each $\varphi\in
C_{l,Lip}(\mathbb{R}^{d\times m})$,
$$
\mathbb{E}[\varphi(B_{t_1},\ldots,B_{t_{m-1}},B_{t_m}-B_{t_{m-1}})]=\mathbb{E}[\psi(B_{t_1},\ldots,B_{t_{m-1}})],
$$
where
$\psi(x_1,\ldots,x_{m-1})=\mathbb{E}[\varphi(x_1,\ldots,x_{m-1},B_{t_m}-B_{t_{m-1}})]$.
\end{definition}
By Peng \cite{P3}, the construction of $G$-expectation is explicit
and natural. We denote by $L^p_G(\mathcal{F}_T)$ (resp.
$L^p_G(\mathcal{F})$) the topological completion of $L_{ip}^0
(\mathcal{F}_T)$ (resp. $L_{ip}^0 (\mathcal{F})$) under the Banach
norm $\mathbb{E}[|\cdot|^p]^\frac{1}{p},\ 1\leq p<\infty.$ We also
denote the extension by $\mathbb{E}$.
\begin{definition}
Let $\mathbb{E}: L_{ip}^0(\mathcal{F}) \rightarrow \mathbb{R}$ be
a $G$-expectation on $L_{ip}^0(\mathcal{F})$, we define the related conditional expectation of $X\in L^0_{ip}(\mathcal{F}_T)$ under $L^0_{ip}(\mathcal{F}_{t_j})$, $0\leq t_1\leq\ldots\leq t_j\leq t_{j+1}\leq\ldots\leq t_n\leq T$:
\begin{align}
\mathbb{E}[X|\mathcal{F}_{t_j}]&=\mathbb{E}[\varphi(B_{t_1},
\ldots,B_{t_n}-B_{t_{n-1}})|\mathcal{F}_{t_j}]\notag\\
&=\mathbb{E}[\psi(B_{t_1},\ldots,B_{t_j}-B_{t_{j-1}})],\notag
\end{align}
where $\psi(x_1,\ldots,x_j)=\mathbb{E}[\varphi(x_1,\ldots,x_j,B_{t_{j+1}}-B_{t_j},\ldots, B_{t_n}-B_{t_{n-1}})].$\\\\
Since, for $X$, $Y\in L_{ip}^0(\mathcal{F}_{t_j})$,
$$
\mathbb{E}[|\mathbb{E}[X|\mathcal{F}_{t_j}]-\mathbb{E}[Y|\mathcal{F}_{t_j}]|]\leq \mathbb{E}[|X-Y|],
$$
the mapping
$\mathbb{E}[\cdot|\mathcal{F}_{t_j}]
:L_{ip}^0(\mathcal{F}_T)\rightarrow L_{ip}^0(\mathcal{F}_{t_j})$
can be continuously extended to $\mathbb{E}[\cdot|\mathcal{F}_{t_j}]:L_G^1(\mathcal{F}_T)\rightarrow L_G^1(\mathcal{F}_{t_j})$.
\end{definition}
From the above definition we know that each $G$-expectation is determined
by the parameter $G$, which is determined by $\Gamma,$ where
$\Gamma$ is some bounded convex closed subset of $\mathbb{R}^{d\times d}.$
Let $P$ be the Wiener measure on $\Omega.$ The filtration
generated by the canonical process $(B_t)_{t\geq0}$ is denoted by
$$\mathcal {F}_t:=\sigma\{B_u,\ 0\leq u\leq t\},\ \mathcal{F}=\{\mathcal {F}_t\}_{t\geq0}.$$
Let $\mathcal {A}_{0,\infty}^\Gamma$ be the collection of all
$\Gamma-$valued $\{\mathcal {F}_t,\ t\geq0\}$ adapted processes on
the interval $[0,\infty),\ i.e.,$ $\theta\in\mathcal
{A}_{0,\infty}^\Gamma$ if and only if $\theta_t$ is $\mathcal {F}_t$
measurable and $\theta_t\in\Gamma$, for each $t\geq0$. For each fixed $\theta\in\mathcal{A}_{0,\infty}^\Gamma$, let
$P_\theta$ be the law of the process $(\int_0^t\theta_sdB_s)_{t\geq0}$ under the Wiener measure
$P.$ \\\\
We denote by $\mathcal {P}=\{P_\theta: \theta\in\mathcal
{A}_{0,\infty}^\Gamma\}$ and define
$$
\bar{C}(A):=\sup_{\theta\in\mathcal
{A}_{0,\infty}^\Gamma}P_\theta(A),\ A\in\mathcal {B}(\Omega).
$$
From Theorem 1 of \cite{DHP}, we know $\mathcal {P}$ is tight and
$\bar{C}$ is a Choquet capacity. For each $X\in \mathcal{B}(\Omega)$, $E_{\theta}(X)$ exists for each $\theta\in \mathcal
{A}_{0,\infty}^\Gamma$. Set
$$
\bar{\mathbb{E}}[X]:=\sup_{\theta\in\mathcal
{A}_{0,\infty}^\Gamma}E_{\theta}(X),
$$
then we can introduce the notion of ``quasi sure"(q.s.).
\begin{definition}\label{de 2.1.6}
A set $A\subset\Omega$ is called polar if $\bar{C}(A)=0.$ A property
is said to hold ``quasi-surely" (q.s.) if it holds outside a polar
set.
\end{definition}
From Theorem 59 of \cite{DHP}, in fact, $L_G^1(\mathcal {F})$ can be
rewritten as the collection of all the q.s. continuous random vectors
$X\in \mathcal{B}(\Omega)$ with $\lim_{n\rightarrow+\infty}\bar{\mathbb{E}}[|X|I_{\{|X|>n\}}]=0$.
Furthermore, for all $X\in L_G^1(\mathcal {F})$,
$\mathbb{E}[X]=\bar{\mathbb{E}}[X]$.\\\\
From Denis, Hu and Peng \cite{DHP} and Gao \cite{G}, we also have
the following monotone convergence theorem:
$$X_n\in L_G^1(\mathcal {F}), \ X_n\downarrow X,\ q.s.\Rightarrow
\mathbb{E}[X_n]\downarrow\bar{\mathbb{E}}[X].
$$
\begin{equation}\label{eq 9}
X_n\in \mathcal{B}(\Omega),\ X_n\uparrow X,\ q.s.,\ E_{\theta}(X_1)>-\infty \
\mbox{for all}\ P_\theta\in\mathcal
{P}\Rightarrow\bar{\mathbb{E}}[X_n]\uparrow\bar{\mathbb{E}}[X].
\end{equation}
In \cite{P3}, a generalized It$\hat{o}$ integral and a generalized
It$\hat{o}$ formula with respect to $G$-Brownian motion are established:
\begin{definition}
For $T\in\mathbb{R}_+,$ a partition of $[0,T]$ is a finite
ordered subset $\pi^N_T=\{t_0, t_1,\ldots,t_N\}$ such that
$0=t_0<t_1<\ldots<t_N=T.$ Let $p\geq1$ be fixed, define
$$
M_G^{p,0}([0,T]):=\{ \eta_t=\sum_{j=0}^{N-1}\xi_{j}I_{[t_{j},t_{j+1})}(t);
\xi_{j}\in L_G^p(\mathcal {F}_{t_{j}})\}.
$$
We set
$$\frac{1}{T}\int_0^T\mathbb{E}(\eta_t(\omega))dt:=\frac{1}{T}\sum_{j=0}^{N-1}\mathbb{E}(\xi_j(\omega))(t_{j+1}-t_j).$$
For each $p\geq1,$ we denote by $M_G^p([0,T])$ the completion of
$M_G^{p,0}([0,T])$ under the norm
$$\|\eta\|_{M_G^p([0,T])}=(\frac{1}{T}\int_0^T\mathbb{E}[|\eta_s|^p]ds)^\frac{1}{p}.$$
\end{definition}
Let $\mathbf{a}=(a_1,\ldots,a_d)^T$ be a given vector in
$\mathbb{R}^d,$ we set
$(B_t^\mathbf{a})_{t\geq0}=(\mathbf{a},B_t)_{t\geq0},$ where
$(\mathbf{a},B_t)$ denotes the
scalar product of $\mathbf{a}$ and $B_t$.
\begin{definition}
For each $\eta\in M_G^{2,0}([0,T])$ with the form
$$\eta_t(\omega)=\sum_{j=0}^{N-1}\xi_{j}I_{[t_{j},t_{j+1})}(t),$$ we
define
$$
\mathcal{I}(\eta)=\int_0^T\eta_sdB_s^\mathbf{a}
:=\sum_{j=0}^{N-1}\xi_j(B_{t_{j+1}}^\mathbf{a}-B_{t_j}^\mathbf{a}),
$$
and the mapping
can be continuously extended to $\mathcal{I}: M_G^2([0,T])\rightarrow L_G^2(\mathcal
{F}_T)$.
Then, for each $\eta\in M_G^2([0,T])$, the stochastic integral is
defined by
$$\int_0^T\eta_sdB_s^\mathbf{a}:=\mathcal{I}(\eta).$$
\end{definition}
We denote by $(\langle B^\mathbf{a}\rangle_t)_{t\geq0}$ the
quadratic variation process of process $(B_t^\mathbf{a})_{t\geq0},$
we know from \cite{P3} that $(\langle
B^\mathbf{a}\rangle_t)_{t\geq0}$ is an increasing process with
$\langle B^\mathbf{a}\rangle_0=0,$ and for each fixed $s\geq0,$
$$\langle B^\mathbf{a}\rangle_{t+s}-\langle B^\mathbf{a}\rangle_{s}=\langle
(B^s)^\mathbf{a}\rangle_t,$$ where $B_t^s=B_{t+s}-B_s, t\geq0,
(B^s)_t^\mathbf{a}=(\mathbf{a},B_t^s).$ \\\\
The mutual variation process of $B^\mathbf{a}$ and
$B^{\bar{\mathbf{a}}}$ is defined by $$\langle
B^\mathbf{a},B^{\bar{\mathbf{a}}} \rangle_t:=\frac{1}{4}(\langle
B^\mathbf{a}+B^{\bar{\mathbf{a}}}\rangle_t-\langle
B^\mathbf{a}-B^{\bar{\mathbf{a}}}\rangle_t).$$
\begin{definition}
Define the mapping $M_G^{1,0}([0,T])\rightarrow L_G^1(\mathcal
{F}_T)$ as follows:
$$\mathcal {Q}(\eta)=\int_0^T\eta(s)d\langle
B^\mathbf{a}\rangle_s:=\sum_{k=0}^{N-1}\xi_k(\langle
B^\mathbf{a}\rangle_{t_{k+1}}-\langle
B^\mathbf{a}\rangle_{t_{k}}).$$ Then $\mathcal {Q}$ can be
uniquely extended to $M_G^1([0,T]).$ We still use $\mathcal
{Q}(\eta)$ to denote the mapping
$\int_0^T\eta(s)d\langle B^\mathbf{a}\rangle_s, \eta\in M_G^1([0,T]).$
\end{definition}
\textbf{Remark:} For any $\mathbf{a}\in \mathbb{R}^d,$ $B_t^\mathbf{a}$ is a
one dimensional $G_\mathbf{a}$-Brownian motion where
$$G_\mathbf{a}(\beta)=\frac{1}{2}\sup_{\gamma\in\Gamma}tr(\beta\gamma\gamma^T\mathbf{a}\mathbf{a}^T)
=\frac{1}{2}(\sigma_{\mathbf{a}\mathbf{a}^T}\beta^+-\sigma_{-\mathbf{a}\mathbf{a}^T}\beta^-),\
\ \ \beta\in\mathbb{R},$$ and
$$\sigma_{\mathbf{a}\mathbf{a}^T}=\sup_{\gamma\in\Gamma}tr(\gamma\gamma^T\mathbf{a}\mathbf{a}^T),\ \
\sigma_{-\mathbf{a}\mathbf{a}^T}=-\sup_{\gamma\in\Gamma}-tr(\gamma\gamma^T\mathbf{a}\mathbf{a}^T).$$
By Corollary 5.3.19 of \cite{P3} we have
$$\langle B\rangle_t\in t\Sigma=\{t\times\gamma\gamma^T,
\gamma\in\Gamma\},$$ therefore, for $0\leq s\leq t,$
$$\langle B^\mathbf{a}\rangle_t-\langle B^\mathbf{a}\rangle_s\leq\sigma_{\mathbf{a}\mathbf{a}^T}(t-s).$$
At the end of the subsection, we give It\^{o}'s formula for the $G$-stochastic calculus.
\begin{theorem}(Proposition 6.3 of \cite{P3})
Let $\alpha^{\nu}$, $\eta^{\nu ij}$ and $\beta^{\nu j}\in M^2_G([0,T])$, $\nu= 1,\ldots,n$, $i, j= 1,\ldots,d$ be bounded processes and consider
$$
X^{\nu}_t=X^{\nu}_0+\int^t_0\alpha^{\nu}_sds+\sum^d_{i,j=1}\int^t_0\eta^{\nu ij}_s d\langle B^i,B^j\rangle_s+\sum^d_{j=1}\int^t_0\beta^{\nu j}_s dB^j_s,
$$
where $X^{\nu}_0\in \mathbb{R}$, $\nu=1,\ldots,n$. Let $\Phi\in C^2(\mathbb{R}^n)$ be a real function with bounded derivatives such that $\{\partial^2_{x^\mu x^\nu}\Phi\}^n_{\mu, \nu=1}$ are uniformly Lipschitz. Then for each $s, t\in [0,T]$, in $L^2_G(\mathcal{F}_t)$
\begin{align*}
\Phi(X_t)-\Phi(X_s)&=\int^t_s\partial_{x^\nu}\Phi(X_u)\alpha^{\nu}_{u}du
+\int^t_s\partial_{x^\nu}\Phi(X_u)\eta^{\nu ij}_{u}d\langle B^i, B^j\rangle_u\\
&+\int^t_s\partial_{x^\nu}\Phi(X_u)\beta^{\nu j}_{u}dB^j_u+\frac{1}{2}\int^t_s\partial^2_{x^\mu x^\nu}\Phi(X_u)\beta^{\mu i}_u\beta^{\nu j}_ud\langle B^i, B^j\rangle_u,
\end{align*}
where the repeated indices $\nu$, $\mu$, $i$ and $j$ imply the summation.
\end{theorem}
\subsection{Technical lemmas}
In order to present our main results, we introduce here some technical lemmas which will be needed in the sequel. In the framework of $G$-expectation, by a classical argument, we also have the following Jensen's inequality and Fatou's lemma:
\begin{lemma}\label{le 2.2.2}Let $\rho : \mathbb{R} \rightarrow \mathbb{R}$ be a continuous
increasing, concave function defined on $\mathbb{R}$, then for each
$X\in L_G^1(\mathcal {F}),$ the following inequality holds:
$$
\rho(\bar{\mathbb{E}}[X])\geq \bar{\mathbb{E}}[\rho(X)].$$
\end{lemma}
\begin{lemma}\label{le 2.2.12}
Suppose $\{X^n, n\geq0\}$ is a sequence of random variables in $L^1_G(\mathcal{F})$ and $Y\in L^1_G(\mathcal{F})$, $\bar{\mathbb{E}}[|Y|]<+\infty$, and for all $n\geq 0$, $X^n \geq Y$, q.s., we have
$$
\bar{\mathbb{E}}[\liminf_{n\rightarrow+\infty}X^n]\leq\liminf_{n\rightarrow+\infty}\bar{\mathbb{E}}[X^n];\\
$$
furthermore, if there exists a random variable $Y\in L^1_G(\mathcal{F})$, $\bar{\mathbb{E}}[|Y|]<+\infty$, and for all $n\geq 0$, $X^n \leq Y$, q.s., then
$$
\bar{\mathbb{E}}[\limsup_{n\rightarrow+\infty}X^n]\geq\limsup_{n\rightarrow+\infty}\bar{\mathbb{E}}[X^n].
$$
\end{lemma}
\textbf{Proof:} By the representation theorem of $G$-expectation in \cite{P1}, we have
\begin{align*}
&E_\theta(\liminf_{n\rightarrow +\infty}X^{n})\leq\liminf_{n\rightarrow +\infty}E_\theta(X^{n})\\
\leq&\liminf_{n\rightarrow +\infty}\sup_{P_\theta\in\mathcal{P}}E_\theta(X^{n})
=\liminf_{n\rightarrow +\infty}\bar{\mathbb{E}}[X^{n}].
\end{align*}
Taking supremum on the left side, we can easily get the result. And the other part of the lemma can be proved in a similar way.\\\\
Then we introduce two important BDG type inequalities for $G$-stochastic integrals.
\begin{lemma}\label{th 2.1.1}(Theorem 2.1 of \cite{G})
Let $p\geq2$ and $\eta=\{\eta_s, s\in[0,T]\}\in M_G^p([0,T]).$ For
$\mathbf{a}\in \mathbb{R}^d,$ set
$X_t=\int_0^t\eta_sdB^\mathbf{a}_s$. Then there exists a continuous
modification $\tilde{X}$ of X, i.e., on some
$\tilde{\Omega}\subset\Omega,$ with $\bar{C}(\tilde{\Omega}^c)=0,$
$\tilde{X}_\cdot(\omega)\in C_0[0,T]$ and
$\bar{C}(|X_t-\tilde{X}_t|\neq0)=0$ for all $t\in[0,T]$, such that
$$
\bar{\mathbb{E}}[\sup_{s\leq u\leq
t}|\tilde{X}_u-\tilde{X}_s|^p]\leq
C_p\sigma^{p/2}_{\mathbf{a}\mathbf{a}^T}\mathbb{E}[\bigg(\int_s^t|\eta_u|^2du\bigg)^{p/2}],
$$
where $0<C_p<\infty$ is a positive constant independent of $\mathbf{a}, \
\eta$ and $\Gamma.$
\end{lemma}
\begin{lemma}\label{th 2.1.2}(Theorem 2.2 of \cite{G})
Let $p\geq1$ and $\mathbf{a},\bar{\mathbf{a}}\in\mathbb{R}^d.$ Let
$\eta\in M_G^p([0,T])$. Then there exists a continuous modification
$\tilde{X}^{\mathbf{a},\bar{\mathbf{a}}}_t$ of
$X^{\mathbf{a},\bar{\mathbf{a}}}_t:=\int_0^t\eta_sd\langle
B^\mathbf{a},B^{\bar{\mathbf{a}}}\rangle_s$ such that for any $0\leq
s\leq t\leq T,$
\begin{align*}
&\mathbb{E}[\sup_{u\in[s,t]}|\tilde{X}^{\mathbf{a},
\bar{\mathbf{a}}}_u-\tilde{X}^{\mathbf{a},\bar{\mathbf{a}}}_s|^p]\\
\leq&\bigg(\frac{1}{4}
\sigma_{(\mathbf{a}+\bar{\mathbf{a}})(\mathbf{a}+\bar{\mathbf{a}})^T}
+\frac{1}{4}
\sigma_{(\mathbf{a}-\bar{\mathbf{a}})(\mathbf{a}-\bar{\mathbf{a}})^T}\bigg)^p
(t-s)^{p-1}\mathbb{E}[\int_s^t|\eta_u|^pdu].
\end{align*}
\end{lemma}
\textbf{Remark:} By the above two Theorems, we can assume that the
stochastic integrals $\int_0^t\eta_sdB^{\mathbf{a}}_s,\
\int_0^t\eta_sd\langle B^\mathbf{a},B^{\bar{\mathbf{a}}}\rangle_s$
and $\int_0^t\eta_sds$ are continuous in $t$ for all
$\omega\in\Omega.$\\\\
The last two lemmas can be regarded as the starting point of this paper, and the proof of the lemma \ref {le 2.2.1} can be found in \cite{CL}.
\begin{lemma}\label{complex}
Suppose that $g$ is a given function satisfying $g(\cdot, x)\in M^2_G([0,T];\mathbb{R}^n)$ for all $x\in \mathbb{R}^n$ and for all $x$, $x_1$ and $x_2\in \mathbb{R}^n$:\\
(A1)\ $|g(t,x)|\leq \beta_1(t)+\beta_2(t)|x|$;\\
(A2)\ $|g(t,x_1)-g(t,x_2)|\leq \beta(t)\gamma(|x_1-x_2|)$,\\
where $\beta_1\in M^2_G([0,T])$, $\beta$, $\beta_2: [0,T]\rightarrow\mathbb{R}^+$ are Lebesgue integrable, $\gamma: \mathbb{R}^+\rightarrow \mathbb{R}^+$ and $\gamma(0+)=0$. Then for all
$X\in M^2_G([0,T];\mathbb{R}^n)$, $g(\cdot, X_\cdot)\in M^2_G([0,T];\mathbb{R}^n)$.
\end{lemma}
\textbf{Remark:} We shall prove this lemma in the appendix. Based on this lemma, the $G$-stochastic differential equation (\ref{eq 2}) is well-defined under the integral-Lipschitz condition.
\begin{lemma} \label{le 2.2.1}
Let $\rho:(0,+\infty)\rightarrow(0,+\infty)$ be a continuous,
increasing function satisfying
\begin{equation}\label{eq 11}\rho(0+)=0,\ \int_0^1\frac{dr}{\rho(r)}=+\infty
\end{equation}
and let $u$ be a measurable, non-negative function defined on $(0,+\infty)$ satisfying
$$
u(t)\leq a+\int^t_0 \beta(s)\rho(u(s))ds,\ t\in(0,+\infty),
$$
where $a\in [0,+\infty)$, and $\beta: [0,T]\rightarrow\mathbb{R}^+$ is Lebesgue
integrable. We have:\\\\
$\quad i)$ If $a=0$, then $u(t)=0$, for $t\in [0,+\infty)$;\\
$\quad ii)$ If $a>0$, we define $v(t)=\int^t_{t_0} (ds/\rho(s))$,
$t\in[0,+\infty)$, where $t_0\in (0,+\infty)$, then
\begin{equation}\label{eq uv}
u(t)\leq v^{-1} (v(a)+\int^t_0 \beta(s) ds).
\end{equation}
\end{lemma}
\section{Solvability of $G$-stochastic differential equations}
In this section, we give the main result of this paper, that is the
existence and uniqueness of a solution to $G$-stochastic differential
equation with integral-Lipschitz coefficients.\\\\
Consider the following stochastic differential equation (\ref{eq 2})
driven by a $d$-dimensional $G$-Brownian motion, and we rewrite it in an equivalent form:
\begin{equation}\label{eq 12}
X_t=x+\int^t_0 b(s,X_s)ds +\sum^d_{i,j=1}\int^t_0 h_{ij}(s,X_s) d\langle B^i, B^j\rangle_s
+ \sum^d_{j=1}\int^t_0 \sigma _j(s,X_s) dB^j_s, \\
\end{equation}
where $t\in[0,T]$, the initial condition $x\in \mathbb {R}^n$ is a given vector, and $b$, $h_{ij}$, $\sigma_j$ are given functions satisfying
$b(\cdot, x)$, $h_{ij}(\cdot, x)$, $\sigma_j(\cdot, x)\in M^2_G([0,T];\mathbb{R}^n)$ for all $x\in \mathbb{R}^n$ and $i$, $j=1, \ldots, d$.
We assume further that the following conditions are satisfied, for all $x$, $x_1$, $x_2\in \mathbb{R}^n$:\\\\
$(H1)$\ $|b(t,x)|^2+\sum^d_{i,j=1}|h_{ij}(t,x)|^2+\sum^d_{j=1}|\sigma_j(t,x)|^2\leq
\beta_1^2(t)+\beta_2^2(t)|x|^2$;\\
$(H2)$\ $|b(t,x_1)-b(t,x_2)|^2+\sum^d_{i,j=1}|h_{ij}(t,x_1)-h_{ij}(t,x_2)|^2$\\
\hspace*{1cm}$+\sum^d_{j=1}|\sigma_j(t,x_1)-\sigma_j(t,x_2)|^2\leq\beta^2(t)\rho(|x_1-x_2|^2)$,\\\\
where $\beta_1\in M_G^2([0,T])$,
$\beta:[0,T]\rightarrow\mathbb{R}^+$, $\beta_2:[0,T]\rightarrow\mathbb{R}^+$ are square integrable, and $\rho:(0,+\infty)\rightarrow(0,+\infty)$ is continuous,
increasing, concave function satisfying (\ref{eq 11}).
\begin{theorem}\label{th 3.1}
We suppose (H1) and (H2), then there exists a unique continuous
process $X(\cdot;x)\in M^2_G([0,T];\mathbb{R}^n)$.
\end{theorem}
\textbf{Proof:} We begin with the proof of uniqueness. Suppose $X(\cdot;x)$ is a solution of (\ref{eq 12}), we have
\begin{align*}
X(t;x_1)-X(t;x_2)&=x_1-x_2+\int_0^t(b(s,X(s;x_1))-b(s,X(s;x_2)))ds\\
&+\sum^d_{i,j=1}\int_0^t(h_{ij}(s,X(s;x_1))-h_{ij}(s,X(s;x_2)))d\langle B^i,B^j\rangle_s\\
&+\sum^d_{j=1}\int_0^t(\sigma_j(s,X(s;x_1))-\sigma_j(s,X(s;x_2)))dB^j_s
\end{align*}
and
\begin{align*}
|X(t;x_1)-X(t;x_2)|^2&\leq4|x_1-x_2|^2+4|\int_0^t(b(s,X(s;x_1))-b(s,X(s;x_2)))ds|^2\\
&+4|\sum^d_{i,j=1}\int_0^t(h_{ij}(s,X(s;x_1))-h_{ij}(s,X(s;x_2)))d\langle B^i,B^j\rangle_s|^2\\
&+4|\sum^d_{j=1}\int_0^t(\sigma_j(s,X(s;x_1))-\sigma_j(s,X(s;x_2)))dB^j_s|^2.
\end{align*}
From Lemma \ref{th 2.1.1}, Lemma \ref{th 2.1.2} and (H1) we
notice that, for some constants $K_1,\ K_2$ and $K_3>0$:
\begin{align*}
&\bar{\mathbb{E}}[\sup_{0\leq r\leq t}|\int_0^r
(b(s,X(s;x_1))-b(s,X(s;x_2)))ds|^2]\\
\leq\ &K_1t\int_0^t\bar{\mathbb{E}}[|b(s,X(s;x_1))-b(s,X(s;x_2))|^2]ds\\
\leq\ &K_1t\int_0^t\beta^2(s)\bar{\mathbb{E}}[\rho(|X(s,x_1)-X(s,x_2)|^2)]ds,
\end{align*}
\begin{align*}
&\bar{\mathbb{E}}[\sup_{0\leq r\leq t}|\int_0^r
(h_{ij}(s,X(s;x_1))-h_{ij}(s,X(s;x_2)))d\langle B^i,B^j\rangle_s|^2]\\
\leq\ &K_2t\int_0^t\bar{\mathbb{E}}[|h_{ij}(s,X(s;x_1))-h_{ij}(s,X(s;x_2))|^2]ds\\
\leq\ &K_2t\int_0^t\beta^2(s)\bar{\mathbb{E}}[\rho(|X(s,x_1)-X(s,x_2)|^2)]ds\\
\end{align*}
and
\begin{align*}
&\bar{\mathbb{E}}[\sup_{0\leq r\leq t}|\int_0^r
(\sigma_j(s,X(s;x_1))-\sigma_j(s,X(s;x_2)))dB^j_s|^2]\\
\leq\ &K_3\bar{\mathbb{E}}\bigg[\int_0^t|\sigma_j(s,X(s;x_1))-\sigma_j(s,X(s;x_2))|^2ds\bigg]\\
\leq\ &K_3\int_0^t\bar{\mathbb{E}}[|\sigma_j(s,X(s;x_1))-\sigma_j(s,X(s;x_2))|^2]ds\\
\leq\ &K_3\int_0^t\beta^2(s)\bar{\mathbb{E}}[\rho(|X(s;x_1)-X(s;x_2)|^2)]ds.
\end{align*}
Now let us put:
$$u(t)=\sup_{0\leq r\leq t}\bar{\mathbb{E}}[|X(r;x_1)-X(r;x_2)|^2],$$
then we have, due to the sub-additivity property of
$\bar{\mathbb{E}}[\cdot]$, that for some positive constants $C_1$ and $C_2$,
$$u(t)\leq C_1|x_1-x_2|^2+C_2\int_0^t\beta^2(s)\bar{\mathbb{E}}[\rho(|X(s;x_1)-X(s;x_2)|^2)]ds .$$
As $\rho$ is concave and increasing, we deduce from Jensen's inequality (Lemma \ref{le
2.2.2}):
\begin{align*}
u(t)&\leq C_1|x_1-x_2|^2+C_2\int_0^t\beta^2(s)\rho(\bar{\mathbb{E}}[|X(s;x_1)-X(s;x_2)|^2])ds\\
&\leq C_1|x_1-x_2|^2+C_2\int_0^t\beta^2(s)\rho(u(s))ds.
\end{align*}
From (\ref{eq uv}), we obtain:
$$
u(t)\leq v^{-1} (v(C_1|x_1-x_2|^2)+C_2\int^t_0\beta^2(s)ds).
$$
In particular, if $x_1=x_2,$ we obtain the uniqueness of the solution to (\ref{eq 12}).\\\\
Now we give the proof of the existence to (\ref{eq 12}). We
define the Picard sequence of processes $\{X^m(\cdot), m\geq0\}$ as follows:
$$X^0(t)=x,\ t\in[0,T],$$
and
\begin{align}\label{eq m+1}
X^{m+1}(t)=x&+\int_0^tb(s,X^m(s))ds+\sum^d_{i,j=1}\int_0^th_{ij}(s,X^m(s))d\langle
B^i,B^j\rangle_s\notag\\
&+\sum^d_{j=1}\int_0^t\sigma_j(s,X^m(s))dB^j_s,\ t\in[0,T].
\end{align}
Because of the basic assumptions and (H1), the sequence
$\{X^m(\cdot),m\geq0\}\subset M_G^2([0,T];\mathbb{R}^n)$ is well
defined. We first establish a priori estimate for
$\{\bar{\mathbb{E}}[|X^m(t)|^2],m\geq0\}.$ \\\\
From (\ref{eq m+1}), we deduce by Lemma \ref{th 2.1.1} and
Lemma \ref{th 2.1.2} that, for some positive constants $C_1$ and $C_2$,
$$\bar{\mathbb{E}}[|X^{m+1}(t)|^2]\leq
C_1|x|^2+C_2\int_0^t\bar{\mathbb{E}}[\beta_1^2(s)+\beta_2^2(s)|X^m(s)|^2]ds.$$
Hence,
\begin{align*}
\bar{\mathbb{E}}[|X^{m+1}(t)|^2]\leq
C_1|x|^2+C_2\int_0^t\bar{\mathbb{E}}[\beta_1^2(s)]ds+C_2\int_0^t\beta_2^2(s)\bar{\mathbb{E}}[|X^m(s)|^2]ds.
\end{align*}
Set
\begin{equation*}\label{eq 13}
p(t)=\bigg(C_1|x|^2+C_2\int_0^t\bar{\mathbb{E}}[\beta_1^2(s)]ds\bigg)\exp\bigg\{C_2\int_0^t\beta_2^2(s)ds\bigg\},
\end{equation*}
then $p$ is the solution of
\begin{equation*}\label{eq 14}
p(t)=C_1|x|^2+C_2\int_0^t\bar{\mathbb{E}}[\beta_1^2(s)]ds+C_2\int_0^t\beta_2^2(s)p(s)ds.
\end{equation*}
By recurrence, it is easy to prove that for any $m\geq0,$
$$\bar{\mathbb{E}}[|X^m(t)|^2]\leq p(t).$$
Set
$$u_{k+1,m}(t)=\sup_{0\leq r\leq t}\bar{\mathbb{E}}[|X^{k+1+m}(r)-X^{k+1}(r)|^2].$$
From the definition of the sequence $\{X^m(\cdot),m\geq0\},$ we have
\begin{align*}
X^{k+1+m}(t)-X^{k+1}(t)&=\int_0^t(b(s,X^{k+m}(s))-b(s,X^k(s)))ds\\
&+\sum^d_{i,j=1}\int_0^t(h_{ij}(s,X^{k+m}(s))-h_{ij}(s,X^k(s)))d\langle B^i,B^j\rangle_s\\
&+\sum^d_{j=1}\int_0^t(\sigma_j(s,X^{k+m}(s))-\sigma_j(s,X^k(s)))dB^j_s.
\end{align*}
By the same method as in the proof of uniqueness, we deduce that, for some positive constant $C$,
$$u_{k+1,m}(t)\leq C\int_0^t\beta^2(s)\rho(u_{k,m}(s))ds.$$
Set
$$v_k(t)=\sup_{m}u_{k,m}(t),\ 0\leq t\leq T,$$
then,
$$0\leq v_{k+1}(t)\leq C\int_0^t\beta^2(s)\rho(v_{k}(s))ds.$$
Finally, we define:
$$\alpha(t)=\limsup_{k\rightarrow+\infty}v_k(t),\ 0\leq
t\leq T.$$
Since $\rho$ is continuous and $v_k(t)\leq 4p(t),$ we have
$$0\leq\alpha(t)\leq C\int_0^t\beta^2(s)\rho(\alpha(s))ds,\ 0\leq
t\leq T.
$$
Hence, by Lemma \ref{le 2.2.1},
$$\alpha(t)=0,\ 0\leq t\leq T.$$ That is,
$\{X^m(\cdot),m\geq0\}$ is a Cauchy sequence in
$L_G^2([0,T];\mathbb{R}^n)$. Set
$$X(t)=\sum_{m=1}^\infty(X^m(t)-X^{m-1}(t)),$$
we notice that, for some positive constants $K_1$, $K_2$ and $K_3$,
\begin{align*}
&\bar{\mathbb{E}}[\sup_{0\leq t\leq
T}|\int_0^t(b(s,X^{m}(s))-b(s,X(s)))ds|^2]\\
\leq&K_1T\int_0^T\beta^2(s)\rho(\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m}(t)-X(t)|^2])ds,
\end{align*}
\begin{align*}
&\bar{\mathbb{E}}[\sup_{0\leq t\leq
T}|\int_0^t(h_{ij}(s,X^{m}(s))-h_{ij}(s,X(s)))d\langle B^i,B^j\rangle_s|^2]\\
\leq&K_2T\int_0^T\beta^2(s)\rho(\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m}(t)-X(t)|^2])ds
\end{align*}
and
\begin{align*}
&\bar{\mathbb{E}}[\sup_{0\leq t
\leq T}|\int_0^t(\sigma_j(s,X^{m}(s))-\sigma_j(s,X(s)))dB^j_s|^2]\\
\leq&K_3\int_0^T\beta^2(s)\rho(\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m}(t)-X(t)|^2])ds,
\end{align*}
since $\rho$ is continuous and $\rho(0+)=0$, we have $X(\cdot)\in M^2_G([0,T];\mathbb{R}^n)$
satisfies (\ref{eq 12}). The proof of the existence of the
solution to (\ref{eq 12}) is now complete. \hfill{} $\square$\\\\
Furthermore, we consider the existence and uniqueness of a solution to the stochastic differential equation (\ref{eq 12}) under some weaker condition than (H2).
\begin{theorem}\label{th 3.5}
We suppose the following condition: for any $x_1$, $x_2\in\mathbb{R}^n$\\\\
(H1')\ $|b(t,x)|^2+\sum^d_{i,j=1}|h_{ij}(t,x)|^2+\sum^d_{j=1}|\sigma_j(t,x)|^2\leq
\beta^2_1(t)+\beta^2_2(t)|x|^2$,\\
(H2')$\left\{\begin{array}{l}
|b(t,x_1)-b(t,x_2)|\leq\beta(t)\rho_1(|x_1-x_2|);\\
|h_{ij}(t,x_1)-h_{ij}(t,x_2)|\leq\beta(t)\rho_1(|x_1-x_2|);\\
|\sigma_j(t,x_1)-\sigma_j(t,x_2)|^2\leq\beta(t)\rho_2(|x_1-x_2|^2),
\end{array}\right.$\\
where $\beta_1\in M^p_G([0,T]; \mathbb{R})$, $\beta_2: [0,T]\rightarrow \mathbb{R}^+$ is $p$-integrable, $p>2$, $\beta:[0,T]\rightarrow\mathbb{R}^+$ is Lebesgue integrable,
and $\rho_1,\ \rho_2:(0,+\infty)\rightarrow(0,+\infty)$ are continuous,
concave and increasing, and both of them satisfy (\ref{eq 11}).
Furthermore, we assume that $$\rho_3(r)=\frac{\rho_2(r^2)}{r}, \
r\in(0,+\infty)$$ is also continuous, concave and increasing, and
$$\rho_3(0+)=0,\ \int_0^1\frac{dr}{\rho_1(r)+\rho_3(r)}=+\infty.$$
Then there exists a unique solution $X$ in $M^p_G([0,T];\mathbb{R}^n)$ to the equation (\ref{eq 12}).
\end{theorem}
\textbf{Example:} If
\begin{align*}
\rho_1(r)&=r\ln\frac{1}{r},\\
\rho_2(r)&=r\ln\frac{1}{r},
\end{align*}
then the conditions for Theorem \ref{th 3.5} are satisfied but not for Theorem \ref{th 3.1}.\\\\
\textbf{Remark:} Fang and Zhang prove in \cite{FZ} a similar uniqueness result for stochastic differential equations, where $\rho$ need not to be concave by a stopping time technique. They derive existence by the well-known Yamada-Watanabe theorem which says that the existence of weak solution and pathwise uniqueness imply the existence of strong solution. For the stochastic differential equation driven by $G$-Brownian motion, neither the stopping time technique nor the corresponding Yamda-Watanabe result are available.\\\\
\textbf{Proof:} We start with the proof of existence. Firstly we define a sequence of processes $\{X^m(\cdot), m\geq0\}$ as
follows:
$$X^0(t)=x,\ t\in[0,T],$$
and
\begin{align*}
X^{m+1}(t)=&x+\int_0^tb(s,X^m(s))ds\\
+&\sum^d_{i,j=1}\int_0^th_{ij}(s,X^m(s))d\langle
B^i, B^j\rangle_s
+\sum^d_{j=1}\int_0^t\sigma_j(s,X^{m+1}(s))d B^j_s.
\end{align*}
Because of the assumptions of this theorem and thanks to Theorem \ref{th 3.1}, the sequence $\{X^m(\cdot),
m\geq0\}$ is well defined in $L^2_G([0,T]; \mathbb{R}^n)$. \\\\
In order to apply It\^{o}'s fomula, we first define the trunction functions $b^N$, $h^N_{ij}$ and $\sigma_j^N$. For $i$, $j=1, \ldots, d$ and $N\geq 0$, we set
\begin{align*}
b^N(t, x)&=
\left\{\begin{array}{r@{,}l}
b(t, x)&\ if\ |b(t, x)|<N,\\
\frac{Nb(t, x)}{|b(t, x)|}&\ if\ |b(t, x)|\geq N;
\end{array}\right.\\
h_{ij}^N(t, x)&=
\left\{\begin{array}{r@{,}l}
h_{ij}(t, x)&\ if\ |h_{ij}(t, x)|<N,\\
\frac{Nh_{ij}(t, x)}{|h_{ij}(t, x)|}&\ if\ |h_{ij}(t, x)|\geq N;
\end{array}\right.\\
\sigma_j^N(t, x)&=
\left\{\begin{array}{r@{,}l}
\sigma_{j}(t, x)&\ if\ |\sigma_{j}(t, x)|<N,\\
\frac{N\sigma_{j}(t, x)}{|\sigma_{j}(t, x)|}&\ if\ |\sigma_{j}(t, x)|\geq N.
\end{array}\right.
\end{align*}
It is easy to verify that $b^N$, $h_{ij}^N$ and $\sigma^N$ still satisfy (H1) and (H2). Define
\begin{align*}
X^{m+1, N}(t)=&x+\int_0^tb^N(s,X^m(s))ds\\
+&\sum^d_{i,j=1}\int_0^th^N_{ij}(s,X^m(s))d\langle
B^i, B^j\rangle_s
+\sum^d_{j=1}\int_0^t\sigma^N_j(s,X^{m+1}(s))d B^j_s.
\end{align*}
By the definition of the It\^{o} integral, for a fixed $m\geq 0$, the sequence $\{X^{m, N}(\cdot),
N\geq0\}$ is well defined in $M^1_G([0,T]; \mathbb{R}^n)$.\\\\
Let us now establish a priori estimates for $\{\bar{\mathbb{E}}[|X^{m}(t)|^p],
m\geq0\}$ and $\{\bar{\mathbb{E}}[|X^{m, N}(t)|^p], m, N\geq0\}$, $p>2$. By Lemma \ref{th 2.1.1} and Lemma \ref{th 2.1.2}, for some positive constant $C_1$ and $C_2$, we have
\begin{align*}
\bar{\mathbb{E}}[|X^{m}(t)|^p]\leq& C_1|x|^p+C_2\int_0^t\bar{\mathbb{E}}[\beta_1^p(s)]ds\\
+&\frac{C_2}{2}\int_0^t\beta_2^p(s)\bar{\mathbb{E}}[|X^{m}(t)|^p]ds+\frac{C_2}{2}\int_0^t\beta_2^p(s)\bar{\mathbb{E}}[|X^{m+1}(t)|^p]ds.
\end{align*}
Taking into consideration that $\beta_1\in M^p_G([0,T]; \mathbb{R})$, $\beta_2$ is $p$-integrable, by induction, we have $\bar{\mathbb{E}}[|X^m(t)|^p]\leq p'(t)$, $p>2$, where $p'(t)$ is the solution to
\begin{equation*}
p'(t)=C_1|x|^p+C_2\int_0^t\bar{\mathbb{E}}[\beta_1^p(s)]ds+C_2\int_0^t\beta_2^p(s)p'(s)ds.
\end{equation*}
Hence, for some positive constant $M$,
\begin{align*}
\sup_{m\geq 0}\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^m(t)|^p]&\leq M.
\end{align*}
In a similar way, we also have,
$$
\sup_{m, N\geq 0}\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m, N}(t)|^p]\leq M,$$
and
\begin{align*}
&\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m, N}(t)-X^m(t)|^p]\\
\leq&2^{p-1}\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m, N}(t)|^p]+ 2^{p-1}\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^m(t)|^p]\\
\leq&2^{p}M.
\end{align*}
Then, for a fixed $m>0$, we have
\begin{align*}
&\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m, N}(t)-X^m(t)|]\\
\leq& \bar{\mathbb{E}}\bigg[\int_0^T|b^N(s,X^m(s))-b(s,X^m(s))|ds\bigg]\\
+&\sum^d_{i,j=1}\bar{\mathbb{E}}\bigg[\int_0^T|h^N_{ij}(s,X^m(s))-h_{ij}(s,X^m(s))|d\langle
B^i, B^j\rangle_s\bigg]\\
+&\sup_{0\leq t\leq T}\sum^d_{j=1}\bar{\mathbb{E}}\bigg[|\int_0^t(\sigma^N_j(s,X^{m+1}(s))-\sigma_j(s,X^{m+1}(s)))d B^j_s|\bigg]\\
\leq&\int_0^T\bar{\mathbb{E}}\bigg[|b(s,X^m(s))|I_{\{|b(s, X^m(s))|\geq N\}}\bigg]ds\\
+&\sum^d_{i, j=1}\int_0^T\bar{\mathbb{E}}\bigg[|h_{ij}(s,X^m(s))|I_{\{|h_{ij}(s, X^m(s))|\geq N\}}\bigg]ds\\
+&\sum^d_{j=1}\bigg(\int_0^T\bar{\mathbb{E}}[|\sigma_{j}(s,X^{m+1}(s))|^2I_{\{|\sigma_{j}(s, X^{m+1}(s))|^2\geq N\}}]ds\bigg)^{\frac{1}{2}}.
\end{align*}
Since $b(\cdot, X^m_\cdot)$, $h_{ij}(\cdot, X^m_\cdot)$ and $\sigma_j(\cdot, X^m_\cdot)\in M^2_G([0, T]; \mathbb{R}^n)$, from Theorem 59 of \cite{DHP}, as $N\rightarrow+\infty$, the right side converges to 0. Thus, $X^{m, N}$ converges to $X^{m}$ in $M^1_G([0,T];\mathbb{R}^n)$.\\\\
Note that as $|x|$ is not $C^2,$ we approximate $|x|$ by
$F_\varepsilon\in C^2$, where
$$F_\varepsilon(x)=(|x|^2+\varepsilon)^\frac{1}{2},\ x\in\mathbb{R}^n,$$
for a given $\varepsilon>0$.
We notice that
$$|\frac{\partial F_\varepsilon(x)}{\partial x_i}|\leq1,\
|\frac{\partial^2F_\varepsilon(x)}{\partial x_i \partial x_j}|\leq\frac{2}{(|x|^2+\varepsilon)^\frac{1}{2}},$$
and $\frac{\partial F_\varepsilon(x)}{\partial x_i}$, $\frac{\partial^2F_\varepsilon(x)}{\partial x_i \partial x_j}$ are bounded and uniformly Lipschitz for $i, j=1, \ldots, n$.\\\\
Applying
$G$-It$\hat{o}$ formula to $F_\varepsilon(X^{k+1+m, N}(t)-X^{k+1, N}(t))$,
and we take the $G$-expectation on both sides. From Lemma \ref{th 2.1.2}, for some positive constant $K$, we get
\begin{align}\label{eq key}
&\bar{\mathbb{E}}[F_\varepsilon(X^{k+1+m, N}(t)-X^{k+1, N}(t))]\notag\\
\leq&\bar{\mathbb{E}}\bigg[\int_0^t|b^N(s,X^{k+m}(s))-b^N(s,X^{k}(s))|ds\bigg]\notag\\
+&K\sum^d_{i,j=1}\bar{\mathbb{E}}\bigg[\int_0^t|h^N_{ij}(s,X^{k+m}(s))-h^N_{ij}(s,X^{k}(s))|ds\bigg]\notag\\
+&K\sum^d_{j=1}\bar{\mathbb{E}}\bigg[\int_0^t\frac{|\sigma^N_j(s,X^{k+m+1}(s))-\sigma^N_j(s,X^{k+1}(s))|^2}{(|X^{k+m+1, N}(s)-X^{k+1, N}(s)|^2+\varepsilon)^\frac{1}{2}}ds\bigg]\notag\\
\leq&(1+Kd^2)\int_0^t\beta(s)\rho_1(\bar{\mathbb{E}}[|X^{k+m}(s)-X^{k}(s)|])ds\notag\\
+&Kd\int_0^t \beta(s)\bar{\mathbb{E}}\bigg[\frac{\rho_2(|X^{k+m+1}(s)-X^{k+1}(s)|^2)}{(|X^{k+m+1, N}(s)-X^{k+1, N}(s)|^2+\varepsilon)^\frac{1}{2}}\bigg]ds.
\end{align}\\
For a fixed $\varepsilon>0$, define
$$\Delta F_\varepsilon^{k, m, N}(t)= |F_\varepsilon(X^{k+m, N}(t)-X^{k, N}(t))-F_\varepsilon(X^{k+m}(t)-X^{k}(t))|, $$
then
\begin{align*}
&\sup_{0 \leq t\leq T}\bar{\mathbb{E}}[\Delta F_\varepsilon^{k+1, m, N}(t)]\\
\leq&\sup_{0 \leq t\leq T}\bar{\mathbb{E}}
[|(X^{k+1+m, N}(t)-X^{k+1, N}(t))-(X^{k+1+m}(t)-X^{k+1}(t))|]\\
\leq&\sup_{0 \leq t\leq T}\bar{\mathbb{E}}
[|X^{k+1+m, N}(t)-X^{k+1+m}(t)|]+\sup_{0 \leq t\leq T}\bar{\mathbb{E}}
[|X^{k+1, N}(t)-X^{k+1}(t)|].
\end{align*}
Hence, the left side of (\ref{eq key}) uniformly converges to $\bar{\mathbb{E}}
[F_\varepsilon(X^{k+1+m}(t)-X^{k+1}(t))]$ as $N\rightarrow+\infty$. \\\\
On the other hand, $\rho_2:(0,+\infty)\rightarrow(0,+\infty)$ are continuous, concave and increasing, then for arbitrary fixed $\delta>0,$ there exist some positive constants
$K_\delta$, such that $\rho_2(x)\leq K_\delta x$, for
$x>\delta$. Choosing $M>0$ sufficiently large, for some positive constant $C_\varepsilon$ and $\alpha>0$,
\begin{align*}\label{eq fan}
&\limsup_{N\rightarrow+\infty}\int_0^T\beta(s)\bar{\mathbb{E}}[\rho_2(|X^{k+m+1}(s)-X^{k+1}(s)|^2)\notag\\
\times&|\frac{1}{(|X^{k+m+1, N}(s)-X^{k+1, N}(s)|^2+\varepsilon)^\frac{1}{2}}-\frac{1}{(|X^{k+m+1}(s)-X^{k+1}(s)|^2+\varepsilon)^\frac{1}{2}}|]ds\notag\\
\leq&\limsup_{N\rightarrow+\infty}\int_0^T\sup_{0 \leq t\leq T}\bar{\mathbb{E}}[\Delta F_\varepsilon^{k+1, m, N}(t)]C_\varepsilon\beta(s)\rho_2(M)ds\notag\\
+&\int_0^T2\varepsilon^{-\frac{1}{2}} K_\delta\beta(s)\bar{\mathbb{E}}[|X^{k+m+1}(s)-X^{k+1}(s)|^2I_{\{|X^{k+m+1}(s)-X^{k+1}(s)|^2\geq M\}}]\notag\\
=&\int_0^T2\varepsilon^{-\frac{1}{2}} K_\delta\beta(s)
\bar{\mathbb{E}}\bigg[\frac{
|X^{k+m+1}(s)-X^{k+1}(s)|^{2+\alpha}}{M^\alpha}\bigg].
\end{align*}
Since $M$ can be arbitrary large, and $\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{k+m+1}(t)-X^{k+1}(t)|^{2+\alpha}]<+\infty$,
\begin{align*}
&\limsup_{N\rightarrow+\infty}\int_0^T\beta(s)\bar{\mathbb{E}}[\rho_2(|X^{k+m+1}(s)-X^{k+1}(s)|^2)\notag\\
\times&|\frac{1}{(|X^{k+m+1, N}(s)-X^{k+1, N}(s)|^2+\varepsilon)^\frac{1}{2}}-\frac{1}{(|X^{k+m+1}(s)-X^{k+1}(s)|^2+\varepsilon)^\frac{1}{2}}|]ds\notag\\
=&0.
\end{align*}\\
Taking $N\rightarrow+\infty$ and $\varepsilon\rightarrow0$ on both side of (\ref{eq key}), we deduce from monotone convergence theorem in \cite{DHP} and \cite{G} that,
\begin{align*}
&\bar{\mathbb{E}}[|X^{k+1+m}(t)-X^{k+1}(t)|]\\
\leq&(1+Kd^2)\int_0^t\beta(s)\rho_1(\bar{\mathbb{E}}[|X^{k+m}(s)-X^{k}(s)|])ds\notag\\
+&Kd\int_0^t \beta(s)\rho_3(\bar{\mathbb{E}}[|X^{k+m+1}(s)-X^{k+1}(s)|])ds.
\end{align*}\\
Set
\begin{align*}
u_{k,m}(t)=&\sup_{0\leq r\leq
t}\bar{\mathbb{E}}[|X^{k+m}(r)-X^{k}(r)|];\\
v_k(t)&=\sup_{m}u_{k,m}(t),\ 0\leq t\leq T.
\end{align*}
For some positive constant $C$,
$$u_{k+1,m}(t)\leq C\int_0^t\beta(s)(\rho_1(u_{k,m}(s))+\rho_3(u_{k+1,m}(s)))ds.$$
then,
$$0\leq v_{k+1}(t)\leq C \int_0^t\beta(s)(\rho_1(v_k(s))+\rho_3(v_{k+1}(s)))ds.$$
Finally, we define: $$\alpha(t)=\limsup_{k\rightarrow+\infty}v_k(t),\
t\geq0,$$ then
$$0\leq \alpha(t)\leq C\int_0^t\beta(s)(\rho_1(\alpha(s))+\rho_3(\alpha(s)))ds,\ 0\leq t\leq T.
$$
By Lemma \ref{le 2.2.1}, $$\alpha(t)=0,\ t\in[0,T].$$
Hence, $\{X^m(\cdot), m\geq0\}$
is a Cauchy sequence in $M_G^1([0,T],\mathbb{R}^n).$ Then there
exists $X(\cdot)\in M_G^1([0,T],\mathbb{R}^n)$ and a subsequence
$\{X^{m_l}(\cdot),l\geq1\}\subset\{X^m(\cdot),m\geq1\}$ such that
$$X^{m_l}\rightarrow X,\ as\ l\rightarrow +\infty,\ q.s..$$\\
By the priori estimates and Lemma \ref{le 2.2.12}, we have
$$
\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X(t)|^p]\leq M.
$$
Hence,
$$
\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m}(t)-X(t)|^p]\leq2^{p}M.
$$
Consequencely, for a fixed $\varepsilon>0$,
\begin{align*}
&\limsup_{m\rightarrow+\infty}(\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^m(t)-X(t)|^2])\\
\leq&\limsup_{m\rightarrow+\infty} (\varepsilon^2\sup_{0\leq t\leq T}\bar{\mathbb{E}}[I_{\{|X^m(t)-X(t)|< \varepsilon\}}]
+\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^m(t)-X(t)|^2I_{\{|X^m(t)-X(t)|\geq \varepsilon\}}])\\
\leq&\varepsilon^2+\limsup_{m\rightarrow+\infty}(\sup_{0\leq t\leq T}((\bar{\mathbb{E}}[|X^m(t)-X(t)|^p])^\frac{2}{p}(\bar{\mathbb{E}}[|I_{\{|X^m(t)-X(t)|\geq \varepsilon\}}|^\frac{p}{p-2}])^\frac{p-2}{p}))\\
\leq&\varepsilon^2+4M^\frac{2}{p}\limsup_{m\rightarrow+\infty}(\sup_{0\leq t\leq T}\bar{\mathbb{E}}[I_{\{|X^m(t)-X(t)|\geq \varepsilon\}}])^\frac{p-2}{p}\\
=&\varepsilon^2.
\end{align*}
The last step above can be easily deduced from $\lim_{m\rightarrow+\infty}(\sup_{0\leq t\leq T}\bar{\mathbb{E}}|X^m(t)-X(t)|)=0$ and Lemma 37 in \cite{DHP}. Since $\varepsilon$ can be arbitrary small, we have $\lim_{m\rightarrow+\infty}\bar{\mathbb{E}}[|X^m(t)-X(t)|^2]=0$.\\\\
Notice that $\rho_1$, $\rho_2$ are continuous and vanish at 0, and also for arbitrary fixed $\delta>0,$ there exist some positive constants
$K_\delta$, such that $\rho_1(x)\leq K_\delta x$, for $x>\delta$. Hence, for some positive constant $K$, we have
\begin{align*}
&\limsup_{m\rightarrow\infty}\bar{\mathbb{E}}[\sup_{0\leq t\leq
T}|\int_0^t(b(s,X^{m}(s))-b(s,X(s)))ds|^2]\\
\leq&
\limsup_{m\rightarrow\infty}K\int_0^T\beta^2(s)\bar{\mathbb{E}}[\rho^2_1(|X^{m}(s)-X(s)|)]ds\\
\leq&\limsup_{m\rightarrow\infty}\bigg(\int_0^TK\beta^2(s)\bar{\mathbb{E}}[\rho^2_1(|X^{m}(s)-X(s)|)I_{\{|X^{m}(s)-X(s)|>\varepsilon\}}]ds\\
+&K\rho^2_1(\varepsilon^2)\int_0^T\beta^2(s)\bar{\mathbb{E}}[I_{\{|X^{m}(s)-X(s)|\leq\varepsilon\}}]ds\bigg)\\
\leq&\limsup_{m\rightarrow\infty}\bigg(\int_0^TK\beta^2(s)K_\delta^2\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m}(t)-X(t)|^2]]ds\bigg)+K\rho_1^2(\varepsilon^2)\int_0^T\beta^2(s)ds\\
=&K\rho_1^2(\varepsilon^2)\int_0^T\beta^2(s)ds
\end{align*}
Since $\varepsilon$
can arbitrary small, then we have
$$\lim_{m\rightarrow\infty}\bar{\mathbb{E}}[\sup_{0\leq t\leq
T}|\int_0^t(b(s,X^{m}(s))-b(s,X(s)))ds|^2]=0.$$ Similarly we get
$$\lim_{m\rightarrow\infty}\bar{\mathbb{E}}[\sup_{0\leq t\leq
T}|\int_0^t(h_{ij}(s,X^{m}(s))-h_{ij}(s,X(s)))d\langle
B^i,B^j\rangle_s|^2]=0$$ and
\begin{align*}
&\limsup_{m\rightarrow+\infty}\bar{\mathbb{E}}[\sup_{0\leq t\leq
T}|\int_0^t(\sigma_j(s,X^{m}(s))-\sigma_j(s,X(s)))dB^j_s|^2]\\
\leq&\limsup_{m\rightarrow+\infty}C\int_0^T\beta^2(s)\rho_2(\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X^{m}(s)-X(s)|^2])ds\\
=&0.
\end{align*}
Then the proof of the existence of the solution to (\ref{eq 12}) is
complete. \\\\
Now we turn to the proof of uniqueness. Suppose $X_1$, $X_2\in M^2_G([0,T];\mathbb{R}^n)$ are two solutions satisfying (\ref {eq 12}). We define the truncation sequence as follows, for $i$, $j=1, \ldots, d$ and $N\geq 0$,
\begin{align*}
b^N(t, x)&=
\left\{\begin{array}{r@{,}l}
b(t, x)&\ if\ |b(t, x)|<N,\\
\frac{Nb(t, x)}{|b(t, x)|}&\ if\ |b(t, x)|\geq N;
\end{array}\right.\\
h_{ij}^N(t, x)&=
\left\{\begin{array}{r@{,}l}
h_{ij}(t, x)&\ if\ |h_{ij}(t, x)|<N,\\
\frac{Nh_{ij}(t, x)}{|h_{ij}(t, x)|}&\ if\ |h_{ij}(t, x)|\geq N;
\end{array}\right.\\
\sigma_j^N(t, x)&=
\left\{\begin{array}{r@{,}l}
\sigma_{j}(t, x)&\ if\ |\sigma_{j}(t, x)|<N,\\
\frac{N\sigma_{j}(t, x)}{|\sigma_{j}(t, x)|}&\ if\ |\sigma_{j}(t, x)|\geq N,
\end{array}\right.
\end{align*}
and
\begin{align*}
X_1^N(t)=&x+\int_0^tb^N(s,X_1(s))ds\\
+&\sum^d_{i,j=1}\int_0^th^N_{ij}(s,X_1(s))d\langle
B^i, B^j\rangle_s
+\sum^d_{j=1}\int_0^t\sigma^N_j(s,X_1(s))d B^j_s;\\
X_2^N(t)=&x+\int_0^tb^N(s,X_2(s))ds\\
+&\sum^d_{i,j=1}\int_0^th^N_{ij}(s,X_2(s))d\langle
B^i, B^j\rangle_s
+\sum^d_{j=1}\int_0^t\sigma^N_j(s,X_2(s))d B^j_s.
\end{align*}
Similarly to the proof of existence, we have
$X^N_1$, $X^N_2$ converge to $X_1$, $X_2$ in $M^1_G([0,T], \mathbb{R}^n)$ respectively.
For a fixed $\varepsilon>0$, applying $G$-It\^o's formula to
$F_\varepsilon(X^N_1(t)-X^N_2(t))$,
\begin{align*}
\bar{\mathbb{E}}[F_\varepsilon(X_1^{N}(t)-X_2^{N}(t))]
\leq&\bar{\mathbb{E}}\bigg[\int_0^t|b^N(s,X_1(s))-b^N(s,X_2(s))|ds\bigg]\\
+&K\sum^d_{i,j=1}\bar{\mathbb{E}}\bigg[\int_0^t|h^N_{ij}(s,X_1(s))-h^N_{ij}(s,X_2(s))|ds\bigg]\\
+&K\sum^d_{j=1}\bar{\mathbb{E}}\bigg[\int_0^t\frac{|\sigma^N_j(s,X_1(s))-\sigma^N_j(s,X_2(s))|^2}{(|X_1^{N}(s)-X_2^{N}(s)|^2+\varepsilon)^\frac{1}{2}}ds\bigg]\\
\leq&(1+Kd^2)\int_0^t\beta(s)\rho_1(\bar{\mathbb{E}}[|X_1(s)-X_2(s)|])ds\\
+&Kd\int_0^t \beta(s)\bar{\mathbb{E}}\bigg[\frac{\rho_2(|X_1(s)-X_2(s)|^2)}{(|X^N_1(s)-X^N_2(s)|^2+\varepsilon)^\frac{1}{2}}\bigg]ds.
\end{align*}
Letting $N\rightarrow+\infty$ and $\varepsilon\rightarrow 0$, we have
\begin{align*}
\bar{\mathbb{E}}[|X_1(t)-X_2(t)|]
&\leq(1+Kd^2)\int_0^t\beta(s)\rho_1(\bar{\mathbb{E}}[|X_1(s)-X_2(s)|])ds\\
&+Kd\int_0^t \beta(s)\rho_3(\bar{\mathbb{E}}[|X_1(s)-X_2(s)|]ds.
\end{align*}
Thus, for some positive constant $C$,
$$
\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X_1(t)-X_2(t)|]\leq C\int_0^t\beta(s)(\rho_1+\rho_3)(\sup_{0\leq t\leq T}\bar{\mathbb{E}}[|X_1(s)-X_2(s)|])ds.
$$
Finally, Lemma \ref{le 2.2.1} gives the uniqueness result.\hfill{}$\square$
\section{Solvability to $G$-backward stochastic differential equations}
In the last section, we prove the existence and uniqueness of a solution to $G$-backward stochastic differential equation with integral-Lipschitz coefficients.\\\\
Consider the following type of $G$-backward stochastic differential equation ($G$-BSDE):
\begin{equation}\label{eq 15}
Y_t=\mathbb{E}[\xi+\int_t^Tf(s,Y_s)ds+\sum^d_{i,j=1}\int_t^Tg_{ij}(s,Y_s)d\langle
B^i,B^j\rangle_s|\mathcal {F}_t],\ t\in[0,T],
\end{equation}
where
$\xi\in L_G^1(\mathcal {F}_T; \mathbb{R}^n)$, and $f$, $g_{ij}$ are given functions satisfying
$f(\cdot, x)$, $g_{ij}(\cdot, x)\in M^1_G([0,T];\mathbb{R}^n)$ for all $x\in \mathbb{R}^n$ and $i$, $j=1, \ldots, d$.\\\\
We assume further that, for all $y$, $y_1$ and $y_2\in \mathbb{R}^n$,
\begin{eqnarray*}
&&|g(s,y)|+|f(s,y)|\leq\beta(t)+ c|y|,\\
&&|g(s,y_1)-g(s,y_2)|+|f(s,y_1)-f(s,y_2)|\leq\rho(|y_1-y_2|),
\end{eqnarray*}
where $c>0$, $\beta\in M_G^1([0,T];\mathbb{R}_+)$ and $\rho:(0,+\infty)\rightarrow(0,+\infty)$ is a continuous, concave, increasing function satisfying (\ref{eq 11}).
\begin{theorem}\label{th 3.3}
Under the assumptions above, (\ref{eq 15}) admits a unique solution
$Y\in M_G^1([0,T],\mathbb{R}^n).$
\end{theorem}
\textbf{Proof:} Let $Y_1,Y_2\in L_G^1([0,T],\mathbb{R}^n)$ be two solutions
of (\ref{eq 15}), then
\begin{align*}
Y_t^1-Y_t^2=&\mathbb{E}[\xi+\int_t^Tf(s,Y^1_s)ds+\sum^d_{i,j=1}\int_t^Tg_{ij}(s,Y^1_s)d\langle
B^i,B^j\rangle_s|\mathcal
{F}_t]\\
-&\mathbb{E}[\xi+\int_t^Tf(s,Y^2_s)ds+\sum^d_{i,j=1}\int_t^Tg_{ij}(s,Y^2_s)d\langle
B^i,B^j\rangle_s|\mathcal {F}_t].
\end{align*}
Due to the sub-additivity property of $\mathbb{E}[\cdot|\mathcal
{F}_t]$, we obtain:
\begin{align*}
|Y_t^1-Y_t^2|
&\leq\mathbb{E}[|\int_t^T(f(s,Y^1_s)-f(s,Y^2_s))ds|\\
&+\sum^d_{i,j=1}|\int_t^T(g_{ij}(s,Y^1_s)-g_{ij}(s,Y^2_s))d\langle B^i,B^j\rangle_s||\mathcal {F}_t].\\
\end{align*}
Taking the $G$-expectation on both sides, we have from Lemma \ref{th
2.1.2} and Lemma \ref{le 2.2.2} that, for a positive constant $K>0,$
\begin{align*}
\mathbb{E}[|Y_t^1-Y_t^2|]&\leq\mathbb{E}[|\int_t^T(f(s,Y^1_s)-f(s,Y^2_s))ds|\\
&+\sum^d_{i,j=1}|\int_t^T(g_{ij}(s,Y^1_s)-g_{ij}(s,Y^2_s))d\langle
B^i,B^j\rangle_s|]\\
&\leq\mathbb{E}[|\int_t^T(f(s,Y^1_s)-f(s,Y^2_s))ds|]\\
&+\sum^d_{i,j=1}\mathbb{E}[|\int_t^T(g_{ij}(s,Y^1_s)-g_{ij}(s,Y^2_s))d\langle
B^i,B^j\rangle_s|]\\
&\leq K\mathbb{E}\int_t^T\rho(|Y_s^1-Y_s^2|)ds\\
&\leq K\int_t^T \rho(\mathbb{E}[|Y_s^1-Y_s^2|])ds.
\end{align*}
Set
$$u(t)=\mathbb{E}[|Y_t^1-Y_t^2|],$$ then
$$u(t)\leq K\int_t^T \rho(u(s))ds,$$
and we deduce from Lemma \ref{le 2.2.1} that,
$$u(t)=0.$$
Then the uniqueness of the solution can be now easily proved. \\\\
As for the existence of solution, we proceed as in Theorem \ref{th
3.1}: define a sequence of $(Y^m, m\geq0),$ as follows:
$$Y^{m+1}_t=\mathbb{E}[\xi+\int_t^Tf(s,Y^m_s)ds+\sum^{d}_{i,j=1}\int_t^Tg_{ij}(s,Y^m_s)d\langle
B^i,B^j\rangle_s|\mathcal {F}_t],\ \ Y^0=0.$$ Then the rest of the proof
goes in a similar way as that in Theorem \ref{th 3.1}, and we omit
it.\hfill{}$\square$
\section{Appendix}
In this appendix, we give the proof of Lemma \ref{complex}. For simplicity, we assume that $g$ is bounded, and the unbounded case can be proved as in the existence part of Theorem \ref{th 3.5} with a series of truncation function $g^N$, for $N\geq 0$.\\\\
Let $J\in C^\infty_0(\mathbb{R}^n)$ be a nonegative function satisfying supp$(J)\subset B(0,1)$ and
$$
\int_{\mathbb{R}^n}J(x)dx=1,
$$
and for $\lambda>0$, put
$$J_\lambda(x)=\frac{1}{\lambda^n}J(\frac{x}{\lambda}),\ x\in \mathbb{R}^n,$$
and
$$g_\lambda(t, x)=\int_{\mathbb{R}^n} J_\lambda (x-y) g(t,y) dy,\ x\in\mathbb{R}^n, t\in [0,T].$$
By a classic analytic argument, $g_\lambda$ is uniformly Lipschitz. Then, for any
$X\in M^2_G([0,T];\mathbb{R}^n)$, we have $g_\lambda(\cdot, X_\cdot)\in M^2_G([0,T];\mathbb{R}^n)$. We only need to verify that $g(\cdot, X_\cdot)$ is the limit of $g_\lambda(\cdot, X_\cdot)$ in $M^2_G([0,T];\mathbb{R}^n)$.\\\\
For a fixed $\lambda>0$,
\begin{align*}
|g_\lambda(t, x)-g(t, x)|\leq\int_{\mathbb{R}^n}J_\lambda(y)|(g(t,x-y)-g(t,x))|dy.
\end{align*}
Then,
\begin{align*}
&\limsup_{\lambda\rightarrow+\infty}\int^T_0\bar{\mathbb{E}}[|g_\lambda(s, X_s)-g(s, X_s)|^2]ds\\
\leq&\limsup_{\lambda\rightarrow+\infty}\int^T_0\bar{\mathbb{E}}\bigg[\bigg|\int_{\mathbb{R}^n}J_\lambda(y)|g(s,X_s-y)-g(s,X_s)|dy\bigg|^2\bigg]ds\\
\leq&\limsup_{\lambda\rightarrow+\infty}\int^T_0\beta(s)\bigg(\int_{\mathbb{R}^n}J_\lambda(y)\gamma(y)dy\bigg)^2ds\\
\leq&\limsup_{\lambda\rightarrow+\infty}\gamma(\frac{1}{\lambda})\int^T_0\beta(s)ds=0,
\end{align*}
from which we deduce desired result.\hfill{}$\square$\\\\
\textbf{Acknowledgement} The authors express special thanks to Prof. Ying Hu, who provided both the initial inspiration for the work and useful suggestions.
|
\section{Introduction}
After more than a half-century of study, the basic physical
processes that are responsible for heating the million-degree
corona and accelerating the supersonic solar wind are now
beginning to be pinned down.
Different mechanisms are probably dominant for different regions
\citep[see reviews by][]{Mn00,HI02,Lg04,As06}.
For example, it seems increasingly clear that bright EUV and
X-ray loops are heated by small-scale, intermittent magnetic
reconnection that is driven by the continual stressing of their
magnetic footpoints \citep{Kl06,Gu07}.
For the open-field regions that link the corona and the solar wind,
there is still disagreement about the relative contributions of
different processes (see {\S}~4 below).
However, we are rapidly approaching a time when these processes
can be included in self-consistent models that can make testable
predictions.
This paper attempts to summarize some recent work that is helping
us to bring observations and theoretical models to the point of
straightforward comparison and testing.
Because the coronal magnetic field varies so substantially both
as a function of position (on the Sun at any one time) and as a
function of time (over the solar cycle), there is always a broad
range of solar wind source regions available for comparison with
model predictions.
In some ways, these variations give us something approaching
the turnable ``parameter knobs'' of a laboratory experiment.
This paper emphasizes the example of differences between the
previous solar minimum (1996--1997) and the present minimum
(2007--2009).
\section{Coronal Source Regions}
The most definitive link between a particular type of
coronal structure (measured via remote sensing) and a specific
type of quasi-steady solar wind flow (measured \emph{in situ}) is
the connection between large coronal holes and high-speed streams
\citep{Wx68,Ke73}.
Coronal holes are generally interpreted as bundles of open
flux tubes that flare out superradially with increasing distance.
Observations from the UVCS instrument on {\em SOHO} suggest
that the range of heights over which the wind's acceleration
occurs in coronal holes can vary greatly, even when the wind at
1 AU is identically fast \citep{Mi01,Mi06}.
The denser slow-speed solar wind appears to come from many
different coronal sources.
Two regions that are often cited as sources of slow wind are:
(1) boundaries between coronal holes and streamers, and
(2) narrow plasma ``stalks'' that extend out from the tops of
streamer cusps \citep[e.g.,][]{Ha97,Fi98,Wa00,St02}.
However, during active phases
of the solar cycle there is evidence that slow wind also
originates in small coronal holes and active regions
\citep{No76,Ng98,Lw04}.
The remainder of this paper will discuss the fast solar wind that
emerges from polar coronal holes at solar minimum \citep{Cr09}.
However, as the SOHO--23 meeting has demonstrated, not all solar
minima are created equal.
The morphology of the coronal magnetic field exhibited some
interesting differences from the previous minimum to the present
minimum.
Polar coronal holes on the disk in 2007--2009 are smaller in
area (by about 15\%) in comparison to those from 1996--1997,
and their mean photospheric magnetic fields are lower by about
40\% as well \citep[e.g.,][]{Kk09,Wa09}.
The magnetic field measured at 1 AU is also lower, but by only
about 20\% \citep{SB08}.
The streamer belt observed in the extended corona has a broader
latitudinal extent than it did in 1996--1997 \citep{Tk09}.
It is likely that this is the result of two contributing factors.
(1) The weaker polar field does not exert as much transverse
pressure, which acts to confine the streamer belt to low
latitudes \citep[see, e.g.,][]{Vz03}.
(2) There are more small and transient coronal holes at low latitudes
during the present minimum; these can also deform the streamer belt.
Some likely consequences of the above differences in the coronal
magnetic field are discussed below.
\section{A Sm\"{o}rg{\aa}sbord of Measurements}
Low-density coronal holes exhibit a complex array of plasma
parameters due to their nearly collisionless nature.
As a result, every particle species evolves towards having
its own unique temperature, its own type of departure from a
Maxwellian velocity distribution, and its own outflow speed.
Remote-sensing measurements of the low corona
(i.e., $r \approx 1$--1.3 $R_{\odot}$) and the extended corona
($r \approx 1.5$ to 10 $R_{\odot}$), as well as {\em in situ}
particle and field detection in the heliosphere
($r > 60 \, R_{\odot}$), can be combined to follow this evolution.
The extended corona is particularly important to study in this
regard, since it is not only where most of the wind's acceleration
occurs, but it is also where many plasma species undergo their
transition from collisional to collisionless dynamics
\citep[see][]{Ko06}.
\begin{figure}[!t]
\plotone{cranmer_fig1.eps}
\caption{Radial dependence of empirically derived temperatures
in polar coronal holes and fast wind streams. See text
for details.}
\end{figure}
As an example of how different types of measurements can help
paint a more complete picture, Figure 1 shows temperature
measurements in polar coronal holes from the last solar minimum
in 1996--1997.
The one-fluid temperatures shown at the lowest heights come from
semi-empirical \cite[dashed curve]{AL08}
and theoretical \citep[solid curve]{Cr07} models of the
photosphere, chromosphere, and low corona.
Electron temperatures in the low corona were measured by
SUMER/{\em{SOHO}} and reported by
\citet[black bars]{Wi06} and \citet[gray bars]{Ln08}.
UVCS-derived proton temperatures in the extended corona are also
shown \citep[triangles]{Cr09} with an attempt to remove the
model-dependent Alfv\'{e}n wave broadening.
Perpendicular temperatures of a representative minor ion
(O$^{5+}$) are shown from SUMER measurements at low heights
\citep[open circles]{LC09} and UVCS measurements in the extended
corona \citep[filled circles]{Cr08}.
{\em In situ} electron and proton temperatures in the fast wind
(dotted and dot-dashed curves, respectively)
are taken from \citet{Ce09}.
The detailed radial dependences of minor ion temperatures have
not yet been precisely measured {\em in situ,} but the gray region
above shows a likely range of values for the abundant O$^{6+}$ ion
\citep[see, e.g.,][]{Co96}.
Coronal holes tend to exhibit preferential ion heating
(i.e., $T_{\rm ion} \gg T_{p} \ga T_{e}$) that primarily
occurs in the direction perpendicular to the background magnetic
field ($T_{\perp} > T_{\parallel}$).
Because of these departures from thermal equilibrium, the
fast solar wind is an optimal ``proving ground'' for studies of
collisionless kinetic processes that many believe are the ultimate
dissipation mechanisms of coronal heating.
It was noticed several decades ago that the damping of ion
cyclotron resonant Alfv\'{e}n waves could naturally give rise to the
observed plasma properties \citep[see reviews by][]{HI02,Ko06,Cr09}.
However, many other dissipation processes have been proposed as
well, and they often involve multiple steps of energy conversion
between waves, reconnection structures, and other nonlinear
plasma features.
\section{Coronal Heating and Wind Acceleration}
Taking all of the above complexities into account and producing a
self-consistent model of coronal heating and solar wind acceleration
(for all particle species) has still not been accomplished.
However, if an assumption is made to consider only the {\em total}
energy content of the plasma---and not its ``partitioning'' into
protons, electrons, and other ions---then the problem becomes
more tractable.
At this level of detail, there are two general types of
physics-based model that attempt to explain the overall flows
of energy:
\begin{enumerate}
\item
There are {\em wave/turbulence-driven} (WTD) models in
which open magnetic flux tubes rooted to the photosphere are
jostled by convection, leading to waves that propagate up into
the corona.
These waves (usually Alfv\'{e}n waves) are often proposed to
partially reflect back down toward the Sun, develop into strong
MHD turbulence, and dissipate over a range of heights.
These models also tend to explain the differences between fast and
slow solar wind {\em not} by any major differences in the lower
boundary conditions, but instead as an outcome of different rates of
lateral flux-tube expansion over several solar radii
\citep{Ho86,WS91,Mt99,Of05,SI06}.
\item
There are {\em reconnection/loop-opening} (RLO) models
in which the flux tubes feeding the solar wind are influenced by
impulsive bursts of mass, momentum, and energy addition.
There is often assumed to be strong coupling between closed,
loop-like magnetic flux systems (that are in the process of emerging,
fragmenting, and diffusing across the surface) and the open flux
tubes that connect to the solar wind.
These models tend to explain the differences between fast and slow
solar wind as a result of qualitatively different rates of flux
emergence, reconnection, and coronal heating at the basal footpoints
of different regions on the Sun
\citep{AM92,Fi99,Fi03,SM03,SM08}.
\end{enumerate}
\citet{Cr07} presented a set of WTD models in which the
one-fluid equations of mass, momentum, and energy conservation
were solved simultaneously with transport equations for
Alfv\'{e}nic and acoustic wave energy.
The coronal heating rate was computed self-consistently from
a phenomenological description of turbulent dissipation of
partially reflected Alfv\'{e}n waves
\citep[see also][]{Mt99,Dm02,VV07}.
It is important to note that these models were run with lower
boundary conditions in the optically thick solar photosphere,
and {\em not} in the transition region (TR) or low corona.
Thus, the properties of the chromosphere, the height of the TR,
and the mass flux of plasma entering the solar wind were all
computed robustly and consistently in these models.
The \citet{Cr07} models of polar coronal holes produced conditions
at 1 AU that appeared consistent with fast solar wind
measurements at the 1996--1997 solar minimum.
More recently, we computed a new set of WTD models that were
designed to correspond to the 2007--2009 solar minimum.
The {\em only change} made to the models was in the radial
dependence of the imposed magnetic field strength.
The \citet{Cr07} magnetic field profile was multiplied by a
smooth function constructed to produce the following effects:
(1) no change to the $\sim$1 kG field strength in the intergranular
flux tubes in the photosphere;
(2) a 40\% reduction in the field strength at the base of the
corona ($r \approx 1.03 \, R_{\odot}$) to account for the
lower polar field strengths measured by low-resolution magnetograms;
and (3) an 18\% reduction in the field strength in interplanetary
space from one minimum to the next \citep[see, e.g.,][]{SB08}.
\begin{figure}[!t]
\plotone{cranmer_fig2.eps}
\caption{Relative changes in the input magnetic field
strength (thick solid curve) and the output solar wind speed
(dotted curve), density (dashed curve), and one-fluid temperature
(dot-dashed curve), found when comparing the polar coronal model
from 1996--1997 (``old'') to that computed for 2007--2009
(``new'').
The thin horizontal line denotes the level of zero change.
The large variations at $r \approx 1.01 \, R_{\odot}$ are due
to a slight mismatch of the two models in the height of the
sharp TR between the chromosphere and the corona.}
\end{figure}
Figure 2 shows the relative changes in the resulting WTD models
of polar coronal holes.
All quantities are ratios of the form
$(X_{\rm new} - X_{\rm old})/X_{\rm old}$, where the subscript
``old'' refers to the 1996--1997 minimum and ``new'' refers to
the 2007--2009 minimum.
The change in the input magnetic field strength, described above,
is shown as a solid line.
The other quantities are all self-consistent {\em outputs} of
the model.
The relative changes at 1 AU can be compared with measured
changes in the plasma parameters from \citet{Mc08} and
\citet{Eb09}.
Table 1 shows this comparison and confirms that the modeled
solar wind responds in a very similar way to the changing
magnetic field as does the actual solar wind.
In both the models and the {\em Ulysses} polar pass data, the
solar wind speed $u$ is relatively unchanged, but the density $n$
and temperature $T$ decrease by factors that hover around 20\%
and 10\%, respectively.
The decreases in gas pressure (proportional to $nT$) and
dynamic pressure (proportional to $n u^{2}$) are between 20\%
and 30\% for both the observations and models.
\begin{table}[!ht]
\caption{Relative changes in fast solar wind
from 1996--1997 to 2007--2009}
\smallskip
\begin{center}
{\small
\begin{tabular}{lcc}
\tableline
\noalign{\smallskip}
& {\em Ulysses} polar data & WTD model output \\
\noalign{\smallskip}
\tableline
\noalign{\smallskip}
speed & --03\% & +01\% \\
density & --17\% & --22\% \\
temperature & --14\% & --08\% \\
gas pressure & --28\% & --21\% \\
dynamic pressure & --22\% & --27\% \\
\noalign{\smallskip}
\tableline
\end{tabular}
}
\end{center}
\end{table}
Figure 2 also shows that the models predict the present solar
minimum should have a smaller temperature in the low corona
and a slightly higher density.
Nobeyama measurements of the radio ``brightness temperature''
(Yashiro et al., this meeting) do indicate a slightly lower
temperature in the upper chromosphere during 2007--2009, in
comparison to 1996--1997, but there are not yet any firm
comparisons from measurements in the low corona.
At the heights observable by UVCS, the predicted temperature
decline in Figure 2 is about 20\% and the predicted density
change has shifted from an increase to a slight decrease.
(At these heights we expect the proton, electron, and heavy
ion temperatures to be behaving differently from one another.)
Preliminary reports of minimum-to-minimum variations from UVCS
indicate that the H~I Ly$\alpha$ intensities are higher in
2007--2009 than in 1996--1997, and the O~VI intensities are lower
(Gardner et al., this meeting; Miralles et al., this meeting).
Both of these changes may be consistent with lower electron
temperatures and electron densities, but only detailed empirical
modeling of the spectral line properties will reveal whether that
is the case.
\section{Conclusions}
The {\em SOHO} era has seen significant progress toward identifying
and characterizing the physical processes responsible for coronal
heating and solar wind acceleration.
As remote-sensing measurements have become available in the
collisionless extended corona, the traditional gap between solar
physics and \emph{in situ} space physics has become narrower.
However, there are still many unanswered questions:
How and where in the solar atmosphere are the relevant waves and
turbulent motions generated?
Which kinds of fluctuation modes (i.e., linear or nonlinear;
Alfv\'{e}n, fast, or slow; high $k_{\parallel}$ or high
$k_{\perp}$) are most important?
What frequencies dominate the radially evolving power spectrum?
What fraction of the interplanetary solar wind comes from
filamentary structures such as coronal reconnection events
and/or plumes and jets?
Answering the above questions involves moving forward in both the
theoretical and observational directions.
A key step to making further progress is the ability to include
both the WTD and RLO processes in existing 3D numerical simulations
of the Sun-heliosphere system.
Some recent progress in producing computationally efficient
approximations to the rates of WTD wave reflection has been reported
by \citet{CH09} and \citet{Cr10}.
Studies of the connection between the evolving ``magnetic carpet''
and the open flux tubes that feed the solar wind are also ongoing.
The plasma parameters of both the major species (protons,
electrons, and He$^{2+}$) and minor ions are not yet known in
coronal holes to the accuracy required to determine the relative
contributions of the proposed physical processes.
As Figure 1 shows, even quantities as basic as the ratio
$T_{p}/T_{e}$ are not yet known with sufficient accuracy
because measurements of the proton and electron temperatures
have not yet been made over the same ranges of heights.
Minor ion measurements need to be extended to a larger number of
ions (i.e., a wider range of ion cyclotron frequencies) so that
the ultimate kinetic damping mechanisms of waves and turbulence can
be determined \citep[see, e.g.,][]{Cr02}.
Also, we do not yet have a good enough observational ``lower boundary
condition'' on the energetics of waves and turbulence in the photosphere.
Existing measurements of the Sun's convective granulation with
sub-arcsecond spatial resolution need to be matched by sub-second
{\em time resolution,} so that the power spectrum of the motions of
small-scale magnetic flux tubes (e.g., G-band bright points) can
be extended to higher frequencies.
\acknowledgements This work was supported by the National
Aeronautics and Space Administration (NASA) under grants
{NNG\-04\-GE77G} and {NNX\-09\-AB27G}
to the Smithsonian Astrophysical Observatory.
|
\section{Introduction}
Passive particles trajectories in a fluid flow provide an easier way of understanding the nature of the flow than a full-fledged study of the Eulerian field variables of the flow. Recent development in Lagrangian measurement techniques has made it possible to study relative motion of pairs of fluid elements in a turbulent flow \cite{Eberhard,Luthi07}. The understanding of the pair dispersion in turbulence is of importance for comprehending many natural processes such as transport and mixing in natural and engineering flows, formation of warm clouds {\it etc}.
\\
Suppose, $L$ is the integral scale and $\eta$ is Kolmogorov length scale in a turbulent flow, then one can classify the process of dispersion in the flow into three distinct regimes based on the separation of the particles relative to the turbulent scales: $(i)$ dissipation range corresponding to $r(t)\ll\eta$ ($r(t)$ is the separation of a pair of fluid elements or particles at time $t$), $(ii)$ inertial range corresponding to $\eta\ll r(t)\ll L$, and $(iii)$ diffusion range corresponding to $r(t)\gg L$.
From the work of Richardson \cite{Richardson} (and subsequently Obukhov \cite{Obukhov}), in the inertial range of fully-developed three-dimensional turbulence, there exits the well-known Richardson-Obukhov law (R-O law): $\left\langle r(t)^2\right\rangle = g \epsilon t^3$. Here, $\epsilon$ is the energy dissipation rate per unit mass, and $g$ is termed as the Richardson constant. Using similarity arguments, Batchelor \cite{Batchelor} extended the study to both short-time and intermediate-time, and obtained
\begin{equation}
\left\langle|\textbf{r}(t)^2 - \textbf{r}_{0}^2|\right\rangle = \begin{cases} \frac{11}{3} C_2 \epsilon^{2/3} r^{2/3}_0 t^2 & t\ll t_0\\ g \epsilon t^3 & t\gg t_0 \end{cases},\label{eq1}
\end{equation}
where $C_2$ is the Kolmogorov constant for the longitudinal second-order velocity structure function and
\begin{equation}
t_0 \equiv \left(\frac{r^{2}_{0}}{\epsilon}\right)^{1/3},\label{eq2}
\end{equation}
the classical correlation time of an eddy of size $r_0$, is the time for which a particle-pair ``remembers" their initial separation. $t_0$ is also called Batchelor time. In this paper, we investigated numerically this ``cross-over time" between the two scalings, $\sim t^2$ and $\sim t^3$, and study how it is related to the initial separation $r_0$. In what follows, we denote the first relation in equation ({\ref{eq1}}) as Batchelor's scaling law; the second one is obviously R-O law. Even though we shall be focused on three dimensional fully-developed turbulence, it is worth keeping in mind that R-O is omnipresent \cite{salazar}: there are reports of observing it even in two-dimensional, quasi-two-dimensional and rotating turbulent flows \cite{nicolleau}. Batchelor's law is no less robust--- it has been detected even in turbulent flows that are not perfectly homogeneous and isotropic \cite{Ouellette}.
\\
In the past decade, theoretical researchers have understood --- in significant details --- the process of advection of particles by developed turbulent flows in the framework of Kraichnan model\cite{r1,r2,r3,r4,r5,r6}.
Therein, calculations of the anomalous intermittent exponents for structure functions in the inertial range show that pair structure function, which determines the Richardson scaling exponent, is not intermittent.
\\
Until now, laboratory experiments have neither confirmed nor refuted the existence of R-O scaling. This, of course, is due to the fact that these experiments are unable to attain very high Reynolds number due to practical constraints.
Though environmental experiments can achieve the highest Reynolds numbers, the problem with such experiments are many--- inhomogeneity, anisotropy, rather uncontrollable experimental conditions, vastness of the domain to be studied, {\it etc.}.
Direct numerical simulation of Navier-Stokes equation (DNS), are much more successful in this respect since the computational power is on a rise, but cannot yet compete with the value of Reynolds numbers obtained in shell model simulations. Also, there are issues with the commonly adopted periodic boundary conditions in DNS as they may interfere with the homogeneity of the system.
We believe that shell models are well suited to test the R-O law as not only can one achieve very high Reynolds number but also there's no sweeping effect disturbing them. It is worth mentioning that sweeping effect's presence makes it hard to observe Kolmogorov scaling \cite{uriel} with which R-O law is so intimately related being the only choice for pair dispersion consistent with Kolmogorov scaling theory. A technique based on a real space transformations of shell models have recently been applied to study force fields study turbulent relative dispersion \cite{luiza} and we adopt a similar method here.
\section{The model}
To study how a pair of particles diffuse in a 3D turbulent flow, we
consider the kinematics of pair particles advected by the
homogeneous turbulent flow obtained by a real-space transformation
of the GOY shell model \cite{luiza,Jensen99,Bohr98}. This model proposed
originally by Gledzer, Yamada and Ohkitani~\cite{Gledzer73,Yamada87}
provides a description of the turbulent motion
embodied in the Navier-Stokes equations. The GOY model is formulated
on a $N$-discrete set of wavenumbers, $k_n=2^n$, with the associated
Fourier modes $u_n$ evolving according to
\begin{eqnarray}
\label{un}
\left(\frac{d}{ dt}+\nu k_n^2 \right) \ u_n \ & = &
i \,k_n \left(a_n \, u^*_{n+1} u^*_{n+2} \, + \, \frac{b_n}{2}
u^*_{n-1} u^*_{n+1} \, + \, \right.\nonumber \\
& & \left.\frac{c_n }{4} \, u^*_{n-1} u^*_{n-2}\right) \ + \ f \delta_{n,2},
\end{eqnarray}
for $n=1,\cdots, N$. The coefficients of the non-linear terms are
constrained by two conservation laws, namely the total energy, $E =
\sum_n |u_n|^2$, and the helicity (for 3D), $H =
\sum_n(-1)^{n}k_n|u_n|$, or the enstrophy (for 2D), $Z = \sum_n
k_n^2|u_n|^2$, in the inviscid limit, i.e.
$f=\nu=0$~\cite{Biferale95}. Therefore, they may be expressed in
terms of a free parameter only $\delta\in [0,2]$, $ a_n=1,\
b_{n+1}=-\delta,\ c_{n+2}=-(1-\delta)$. As observed by
Kadanoff~\cite{Kadanoff95}, one obtains the canonical value
$\delta = 1/2$, when the helicity is conserved. The set
(\ref{un}) of $N$-coupled ordinary differential equations can be
numerically integrated by standard techniques~\cite{Pisarenko93}. We
have used standard parameters in this paper $N = 27, \nu = 10^{-9},
k_0 = 0.05, f = 5 \times 10^{-3}(1+i)$.
The GOY model is defined in $k$-space but we study particle
dispersion in direct space obtained by an inverse Fourier
transform~\cite{Jensen99} of the form
\begin{equation}
\label{field}
{\vec v} ({\vec r},t)=\sum_{n=1}^{N}
{\vec c}_n [u_n(t) e^{i {\vec k}_n\cdot {\vec r}}
+ c.\, c.].
\end{equation}
Here the wavevectors are ${\vec k}_n ~=~ k_n {\vec e}_n$ where
${\vec e}_n$ is a unit vector in a random direction, for each shell
$n$ and ${\vec c}_n$ are unit vectors in random directions. We
ensure that the velocity field is incompressible, $\nabla\cdot{\vec
v} =0$, by constraining ${\vec c}_n \cdot {\vec e}_n =0,~\forall n
$. In our numerical computations we consider the vectors ${\vec
c}_n$ and ${\vec e}_n$ quenched in time but averaged over many
different realizations of these.
As an example of the motion in this field, Fig.~\ref{goy} shows the
trajectories of two passively advected particles. As the relative
distance diverges in time, the two particles experience different
force fields, which in turn typically increase the difference in the
relative velocities of the two particles. The figure shows the
individual particles as they are advected, first together and later
diverging away from each other when they are encased in different
eddies.
\begin{figure}[hbt]
\includegraphics[width=0.50\textwidth]{fig1.eps}
\caption{({\it Colour Online}) Two particles being advected in a random force
field, generated by the GOY shell model.
\label{goy}}
\end{figure}
\section{Results}
As a main result from our simulations, we find that both Batchelor's and R-O scalings are clearly observed, as shown in Fig. (\ref{fig2}). Please note that we have plotted $\left\langle|\textbf{r}(t) - \textbf{r}_{0}|^2\right\rangle$ and not $\left\langle|\textbf{r}(t)^2 - \textbf{r}_{0}^2|\right\rangle$ on the y-axis. This is to make contact with the experimental results \cite{Ouellette} which demonstrate that the correlation between the initial separation and the relative velocity of the pair may not be neglected when the flow is not perfectly homogeneous.
However, the GOY model--- by construction--- is a shell model of a perfectly homogeneous and isotropic turbulence. So, within the paradigm of shell models, probably it doesn't make much of a difference if one chooses to work with $\left\langle|\textbf{r}(t)^2 - \textbf{r}_{0}^2|\right\rangle$.\\
\begin{figure}[hbt]
\includegraphics[width=0.50\textwidth]{fig2.eps}
\caption{({\it Colour online}) The initial separation $r_0$ is $10^{-4}$ and the data are averaged over 10000 different simulations. One can distinctly note the presence of both the R-O law and the Batchelor's scaling law. The cross-over time $t_0$ is somewhere in the third decade on the x-axis. The deviation of the blue curve (marked $\left\langle|\textbf{r}(t) - \textbf{r}_{0}|^2\right\rangle/t^3$) from R-O scaling at higher times is due to the fact that the pair-separations are no longer in the inertial regime. All the units are in GOY-shell-model units.}\label{fig2}
\end{figure}
\begin{figure}[hbt]
\includegraphics[width=0.50\textwidth]{fig3.eps}
\caption{({\it Colour online}) The cross-over times $t_0$ plotted the initial pair-separations $r_0$. The dashed line has the theoretical slope predicted by Batchelor. All the units are in GOY-shell-model units.}\label{fig3}
\end{figure}
With no knowledge of the exact functional dependence of $\left\langle|\textbf{r}(t) - \textbf{r}_{0}|^2\right\rangle$ on various parameters of the flow, it is quite a tough task to exactly locate the cross over time, $t_0$, between the Batchelor regime and the R-O regime. However, we can use the following rather crude but effective method to see whether equation (\ref{eq2}) is validated by our simulations.
To investigate where the crossover happens, we estimated the slope along the blue curve, $\left\langle|\textbf{r}(t) - \textbf{r}_{0}|^2\right\rangle/t^3$ in Fig. (\ref{fig2}). Obviously, when the slope is close to zero, we are in R-O scaling regime. As the crossover occurs before the slope is zero, we decide on the following algorithm: determine the time corresponding to the point where the slope is equal to $-0.5$\footnote{It may be mentioned that while experimenting with this approximate method of locating crossover time, we noted that results look more or less same for other values for ``cut-off" slope around $-0.5$; thus we report herein results only for the cut-off slope$=-0.5$.} and simply denote that $t_0$; repeat the procedure for the curves with various different $r_0$. We have used the data generated this way to obtain Fig. (\ref{fig3}) where we plot the cross-over times $t_0$ versus the initial pair-separations $r_0$. One can note the reasonable agreement of the data with the theoretical Batchelor prediction (the dashed line) when the initial separation
is not too small. What has been shown is that our data follow the expected trend; of course, data points are scattered around the theoretical slope.
The deviation at smaller initial separations --- between $r_0=10^{-3} \textrm{ to } r_0=5 \times 10^{-3}$ --- probably has the following explanation.
In the case of our numerics, the Kolmogorov scale, $\eta$, is of the order of $10^{-5}$.
The Reynolds number, $\textrm{Re}\sim 1/\nu$, is of the order of $10^{9}$.
One may thus note that the Taylor microscale, $\lambda\sim \textrm{Re}^{1/4}\eta$ is more than $100$ times greater than the Kolmogorov dissipation scale.
This leads to the fact that in our model $\lambda\sim 10^{-3}$, implying that the data points in the flat region of Fig. (\ref{fig3}) do not quite correspond to the initial separations which are {\textit {deep}} inside the inertial regime; and hence, it is not surprising that they do not follow the Batchelor prediction.
Nevertheless, we feel it is quite remarkable that the GOY shell-model of turbulence can reproduce this law very well for larger separations.
Similar result is obtained if one repeats the aforementioned algorithm for the red curve,$\left\langle|\textbf{r}(t) - \textbf{r}_{0}|^2\right\rangle/t^2$,
in Fig. (\ref{fig2}) focusing on the Batchelor's scaling regime.
\\
Another way to find the relation between the $r_0$ and $t_0$ comes from the fact that R-O scaling is $\left\langle r^2\right\rangle = g \epsilon t^3$ while Batchelor considered $\left\langle|\textbf{r}(t)^2 - \textbf{r}_{0}^2|\right\rangle$. So, if we plot $\left\langle r^2\right\rangle$ and $\left\langle|\textbf{r}(t)^2 - \textbf{r}_{0}^2|\right\rangle$ versus time, they will become indistinguishable when R-O scaling sets in and this will happen around $t_0$ for the corresponding $r_0$. Thus, one can conveniently devise an algorithm to find $t_0$ in this case. We may mention that also from applying this indirect method, Eq. (\ref{eq2}) stands validated.
\\
In the passing, it may be mentioned that within the framework adopted in this paper, one can also study exit-time statistics\cite{boffetta1,boffetta2} of the pair dispersion phenomenon to good effect. The corresponding results will be reported elsewhere.
\section{Conclusion}
We have studied pair dispersion in a turbulent flow, applying the GOY shell model (Eq. (\ref{un})) and Fourier transforming the complex shell-velocities back into real space (Eq. (\ref{field})). This procedure results in strongly turbulent velocity field where the dispersion of pair of particles can be easily studied by advecting the passive particles in the velocity field. In particular, we have investigated how the dispersion depends on the initial separations.\\
We may mention that this is probably for the first time that a shell-model of turbulence has shown the simultaneous existence of both the R-O and the Batchelor regimes in turbulent pair dispersion.
Whereas devising a better method of finding $t_0$ would be worth investigating, the temporal extent over which these laws are valid can easily be increased if one uses a larger number of shells.\\
In closing, we hope that our results may spark interest in implementing GOY model to study pair dispersion in the enstrophy cascade dominated regime in two-dimensional flows. Also, it might be fruitful to use GOY model modified appropriately \cite{Sagar4} for investigating turbulent relative dispersion in more realistic case of rotating turbulence which is of profound geophysical and astrophysical interest.
\begin{acknowledgements}
We wish to thank Prof. R. Benzi for pointing out the importance of Taylor microscale in explaining our numerical results.
\end{acknowledgements}
|
\section{Introduction}
In past decades, we have witnessed a rapid growth in the number of
radio selected Active Galactic Nuclei (AGNs) resulting from large
and deep radio surveys such as the Faint Images of Radio Sky at
Twenty centimeters (FIRST, Becker, White \& Helfand 1995) and the
NRAO VLA Sky Survey (NVSS, Condon et al. 1998) coupled with optical
spectroscopy follow-ups or dedicated spectroscopic surveys such as
the Two degree Field (2dF, Maddox et al. 1998; Boyle et al. 2001)
and the Sloan Digital Sky Survey (SDSS, York et al. 2000; Stoughton
et al. 2002). However, some fundamental issues regarding the origin
of radio emission remain hotly debated: is the radio loudness
dichotomy true or not (Kellermann et al. 1989; Miller, Peacock \&
Mead 1990; Hewett et al. 2001; Ivezic et al. 2002)? Are radio jets
in radio quiet/intermediate quasars relativistic (Readhead et al.
1988; Wilson \& Colbert 1995; Meier et al. 2001)? Which physical
parameters control the large range of radio strength despite their
great similarity in the SED at other wavelengths for various quasars
(Barthel 1989; Urry \& Padovani 1995; Jackson \& Wall 1999; Boroson
2002; Aars et al. 2005)?
Concerning the first question, the ambiguity is caused largely by
various selection biases introduced by the survey limits, the
incompleteness in the radio sample due to their complex morphologies
or in the optical sample due to various color selections (Ivezic et
al. 2002; Cirasuolo et al. 2003a; Best et al. 2005). The traditional
radio surveys were carried out at shallow flux density limits (0.1-1
Jy) and primarily radio-loud quasars were detected (e.g., Bennett
1962; Smith \& Spinrad 1976; Colla et al. 1972; Fanti et al. 1974;
Large et al. 1981). Only the two latest radio surveys NVSS and FIRST
have enough sensitivity and positional accuracy to allow for the
detection of large number of radio intermediate and radio quiet
quasars in conjunction with moderately deep, large area optical
surveys such as 2dF and SDSS. However, by taking advantage of the
accuracy in the position of radio sources, most authors constructed
the radio quasar or quasar candidate sample using solely the
position match between radio and optical sources (Gregg et al. 1996;
White 1999; Lacy \& Ridgway 2001; McMahon et al 2002; Richards
et.al. 2002; Cirasuolo et al.2003b). This process introduces a bias
against lobe-dominated radio quasars: either they were missed or
their radio flux underestimated.
White, Becker \& Gregg (1999, 2000) argued that this incompleteness
is not severe in FIRST Bright Quasar Survey (FBQS), which was
selected by matching of optical counterparts within a $1\farcs2$
position offset to the radio sources in the FIRST catalog. Using a
matching radius of $2^{''}$, Ivezic et al. (2002) estimated that
less than $10\%$ SDSS-FIRST associations have complex radio
morphology, and core-lobe and double-lobe sources together represent
about only $5\%$ in the radio quasars and galaxy sample. Using a
novel technique, de Vries et al. (2006) constructed a Fanaroff-Riley
type 2 (FR-II) quasar sample, and found that $27\%$ FR-II quasars do
not show cores at the FIRST flux limit. These authors also compared
the emission line properties and optical colors of these FR II
quasars with radio quiet quasars. It should be noted that these
missing quasars are not random, but are all extended sources and
tend to be more radio loud (Falcke et al 1996; Ivezic et al. 2002;
Best et al 2005). As a result, the statistical properties of the
sample, such as radio loudness and radio luminosity distribution,
will be affected by this selection effect.
In this paper we study in detail of selection effects in the SDSS
radio quasar samples. We identified from SDSS and FIRST a large
sample of radio quasars, including those with complex radio
morphology. Besides using positional coincidence as a primary
selection criterion, we manually examined the FIRST images for all
of the candidates with extended radio morphology. Through this less
efficient process, we obtained a sample of $3641$ spectroscopically
confirmed quasars with secure radio identification. A detail
comparison of this sample with other radio quasars sample is given.
Various selection effects are quantified. Throughout this paper, we
will adopt a concordant cosmology with $H_0 = 70$
km~s$^{-1}$~Mpc$^{-1}$, $\Omega_m = 0.3$, and $\Omega_{\Lambda}=
0.7$.
\section{The Radio Quasar Sample}
\subsection{Optical data}
Our starting point is the SDSS quasar catalog constructed by Schneider
et al (2005). The catalog consists of quasars that contain at
least one broad emission line ($FWHM \gtrsim 1000$ km~s$^{-1}$
), or are unambiguously broad absorption line quasars. These quasars
selected from SDSS photometric catalog either by their colors or for
their positional coincidence with radio sources in the FIRST catalog
(within 2\texttt{"} of a FIRST source) or ROSAT X-ray sources
(within 10-20\texttt{"} of a ROSAT source), were spectroscopically
confirmed to meet above criteria and their absolute optical
magnitudes at $i$-band $M_i\le -22.0$. The sample also includes some
supplementary quasars that meet the above criteria, but were
selected initially as galaxy targets (Eisenstein et al. 2001;
Strauss et al. 2002). Note that the magnitude limits for various
candidates are different, $i<19.1$ for color selected low $z$
($z<3.0$) quasars, $i<20.2$ for color-selected high $z$ quasars,
$i<19.1$ for FIRST and ROSAT sources, and $i<17.7$ for the main
galaxy sample. The final sample contains $46,420$ quasars in the
redshift range $0.078 \le z \le 5.414$, absolute magnitude range
$-30.2 \le M_{i} \le -22.0$, and $i$-band optical magnitude range
$15.10\le i \le 21.78$.
We use a subsample of this catalog in order to minimize the bias
introduced in the selection of quasar candidates. The SDSS sources
with their target-flags as ``QSO\_FIRST\_CAP'' or
``QSO\_FIRST\_SKIRT'' only (hereafter FIRST-only sources) will bias
against the lobe-dominated radio sources, and are redder than the
color-selected quasars (Richards et al. 2002). Therefore, they will
be excluded from the sample. Since radio and X-ray emission from
quasars are well-correlated (Shastri et al. 1993, Brinkmann et al.
2000; Padovani et al. 2003), the ROSAT-only selected quasars will be
biased toward radio strong sources, and as such are excluded from
the sub-sample.
\subsection{Radio data}
Radio counterparts to SDSS quasars are found using the FIRST survey.
The survey covers about 10,000 $deg^2$ and is $95$\% complete to 2
mJy, $80$\% complete to 1mJy (Becker et al 1995). The source surface
density in this survey is $\sim90$ deg$^{-2}$. At the 1 mJy source
detection threshold, the resolution is better than $5^{''}$.
Individual sources down to 1 mJy threshold have $90$\% confidence
error circles with radii of $\le 1^{''}$.
The FIRST catalog was produced by fitting a two-dimensional Gaussian
to each source to generate major axes, minor axes, peak and
integrated flux densities from the co-added images. The major axes
have been deconvolved to remove blurring by the elliptical Guassian
fitting. For bright sources, the size was determined down to $\sim
1/3$ of the beam size $5\farcs4$. The FIRST survey also provides
clean radio images.
\subsection{Compact radio quasars}
Most radio quasars are compact sources in the FIRST images and as
such, they can be identified through cross-correlation of the FIRST
catalog with the SDSS quasar catalog by adopting a small matching
radius in the position offset.
The matching radius is a trade-off between the completeness and
random association, i.e., the higher completeness necessarily
implies higher random contamination. Knapp et al. (2002) showed that
the random association increases with matching radius at $\leqslant
2\farcs5$. Magliocchetti \& Maddox (2002) found that $2^{''}$
matching radius could include $\sim97\%$ of the true matches above 1
mJy level. Gregg \& Becker (1996) estimated that more than $95\%$
FBQS-I quasars with magnitude $E\leqslant 17.8$ are within
$1\farcs1$ offset between the POSS-I and FIRST positions, while
White, Becker et al. (1999) showed that $\sim 1^{''}-1\farcs1$
matching radius will eliminate most of false quasar candidates at
the expense of only $5\%$ of incompleteness, and the fraction of
optical candidates found to be quasars declines steadily from $80\%$
near $0^{''}$ offset to $\sim 20\%$ at $1\farcs2$, and is constant
farther out. Ivezic et al. (2002) estimated that $1^{''}$ matching
radius produces $72\%$ completeness and $1.5\%$ random
contamination, and $1\farcs5$ matching radius with $85\%$
completeness and a contamination of 3\%. For their photometric SDSS
quasar sample, they estimate $3^{''}$ matching radius will cover
almost all counterpart but with 9\% random contamination.
Following Richards et al. (2002), we used a matching radius of
$2^{''}$ for compact FIRST sources and obtained 2782 matches with a
false rate\footnote{The fraction of chance coincidence is estimated
as $\rho \pi r^{2} N_{s}/N_{m}$, where $\rho $ is the surface
density of FIRST sources, $r$ the matching radius, $N_{s}$ the
number of quasars in the catalog of SDSS DR3, and $N_{m}$ the number
of matches.} of $\sim 0.15\%$. In addition, we found 71 SDSS quasars
locate within the ellipses of FIRST sources with core-jet or diffuse
structure, but with their optical-radio offset larger than the
$2^{''}$ matching radius. We added these quasars to our sample after
visual confirmation of the true association.
\subsection{Extended radio quasars}
The two point angular correlation function (Cress et al. 1995) can
be used to define an appropriate scale for searching for radio
matches with complex radio morphologies. The correlation function in
$0.02-2^{\circ}$ can be well fitted by a power law $A \times
{\theta}^{\gamma}$, where the angle $\theta$ in units of degree,
$A\sim3 \times 10^{-3}$ and ${\gamma}\sim-1.1$. At the ${\theta}
\sim 0.1^{\circ}$ ($6^{\prime}$), it drops to a value of only 0.038
which means that double and multi-component FIRST sources are shown
to have a little clustering amplitude beyond that angular limit,
i.e., intrinsic correlated double and multi components mostly
clustered under that scale. We also note that the physical size for
the $6^{\prime}$ angle at z=0.05 is $\sim 353$ kpc, approximately to
the scale of most radio jet and lobe ($\sim100$~kpc, Readhead et al.
1988; Jackson 1999, krolik 1999 ). All quasars in the SDSS DR3 have
redshift above 0.05. So we use the $6^{\prime}\times6^{\prime}$
FIRST map surrounding the quasar to determine possible complex
structure.
The extended radio quasars are extracted in two steps. First, the
candidates of radio quasars are selected with one of the following
two simple criteria: (1) two radio sources locate nearly
symmetrically around the quasar position, i.e., the angle between
optical-radio connections lies $150^{\circ} \lesssim \theta \lesssim
210^{\circ}$, and the ratio of their distances to the quasar is $1/3
\lesssim d_1/d_2 \lesssim 3$; (2) more than two radio sources
scatter around the SDSS quasars. The first criterion allows us to
detect radio sources with symmetric lobes, either of FR-II or FR-I
type (Fanaroff \& Riley 1974), which is similar to that used by de
Vries et al. (2006). The FR-I sources are core-dominated sources
with a bright nucleus and two extended lobes with the surface
brightness decreasing towards the edges. In contrast, the FR-II
sources are generally lobe-dominated sources, and always show
brighter lobes, usually with a hot spot, and with or without a weak
core (see Fig. \ref{FIRST_image}). However, when viewed at an
extreme angle, the FR-II sources may be dominated by the brighter
core (Barthel 1989; Hoekstra et al. 1997; Hardcastle et al.1998). By
the first criterion, sources with distorted asymmetric lobes may be
missed. The second criterion is designed for more complex radio
morphologies, such as sources with distorted asymmetric lobes and a
compact core or for cases in which extended lobes are resolved into
complex structure in the FIRST image. (see Fig. \ref{FIRST_image})
With these criteria, we selected 3115 radio quasar candidates. Among
them, 1035 sources are selected by the first criterion, and 2080 by
the second one. A $6^{\prime} \times6^{\prime}$ cutout of FIRST
image\footnote
{$http://archive.stsci.edu/vlafirst/getting\_started.html$} centered
at the quasar candidate was extracted for each SDSS source, and
visually checked. We use the radio morphologies in the 3CR radio
sources \footnote{$http:/www.jb.man.ac.uk/atlas/sample.html $} as
the reference for the true matches. We found about 70\% of these
radio components in $6^{\prime} \times6^{\prime}$ cutouts are likely
not related to the SDSS quasars, i.e., they are isolated components
or radio components related to other SDSS sources, or they have no
convincing evidence for their connection to the quasar.
They will be excluded either from the radio flux estimation
or from the sample.
In the end, 859 extended sources are selected (with unambiguous
radio lobes, jets) from initial 3115 candidates. Among them, about
half (409) show FR-II morphologies, and 564 have radio core
components (within 2\texttt{"} circle of optical quasar position) in
the FIRST catalog. In comparison, de Vries et al. found 422 FR-II
quasars using Abazajian et al. (2005) DR3 quasar sample (44984
sources),
of which 359 are in common. The members of radio counterparts in
the de Vries sample that are not included in our sample are 63. Most
these ``lost'' objects (41 out of 63) are not included in the SDSS
DR3 quasar catalog of Schneider et al. (2005). The other 18 were
excluded by us due to their "FIRST-only" or "ROSAT-only"
target-flag. The remaining 4 were excluded by us for their
inconvincible connection between the radio and optical sources. We
also notice that, for 83 of the 359 common quasars in both sample
that show more complex radio structure than simply "double-lobes",
i.e., the lobes resolved in the FIRST image, we account for the
radio fluxes of all components. As a comparison, Vries et al. (2006)
only considered the radio fluxes of the two components located
symmetrically around the quasar, thus the total radio fluxes are
systematically underestimated for these sources. The flux difference
between the two ranges from $2\%$ to $86\%$ , with a typical value
of $31.2\%$.
The maximum projected physical distance of the lobe component is
independent of the radio power (see left panel of Fig.
\ref{lobe-distrib}) while the rest-frame peak intensity increases
with radio power as shown in right panel of Fig. \ref{lobe-distrib}.
The former is expected if the radio power does not vary significant
during the quasar activity phase while the lobe to push further
outward. The latter may result from the interaction of powerful
radio jet with medium, which will produce more relativistic
electrons, thus high intrinsic brightness.
By merging the compact and extended radio quasar sample after
eliminating duplicate sources, we obtain a sample of 3641 radio
quasars. Among them, 859 are multi-component sources (extended
sources, hereafter) and the remain 2782 are single component sources
(hereafter compact sources). The average apparent magnitude of this
sample is $m_{i} \sim 18.78$, the median redshift is $\sim 1.36$
(see Fig.\ref{redshift}). The distributions of optical selected and
optical+radio selected sub-sample are significant difference. As
noticed by Richards et al. (2006), radio-selected and
optically-selected quasars show different redshift distribution in
the sense that the optical selected quasars show deeper deficit at
$z\sim 2.7$, while the optical+radio quasars distribute smoother at
$z>2$. This is due to the selection effect of the optically-selected
SDSS quasars based on optical colors. We can also see that radio
quasars peak at a slightly lower redshift and are more abundant
between redshifts of 2.2 and 3.0. This may be due to the
comparatively shallower survey depth of FIRST, so that we can find
more radio counterparts of optical-selected quasars at lower
redshift.
Note that the redshift distribution of extended sources peaks at
lower redshift than the compact quasars, i.e., the fraction of
extended sources decreases with redshift. The median radio flux is
5.49 mJy for compact quasars and 48.31 mJy for extended quasars.
\section{On the selection effect of radio quasar sample}
\subsection{The incompleteness of extended sources caused by the surface
brightness limit}
FIRST is not sensitive to low surface brightness emission due to its small
beam size. As such diffuse emission, if present, will be missed in the FIRST
survey. It should be noted that the surface brightness are lowered due to
cosmological expansion ($I\propto(1+z)^{-4-\alpha_r}$), as such lobes may
fall below the detection limit (0.75 mJy per beam) of the FIRST survey at
large redshift (see below). In the worst case, if the source is dominated
by a diffuse component, they may escape the detection at all in the FIRST
survey. We use cross correlation between SDSS and NVSS to constrain this.
The 1.4GHz NRAO VLA Sky Survey (NVSS; Condon etal 1998 ) provided a
catalog contain $2\times 10^6 $ sources stronger than $2.5$ mJy. It
is $90$ percent complete at integrated flux density $S_{1.4GHz}=3$
mJy and $99$ percent complete at $S_{1.4GHz}=3.5$ mJy. With a
synthesized beam of $45^{''}$ (FWHM), NVSS is much more sensitive to
lower surface brightness components than FIRST. Therefore, it can be
used to check the fraction of quasars with only diffuse emission.
For typical redshift of quasars in this sample, the radio emission
should be unresolved by NVSS.
In the following we will consider radio sources with NVSS flux
density larger than $3$ mJy. The matching radius used in the
cross-correlation of DR3 quasars and NVSS sources is trade-off
between random contamination and completeness. First, we estimate
that the fractions of chance coincidence corresponding to $15^{''}$,
$20^{''}$, $25^{''}$ matching radius are $\sim 3.95$\%,
$\sim6.68\%$, $\sim9.64\%$ respectively. Second, we checked 775
extended sources with FIRST flux above 3~mJy in our FIRST-DR3 quasar
sample. Among them, NVSS counterparts are found for 719 quasars ($
\sim 92.8$\%) within $20^{''}$ offset of the quasars. The fraction
decreases to $87.65\%$ with a $15^{''}$ matching radius, and
increase to $95.6\%$ with $25^{''}$ matching radius. We take
$20^{''}$ as the matching radius for the NVSS-SDSS match.
With $20^{''}$ matching radius, we extracted $3029$ NVSS radio
counterparts to SDSS DR3 quasars in the overlapping area of FIRST,
SDSS DR3 and NVSS, with the NVSS flux density limit of 3 mJy. In
this DR3-NVSS sample, $227$ sources are not present in our DR3-FIRST
radio quasar sample. $208$ of these "lost" sources
locate at a distance larger than the NVSS major axis from the quasar.
We found the NVSS-SDSS offset distribution of these "lost" sources
are significantly different from the "found" ones
(see the bottom panel of Fig. \ref{NVSS-FIRST_flux}),
indicating most of them are due to chance coincidence.
This number is close to the expected rate of chance coincidences
($\sim 6.68\%$). Note in passing four sources are found in our
DR3-FIRST sample, but their NVSS flux density are less than 3 mJy.
Therefore, no more than 19 extended radio quasars may be missed due
to the higher FIRST resolution at the NVSS flux density limit of 3
mJy. This lost fraction is $\sim 19/(775+19)\sim2.3\%$. Therefore,
above 3 mJy flux density limit, the radio quasars with only diffuse
radio sources are rare.
However, weak diffuse emission is likely to be overlooked by FIRST.
We plotted the distribution of $k=\log(f_{NVSS}/f_{FIRST})$ at FIRST
flux $>$ 3~mJy and $<$ 3~mJy (left panel of
Fig.\ref{NVSS-FIRST_flux}), where $f_{FIRST}$ is the sum of all
radio sources within the NVSS beam size. The $k$ is peaked at zero
with a tail toward $<k>\sim 0.04$ above 3 mJy, which suggests that
most of the NVSS flux had been detected by the FIRST. But below 3
mJy limit, $<k>$ increased to 0.24. This indicates that some diffuse
radio flux such as weak lobes of extended sources can be
underestimated by the FIRST as flux density decreases. As a result,
some extended sources with low surface brightness may be
mis-identified as compact sources. This may explain why most of the
extended sources have flux greater than 3 ~mJy (see the right panel
of Fig.\ref{NVSS-FIRST_flux}).
\subsection{The selection effects of a radio quasar sample}
In this section, we will address the selection effects introduced by
the positional coincidence alone using our radio quasar sample. This
includes the lost fraction of lobe-only objects and the
underestimate of the extended flux. In the following, we will divide
the radio quasars into extended and compact sources according to
whether one or more extra-core components are present or not. The
core component is defined as the single component within the
$2^{''}$ radius of the optical quasar.
The distribution of position offsets of the closest radio component
to the SDSS quasar is shown in the left panel of
Fig.\ref{obs-e-c-chars} for our radio quasar sample. We find that
13.0\% of quasars will be lost with a matching radius in the
position offset of $1\farcs2$, 10.4\% at $1\farcs5$ and $8.1\%$ at
$2^{''}$. With these numbers, we conclude that the fraction of radio
quasars missed due to lack of detectable radio cores is low.
Using the positional coincidence will underestimate the radio flux
density if there is one or more off-core components even if a radio
core is present. This will be important particularly in
lobe-dominated quasars. To quantify this bias, we calculate $q$ as
the flux ratio of the core component to the total radio flux (the
summation over all radio counterparts that are associated with the
radio quasar). The distribution of $q$ for 564 quasars with
core-lobe structure and 295 objects without cores ($\sim 35\%$ of
all extended sources) are shown in Fig.\ref{obs-e-c-chars}. It is
interesting to note that the distribution of $q$ keep nearly
constant between $0 < q < 1.0$, i.e., the number of strong-lobe
sources ($q < 0.5$) and weak-lobe sources ($q > 0.5$) are similar.
We found that the radio fluxes are either underestimated by a factor
more than two or complete missed in about 16.8\% of all radio
quasars. The sources with two or more components are significantly
more numerous than estimated by Ivezic et al. (2002).
Note that with our definitions whether a quasar is compact or
extended depends also on the redshift of the quasar.
Fig.\ref{redshift} shows that the fraction of extended sources peaks
at redshift $\sim 1.0$ and decreases dramatically at redshift
$\gtrsim 2.0$. Note that this is not due to the increase of the
angular distance because it is peaked around 1.5 for the
cosmological model adopt here, and most extended sources at redshift
less than 0.5 have sizes that should be resolved at redshift around
2-4 ( Fig.\ref{lobe-z}). This can be due to either an evolution in
the radio structure of quasars or the decrease of brightness caused
by cosmology expansion ($I\propto(1+z)^{-4-\alpha_r}$)
(Fig.\ref{lobe-z}).
We show the optical color distributions in Fig. \ref{color}. In
order to eliminate the dependence on redshift, we divide our sample
into 20 redshift bins. For each bin, we subtract from $g-i$ the
median value of $(g-i)_{median}$ in that bin. Surprisingly, the
color distribution of extended radio quasars (with and without core)
does not appear redder, but slightly bluer than the compact radio
quasars. Kolmogorov-Smirnoff test deems the significant difference
between the color distribution of extended and compact sources at
the confidence level $\gtrsim 97\%$. Also, student-t test suggests
that the mean color of the extended radio quasars is bluer than that
of the compact quasars at a confidence level of $\gtrsim 99\%$.
The result remains the same for the core-only and lobe-only sources
and for quasars with unresolved and resolved cores (resolved core:
log$ (f_{int}/f_{peak})^2>0.1$, with $f_{int}>3.6$ mJy in order to
gain sufficient S/N ratio, (Ivezic et al. 2002). (see
Fig.\ref{color} ). But, the core-lobe sources and the only-lobe
sources are indistinguishable. These results are in line with that
of de Vries et al.(2006), who found that the composite spectrum of
FR-II quasars is flatter than that of radio compact quasars.
Since most quasars have redshift of $z\sim 2$, we calculate the
$k$-corrected UV flux at $2500\mathring{A}$ at the rest frame of
quasars by interpolating or extrapolating the SDSS five apparent
magnitudes using a $Spline$ function. Extrapolation is required for
only a small number of quasars at low redshift ($z<0.5$). The radio
loudness is defined as the flux ratio of $k$-corrected radio flux at
20~cm and the UV flux at $2500~\mathring{A}$. An average of radio
spectral index $\alpha_r=-0.5$ for quasars is assumed. We plot the
average UV absolute magnitudes versus redshift in left panel of the
Fig.\ref{R_O_lum}. We find that the average UV absolute magnitudes
of extended and compact sources are similar, as they are for
resolved and unresolved sources.
Next, we calculated the $k$-corrected radio power at 20 cm as
$P_{radio}=4 \pi D^{2}_{L} f_{int}/ (1+z)^{1+\alpha _{r}}$, where
the radio spectral index $\alpha_{r}$ is assume to be $-0.5$ for all
objects. The radio power and the radio loudness distribution is
shown in Fig.\ref{P-R_distribution}, and the radio power as a
function $z$ is shown in right panel of Fig.\ref{R_O_lum}.
Evidently, extended radio quasars are much more powerful in the
radio than compact radio quasars despite their similar optical
luminosity.
The difference in the radio power between the extended and compact
sources decreases with increasing redshift, a factor of more than 10
at redshifts less than 0.5 to a factor $\simeq$2 or so at redshift
larger than 2 (Fig.\ref{R_O_lum}). This might be caused by a
combination of the survey limit, with which only powerful radio
sources can be detected at high $z$, and an increase in the
detecting limit of intrinsic brightness for the extended lobes at
high redshift, i.e., only very bright lobes are detected, and thus
only very powerful FR II sources. At redshift less than 0.5, the
radio power of compact radio quasars is close to the border of
FR-I/FR-II division.
It was proposed that strong radio emission from extended quasars may
be enhanced by interaction of powerful radio jets with the
interstellar medium (Bridle et al. 1995; Wills \& Brotherton 1995).
As such the difference in the radio power of core-dominated and
extended quasars is due to their different environment, rather than
their different central engine. To check this point, we compare the
core radio power for those quasars with detected cores, and found
that extended quasars have more powerful cores. Therefore,
our result does not support this interpretation.
The radio and optical flux limits introduce another selection effect
on the radio loudness distribution of quasars. At a given optical
magnitude limit, only quasars with their radio loudness above
certain limit can be detected in the FIRST survey due its flux
limit, i.e., the sample is complete above certain radio loudness.
Using a similar strip in the log$(RL)-i$ plane as Ivezic et al.
(2004), we estimate the conditional radio loudness distribution
under different photometric limit. As shown in Fig.\ref{cond_prob},
the distribution of log$(RL)$ for compact sources peak at $\sim 2$
for $19<i<21$, consistent with Ivezic et al. (2004). Adding the core
flux of extended sources, the radio loud peak becomes more
significant with the amplitude of the dip at log$(RL)\sim2$
increasing from 24\% to 37\% due to their contribution to large
radio loudness portion. As shown in Fig.\ref{cond_prob}, the
distribution for extended quasars keeps to peak at log$(RL)\gtrsim
2.7$ at different i-magnitude bins, while the compact sources shown
peak at log$(RL)\lesssim 2.3$. Furthermore, the compact sources
distribution peaks at lower radio loudness when i decreased.
By plotting the radio loudness under different redshift, we find
that the peak of the compact radio quasars moves to small RL as
redshift decreases, while the distribution of the extended radio
quasars remains the same (peaked at $\sim 3$) for all redshift (see
Fig.\ref{e-c-rl2500-z}) This radio loudness distribution of the
extended (also more radio loud) radio quasars is consistent with
Cirasuolo et al. (2003b), who modeled the radio loudness
distribution with double-Gaussian function by fitting the FIRST
selected 2dF quasars, and found that the intrinsic radio loudness of
radio quasars peaks at log$(RL)=2.7\pm 0.2$ and log$(RL)=-0.5\pm
0.3$.
\section{Conclusion and Discussion}
We have constructed a relatively unbiased large radio quasar sample
using the SDSS quasar catalog (Schneider et al. 2005), and the FIRST
catalog and images. Apart from positional coincidence of radio
sources within $2^{''}$ of quasars, we also identify the radio
counterpart of quasars with complex radio morphologies such as lobe
-dominated quasars by visual inspection of their radio images. We
find that using the positional coincidence alone will miss $\sim
8\%$ radio counterpart that do not show radio core at FIRST flux
limit of 1 mJy, and under-estimates the radio flux by a factor of
more than two in another $\sim 9\%$ objects. By comparing the radio
flux from FIRST survey with that from NVSS, we found that lobes in
weak radio sources tend to be missed in this sample. So these
numbers are only lower limits.
Quasars with extended radio emission show both larger radio powers
and radio loudness, and appear somewhat bluer than radio compact
quasars despite their indistinguishable optical luminosity. As such,
the radio extended quasars account for nearly one third of radio
loud quasars at log$(RL)>2.2$. Naturally, including the extended
emission and weak core radio sources increases the fraction of the
radio loud objects and the significance of radio loud peak in the
distribution of Hough \& Readhead radio loudness.
Our results in the first glance are not consistent with the simple
unification scheme in which radio compact quasars are extended ones
viewed along radio jet, for which the relativistic beaming enhances
the core radio emission and as such the total radio power (Wills \&
Browne 1986; Hough \& Readhead 1988; Barthel 1989; Falcke et al.
2004) when the projection effect would make the apparent size
smaller. Within such a scheme, the unresolved core is enhanced
because of the beaming effect. That the lobe-dominated quasars are
more luminous in radio seems to contradict this model. However,
there are at least two selection effects that make the average radio
power in the core-dominated sources smaller.
First, as we showed in the last section that the peak brightness of radio
lobe component is correlated with radio power, and the FIRST will not be able
to detect the lobe component in less radio power sources, especially at
high redshift (see also Fig.\ref {lobe-distrib}). Second, if most of
core-dominated quasars are intrinsically radio weaker (Wang et al. 2006)
and if the radio luminosity function of them is steep, their average
apparent luminosity can be lower even if the radio power had been boosted.
The relative number density of the extended and compact sources also
suggests that majority of compact sources are either of beamed,
intrinsically much weak radio quasars or intrinsically compact radio
sources, such as CSS (Compact Steep Spectrum Objects) or GPS
(GigaHertz Peaked Sources). With typical Lorentz factor of 10-15 for
jets in FR-II radio quasars, the boosted emission can be viewed in
only relatively small fraction of solid angle between the line of
sight and the jet, $\theta <7^{\circ} $, whereas at other angles the
core is weakened due to Doppler effect. If the power of the
un-beamed radio cores is near 0.005 of that of the lobes, as
determined for 3CR sources (Urry \& Padovani 1995), and if the
intrinsic radio loudness distribution following Cirasuolo et al.
(2003b) , and with the luminosity distribution as DR3 quasar sample,
we can estimate that $\sim 62\%$ beamed intrinsic radio quiet
sources could be detected by FIRST. Since GPS and CSS are all
powerful radio sources, it is likely most of these compact quasars
are beamed radio intermediate quasars as proposed by Falcke et al.
(1996)
Our results suggest that lobe-dominated sources are not particularly reddened,
in agreement with the finding by de Vries et al. (2005). This is valid even
for quasars without detectable radio core at flux down to the FIRST limit. The line of
sight does not intercept the dusty torus in those lobe-dominated quasars.
However, Backer et al. (1997) found that most lobe-dominated quasars in the Molonglo quasars are reddened by $A_V\simeq 2-4$, and CSSs are most
reddened. It should be pointed out that quasars reddened by this amount cannot
be found in the `color' selected sample, in particularly at high redshift due
to strong attenuation in ultraviolet. The slightly bluer color for
lobe-dominated quasars in this sample could be due to inclusion of
CSS-like objects in the core-dominant objects or due to a selection
effect by which reddened "extended" radio quasars are lost.
If Baker (1997) is correct, we might miss a large number of heavily
reddened lobe-dominated quasars. Although most such quasars are likely
below the magnitude limit of spectroscopic quasar sample, in principle
the FIRST selected sample is able to detect some of such reddened quasars,
particularly in the low redshift if a weak core is present. We look at the
spectroscopic sources that selected as FIRST sources only, and find
that the FIRST-only selected spectroscopic sources are indeed much
redder. But the fraction of quasars with extended lobes in FIRST-only sources
is very low($\sim 0.2\%$) probably due to large extinction. Thus it is not
conclusive whether a large number of such reddened lobe-dominated quasars
do exist.
\acknowledgements We thank the anonymous referee for the constructive comments.
This work is supported by Chinese Natural Science Foundation through
CNSF-10233030 and CNSF-10573015.
\begin{figure}[Ht]
\centering\includegraphics[height=6cm,width=8cm]{lu.fig1a.eps}
\vspace{5mm}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig1b.eps}
\vspace{5mm}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig1c.eps}
\vspace{5mm}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig1d.eps}
\vspace{5mm}
\caption{\label{FIRST_image} Top two panels: FIRST
images of typical FRII-type radio quasars. Bottom two panels: the
FIRST images of the multi-component radio quasars.}
\end{figure}
\begin{figure}
\plottwo{lu.fig2a.eps}{lu.fig2b.eps}
\caption{\label{lobe-distrib}
Left panel: The distance of the furthest radio lobe to the quasar
versus the total radio power for extended radio quasars. There is no
apparent correlation. Right panel: The intrinsic peak brightness of
resolved lobes versus the total radio power (defined as $P_{tot}$)
for extended radio quasars.}
\end{figure}
\begin{figure}
\plottwo{lu.fig3a.eps}{lu.fig3b.eps}
\caption{\label{redshift} Left
panel: The redshift distribution of optical selected radio quasars
in our sample (green line) and of all optical selected DR3 quasars
in Schneider et al. (2005) (red line). The dot-dashed line shown the
ratio of optical+radio quasar to all optical selected quasars. Right
panel: The redshift distribution of extended radio quasars (green
line) and compact quasars (red line). And the dot-dashed line shown
the fraction of extended quasars to all radio quasars. Obviously the
ratio drops as redshift increases up to z=2.5. }
\end{figure}
\begin{figure}
\plottwo{lu.fig4a.eps}{lu.fig4b.eps}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig4c.eps}
\linespread {0.3} \caption{\label{NVSS-FIRST_flux} Top left panel:
The distribution of log$(f_{NVSS}/f_{FIRST})$ . The green line
represents the sources with FIRST flux above the 3 mJy limit, and
the red line with FIRST flux below the 3 mJy limit. The FIRST flux
is the sum of all radio sources within NVSS beam size. Top right
panel: The flux distribution of extended sources (green line) and
compact sources (red line). The vertical dot-dash line marks the 3
mJy flux. Bottom panel: The NVSS-SDSS offset distribution of the
"selected" matches and "rejected" matches in 20\texttt{"} matching
radius.}
\end{figure}
\begin{figure}
\plottwo{lu.fig5a.eps}{lu.fig5b.eps} \caption{\label{obs-e-c-chars}
Left panel: The accumulated fraction of sources with nearest FIRST
counterpart within certain matching radius. The solid line
represents the lobe-dominant sources and the dashed one all the
radio quasars. Right panel: The distribution of flux ratio $q$. $q$
defined as the flux ratio of "core" component (the radio counterpart
within 2\texttt{"}) to the total radio flux (summation over all
radio counterparts that associate to the radio quasar). }
\end{figure}
\begin{figure}
\plottwo{lu.fig6a.eps}{lu.fig6b.eps} \caption{\label{lobe-z} Left
panel: The intrinsic brightness of the lobes versus redshift. The
solid line represents the surface brightness at the detection limit
as 0.75 mJy/beam. Right panel: The physical size of extended sources
at different redshift. The size is the physical distance of the
furthest radio component associated with the quasar. The solid line
marks the minimum distance that can
be resolved in the FIRST survey at the corresponding redshift (on scales down
to $\sim 1/3$ the beam size of 5\farcs4.)}
\end{figure}
\begin{figure}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig7a.eps}
\linespread {0.3}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig7b.eps}
\linespread {0.3}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig7c.eps}
\linespread {0.3}
\centering\includegraphics[height=6cm,width=8cm]{lu.fig7d.eps}
\linespread {0.3} \caption{\label{color} The median-subtracted $g-i$
color distribution for quasars with redshift $0<z<3$ (refer to the
text for definition). A comparison between extended (green lines)
and compact radio quasars (red line) is shown in the top left panel;
between resolved (green line) and unresolved compact radio quasars
(red line) in the bottom left panel; between extended quasars with
(red line) and without (green line) core in the top right panel; and
between only-lobe (green line) and only-core (red line) quasars. }
\end{figure}
\begin{figure}[Ht]
\plottwo{lu.fig8a.eps}{lu.fig8b.eps} \linespread {0.3}
\caption{\label{R_O_lum} Left panel: The average absolute magnitude
at 2500A versus redshift. The green line represents the extended
sources, and the red one the compact sources. Right panel: The
average radio power of extended (green line) sources and compact
(red line) sources at 1.4GHz versus redshift. The thin dashed line
indicates the FIRST detection limit of 1mJy. Error bars are shown at
1$\sigma$ scatter around the average.}
\end{figure}
\begin{figure}
\plottwo{lu.fig9a.eps}{lu.fig9b.eps} \linespread {0.3}
\caption{\label{P-R_distribution} The radio luminosity (left panel)
and the radio loudness ($f_{1.4GHz}/ f_{2500A}$ in rest frame, right
panel) distributions. The green line represents for extended radio
quasars, and the red line for compact radio quasars, and the black
line for all radio quasars.}
\end{figure}
\begin{figure}[Ht]
\plottwo{lu.fig10a.eps}{lu.fig10b.eps} \linespread {0.3}
\caption{\label{cond_prob} The left panel: The conditional radio
loudness distribution for compact sources and for all cores
(component within 2 arcesec offset from quasar. the orange line and
red line represent compact sources and cores with $15<m_{2500}<19$,
the blue and purple line represent compact sources and cores with
$19<m_{2500}<21$. The right panel: conditional distribution of
extended sources with $15<m_{2500}<19$ (dashed line) and
$19<m_{2500}<21$ (solid line).}
\end{figure}
\begin{figure}[Ht]
\plottwo{lu.fig11a.eps}{lu.fig11b.eps} \linespread {0.3}
\caption{\label{e-c-rl2500-z} The radio loudness distribution of the
compact (left panel) and extended (right panel)
sources below certain redshifts, the descending lines for $z<3,z<2,z<1.5,$ and $z<0.8$
, respectively.}
\end{figure}
|
\section{Introduction}
In this article we
study exit times of continuous-time random walks (CTRWs) with
drift. By this we understand a
random process $X_t$ whose evolution in
time
can be thought of as the result of the combined effect of a constant
drift and the occurrence of random
jumps. Thus, for $t\geqslant t_0$ we define
\begin{equation}
X_t=X_{t_0}+v(t-t_0)+S_{t}, \quad S_t=\sum_{n=1}^\infty J_n\
\theta(t-t_n), \label{process
\end{equation}
where $\theta(u)=1$ for $u \geqslant 0$,
$t_n=\tau_1+\cdots+\tau_n$ are the jump times, $J_n$ the
jump magnitude and $v>0$ by convention. Note that as the notation
suggests, $t_0$ coincides with a jump. The process $X_t$ may
represent physically the position of a diffusing particle, for
instance. All through this paper we will assume that (i) the
sojourn times $\tau_n>0$
are independent and identically distributed (i.i.d.) random variables with
probability density and cumulative distribution function (PDF and,
respectively, CDF) $\psi(t)$ and $\Psi(t)= \int_0^t \psi(t')d
t'$; (ii) $\{J_n\}$ is a sequence of i.i.d. random variables with
common PDF $h(\cdot)$; and (iii), that $ J_m$ is independent of $\tau_n $ for any $n,m$.
In absence of drift, and when the holding times $\tau_n$ are exponentially distributed, $\tau_n\sim \mathcal E(\lambda)$ for some $\lambda >0$, the jump process $S_t$ is a classical compound Poisson process (CPP)
characterized by having independent increments $S_{t+h}- S_t$; in addition the
associated ``arrival-process'' $N_t\equiv\sum_{n=1}^\infty \theta(t-t_n)$ is
Poisson
distributed: $N_t \sim
\mathcal P(\lambda
t)$. Therefore, drift-less CTRWs can be considered as a generalization of CPPs where the holding times of the
processes $S_t$ are arbitrarily distributed. In statistical physics such drift-less CTRWs have been widely
used after the work of Montroll $\&$ Weiss \cite{mw65,w94} and
their relevant statistical properties, like correlation functions
and the behavior in the continuum limit,~\footnote{This limit corresponds to the assumption that the sojourn time between consecutive jumps and their characteristic size go to zero in an appropriate and coordinated way. We will show an explicit example of this limit in the Appendix.}
a subject of intense interest
\cite{JT74,JT76,Go,Sch}. Applications of CTRWs can be found in the study of transport in
disordered media (e.g., \cite{Sch,ms84,Weiss-porra,Margolin1}),
anomalous relaxation in polymer chains~\cite{Hu}, sandpile and earthquake
modeling (e.g., \cite{hs02,Me03}), random networks \cite{Be},
self-organized
criticality in granular systems~\cite{Bo}, scaling properties of L\'evy walks~\cite{Ku}, electron tunneling~\cite{Gu}, transmission
tomography~\cite{OW,OW2},
distribution of matter in the universe~\cite{Os71} and changes in stock
markets due to unexpected catastrophes~\cite{m76}. More
recently,
the use of CTRWs has been advocated to give a microscopic, tick-by-tick,
description of financial markets: see~\cite{mmw03,mmp05,mmpw06,mpmlmm05,s09}. A comprehensive review of CTRW applications in finance and economics is given in \cite{s06}.
Physically,
the introduction of these general CTRWs stems from the fact that
in many settings the exponential holding-time assumption may
be inadequate to describe the physical situation |see \cite{Me03,mmw03}. Additional motivation arises from
the observation that there is an extensive number of different physical systems that show some sort of anomalous diffusion |a subject of great interest from the viewpoint of statistical physics in the last years|, and that this anomalous behavior can well described by fractional Fokker-Planck equations obtained after imposing the continuum limit on CTRW models~\cite
Metzler,s07,m07}.
The further addition of the drift term to a CTRW, as we do in Eq.~(\ref{process}), is a natural and significant incorporation. The resulting process |which can be viewed as the discrete analogue of a (fractional) diffusion with drift| is known for playing a fundamental role in
the modeling of
the cash flow at an insurance company \cite{j03,zyl10} and, more recently, it has been shown that it
also rules the rate of energy dissipation in nonlinear optical
fibers~\cite{VM1}. Indeed, present and forthcoming results can be also of interest in
transport in amorphous media~\cite{Ba},
models of decision
and response time in psychology~\cite{Sm} and neuron dynamics~\cite{Sm2}.
In all these scenarios one is faced with the basic problem of determining the first-passage time for a CTRW with drift, a question that has been theoretically considered in the past~\cite{m87,m92,c97,cmc97,rd00a,rd00b,MB,is07}. The usual approach taken there entails the computation of transition and first-passage time distributions of the process, e.g. \cite{m92,MB}, and the results are typically obtained under the assumption of the continuum limit~\cite{rd00a,rd00b,MB,is07}.
This procedure has incontestable advantages for obtaining the leading-order behavior, but it is not adequate for a detailed analysis of the statistical properties of the process at the inter-jump timescale, as in~\cite{mmw03,mpmlmm05,
j03,zyl10,VM1}.
Motivated by the above,
here we pose the problem of evaluating the mean exit time from
the interval
$(0,b)$ of a drifting CTRW $X_t$ when {\it the only available information is the present state\/} $X_r=x$, where $r \geqslant 0$ is the present chronological time
|note that, by
adjusting the time clock and spatial scale, the results carry
over to any interval $(a,b)$ and initial time $t_0$. Here
$X_t$ is given by Eq.~(\ref{process}) where
$S_t$ is the jump part and $N_t$
the associated counting process (a renewal
process). In the drift-less case previous work in this regard includes that
of \cite{mmp05,mmpw06} where a linear integral equation for the
mean escape time after a jump off a given interval is derived. We
note however that these
results do not cover a generic situation. Indeed, while for
CPPs (as for the general L\'evy processes) Markov property
implies that results derived starting at a jump time carry over
to arbitrary present, no such inference is possible for a generic
CTRW due to its non-Markovian nature. Thus, escape times \it
depend on the actual state and available information \rm and
hence the question as to how to generalize the former results to
\it general present time \rm $r$ appears naturally. We remark that
implications ensuing from the lack of
Markovianess
have been ignored at large in the literature, a gap that we
have intended to fill in |see \cite{VM2}. In particular, it remains
an open
question to what extent dropping the assumption that ``the
present is a jump
time'' affects the relevant probabilities.
Here we address these issues and
generalize the results of \cite{mmp05,mmpw06} in a twofold way
by assuming that (i) a drift $vt>0$ operates on the system and
(ii) the present $r$ is an arbitrary time, not necessarily a
jump instant, and the observer \it has knowledge of the present,
but not of the history, of the system\rm.
The interest of this problem goes far beyond the purely academic
since such a situation may appear in
several different physical contexts. For example, one might be
interested in predicting the mean time for an
insurance/financial company to go bankrupt from the knowledge
of just the actual company budget, i.e., when the information
regarding the company's past performance has not been disclosed.
A second example is provided by the study of the distribution
of inhomogeneities in an optical fiber; it was found \cite{VM1} that the signal's energy amplitude at a point $t$
($t$ is the spatial variable in this setting) involves a CTRW
with drift $X_t$; in this context~\footnote{This situation where
$t$ represents the {\it the space variable } may occur in
different physical systems,
with $X$ standing for some physical observable of interest,
like the energy. In this connection, $r$ might represent the
location of a detector, or a sensitive part of the appliance at
which measures are taken.} one typically knows only the value of
the energy at the observation point $r$, not on the whole fiber.
More generally the approach will be relevant in situations where
either the elapsed time between events is ``large'' (it might be as large as
years, in a context of catastrophes observation)
or when the event's times $t_n$ are not physically
measurable observables and only mean escape times are. (Note that
both the mean exit time and initial time are typically {\it
macroscopic\/} magnitudes.)
The article is structured as follows. In Section~\ref{S2} we show how the solution to these problems involves ideas drawn from renewal theory
and solve the simpler case when $v=0$. The case $v>0$ is
considered in the next sections where it is found that
key properties of the obtained equation depend on the sign of the
jumps. In Section~\ref{S_up} it is shown that if this sign is positive the solution can be
given in closed form by Laplace transformation |cf.
Eqs.~(\ref{FF_sol}) and~(\ref{hattime}). For the case when drift and jumps have opposite signs
we find integral equations that the relevant objects satisfy, see Section~\ref{S_down}, but no closed
solution can be given in a general situation. Section~\ref{S_two-sided}
addresses the most general scenario in which
jumps $J_n$ can take both signs. Solvable cases are
discussed there.
In all cases we exemplify our results by considering the particular instance when
sojourn times have
Erlang distribution, $\mathcal Er(\lambda,2)$.
$\mathcal Er(\lambda,n)$ corresponds to having a sum of $n$ independent exponential
variables and hence generalizes the exponential density in a
natural way,
\begin{equation}
\psi(t) = \frac{\lambda^n t^{n-1}}{(n-1)!} e^{-\lambda t},\, n\in \Bbb N,\, t\geqslant 0,
\label{psi_E2
\end{equation}
while it maintains an adequate capability to fit
measured data. From a physical perspective these facts make
this density a natural candidate to describe multi-component systems which
operate only when several independent, exponentially distributed
operations have been completed or whenever there is a hidden
Poissonian flux of information and jumps only appear as the
outcome of two or more consecutive arrivals.
This explains the interest that
it has drawn in the field of information traffic
\cite{sb00,fc02}. Similarly the appearance of this distribution
to model transaction orders in financial markets can also be
expected since
it takes, at least, two arrivals (buy and sell orders) for a transaction
to be completed. For further applications to ruin problems and
insurance see \cite{dh01,lg04}.
\section{\label{S2}The problem}
Recall that we aim to study exit times of a drifting CTRW
$X_t$
given the present state $ X_r=x$. To this end let $r+\mathbf{t}_b^{x,r}$ be the first time past
$r$ at which $X_t$ exits $(0,b)$; $T_{b}(x,r)$ be its expected
value: $T_{b}(x,r) =\EE[\mathbf{t}_b^{x,r}]$; finally let
$\TT_{b}(x) $ denote the mean exit time off $(0,b)$ after
a jump $t_n$ (loosely one has $\TT_{b}(x)=T_{b}(x,t_n), n=0,\dots,
\infty$; however the relation between both quantities is not
trivial, as we see below in Section~\ref{S_up}). Note also that here and
elsewhere we use $\Bbb E[\cdot]$ to denote expectation.
In the exponential Markov case $
\mathbf{t}_b^{x,r}$ is independent of $r$, $T_{b}(x,r)=\TT_{b}(x)$
and it \it only \rm remains to formulate (and solve) the equation
that this object satisfies. However this situation no longer
holds in the generic, non-Markovian case where $
\mathbf{t}_b^{x,r}$ does depend on $r$. We find (see below and
Section~\ref{S_up}) that the relation between $T_{b}(x,r)$ and
$\TT_{b}(x)$ involves the ``excess life'' $E_r\equiv
t_{N_r+1}-r$, or time elapsed until the next arrival occurs.
We now sketch classical renewal theory (see \cite{kt81,c65}
and \cite{Go}) that
shows how to construct the CDF $\Phi(t|r)\equiv
\Bbb P\{E_r\leqslant t\}$ of $E_r$.
Let $m(t)\equiv \EE[N_t]$ be the mean
number of jumps up to $t$: the renewal function. It satisfies
the integral renewal equation
\begin{equation}
m(t)= \Psi(t) +\int_0^t m(t-t')\psi(t')d t'. \label{re
\end{equation}
Then, by using the total
probability theorem it can be proved that
\begin{equation}
\Phi(t|r) = \int^{r+ t }_{r}\left[1- \Psi(r+ t-t')\right]d m(t').
\label{repp
\end{equation} Upon solution of the above
integral equations we obtain $\Phi(t|r)$. Actually, they can be solved
with all generality
by recourse to Laplace
transformation. Let $\hat g(s)$ be the Laplace
transform of a function $g(t)$ so that
\begin{equation} g(t) =\frac{1}{2\pi i}
\int_{c-i\infty}^{c+i\infty} e^{ st} \hat g(s) d s,\ c>0.
\label{def_m
\end{equation}
Then, Eqs. (\ref{re}) and (\ref{repp}) allow to recover the
distribution of $E_r$ in closed form via
\begin{subequations}
\label{hat_phi}
\begin{eqnarray}
\hat m(s)&=&\frac{1}{s} \frac{\hat\psi(s)}{1-\hat \psi(s)}, \\
\hat \phi(s|r)&=& e^{s r } \left[1-\hat \psi(s)\right] \int_r^\infty e^{-sl} \hat m(l)l d l,
\end{eqnarray}
\end{subequations}
where $ \hat m(s) $ and $ \hat \phi(s|r)$ are the Laplace transforms of $m(t)$ and $\phi(t|r) \equiv \p_t \Phi(t|r)$.
If $v=0$ these expressions can be used to relate $T_{b}(x,r) $ and
$\TT_{b}(x)$. Indeed let $t_{N_r}\eq t_n$, say, be the ``last'' jump time
and $E_r^- \equiv r- t_n$ the time elapsed
from $t_n$ to the present. Then, with $E_r^+\equiv
E_r$ one obviously has that
$E_r^-+E_r^+=t_{n+1}-t_n\equiv\tau_{n+1}$, and that the exit time right
after $t_n$ is that after $r$, $\mathbf{t}_b^{x,r}$, plus $E_r^-$, and hence
\begin{equation}\TT_{b}(x)=\EE [E_r^-]+
\EE[\mathbf{t}_b^{x,r}]=\EE[\tau_{n+1}]-\EE[E_r^+]+
T_{b}(x,r).\label{v=0
\end{equation} Thus,
$T_{b}(x,r)
$ follows adding a correction term to $\TT_{b}(x)$ which depends only on $r$. Finally $\TT_{b}(x)$
is obtained by solving a linear
integral equation |see \cite{mmp05,mmpw06}.
Unfortunately when $v\ne 0$ this simple argument fails as
then \it knowledge of the present position does not entail its knowledge
at $t_n$\rm.
In the next sections we derive the relevant correction
to
the mean
exit time. This correction depends now in all parameters $r$, $x$ and $b$ |see Eqs.~(\ref{Tb_up}) and~(\ref{Tb_down}) below.
\section{\label{S_up}Jump process with favorable drift
In this section we consider the case when both drift and jumps
have a positive sign, i.e. when $X_t$ is increasing. As a result, the process can only leave the interval through the upper
boundary $b$. Let us assume that at time $t=r$ the system is in $X_r=x\in(0,b)$, and that the excess life $E_r$ is known in advance, $E_r=l$. If the excess life is longer than $\varrho \equiv {b-x\over v}< l$, the drift will drive the process
out of the region at time $r+\varrho$, before the next jump takes place at $t_{N_r+1}$. Conversely, if $l \leqslant \varrho$ at least a jump of size $J_{N_r+1}=u$ will occur prior to exiting the interval. Note that just before $t_{N_r+1}$ the process $X_t$ is no longer at $x$, but at $x+v l$. Therefore, two possible scenarios appear: either
the jump size is larger than the remaining distance up to the upper boundary, $u\geqslant b-x-v l$, and the process leaves the
interval at $r+E_r$,
or it does not. In the latter case the problem renews from $t=r+l$, $X_{r+l}=x+v
l+u$, so the mean escape time will be increased by an amount $\TT_b\left(x+v
l+u\right)$. It can be proven that these considerations imply
that $T_b(x,r)$ must satisfy |recall that $\psi(\cdot)$, $h(\cdot)$ are the
waiting-time and, respectively, jump PDFs| the following equation
\begin{eqnarray}
T_b(x,r)=\left[1-\Phi\left(\varrho|r\right)\right] \varrho +
\int^{\varrho}_{0} l
\phi(l|r) \int^{\infty}_{b-x-v l} h(u) d ud l\nonumber \\
+\int^{\varrho}_{0} \phi(l|r) \int^{b-x-v l}_{0} h(u)
\left[l+\TT_b\left(x+vl+u\right)\right]d ud l,
\label{Tb_up0}
\end{eqnarray} an expression that relates $T_b(x,r)$ and $\TT_b(x)$.
Note that since $\Phi\left(l|r\right)$ depends on $r$ so it
does the mean time $T_b(x,r)$. Finally, after some rearrangement, Eq.~(\ref{Tb_up0}) can be conveniently written as
\begin{eqnarray}
T_b(x,r)= \int^{\varrho}_{0}\left[1-\Phi\left(l|r\right)\right]d
l \nonumber \\+ \frac{1}{v}\int^{v \varrho}_{0}
\phi\left(\left.\varrho-\frac{z}{v}\right|r\right) \int^{z}_{0} h(u)
\TT_b\left(b-z+u\right)d u d z. \label{Tb_up
\end{eqnarray}
Similarly, by
letting
$r\to 0$, we find that
$\TT_b(x)$ must satisfy
\begin{eqnarray}
\TT_b(x)= \int^{\varrho}_{0}\left[1-\Psi(l)\right]d l \nonumber \\
+\frac{1}{v}\int^{v \varrho}_{0}
\psi\left(\varrho-\frac{z}{v}\right) \int^{z}_{0} h(u)
\TT_b\left(b-z+u\right)d ud z, \label{TTb_up
\end{eqnarray}
$x\in(0,b)$. Equation~(\ref{Tb_up}) along with~(\ref{TTb_up}) allow us to solve
the posed problem. The second of these
defines an integral equation for $\TT_b(x)$ which, if $v=0$,
reduces upon appropriate change to that of \cite{mmp05,mmpw06}.
It is remarkable that Eq.~(\ref{TTb_up}) \it can be solved in a fully
explicit way. \rm To this end we define the allied object
$\FF(y)$,
as the solution of the following integral
equation
\begin{eqnarray}
\FF(y)= \int^{y/v}_{0}\left[1-\Psi(l)\right]d l \nonumber \\
+\frac{1}{v}\int^{y}_{0} \psi\left(\frac{y-z}{v}\right)
\int^{z}_{0} h(u) \FF\left(z-u\right)d ud z, \label{FF_eq
\end{eqnarray}
for $y\in \RR^{+}$. Then it follows that $\TT_b(x) =\FF(b-x)$, for
$x\in (0,b)$. We note further that taking a Laplace
transformation in Eq.~(\ref{FF_eq}) we find that
\begin{equation}
\hat{\FF}(s)= \frac{1}{v s^2}\frac{1-\hat{\psi}(s
v)}{1-\hat{\psi}(s v)\hat{h}(s)}, \label{FF_sol}
\end{equation}
where $\hat{\FF}(s)\equiv \int^{\infty}_{0} \FF(y) e^{-s y}d y$,
$\hat{h}(s)\equiv \int^{\infty}_{0} h(u) e^{-s u}d u$, and
$\hat{\psi}(s)\equiv \int^{\infty}_{0} \psi(s) e^{-s t}d t$. Here
$s=s_R+is_I$ is complex and $s_R\geqslant 0$. Further, we also have
\begin{equation}
\hat{J}(s|r) = \hat{\FF}(s)- \frac{ 1- \hat{h}(s) }{v
s^2\left[1-\hat{\psi}(s v)\hat{h}(s)\right]}\left[\hat{\phi}(s
v|r)-\hat{\psi}(s v )\right],\label{hattime}
\end{equation}
where the correction to the mean time after a jump is clearly
displayed and again, for convenience, we defined $J(y|r)\eq T_b(b-y|r)$, and extended (\ref{Tb_up}) to $y\geqslant 0$. Thus ${\FF}(y) $ and $J(y|r)$ can be recovered by
Laplace inversion, cf. Eq.~(\ref{def_m}), and $\TT_b(x)$ and $T_b(x|r)$ will eventually follow.
Equation~(\ref{hattime}) has several mathematical limits of applied interest. We first consider the situation when $v$ is small. Let $\mu\equiv~\EE[\tau_{n+1}]$, $\mu_r\equiv~\EE[E_{r}]$; then, using that $ \hat{\psi}(s v )=1-\mu sv
+O(v^2)$ and so forth we see that for small $v$, $\hat{ J}(s|r)$
has an expansion in powers of $v$ as
\begin{equation}
\hat{J}(s|r) = \hat{\FF}(s) - \frac{\mu-\mu_r}{s} - \frac{
\hat{h}(s) }{ 1- \hat{h}(s) } (\mu-\mu_r)\mu v + O(v^2),
\end{equation}
which implies, in particular, Eq.~(\ref{v=0}). Another interesting
case is obtained letting $r\to\infty$: the steady-state solution.
This limit is relevant since it can be associated to a
situation in which \it the only information available to the
observer is the present value of the stochastic process, not even
the starting point\rm. Recalling that by the renewal theorem
$\un{t\to\infty}\lim m(t)/t = 1/\mu$, Eqs.~(\ref{hat_phi}) and~(\ref{hattime}) yield
\begin{eqnarray} \hat\phi (s|\infty)
&=& \frac{1-\hat \psi(s)}{s\mu}, \\
\hat{J}(s|\infty) &=&
\hat{\FF}(s)- \frac{ 1- \hat{h}(s) }{v^2 s^3 \mu}
\frac{1-(1+s v\mu)\hat \psi(sv)}{ 1-\hat{\psi}(s
v)\hat{h}(s)}.
\end{eqnarray}
We illustrate the ideas above by detailing the case when jump magnitudes
have
an exponential distribution
$h(u)=\gamma e^{-\gamma u}$, where $\gamma>0$ is a real parameter, and sojourn times
an Erlang distribution $\tau_n\sim\mathcal Er(\lambda,2)$.
Hence
\begin{equation}
\hat{h}(s)= \frac{\gamma}{\gamma +s},\quad \hat{\psi}(s) = \frac{\lambda^2}{(\lambda +s)^2}.
\label{h_hat}
\end{equation}
In this case Eq.~(\ref{FF_eq}) yields that $\hat{\FF}(s)$ is the
following rational function
\begin{equation}
\hat{\FF}(s)=\frac{1}{v s^2}\frac{2\lambda \gamma v+ v(2 \lambda
+\gamma v) s + v^2 s^2}{\lambda (\lambda + 2\gamma v) + v(2
\lambda +\gamma v) s+v^2 s^2}.
\end{equation}
Hence, the mean escape time of $(0,b)$ after a jump is
\begin{eqnarray}
\TT_b(x)= \frac{2 \gamma v}{\lambda + 2 \gamma v} \varrho \nonumber\\
+\frac{\lambda^2}{\zrp-\zrm}\left[\frac{1-e^{-\zrm \varrho}}{\zrm^2}-\frac{1-e^{-\zrp
\varrho}}{\zrp^2}\right],
\end{eqnarray}
where we recall that $v \varrho=b-x$ is the initial
distance to the boundary $b$
and
\begin{equation}
z_{\pm}\equiv
\lambda +\frac{\gamma v}{2}\left(1\pm\sqrt{1-\frac{4\lambda}{\gamma
v}}\right),
\end{equation}
with $\mbox{Re}[z_{\pm}]>0$.
The evaluation of the mean exit time starting at $r$,
$T_b(x,r)$, involves the renewal function and excess-life
distribution. We first obtain from Eq.~(\ref{hat_phi})
\begin{equation}
\hat m(s)=\frac{\lambda^2}{s^2(2\lambda +s)}, \quad \hat
\phi(s|r)=\hat \psi(s)+ \frac{ (1-e^{-2 \lambda r})s\lambda }{ 2
(\lambda +s)^2}.
\end{equation}
By inversion we get $m(t)= \big(2 \lambda t + e^{-2 \lambda
t}-1\big)/4$ and
\begin{equation} \ \Phi(t|r)
=
1-e^{-\lambda t} \left[ 1+ \left(\frac{ 1+{e^{-2 \lambda r
}}}{2}\right)\lambda t \right].
\end{equation}
Then by using Eq.~(\ref{hattime}) we find at last that
\begin{eqnarray} T_b(x,r) =\TT_b(x) \nonumber
\\
- \frac{\lambda}{2}\frac{1-e^{-2\lambda r}}{\zrp-\zrm}\left[\frac{1-e^{-\zrm \varrho}}{\zrm}-\frac{1-e^{-\zrp \varrho}}{\zrp}\right]
\end{eqnarray}
Plots of this function in terms of $x$ are given in Fig.~\ref{E2_up} for several values of $r$ and a certain choice of the rest of parameters.
\begin{figure}[htb]
{
\includegraphics[width=0.45\textwidth,keepaspectratio=true]{Fig1.eps}
} \caption{(Color online) Sample representations of $T_b(x,r)$ for different values of $r$: $r=0$, i.e. $\TT_b(x)$, solid (black) line, $r=0.4$, dashed (red) line and $r=10.0$, dot-dashed (blue) line. The rest of the parameters were chosen as follows: $b=1.0$, $v=0.1$, $\lambda=1.0$ and $\gamma=0.1$.} \label{E2_up}
\end{figure}
Finally note how, in particular, if $v=0$ then
\begin{eqnarray}
\TT_b(x)=\frac{2}{\lambda}\left[1+\gamma (b-x)\right] &=& \EE[\tau_n]
\left(1+ \frac{b-x}{\EE[J_n]}\right), \\
T_{b}(x,r)&=&\TT_b(x)
-\frac{1- e^{-2\lambda r}}{2\la}.\end{eqnarray}
\section{\label{S_down}The case of opposite drift and jumps}
We now consider the case when the sign of the drift $vt>0$ is
opposite to that of the jumps.
In this case the process can leave the
interval through both the upper and the lower boundaries: the
drift pushes steadily the process up, whereas the jumps threaten
the system with a downside exit. The resulting process is
a prototype model in risk management to describe the
dynamics of the cahsflow $X_t$ at an insurance company under
the assumption that premiums are received at a constant rate
$v>0$ and that the company incurs in losses $J_n<0$
from
claims reported at times $t_n, n=1,\dots\infty$ (the
Cramer-Lundberg model).
As before, we analyze the evolution starting at $t=r$ with $X_r=x$. Then if
$E_r=l$, $l > \varrho$,
the drift will drive the
process out of the region through the {\it upper boundary\/} at time
$r+\varrho$. Otherwise ($l \leqslant \varrho$) at least a jump
$J_{N_r+1}=-u$, say, occurs at time $t_{N_r+1}$ before escape, and two possible scenarios appear depending on the relative magnitudes of the jump size and the location of the process right before the jump, $X_t=x+v l$: for $u\geqslant x+v l$ the
process will leave the interval through the {\it lower boundary\/} at $r+E_r$, but when $u<x+v l$ the process after the jump will remain inside the interval, $X_{r+l}=x+v l-u$, the exit problem will start afresh, and the mean escape time will be increased by $\TT_b\left(x+v l-u\right)$. Again these considerations imply
that $T_b(x,r)$ and $\TT_b(x)$ must satisfy for $0< x< b$
\begin{eqnarray}
T_b(x,r)=\left[1-\Phi\left(\varrho|r\right)\right] \varrho +
\int^{\varrho}_{0} l
\phi(l|r) \int^{\infty}_{x+v l} h(u) d ud l \nonumber \\
+\int^{\varrho}_{0} \phi(l|r) \int^{x+vl}_{0} h(u)
\left[l+\TT_b\left(x+vl-u\right)\right]d ud l\nonumber \\
=
\int^{\varrho}_{0}\left[1-\Phi\left(l|r\right)\right]d l \nonumber \\
+
\frac{1}{v}\int^{b}_{x}
\phi\left(\left.\frac{z-x}{v}\right|r\right) \int^{z}_{0} h(u)
\TT_b\left(z-u\right)d ud z,\nonumber \\ \label{Tb_down}
\end{eqnarray}
\begin{eqnarray}
\TT_b(x)= \int^{\varrho}_{0}\left[1-\Psi(l)\right]d l
\nonumber \\
+\frac{1}{v} \int^{b}_{x} \psi\left(\frac{z-x}{v}\right)
\int^{z}_{0} h(u) \TT_b\left(z-u\right)d ud z. \nonumber \\ \label{TTb_down}
\end{eqnarray}
Hence
$T_b(x,r)$ follows in terms of quadratures also in this case, given $\TT_b(x)$.
Unfortunately, unlike what happens for the case considered in the previous Section,
Eq.~(\ref{TTb_down}) can not be solved in closed form
for arbitrary PDFs $\psi(\cdot)$ and
$h(\cdot)$. Further progress can be made for Erlang times, $\mathcal Er(\lambda,2)$.
Indeed, in this case (\ref{TTb_down}) reads
\begin{eqnarray}
\TT_b(x)= \frac{2-\left(2+\lambda \varrho\right)e^{-\lambda
\varrho}}{\lambda} \nonumber \\+\frac{\lambda^2}{v^2} \int^{b}_{x}
\left(z-x\right)e^{-\lambda \left(z-x\right)/v} \int^{z}_{0} h(u)
\TT_b\left(z-u\right)d ud z,
\end{eqnarray}
and hence, by differentiation we find that
$ \TT_b(x)$, for $0< x< b$, also satisfies the following
integral-differential equation:
\begin{eqnarray}
\TT_b''(x)-\frac{2\lambda}{v}\TT_b'(x)+\frac{\lambda^2}{v^2}\TT_b(x)=
\frac{2\lambda}{v^2} \nonumber \\+\frac{\lambda^2}{v^2}\int^{x}_{0}h(u) \TT_b\left(x-u\right)d u,
\label{IDb_down}
\end{eqnarray}
subject to the following boundary conditions:
\begin{equation}
\lim_{x\to b}\TT_b(x)=0,\mbox{ and }\lim_{x\to b}\TT_b'(x)=-1/v.
\label{boundary}
\end{equation}
We first consider a general solution to this equation extending
it to the full real axis, so we will drop the subscript in $\TT_b(x)$.
We find a solution by Laplace transformation as
\begin{eqnarray}
\hat{ \TT }(s)= \frac{2\lambda/s + (B v-2 \lambda A)v + A
v^2s}{\lambda^2[1-\hat{h}(s)]-2\lambda v s+ v^2s^2},
\end{eqnarray}
where $A $ and $B$ are
$\TT(0)$ and $\TT'(0)$ respectively. By
inversion, cf. Eq.~(\ref{def_m}), $\TT(x)$ follows in terms of
$A$ and $B$. By requiring~(\ref{boundary})
we obtain a linear algebraic system for $A$ and $B$,
which upon solution yields $\TT_b(x)$ in closed form.
To be specific we consider the case when jumps are also exponentially
distributed: $h(x)=\gamma e^{-\gamma x}$. Then we have~(\ref{h_hat})
and $\hat{\TT}(s)$ is the rational
function
\begin{eqnarray}
\hat{\TT}(s)= \frac{(\gamma+s)[2\lambda + (B v-2 \lambda A)v s+ A
v^2s^2]}{s^2[\lambda(\lambda-2 \gamma v)+(\gamma v-2\lambda) v s+
v^2s^2]}. \label{LUps}
\end{eqnarray}
Upon re-scale of constants the inverse Laplace transform
of~(\ref{LUps}) reads
\begin{eqnarray}
\left(\frac{\lambda}{2}-\gamma v\right)\TT(x)\nonumber \\
=\tilde A + \gamma x +
\left[\frac{\lambda}{v^2}(\lambda-\lambda
\tilde A+ v\tilde B) -\frac{\tilde B}{2}
\xi_{-}\right]\frac{e^{\xi_+ x}-1}{\xi_+(\xi_+-\xi_-)} \nonumber \\
- \left[\frac{\lambda}{v^2}(\lambda-\lambda\tilde A+ v
\tilde B) -\frac{\tilde B}{2}\xi_{+}\right]\frac{e^{\xi_-
x}-1}{\xi_-(\xi_+-\xi_-)},\nonumber \\
\end{eqnarray}
with
\begin{equation}
\xi_{\pm}\equiv
\frac{\lambda}{v}-\frac{\gamma}{2}\pm\frac{\gamma}{2}\sqrt{1+\frac{4\lambda}{\gamma
v}}.\end{equation}
Unfortunately the final expressions for $\tilde A$ and $\tilde B$ after imposing~(\ref{boundary}) are not
very illuminating so
we do not transcribe them here. Sampling values for different
parameter specifications can be found in Fig.~\ref{E2_down}.
The limit $b\to \infty$ is interesting as $\TT_\infty(x)$ gives
the probability that $X_t$ ever hits $0$. This corresponds to the
classical ruin probability in an insurance context. It turns
out that $\TT_\infty(x)$ can be determined in a direct way that
avoids solving the aforementioned linear system. Without
proof~\footnote{We elaborate on a similar expression in the next
section.} we note that if $\lambda
> 2 \gamma
v $ then
\begin{equation}
\TT_\infty(x)=\frac{2(1+\gamma x)}{\lambda-2 \gamma v}= \left( 1+\gamma
x \right)\Big/\left( \frac{1}{ \Bbb E[\tau_n]}- \frac{v}{ \Bbb E[J_n]}
\right),
\end{equation}
while $\TT_\infty(x)=\infty$ otherwise. Once $\TT_b(x)$ is
known, $T_b(x,t)$ follows again by integration |see Fig.~\ref{E2_down}.
\begin{figure}[htbp]
{
\begin{tabular}{rl}
(a)&\includegraphics[width=0.45\textwidth,keepaspectratio=true]{Fig2a.eps}\\
(b)&\includegraphics[width=0.45\textwidth,keepaspectratio=true]{Fig2b.eps}\\
\end{tabular}
}
\caption{(Color online) Sample representations of $T_b(x,r)$ for different values of $r$: $r=0$, i.e. $\TT_b(x)$, solid (black) line, $r=0.4$, dashed (red) line and $r=10.0$, dot-dashed (blue) line. The rest of the parameters were chosen as follows, $b=1.0$, $v=0.1$, $\lambda=1.0$ and (a) $\gamma=0.1$; (b) $\gamma=4.0$.
We see how, when drift and jumps have a different sign, an interesting behavior may appear in $T_b(x,r)$.
In the upper panel we observe that this function is no longer decreasing with $x$ and has a maximum in the interior of the interval, cf. Fig.~\ref{E2_up}. In the lower panel we observe how a cross-over phenomenon may eventually take place reflecting the fact that $\TT_b(x)$ need not be greater than $T_b(x, r)$: this behavior can be traced to the fact that as $r$ increases the jump probability increases, which results in a reduction of the escape time if the process is in the vicinity of the lower boundary but in an increment when close to the upper boundary. The maxima position moves toward greater values of $x$ for larger choices of $r$, in both cases.
} \label{E2_down}
\end{figure}
\section{\label{S_two-sided}Two-sided jump process with drift}
We finally consider the general case corresponding to a jump
process where $J_n$ can take both signs and hence can exit
$(0,b)$ through either of the boundaries. The relevant analysis
is similar to that of the last section if one incorporates the
possibility of an upper exit due to a jump: If $E_r=l>\varrho$ the drift drives directly the process through the upper boundary. If $l\leqslant \varrho$ at least a jump occurs before the exit of the process. When the size $u$ of this jump is positive and $u\geqslant b-x-vl$ the process leaves the interval at time $r+E_r$ through the upper boundary, when $u\leqslant -(x+vl)<0$ the exit takes place through the lower boundary, otherwise the process remains inside the interval $(0,b)$ and the problem restarts.
Skipping minor details
we obtain that $\TT_b(x)$ and $T_b(x,r)$ satisfy respectively
\begin{eqnarray}
T_b(x,r)=
\int^{\varrho}_{0}\left[1-\Phi\left(l|r\right)\right]d l \nonumber \\+
\frac{1}{v}\int^{b}_{x}
\phi\left(\left.\frac{z-x}{v}\right|r\right) \int^{b-z}_{-z} h(u)
\TT_b\left(z+u\right)d ud z,
\end{eqnarray}
and
\begin{eqnarray}
\TT_b(x)= \int^{\varrho}_{0}\left[1-\Psi\left(l\right)\right]d
l \nonumber \\+ \frac{1}{v}\int^{b}_{x} \psi\left(\frac{z-x}{v}\right)
\int^{b-z}_{-z} h(u) \TT_b\left(z+u\right)dud z.
\end{eqnarray}
In a general situation, the latter integral equation is not
solvable in closed form. To gain some insight we use the
decomposition
$ h(x)= q h_+(x) + p h_-(x)$ where $p$ is the probability that
a given jump be negative, $q\equiv 1-p$ and
$h_{\pm}(\cdot)$ are the jump PDF
in the positive/negative regions, i.e. $h_+(x)\equiv h(x)
\theta(x) \big/ q$, say.
Note that if $h_-(x)=0, \fo x\leqslant 0$, we recover the case considered
in Section~\ref{S_up}, solvable via Laplace
transformation. It turns out that we can still \it construct an
analytic closed solution \rm in
the more general case when $h_-(x) $ vanishes \it only \rm on $(-b,0)$ |but not on
$(-\infty,-b]$.
In such
a case a negative jump will drive the process out of the
interval $(0,b)$ through the lower boundary. Thus $p$ is related
to the ruin risk in a economic scenario or to the breakdown
probability in a physical system.
The equation for
$\TT_b(x)$ reads in this case
\begin{eqnarray}
\TT_b(x)= \int^{\varrho}_{0}\left[1-\Psi\left(l\right)\right]d
l \nonumber\\+ \frac{q}{v}\int^{b}_{x} \psi\left(\frac{z-x}{v}\right)
\int^{b-z}_{0} h_+(u) \TT_b\left(z+u\right)d ud z,
\label{TTb_ruin}
\end{eqnarray}
while a similar expression, with $\phi(\cdot|r)$ in place of $\psi(\cdot)$, holds for $T_b(x,r)$.
Note that these equations are independent of the form of $h_-(\cdot)$
and apart from the factor $q$ in front of the second term they
resemble Eqs.~(\ref{Tb_up}) and~(\ref{TTb_up}); it follows that
we can resort to the same technique used in Section~\ref{S_up}:
We consider
\begin{eqnarray}
\FF(y)= \int^{y/v}_{0}\left[1-\Psi(l)\right]d l \nonumber \\+
\frac{q}{v}\int^{y}_{0} \psi\left(\frac{y-z}{v}\right)
\int^{z}_{0} h(u) \FF\left(z-u\right)d ud z,
\end{eqnarray}
for $y\in \RR^{+}$ which is again solvable by means of a Laplace
transform; then $\TT_b(x)$ follows from $\TT_b(x) =\FF(b-x)$ for
$x\in (0,b)$. With the previous selection for $h(\cdot)$ and
$\psi(\cdot)$, Eq.~(\ref{h_hat}), and in terms of $z=sv $
we find
\begin{eqnarray}
\hat{\FF}(s)=\frac{v}{z} \frac{2\lambda \gamma v + (2 \lambda+\gamma v ) z + z^2}{p \lambda^2 \gamma v + \lambda (\lambda +2\gamma v ) z+ (2 \lambda +\gamma v ) z^2+z^3 }.\label{FF}
\end{eqnarray}
We first consider the case when $\lambda
=\gamma v$. Under this assumption $\hat{\FF}(s)$ has poles at
$z=z_j\equiv \lambda \left(q^{1/3} e^{2\pi ij/3}-1\right)$, $j\in\{1,2,3\}$, and $z=0$.
Inverse Laplace transformation yields the mean exit time as
\begin{equation}
\TT_b(x)= \frac{2}{ p\lambda}+ \sum_{j=1}^3 \frac{ 1+q^{-1/3} e^{-2\pi ij/3}}{ 3 z_j} e^{z_j\varrho}.
\end{equation}
Returning to the general case we see that
the inversion of the
Laplace transform involves solving a cubic
equation, and though explicit formulas are available the resulting expression is
awkward. Still, the large $b$ limit
can be discerned with all generality. To this end we note that by
appealing to Hurwitz's stability criteria it can be proven that
all three roots $z_j$, $j\in\{1,2,3\}$, of the denominator in expression
(\ref{FF}) |apart from $z=0$| have negative real parts. Hence,
evaluating the inversion integral by residues we find
\begin{eqnarray}
\TT_b(x)=\frac{2}{ p\lambda} + \sum_{j=1}^3C_j
e^{z_j\varrho},
\end{eqnarray}
where $C_j$ are certain constants. Thus, letting $b\to\infty$ we
see that $\TT_b(x) \underset{b \to\infty}\to \frac{2 }{ p\lambda} =\frac{\EE[\tau_n]}{1-\EE[\theta(J_n)]}$.
The evaluation of the correction to $T_b(x,r)$ does not present
particular difficulties. We leave it as an exercise to the
interested reader.
\section{Conclusions}
We have analyzed the mean exit time for a
general CTRW with drift. If the present coincides with a jump time we find that it satisfies a
certain integral equation whose solvability is analyzed. We consider next the generic case when the present is an
arbitrary instant and
the history of the system is not available to the observer and only the present state is.
It turns out that the corresponding escape time can be obtained by incorporating an appropriate correction, which can be described
in terms of the ``excess life'', a familiar object in renewal
theory. We
find that when the drift and jump components have the same sign the equations that these objects satisfy can be solved in closed form
via Laplace transformation, irrespective of the distribution;
otherwise, one must restrict to particular choices of the
sojourn-time distribution. The case corresponding to the
classical Erlang distribution is analyzed in detail. The more general case when jumps take both positive and negative signs is
also considered and solved under certain severe conditions. We plan to generalize these ideas to a
more general class of waiting-time distributions and pinpoint
conditions that guarantee the reducibility of the original formulation to simpler differential
equations.
The relevance of these results from a physical perspective is discussed in several
connections of interest including possible
applications to risk, finance and distribution of energy in optical systems, which will be the matter of future publications. We also point
out the relevance of the approach whenever the time between events is
``large''
or when the arrival times are not physical
observables.
\acknowledgments
The authors acknowledge support from MICINN under contracts No.
FIS2008-01155-E, FIS2009-09689, and MTM2009-09676; from Junta de Castilla y Le\'on, SA034A08; and Generalitat de Catalunya, 2009SGR417.
|
Subsets and Splits